Asset Pricing
John H. Cochrane
June 12, 2000
1
Acknowledgments This book owes an enormous intellectual debt to Lars...

Author:
John H. Cochrane

This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!

Asset Pricing

John H. Cochrane

June 12, 2000

1

Acknowledgments This book owes an enormous intellectual debt to Lars Hansen and Gene Fama. Most of the ideas in the book developed from long discussions with each of them, and trying to make sense of what each was saying in the language of the other. I am also grateful to all my colleagues in Finance and Economics at the University of Chicago, and to George Constantinides especially, for many discussions about the ideas in this book. I thank George Constantinides, Andrea Eisfeldt, Gene Fama, Wayne Ferson, Owen Lamont, Anthony Lynch, Dan Nelson, Alberto Pozzolo, Michael Roberts, Juha Seppala, Mike Stutzer, Pietro Veronesi, an anonymous reviewer, and several generations of Ph.D. students at the University of Chicago for many useful comments. I thank the NSF and the Graduate School of Business for research support. Additional material and both substantive and typographical corrections will be maintained at http://www-gsb.uchicago.edu/fac/john.cochrane/research/papers c John H. Comments and suggestions are most welcome This book draft is copyright ° Cochrane 1997, 1998, 1999, 2000 John H. Cochrane Graduate School of Business University of Chicago 1101 E. 58th St. Chicago IL 60637 773 702 3059 [email protected] June 12, 2000

2

Contents Acknowledgments

2

Preface

8

Part I.

Asset pricing theory

12

1 Consumption-based model and overview

13

1.1

Basic pricing equation

14

1.2

Marginal rate of substitution/stochastic discount factor

16

1.3

Prices, payoffs and notation

17

1.4

Classic issues in finance

20

1.5

Discount factors in continuous time

33

1.6

Problems

38

2 Applying the basic model

41

2.1

Assumptions and applicability

41

2.2

General Equilibrium

43

2.3

Consumption-based model in practice

47

2.4

Alternative asset pricing models: Overview

49

2.5

Problems

51

3 Contingent Claims Markets

54

3.1

Contingent claims

54

3.2

Risk neutral probabilities

55

3.3

Investors again

57

3.4

Risk sharing

59

3.5

State diagram and price function

60

4 The discount factor

64

4.1

Law of one price and existence of a discount factor

64

4.2

No-Arbitrage and positive discount factors

69

3

4.3

An alternative formula, and x∗ in continuous time

74

4.4

Problems

76

5 Mean-variance frontier and beta representations

77

5.1

Expected return - Beta representations

77

5.2

Mean-variance frontier: Intuition and Lagrangian characterization

80

5.3

An orthogonal characterization of the mean-variance frontier

83

5.4

Spanning the mean-variance frontier

88

5.5

A compilation of properties of R , R

5.6

Mean-variance frontiers for m: the Hansen-Jagannathan bounds

92

5.7

Problems

97

∗

e∗

and x

∗

6 Relation between discount factors, betas, and mean-variance frontiers

89

98

6.1

From discount factors to beta representations

98

6.2

From mean-variance frontier to a discount factor and beta representation

101

6.3

Factor models and discount factors

104

6.4

Discount factors and beta models to mean - variance frontier

108

6.5

Three riskfree rate analogues

109

6.6

Mean-variance special cases with no riskfree rate

115

6.7

Problems

118

7 Implications of existence and equivalence theorems

120

8 Conditioning information

128

8.1

Scaled payoffs

129

8.2

Sufficiency of adding scaled returns

131

8.3

Conditional and unconditional models

133

8.4

Scaled factors: a partial solution

140

8.5

Summary

141

8.6

Problems

142

9 Factor pricing models 9.1

143

Capital Asset Pricing Model (CAPM) 4

145

9.2

Intertemporal Capital Asset Pricing Model (ICAPM)

156

9.3

Comments on the CAPM and ICAPM

158

9.4

Arbitrage Pricing Theory (APT)

162

9.5

APT vs. ICAPM

171

9.6

Problems

172

Part II.

Estimating and evaluating asset pricing models

10 GMM in explicit discount factor models

174 177

10.1 The Recipe

177

10.2 Interpreting the GMM procedure

180

10.3 Applying GMM

184

11 GMM: general formulas and applications

188

11.1 General GMM formulas

188

11.2 Testing moments

192

11.3 Standard errors of anything by delta method

193

11.4 Using GMM for regressions

194

11.5 Prespecified weighting matrices and moment conditions

196

11.6 Estimating on one group of moments, testing on another.

205

11.7 Estimating the spectral density matrix

205

11.8 Problems

212

12 Regression-based tests of linear factor models

214

12.1 Time-series regressions

214

12.2 Cross-sectional regressions

219

12.3 Fama-MacBeth Procedure

228

12.4 Problems

234

13 GMM for linear factor models in discount factor form

235

13.1 GMM on the pricing errors gives a cross-sectional regression

235

13.2 The case of excess returns

237

13.3 Horse Races

239 5

13.4 Testing for characteristics

240

13.5 Testing for priced factors: lambdas or b’s?

241

13.6 Problems

245

14 Maximum likelihood

247

14.1 Maximum likelihood

247

14.2 ML is GMM on the scores

249

14.3 When factors are returns, ML prescribes a time-series regression

251

14.4 When factors are not excess returns, ML prescribes a cross-sectional regression

255

14.5 Problems

256

15 Time series, cross-section, and GMM/DF tests of linear factor models

258

15.1 Three approaches to the CAPM in size portfolios

259

15.2 Monte Carlo and Bootstrap

265

16 Which method?

271

Part III.

284

Bonds and options

17 Option pricing

286

17.1 Background

286

17.2 Black-Scholes formula

293

17.3 Problems

299

18 Option pricing without perfect replication

300

18.1 On the edges of arbitrage

300

18.2 One-period good deal bounds

301

18.3 Multiple periods and continuous time

309

18.4 Extensions, other approaches, and bibliography

317

18.5 Problems

319

19 Term structure of interest rates

320

19.1 Definitions and notation

320 6

19.2 Yield curve and expectations hypothesis

325

19.3 Term structure models – a discrete-time introduction

327

19.4 Continuous time term structure models

332

19.5 Three linear term structure models

337

19.6 Bibliography and comments

348

19.7 Problems

351

Part IV.

Empirical survey

352

20 Expected returns in the time-series and cross-section

354

20.1 Time-series predictability

356

20.2 The Cross-section: CAPM and Multifactor Models

396

20.3 Summary and interpretation

409

20.4 Problems

413

21 Equity premium puzzle and consumption-based models

414

21.1 Equity premium puzzles

414

21.2 New models

423

21.3 Bibliography

437

21.4 Problems

440

22 References

Part V.

442

Appendix

455

23 Continuous time

456

23.1 Brownian Motion

456

23.2 Diffusion model

457

23.3 Ito’s lemma

460

23.4 Problems

462

7

Preface Asset pricing theory tries to understand the prices or values of claims to uncertain payments. A low price implies a high rate of return, so one can also think of the theory as explaining why some assets pay higher average returns than others. To value an asset, we have to account for the delay and for the risk of its payments. The effects of time are not too difficult to work out. However, corrections for risk are much more important determinants of an many assets’ values. For example, over the last 50 years U.S. stocks have given a real return of about 9% on average. Of this, only about 1% is due to interest rates; the remaining 8% is a premium earned for holding risk. Uncertainty, or corrections for risk make asset pricing interesting and challenging. Asset pricing theory shares the positive vs. normative tension present in the rest of economics. Does it describe the way the world does work or the way the world should work? We observe the prices or returns of many assets. We can use the theory positively, to try to understand why prices or returns are what they are. If the world does not obey a model’s predictions, we can decide that the model needs improvement. However, we can also decide that the world is wrong, that some assets are “mis-priced” and present trading opportunities for the shrewd investor. This latter use of asset pricing theory accounts for much of its popularity and practical application. Also, and perhaps most importantly, the prices of many assets or claims to uncertain cash flows are not observed, such as potential public or private investment projects, new financial securities, buyout prospects, and complex derivatives. We can apply the theory to establish what the prices of these claims should be as well; the answers are important guides to public and private decisions. Asset pricing theory all stems from one simple concept, derived in the first page of the first Chapter of this book: price equals expected discounted payoff. The rest is elaboration, special cases, and a closet full of tricks that make the central equation useful for one or another application. There are two polar approaches to this elaboration. I will call them absolute pricing and relative pricing. In absolute pricing, we price each asset by reference to its exposure to fundamental sources of macroeconomic risk. The consumption-based and general equilibrium models described below are the purest examples of this approach. The absolute approach is most common in academic settings, in which we use asset pricing theory positively to give an economic explanation for why prices are what they are, or in order to predict how prices might change if policy or economic structure changed. In relative pricing, we ask a less ambitious question. We ask what we can learn about an asset’s value given the prices of some other assets. We do not ask where the price of the other set of assets came from, and we use as little information about fundamental risk factors as possible. Black-Scholes option pricing is the classic example of this approach. While limited in scope, this approach offers precision in many applications. 8

Asset pricing problems are solved by judiciously choosing how much absolute and how much relative pricing one will do, depending on the assets in question and the purpose of the calculation. Almost no problems are solved by the pure extremes. For example, the CAPM and its successor factor models are paradigms of the absolute approach. Yet in applications, they price assets “relative” to the market or other risk factors, without answering what determines the market or factor risk premia and betas. The latter are treated as free parameters. On the other end of the spectrum, most practical financial engineering questions involve assumptions beyond pure lack of arbitrage, assumptions about equilibrium “market prices of risk.” The central and unfinished task of absolute asset pricing is to understand and measure the sources of aggregate or macroeconomic risk that drive asset prices. Of course, this is also the central question of macroeconomics, and this is a particularly exciting time for researchers who want to answer these fundamental questions in macroeconomics and finance. A lot of empirical work has documented tantalizing stylized facts and links between macroeconomics and finance. For example, expected returns vary across time and across assets in ways that are linked to macroeconomic variables, or variables that also forecast macroeconomic events; a wide class of models suggests that a “recession” or “financial distress” factor lies behind many asset prices. Yet theory lags behind; we do not yet have a well-described model that explains these interesting correlations. In turn, I think that what we are learning about finance must feed back on macroeconomics. To take a simple example, we have learned that the risk premium on stocks – the expected stock return less interest rates – is much larger than the interest rate, and varies a good deal more than interest rates. This means that attempts to line investment up with interest rates are pretty hopeless – most variation in the cost of capital comes from the varying risk premium. Similarly, we have learned that some measure of risk aversion must be quite high, or people would all borrow like crazy to buy stocks. Most macroeconomics pursues small deviations about perfect foresight equilibria, but the large equity premium means that volatility is a first-order effect, not a second-order effect. Standard macroeconomic models predict that people really don’t care much about business cycles (Lucas 1987). Asset prices are beginning to reveal that they do – that they forego substantial return premia to avoid assets that fall in recessions. This fact ought to tell us something about recessions!

This book advocates a discount factor / generalized method of moments view of asset pricing theory and associated empirical procedures. I summarize asset pricing by two equations: pt = E(mt+1 xt+1 ) mt+1 = f(data, parameters). where pt = asset price, xt+1 = asset payoff, mt+1 = stochastic discount factor. 9

The major advantage of the discount factor / moment condition approach are its simplicity and universality. Where once there were three apparently different theories for stocks, bonds, and options, now we see each as just special cases of the same theory. The common language also allows us to use insights from each field of application in other fields. This approach also allows us to conveniently separate the step of specifying economic assumptions of the model (second equation) from the step of deciding which kind of empirical representation to pursue or understand. For a given model – choice of f (·) – we will see how the first equation can lead to predictions stated in terms of returns, price-dividend ratios, expected return-beta representations, moment conditions, continuous vs. discrete time implications and so forth. The ability to translate between such representations is also very helpful in digesting the results of empirical work, which uses a number of apparently distinct but fundamentally connected representations. Thinking in terms of discount factors often turns out to be much simpler than thinking in terms of portfolios. For example, it is easier to insist that there is a positive discount factor than to check that every possible portfolio that dominates every other portfolio has a larger price, and the long arguments over the APT stated in terms of portfolios are easy to digest when stated in terms of discount factors. The discount factor approach is also associated with a state-space geometry in place of the usual mean-variance geometry, and this book emphasizes the state-space intuition behind many classic results. For these reasons, the discount factor language and the associated state-space geometry is common in academic research and high-tech practice. It is not yet common in textbooks, and that is the niche that this book tries to fill. I also diverge from the usual order of presentation. Most books are structured following the history of thought: portfolio theory, mean-variance frontiers, spanning theorems, CAPM, ICAPM, APT, option pricing, and finally consumption-based model. Contingent claims are an esoteric extension of option-pricing theory. I go the other way around: contingent claims and the consumption-based model are the basic and simplest models around; the others are specializations. Just because they were discovered in the opposite order is no reason to present them that way. I also try to unify the treatment of empirical methods. A wide variety of methods are popular, including time-series and cross-sectional regressions, and methods based on generalized method of moments (GMM) and maximum likelihood. However, in the end all of these apparently different approaches do the same thing: they pick free parameters of the model to make it fit best, which usually means to minimize pricing errors; and they evaluate the model by examining how big those pricing errors are. As with the theory, I do not attempt an encyclopedic compilation of empirical procedures. The literature on econometric methods contains lots of methods and special cases (likelihood ratio analogues of common Wald tests; cases with and without riskfree assets and when factors do and don’t span the mean variance frontier, etc.) that are seldom used in practice. I 10

try to focus on the basic ideas and on methods that are actually used in practice. The accent in this book is on understanding statements of theory, and working with that theory to applications, rather than rigorous or general proofs. Also, I skip very lightly over many parts of asset pricing theory that have faded from current applications, although they occupied large amounts of the attention in the past. Some examples are portfolio separation theorems, properties of various distributions, or asymptotic APT. While portfolio theory is still interesting and useful, it is no longer a cornerstone of pricing. Rather than use portfolio theory to find a demand curve for assets, which intersected with a supply curve gives prices, we now go to prices directly. One can then find optimal portfolios, but it is a side issue for the asset pricing question. My presentation is consciously informal. I like to see an idea in its simplest form and learn to use it before going back and understanding all the foundations of the ideas. I have organized the book for similarly minded readers. If you are hungry for more formal definitions and background, keep going, they usually show up later on in the chapter. Again, my organizing principle is that everything can be traced back to specializations of the basic pricing equation p = E(mx). Therefore, after reading the first chapter, one can pretty much skip around and read topics in as much depth or order as one likes. Each major subject always starts back at the same pricing equation. The target audience for this book is economics and finance Ph.D. students, advanced MBA students or professionals with similar background. I hope the book will also be useful to fellow researchers and finance professionals, by clarifying, relating and simplifying the set of tools we have all learned in a hodgepodge manner. I presume some exposure to undergraduate economics and statistics. A reader should have seen a utility function, a random variable, a standard error and a time series, should have some basic linear algebra and calculus and should have solved a maximum problem by setting derivatives to zero. The hurdles in asset pricing are really conceptual rather than mathematical.

11

PART I Asset pricing theory

12

Chapter 1. Consumption-based model and overview I start by thinking of an investor who thinks about how much to save and consume, and what portfolio of assets to hold. The most basic pricing equation comes from the first-order conditions to that problem, and say that price should be the expected discounted payoff, using the investor’s marginal utility to discount the payoff. The marginal utility loss of consuming a little less today and investing the result should equal the marginal utility gain of selling the investment at some point in the future and eating the proceeds. If the price does not satisfy this relation, the investor should buy more of the asset. From this simple idea, I can discuss the classic issues in finance. The interest rate is related to the average future marginal utility, and hence to the expected path of consumption. High real interest rates should be associated with an expectation of growing consumption. In a time of high real interest rates, it makes sense to save, buy bonds, and then consume more tomorrow. Most importantly, risk corrections to asset prices should be driven by the covariance of asset payoffs with consumption or marginal utility. For a given expected payoff of an asset, an asset that does badly in states like a recession, in which the investor feels poor and is consuming little, is less desirable than an asset that does badly in states of nature like a boom when the investor feels wealthy and is consuming a great deal. The former assets will sell for lower prices; their prices will reflect a discount for their riskiness, and this riskiness depends on a co-variance. This is the fundamental point of the whole book. Of course, the fundamental measure of how you feel is marginal utility; given that assets must pay off well in some states and poorly in others, you want assets that pay off poorly in states of low marginal utility, when an extra dollar doesn’t really seem all that important, and you’d rather that they pay off well in states of high marginal utility, when you’re hungry and really anxious to have an extra dollar. Most of the book is about how to go from marginal utility to observable indicators. Consumption is low when marginal utility is high, of course, so consumption may be a useful indicator. Consumption is also low and marginal utility is high when the investor’s other assets have done poorly; thus we may expect that prices are

13

C HAPTER 1

C ONSUMPTION - BASED MODEL AND OVERVIEW

low for assets that covary positively with a large index such as the market portfolio. This is the Capital Asset Pricing Model. The rest of the book comes down to useful indicators for marginal utility, things against which to compute a covariance in order to predict the risk-adjustment for prices.

1.1

Basic pricing equation

An investor’s first order conditions give the basic consumption-based model, · 0 ¸ u (ct+1 ) pt = Et β 0 xt+1 . u (ct ) Our basic objective is to figure out the value of any stream of uncertain cash flows. I start with an apparently simple case, which turns out to capture very general situations. Let us find the value at time t of a payoff xt+1 . For example, if one buys a stock today, the payoff next period is the stock price plus dividend, xt+1 = pt+1 +dt+1 . xt+1 is a random variable: an investor does not know exactly how much he will get from his investment, but he can assess the probability of various possible outcomes. Don’t confuse the payoff xt+1 with the profit or return; xt+1 is the value of the investment at time t + 1, without subtracting or dividing by the cost of the investment. We find the value of this payoff by asking what it is worth to a typical investor. To do this, we need a convenient mathematical formalism to capture what an investor wants. We model investors by a utility function defined over current and future values of consumption, U(ct , ct+1 ) = u(ct ) + βEt [u(ct+1 )] , where ct denotes consumption at date t. We will often use a convenient power utility form, u(ct ) = The limit as γ → 1 is

1 1−γ c . 1−γ t

u(c) = ln(c). The utility function captures the fundamental desire for more consumption, rather than posit a desire for intermediate objectives such as means and variance of portfolio returns. Consumption ct+1 is also random; the investor does not know his wealth tomorrow, and hence how much he will decide to consume. The period utility function u(·) is increasing, reflecting a desire for more consumption, and concave, reflecting the declining marginal value of additional consumption. The last bite is never as satisfying as the first. 14

S ECTION 1.1

BASIC PRICING EQUATION

This formalism captures investors’ impatience and their aversion to risk, so we can quantitatively correct for the risk and delay of cash flows. Discounting the future by β captures impatience, and β is called the subjective discount factor. The curvature of the utility function also generates aversion to risk and to intertemporal substitution: The consumer prefers a consumption stream that is steady over time and across states of nature. Now, assume that the investor can freely buy or sell as much of the payoff xt+1 as he wishes, at a price pt . How much will he buy or sell? To find the answer, denote by e the original consumption level (if the investor bought none of the asset), and denote by ξ the amount of the asset he chooses to buy. Then, his problem is, max u(ct ) + Et βu(ct+1 ) s.t. {ξ}

ct ct+1

= et − pt ξ = et+1 + xt+1 ξ

Substituting the constraints into the objective, and setting the derivative with respect to ξ equal to zero, we obtain the first-order condition for an optimal consumption and portfolio choice, pt u0 (ct ) = Et [βu0 (ct+1 )xt+1 ]

(1)

· 0 ¸ u (ct+1 ) pt = Et β 0 xt+1 . u (ct )

(2)

or,

The investor buys more or less of the asset until this first order condition holds. Equation (1.1) expresses the standard marginal condition for an optimum: pt u0 (ct ) is the loss in utility if the investor buys another unit of the asset; Et [βu0 (ct+1 )xt+1 ] is the increase in (discounted, expected) utility he obtains from the extra payoff at t+1. The investor continues to buy or sell the asset until the marginal loss equals the marginal gain. Equation (1.2) is the central asset-pricing formula. Given the payoff xt+1 and given the investor’s consumption choice ct , ct+1 , it tells you what market price pt to expect. Its economic content is simply the first order conditions for optimal consumption and portfolio formation. Most of the theory of asset pricing just consists of specializations and manipulations of this formula. Notice that we have stopped short of a complete solution to the model, i.e. an expression with exogenous items on the right hand side. We relate one endogenous variable, price, to two other endogenous variables, consumption and payoffs. One can continue to solve this model and derive the optimal consumption choice ct , ct+1 in terms of the givens of the model. In the model I have sketched so far, those givens are the income sequence et , et+1 and a specification of the full set of assets that the investor may buy and sell. We will in fact study 15

C HAPTER 1

C ONSUMPTION - BASED MODEL AND OVERVIEW

such fuller solutions below. However, for many purposes one can stop short of specifying (possibly wrongly) all this extra structure, and obtain very useful predictions about asset prices from (1.2), even though consumption is an endogenous variable.

1.2

Marginal rate of substitution/stochastic discount factor

We break up the basic consumption-based pricing equation into p = E(mx)

m=β

u0 (ct+1 ) u0 (ct )

where mt+1 is the stochastic discount factor. A convenient way to break up the basic pricing equation (1.2) is to define the stochastic discount factor mt+1 mt+1 ≡ β

u0 (ct+1 ) u0 (ct )

(3)

Then, the basic pricing formula (1.2) can simply be expressed as pt = Et (mt+1 xt+1 ).

(4)

When it isn’t necessary to be explicit about time subscripts or the difference between conditional and unconditional expectation, I’ll suppress the subscripts and just write p = E(mx). The price always comes at t, the payoff at t + 1, and the expectation is conditional on time t information. The term stochastic discount factor refers to the way m generalizes standard discount factor ideas. If there is no uncertainty, we can express prices via the standard present value formula pt =

1 xt+1 Rf

(5)

where Rf is the gross risk-free rate. 1/Rf is the discount factor. Since gross interest rates are typically greater than one, the payoff xt+1 sells “at a discount.” Riskier assets have lower prices than equivalent risk-free assets, so they are often valued by using risk-adjusted 16

S ECTION 1.3

P RICES , PAYOFFS AND NOTATION

discount factors,

pit =

1 Et (xit+1 ). Ri

Here, I have added the i superscript to emphasize that each risky asset i must be discounted by an asset-specific risk-adjusted discount factor 1/Ri . In this context, equation (1.4) is obviously a generalization, and it says something deep: one can incorporate all risk-corrections by defining a single stochastic discount factor – the same one for each asset – and putting it inside the expectation. mt+1 is stochastic or random because it is not known with certainty at time t. As we will see, the correlation between the random components of m and xi generate asset-specific risk corrections. mt+1 is also often called the marginal rate of substitution after (1.3). In that equation, mt+1 is the rate at which the investor is willing to substitute consumption at time t + 1 for consumption at time t. mt+1 is sometimes also called the pricing kernel. If you know what a kernel is and express the expectation as an integral, you can see where the name comes from. It is sometimes called a change of measure or a state-price density for reasons that we will see below. For the moment, introducing the discount factor m and breaking the basic pricing equation (1.2) into (1.3) and (1.4) is just a notational convenience. As we will see, however, it represents a much deeper and more useful separation. For example, notice that p = E(mx) would still be valid if we changed the utility function, but we would have a different function connecting m to data. As we will see, all asset pricing models amount to alternative models connecting the stochastic discount factor to data, while p = E(mx) is a convenient accounting identity with almost no content. At the same time, we will study lots of alternative expressions of p = E(mx), and we can summarize many empirical approaches to p = E(mx). By separating our models into these two components, we don’t have to redo all that elaboration for each asset pricing model.

1.3

Prices, payoffs and notation

The price pt gives rights to a payoff xt+1 . In practice, this notation covers a variety of cases, including the following: 17

C HAPTER 1

C ONSUMPTION - BASED MODEL AND OVERVIEW

Stock Return Price-dividend ratio Excess return Managed portfolio Moment condition One-period bond Risk free rate Option

Price pt pt 1 pt dt

0 zt E(pt zt ) pt 1 C

Payoff xt+1 pt+1 + dt+1 Rt+1´ ³ pt+1 dt+1 dt+1 + 1 dt b Ret+1 = Rat+1 − Rt+1 zt Rt+1 xt+1 zt 1 Rf max(ST − K, 0)

The price pt and payoff xt+1 seem like a very restrictive kind of security. In fact, this notation is quite general and allows us easily to accommodate many different asset pricing questions. In particular, we can cover stocks, bonds and options and make clear that there is one theory for all asset pricing. For stocks, the one period payoff is of course the next price plus dividend, xt+1 = pt+1 + dt+1 . We frequently divide the payoff xt+1 by the price pt to obtain a gross return Rt+1 ≡

xt+1 pt

We can think of a return as a payoff with price one. If you pay one dollar today, the return is how many dollars or units of consumption you get tomorrow. Thus, returns obey 1 = E(mR) which is by far the most important special case of the basic formula p = E(mx). I use capital letters to denote gross returns R, which have a numerical value like 1.05. I use lowercase letters to denote net returns r = R − 1 or log (continuously compounded) returns ln(R), both of which have numerical values like 0.05. One may also quote percent returns 100 × r.

Returns are often used in empirical work because they are typically stationary over time. (Stationary in the statistical sense; they don’t have trends and you can meaningfully take an average. “Stationary” does not mean constant.) However, thinking in terms of returns takes us away from the central task of finding asset prices. Dividing by dividends and creating a payoff µ ¶ pt+1 dt+1 xt+1 = 1 + dt+1 dt

corresponding to a price pt /dt is a way to look at prices but still to examine stationary variables. Not everything can be reduced to a return. If you borrow a dollar at the interest rate Rf and invest it in an asset with return R, you pay no money out-of-pocket today, and get the 18

S ECTION 1.4

C LASSIC ISSUES IN FINANCE

payoff R − Rf . This is a payoff with a zero price, so you obviously can’t divide payoff by price to get a return. Zero price does not imply zero payoff. It is a bet in which the chance of losing exactly balances its chance of winning, so that it is not worth paying extra to take the bet. It is common to study equity strategies in which one short sells one stock or portfolio and invests the proceeds in another stock or portfolio, generating an excess return. I denote any such difference between returns as an excess return, Re . It is also called a zero-cost portfolio or a self-financing portfolio. In fact, much asset pricing focuses on excess returns. Our economic understanding of interest rate variation turns out to have little to do with our understanding of risk premia, so it is convenient to separate the two exercises by looking at interest rates and excess returns separately. We also want to think about the managed portfolios, in which one invests more or less in an asset according to some signal. The “price” of such a strategy is the amount invested at time t, say zt , and the payoff is zt Rt+1 . For example a market timing strategy might put a weight in stocks proportional to the price-dividend ratio, investing less when prices are higher. We could represent such a strategy as a payoff using zt = a − b(pt /dt ).

When we think about conditioning information below, we will think of objects like zt as instruments. Then we take an unconditional expectation of pt zt = Et (mt+1 xt+1 )zt , yielding E(pt zt ) = E(mt+1 xt+1 zt ). We can think of this operation as creating a “security” with payoff xt+1 zt+1 , and “price” E(pt zt ) represented with unconditional expectations.

A one period bond is of course a claim to a unit payoff. Bonds, options, investment projects are all examples in which it is often more useful to think of prices and payoffs rather than returns. Prices and returns can be real (denominated in goods) or nominal (denominated in dollars); p = E(mx) can refer to either case. The only difference is whether we use a real or nominal discount factor. If prices, returns and payoffs are nominal, we should use a nominal discount factor. For example, if p and x denote nominal values, then we can create real prices and payoffs to write pt = Et Πt

·µ 0 ¸ ¶ u (ct+1 ) xt+1 β 0 u (ct ) Πt+1

where Π denotes the price level (cpi). Obviously, this is the same as defining a nominal discount factor by ·µ 0 ¶ ¸ u (ct+1 ) Πt pt = Et β 0 xt+1 u (ct ) Πt+1 To accommodate all these cases, I will simply use the notation price pt and payoff xt+1 . These symbols can denote 0, 1, or zt and Rte , rt+1 , or zt Rt+1 respectively, according to the case. Lots of other definitions of p and x are useful as well. 19

C HAPTER 1

1.4

C ONSUMPTION - BASED MODEL AND OVERVIEW

Classic issues in finance

I use simple manipulations of the basic pricing equation to introduce classic issues in finance: the economics of interest rates, risk adjustments, systematic vs. idiosyncratic risk, expected return-beta representations, the mean-variance frontier, the slope of the mean-variance frontier, time-varying expected returns, and present value relations. A few simple rearrangements and manipulations of the basic pricing equation p = E(mx) give a lot of intuition and introduce some classic issues in finance, including determinants of the interest rate, risk corrections, idiosyncratic vs. systematic risk, beta pricing models, and mean variance frontiers. 1.4.1

Risk free rate

Rf = 1/E(m). With lognormal consumption growth, rtf = δ + γEt ∆ ln ct+1 −

γ2 2 σ (∆ ln ct+1 ) 2 t

Real interest rates are high when people are impatient (δ), when expected consumption growth is high (intertemporal substitution), or when risk is low (precautionary saving). A more curved utility function (γ) or a lower elasticity of intertemporal substitution (1/γ) means that interest rates are more sensitive to changes in expected consumption growth. The risk free rate is given by Rf = 1/E(m).

(6)

The risk free rate is known ahead of time, so p = E(mx) becomes 1 = E(mRf ) = E(m)Rf . If a risk free security is not traded, we can define Rf = 1/E(m) as the “shadow” risk-free rate. (In some models it is called the “zero-beta” rate.) If one introduced a risk free security with return Rf = 1/E(m), investors would be just indifferent to buying or selling it. I use Rf to simplify formulas below with this understanding. To think about the economics behind real interest rates in a simple setup, use power utility 20

S ECTION 1.4

C LASSIC ISSUES IN FINANCE

u0 (c) = c−γ . Start by turning off uncertainty, in which case Rf =

1 β

µ

ct+1 ct

¶γ

.

We can see three effects right away: 1. 2.

3.

Real interest rates are high when people are impatient, when β is low. If everyone wants to consume now, it takes a high interest rate to convince them to save. Real interest rates are high when consumption growth is high. In times of high interest rates, it pays investors to consume less now, invest more, and consume more in the future. Thus, high interest rates lower the level of consumption today, while raising its growth rate from today to tomorrow. Real interest rates are more sensitive to consumption growth if the power parameter γ is large. If utility is highly curved, the investor cares more about maintaining a consumption profile that is smooth over time, and is less willing to rearrange consumption over time in response to interest rate incentives. Thus it takes a larger interest rate change to induce him to a given consumption growth. To understand how interest rates behave when there is some uncertainty, I specify that consumption growth is lognormally distributed. In this case, the real riskfree rate equation becomes rtf = δ + γEt ∆ ln ct+1 −

γ2 2 σ (∆ ln ct+1 ) 2 t

where I have defined the log riskfree rate rtf and subjective discount rate δ by rtf = ln Rtf ; β = e−δ , and ∆ denotes the first difference operator, ∆ ln ct+1 = ln ct+1 − ln ct . To derive expression (1.7) for the riskfree rate, start with " µ ¶−γ # ct+1 = 1/Et β . ct

Rtf Using the fact that normal z means

1

E (ez ) = eE(z)+ 2 σ 21

2

(z)

(7)

C HAPTER 1

C ONSUMPTION - BASED MODEL AND OVERVIEW

(you can check this by writing out the integral that defines the expectation), we have · ¸−1 γ2 2 . Rft = e−δ e−γEt (∆ ln ct+1 )+ 2 σt (∆ ln ct+1 ) Then take logarithms. The combination of lognormal distributions and power utility is one of the basic tricks to getting analytical solutions in this kind of model. Section 1.5 shows how to get the same result in continuous time. Looking at (1.7), we see the same results as we had with the deterministic case. Real interest rates are high when impatience δ is high and when consumption growth is high; higher γ makes interest rates more sensitive to consumption growth. The new σ2 term captures precautionary savings. When consumption is more volatile, people with this utility function are more worried about the low consumption states than they are pleased by the high consumption states. Therefore, people want to save more, driving down interest rates. We can also read the same terms backwards: consumption growth is high when real interest rates are high, since people save more now and spend it in the future, and consumption is less sensitive to interest rates as the desire for a smooth consumption stream, captured by γ, rises. . Section 2.2 below takes up the question of which way we should read this equation – as consumption determining interest rates, or as interest rates determining consumption. For the power utility function, the curvature parameter γ simultaneously controls intertemporal substitution – aversion to a consumption stream that varies over time, risk aversion – aversion to a consumption stream that varies across states of nature, and precautionary savings, which turns out to depend on the third derivative of the utility function. This link is particular to the power utility function. We will study utility functions below that loosen the links between these three quantities. 1.4.2

Risk corrections

p=

E(x) + cov(m, x) Rf

¢ ¡ E(Ri ) − Rf = −Rf cov m, Ri .

Payoffs that are positively correlated with consumption growth have lower prices, to compensate investors for risk. Expected returns are proportional to the covariance of returns with discount factors. Using the definition of covariance cov(m, x) = E(mx) − E(m)E(x), we can write 22

S ECTION 1.4

C LASSIC ISSUES IN FINANCE

equation (1.2) as p = E(m)E(x) + cov(m, x).

(8)

Substituting the riskfree rate equation (1.6), we obtain p=

E(x) + cov(m, x) Rf

(9)

The first term in (1.9) is the standard discounted present value formula. This is the asset’s price in a risk-neutral world – if consumption is constant or if utility is linear. The second term is a risk adjustment. An asset whose payoff covaries positively with the discount factor has its price raised and vice-versa. To understand the risk adjustment, substitute back for m in terms of consumption, to obtain p=

E(x) cov [βu0 (ct+1 ), xt+1 ] + Rf u0 (ct )

(10)

Marginal utility u0 (c) declines as c rises. Thus, an asset’s price is lowered if its payoff covaries positively with consumption. Conversely, an asset’s price is raised if it covaries negatively with consumption. Why? Investors do not like uncertainty about consumption. If you buy an asset whose payoff covaries positively with consumption, one that pays off well when you are already feeling wealthy, and pays off badly when you are already feeling poor, that asset will make your consumption stream more volatile. You will require a low price to induce you to buy such an asset. If you buy an asset whose payoff covaries negatively with consumption, it helps to smooth consumption and so is more valuable than its expected payoff might indicate. Insurance is an extreme example. Insurance pays off exactly when wealth and consumption would otherwise be low–you get a check when your house burns down. For this reason, you are happy to hold insurance, even though you expect to lose money—even though the price of insurance is greater than its expected payoff discounted at the risk free rate. To emphasize why the covariance of a payoff with the discount factor rather than its variance determines its riskiness, keep in mind that the investor cares about the volatility of consumption. He does not care about the volatility of his individual assets or of his portfolio, if he can keep a steady consumption. Consider what happens to the volatility of consumption if the investor buys a little more ξ of payoff x: σ2 (c) becomes σ 2 (c + ξx) = σ 2 (c) + 2ξcov(c, x) + ξ 2 σ2 (x) For small (marginal) portfolio changes, the covariance between consumption and payoff determines the effect of adding a bit more of each payoff on the volatility of consumption. We use returns so often that it is worth restating the same intuition for the special case that 23

C HAPTER 1

C ONSUMPTION - BASED MODEL AND OVERVIEW

the price is one and the payoff is a return. Start with the basic pricing equation for returns, 1 = E(mRi ). I denote the return Ri to emphasize that the point of the theory is to distinguish the behavior of one asset Ri from another Rj . The asset pricing model says that, although expected returns can vary across time and assets, expected discounted returns should always be the same, 1. Applying the covariance decomposition, 1 = E(m)E(Ri ) + cov(m, Ri )

(11)

E(Ri ) − Rf = −Rf cov(m, Ri )

(12)

and, using Rf = 1/E(m),

or E(Ri ) − Rf = −

cov[u0 (ct+1 ), Rit+1 ] . E[u0 (ct+1 )]

(13)

All assets have an expected return equal to the risk-free rate, plus a risk adjustment. Assets whose returns covary positively with consumption make consumption more volatile, and so must promise higher expected returns to induce investors to hold them. Conversely, assets that covary negatively with consumption, such as insurance, can offer expected rates of return that are lower than the risk-free rate, or even negative (net) expected returns. Much of finance focuses on expected returns. We think of expected returns increasing or decreasing to clear markets; we offer intuition that “riskier” securities must offer higher expected returns to get investors to hold them, rather than saying “riskier” securities trade for lower prices so that investors will hold them. Of course, a low initial price for a given payoff corresponds to a high expected return, so this is no more than a different language for the same phenomenon. 1.4.3

Idiosyncratic risk does not affect prices

Only the component of a payoff perfectly correlated with the discount factor generates an extra return. Idiosyncratic risk, uncorrelated with the discount factor, generates no premium. You might think that an asset with a high payoff variance is “risky” and thus should have a large risk correction. However, if the payoff is uncorrelated with the discount factor m, the asset receives no risk-correction to its price, and pays an expected return equal to the risk-free 24

S ECTION 1.4

C LASSIC ISSUES IN FINANCE

rate! In equations, if cov(m, x) = 0 then p=

E(x) . Rf

This prediction holds even if the payoff x is highly volatile and investors are highly risk averse. The reason is simple: if you buy a little bit more of such an asset, it has no first-order effect on the variance of your consumption stream. More generally, one gets no compensation or risk adjustment for holding idiosyncratic risk. Only systematic risk generates a risk correction. To give meaning to these words, we can decompose any payoff x into a part correlated with the discount factor and an idiosyncratic part uncorrelated with the discount factor by running a regression, x = proj(x|m) + ε. Then, the price of the residual or idiosyncratic risk ε is zero, and the price of x is the same as the price of its projection on m. The projection of x on m is of course that part of x which is perfectly correlated with m. The idiosyncratic component of any payoff is that part uncorrelated with m. Thus only the systematic part of a payoff accounts for its price. Projection means linear regression without a constant, proj(x|m) =

E(mx) m. E(m2 )

You can verify that regression residuals are orthogonal to right hand variables E(mε) = 0 from this definition. E(mε) = 0 of course means that the price of ε is zero. µ ¶ µ ¶ E(mx) 2 E(mx) p (proj(x|m)) = p m = E m = E(mx) = p(x). E(m2 ) E(m2 ) The words “systematic” and “idiosyncratic” are defined differently in different contexts, which can lead to some confusion. In this decomposition, the residuals ε can be correlated with each other, though they are not correlated with the discount factor. The APT starts with a factor-analytic decomposition of the covariance of payoffs, and the word “idiosyncratic” there is reserved for the component of payoffs uncorrelated with all of the other payoffs. 1.4.4

Expected return-beta representation

25

C HAPTER 1

C ONSUMPTION - BASED MODEL AND OVERVIEW

We can write p = E(mx) as E(Ri ) = Rf + β i,m λm

or

We can express the expected return equation (1.12), for a return Ri , as µ ¶µ ¶ cov(Ri , m) var(m) i f E(R ) = R + − var(m) E(m) E(Ri ) = Rf + β i,m λm

(14)

(15)

where β im is the regression coefficient of the return Ri on m. This is a beta pricing model. It says that expected returns on assets i = 1, 2, ...N should be proportional to their betas in a regression of returns on the discount factor. Notice that the coefficient λm is the same for all assets i,while the β i,m varies from asset to asset. The λm is often interpreted as the price of risk and the β as the quantity of risk in each asset. Obviously, there is nothing deep about saying that expected returns are proportional to betas rather than to covariances. There is a long historical tradition and some minor convenience in favor of betas. The betas refer to the projection of R on m that we studied above, so you see again a sense in which only the systematic component of risk matters. −γ With m = β (ct+1 /ct ) , we can take a Taylor approximation of equation (1.14) to express betas in terms of a more concrete variable, consumption growth, rather than marginal utility. The result, which I derive more explicitly and conveniently in the continuous time limit below, is

E(Ri ) = Rf + β i,∆c λ∆c λ∆c = γvar(∆c).

(1.16)

Expected returns should increase linearly with their betas on consumption growth itself. In addition, though it is treated as a free parameter in many applications, the factor risk premium λ∆c is determined by risk aversion and the volatility of consumption. The more risk averse people are, or the riskier their environment, the larger an expected return premium one must pay to get investors to hold risky (high beta) assets. 1.4.5

Mean-variance frontier

All asset returns lie inside a mean-variance frontier. Assets on the frontier are perfectly correlated with each other and the discount factor. Returns on the frontier can be generated as portfolios of any two frontier returns. We can construct a discount factor from any frontier 26

S ECTION 1.4

C LASSIC ISSUES IN FINANCE

return (except Rf ), and an expected return-beta representation holds using any frontier return (except Rf ) as the factor. Asset pricing theory has focused a lot on the means and variances of asset returns. Interestingly, the set of means and variances of returns is limited. All assets priced by the discount factor m must obey ¯ ¯ ¯E(Ri ) − Rf ¯ ≤ σ(m) σ(Ri ). E(m)

(17)

To derive (1.17) write for a given asset return Ri

1 = E(mRi ) = E(m)E(Ri ) + ρm,Ri σ(Ri )σ(m) and hence E(Ri ) = Rf − ρm,Ri

σ(m) σ(Ri ). E(m)

(18)

Correlation coefficients can’t be greater than one in magnitude, leading to (1.17). This simple calculation has many interesting and classic implications. 1. Means and variances of asset returns must lie in the wedge-shaped region illustrated in Figure 1. The boundary of the mean-variance region in which assets can lie is called the mean-variance frontier. It answers a naturally interesting question, “how much mean return can you get for a given level of variance?” 2. All returns ¯ frontier are perfectly correlated with the discount factor: the frontier ¯ on the is generated by ¯ρm,Ri ¯ = 1. Returns on the upper part of the frontier are perfectly negatively correlated with the discount factor and hence positively correlated with consumption. They are “maximally risky” and thus get the highest expected returns. Returns on the lower part of the frontier are perfectly positively correlated with the discount factor and hence negatively correlated with consumption. They thus provide the best insurance against consumption fluctuations. 3. All frontier returns are also perfectly correlated with each other, since they are all perfectly correlated with the discount factor. This fact implies that we can span or synthesize any frontier return from two such returns. For example if you pick any single frontier return Rm then all frontier returns Rmv must be expressible as

for some number a.

¢ ¡ Rmv = Rf + a Rm − Rf

4. Since each point on the mean-variance frontier is perfectly correlated with the discount 27

C HAPTER 1

C ONSUMPTION - BASED MODEL AND OVERVIEW

E(R)

Mean-variance frontier Slope σ(m)/E(m) Idiosyncratic risk Ri Rf Some asset returns

σ(R)

Figure 1. Mean-variance frontier. The mean and standard deviation of all assets priced by a discount factor m must line in the wedge-shaped region factor, we must be able to pick constants a, b, d, e such that m = a + bRmv Rmv = d + em. Thus, any mean-variance efficient return carries all pricing information. Given a meanvariance efficient return and the risk free rate, we can find a discount factor that prices all assets and vice versa. 5. Given a discount factor, we can also construct a single-beta representation, so expected returns can be described in a single - beta representation using any mean-variance efficient return (except the riskfree rate), £ ¤ E(Ri ) = Rf + β i,mv E(Rmv ) − Rf .

The essence of the β pricing model is that, even though the means and standard deviations of returns fill out the space inside the mean-variance frontier, a graph of mean returns versus betas should yield a straight line. Since the beta model applies to every return including Rmv itself, and Rmv has a beta of one on itself, we can identify the factor risk premium as λ = E(Rmv − Rf ). The last two points suggest an intimate relationship between discount factors, beta models and mean-variance frontiers. I explore this relation in detail in Chapter 6. A problem at the end of this chapter guides you through the algebra to demonstrate points 4 and 5 explicitly. 28

S ECTION 1.4

C LASSIC ISSUES IN FINANCE

6. We can plot the decomposition of a return into a “priced” or “systematic” component and a “residual,” or “idiosyncratic” component as shown in Figure 1. The priced part is perfectly correlated with the discount factor, and hence perfectly correlated with any frontier asset. The residual or idiosyncratic part generates no expected return, so it lies flat as shown in the figure, and it is uncorrelated with the discount factor or any frontier asset.. 1.4.6

Slope of the mean-standard deviation frontier and equity premium puzzle

The Sharpe ratio is limited by the volatility of the discount factor. The maximal risk-return tradeoff is steeper if there is more risk or more risk aversion ¯ ¯ ¯ E(R) − Rf ¯ σ(m) ¯ ¯ ¯ σ(R) ¯ ≤ E(m) ≈ γσ(∆ ln c) This formula captures the equity premium puzzle, which suggests that either people are very risk averse, or the stock returns of the last 50 years were good luck which will not continue. The ratio of mean excess return to standard deviation E(Ri ) − Rf = Sharpe ratio σ(Ri ) is known as the Sharpe ratio. It is a more interesting characterization of any security than the mean return alone. If you borrow and put more money into a security, you can increase the mean return of your position, but you do not increase the Sharpe ratio, since the standard deviation increases at the same rate as the mean. The slope of the mean-standard deviation frontier is the largest available Sharpe ratio, and thus is naturally interesting. It answers “how much more mean return can I get by shouldering a bit more volatility in my portfolio?” Let Rmv denote the return of a portfolio on the frontier. From equation (1.17), the slope of the frontier is ¯ ¯ ¯ E(Rmv ) − Rf ¯ σ(m) f ¯ ¯ ¯ σ(Rmv ) ¯ = E(m) = σ(m)R . Thus, the slope of the frontier is governed by the volatility of the discount factor.

For an economic interpretation, again consider the power utility function, u0 (c) = c−γ , ¯ ¯ ¯ E(Rmv ) − Rf ¯ σ [(ct+1 /ct )−γ ] ¯ ¯ i. (19) ¯ σ(Rmv ) ¯ = h E (ct+1 /ct )−γ

The standard deviation is large if consumption is volatile or if γ is large. We can state this 29

C HAPTER 1

C ONSUMPTION - BASED MODEL AND OVERVIEW

approximation again using the lognormal assumption. If consumption growth is lognormal, ¯ ¯ ¯ E(Rmv ) − Rf ¯ p 2 2 γ σ (∆ ln ct+1 ) − 1 ≈ γσ(∆ ln c). ¯ ¯ ¯ σ(Rmv ) ¯ = e

(20)

(A problem at the end of the chapter guides you though the algebra of the first equality. The relation is exact in continuous time, and thus the approximation is easiest to derive by reference to the continuous time result; see section 1.5.) Reading the equation, the slope of the mean-standard deviation frontier is higher if the economy is riskier – if consumption is more volatile – or if investors are more risk averse. Both situations naturally make investors more reluctant to take on the extra risk of holding risky assets. This expression is also the slope of the expected return beta line of the consumption beta model, (1.16). (Or, conversely, in an economy with a high Sharpe ratio, low risk aversion investors should take on so much risk that their consumption becomes volatile.) In postwar US data, the slope of the mean-standard deviation frontier, or of expected return-beta lines is much higher than reasonable risk aversion and consumption volatility estimates suggest. This is the “equity premium puzzle.” Over the last 50 years in the U.S., real stock returns have averaged 9% with a standard deviation of about 16%, while the real return on treasury bills has been about 1%. Thus, the historical annual market Sharpe ratio has been about 0.5. Aggregate consumption growth has been about 1%. Thus, we can only reconcile these facts with (1.20) if investors have a risk aversion coefficient of 50! Obvious ways of generalizing the calculation just make matters worse. Equation (1.20) relates consumption growth to the mean-variance frontier of all contingent claims. The market indices with 0.5 Sharpe ratios are if anything inside that frontier, so recognizing market incompleteness will only make matters worse. Aggregate consumption has about 0.2 correlation with the market return, while the equality (1.20) takes the worst possible case that consumption growth and asset returns are perfectly correlated. If you add this fact, you need risk aversion of 250 to explain the market Sharpe ratio in the face of 1% consumption volatility! Individuals have riskier consumption streams than aggregate, but as their risk goes up their correlation with any aggregate must decrease proportionally, so to first order recognizing individual risk will not help either. Clearly, either 1) people are a lot more risk averse than we might have thought 2) the stock returns of the last 50 years were largely good luck rather than an equilibrium compensation for risk, or 3) something is deeply wrong with the model, including the utility function and use of aggregate consumption data. This “equity premium puzzle” has attracted the attention of a lot of research in finance, especially on the last item. I return to the equity premium in more detail in Chapter 21. 1.4.7

Random walks and time-varying expected returns

30

S ECTION 1.4

C LASSIC ISSUES IN FINANCE

If investors are risk neutral, returns are unpredictable, and prices follow martingales. In general, prices scaled by marginal utility are martingales, and returns can be predictable if investors are risk averse and if the conditional second moments of returns and discount factors vary over time. This is more plausible at long horizons. So far, we have concentrated on the behavior of prices or expected returns across assets. We should also consider the behavior of the price or return of a given asset over time. Going back to the basic first order condition, pt u0 (ct ) = Et [βu0 (ct+1 )(pt+1 + dt+1 )].

(21)

If investors are risk neutral, i.e. if u(c) is linear or there is no variation in consumption, if the security pays no dividends between t and t + 1, and for short time horizons where β is close to one, this equation reduces to pt = Et (pt+1 ). Equivalently, prices follow a time-series process of the form pt+1 = pt + εt+1 . If the variance σ2t (εt+1 ) is constant, prices follow a random walk. More generally, prices follow a martingale. Intuitively, if the price today is a lot lower than investor’s expectations of the price tomorrow, then people will try to buy the security. But this action will drive up the price of the security until the price today does equal the expected price tomorrow. Another way of saying the same thing is that returns should not be predictable; dividing by pt , expected returns Et (pt+1 /pt ) = 1 should be constant; returns should be like coin flips. The more general equation (1.21) says that prices should follow a martingale after adjusting for dividends and scaling by marginal utility. Since martingales have useful mathematical properties, and since risk-neutrality is such a simple economic environment, many asset pricing results are easily derived by scaling prices and dividends by marginal utility first, and then using “risk-neutral” formulas and economic arguments. Since consumption and risk aversion don’t change much day to day, we might expect the random walk view to hold pretty well on a day-to-day basis. This idea contradicts the still popular notion that there are “systems” or “technical analysis” by which one can predict where stock prices are going on any given day. It has been remarkably successful. Despite decades of dredging the data, and the popularity of television and radio reports that purport to explain where markets are going, trading rules that reliably survive transactions costs and do not implicitly expose the investor to risk have not yet been reliably demonstrated. However, more recently, evidence has accumulated that long-horizon excess returns are quite predictable, and to some this indicates that the whole enterprise of economic explanation of asset returns is flawed. To think about this issue, write our basic equation for expected 31

C HAPTER 1

C ONSUMPTION - BASED MODEL AND OVERVIEW

returns as Et (Rt+1 ) − Rtf

covt (mt+1 , Rt+1 ) Et (mt+1 ) σt (mt+1 ) σt (Rt+1 )ρt (mt+1 , Rt+1 ) = Et (mt+1 ) ≈ γ t σ t (∆ct+1 )σt (Rt+1 )ρt (mt+1 , Rt+1 ). = −

(1.22)

I include the t subscripts to emphasize that the relation applies to conditional moments. Sometimes, the conditional mean or other moment of a random variable is different from its unconditional moment. Conditional on tonight’s weather forecast, you can better predict rain tomorrow than just knowing the average rain for that date. In the special case that random variables are i.i.d. (independent and identically distributed), like coin flips, the conditional and unconditional moments are the same, but that is a special case and not likely to be true of asset prices, returns, and macroeconomic variables. In the theory so far, we have thought of an investor, today, forming expectations of payoffs, consumption, and other variables tomorrow. Thus, the moments are really all conditional, and if we want to be precise we should include some notation to express this fact. I use subscripts Et (xt+1 ) to denote conditional expectation; the notation E(xt+1 |It ) where It is the information set at time t is more precise but a little more cumbersome. Examining equation (1.22), we see that returns can be somewhat predictable. First, if the conditional variance of returns changes over time, we might expect the conditional mean return to vary as well – the return can just move in and out a line of constant Sharpe ratio. This explanation does not seem to help much in the data; variables that forecast means do not seem to forecast variances and vice versa. Unless we want to probe the conditional correlation, predictable excess returns have to be explained by changing risk – σt (∆ct+1 ) – or changing risk aversion γ. It is not plausible that risk or risk aversion change at daily frequencies, but fortunately returns are not predictable at daily frequencies. It is much more plausible that risk and risk aversion change over the business cycle, and this is exactly the horizon at which we see predictable excess returns. Models that make this connection precise are a very active area of current research. 1.4.8

Present value statement

pt = Et

∞ X

mt,t+j dt+j .

j=0

It is convenient to use only the two period valuation, thinking of a price pt and a payoff 32

S ECTION 1.5

D ISCOUNT FACTORS IN CONTINUOUS TIME

xt+1 . But there are times when we want to relate a price to the entire cash flow stream, rather than just to one dividend and next period’s price. The most straightforward way to do this is to write out a longer term objective, Et

∞ X

β j u(ct+j ).

j=0

Now suppose an investor can purchase a stream {dt+j } at price pt . As with the two-period model, his first order condition gives us the pricing formula directly, pt = Et

∞ X

βj

j=0

∞ X u0 (ct+j ) d = E mt,t+j dt+j . t+j t u0 (ct ) j=0

(23)

You can see that if this equation holds at time t and time t + 1, then we can derive the two-period version pt = Et [mt+1 (pt+1 + dt+1 )]

(24)

Thus, the infinite period and two period models are equivalent. (Going in the other direction is a little tougher. If you chain together (1.24), you get (1.23) plus an extra term. To get (1.23) you also need the “transversality condition” limt→∞ Et mt,t+j pt+j = 0. This is an extra first order condition of the infinite period investor, which is not present with overlapping generations of two-period investors. It rules out “bubbles” in which prices grow so fast that people will buy now just to resell at higher prices later, even if there are no dividends.) From (1.23) we can write a risk-adjustment to prices, as we did with one period payoffs, pt =

∞ X Et dt+j j=1

Rft,t+j

+

∞ X

covt (dt+j , mt,t+j )

j=1

f where Rt,t+j ≡ Et (mt,t+j )−1 is the j period interest rate. Again, assets whose dividend streams covary negatively with marginal utility, and positively with consumption, have lower prices, since holding those assets gives the investor a more volatile consumption stream. (It is commonP instead to write prices as a discounted value using a risk adjusted discount factor, ∞ e.g. pt = j=1 Et dt+j /Rt,t+j but this approach is difficult to use correctly for multiperiod problems, especially when expected returns can vary over time.)

1.5

Discount factors in continuous time

Continuous time versions of the basic pricing equations. 33

C HAPTER 1

C ONSUMPTION - BASED MODEL AND OVERVIEW

Discrete 0 P j u (ct+j ) pt = Et ∞ j=1 β u0 (ct ) Dt+j 0

t+1 ) mt+1 = β uu(c 0 (c ) t p = E(mx)

E(R) = Rf − Rf cov(m, R)

Continuous R∞ pt u0 (ct ) = Et s=0 e−δs u0 (ct+s )Dt+s ds Λt = e−δt u0 (ct ) 0 =³ΛD´ dt + Et [d(Λp)] h i f D dΛ dp + Et dp dt = r dt − E t t p p Λ p

It is often convenient to express asset pricing ideas in the language of continuous time stochastic differential equations rather than discrete time stochastic difference equations as I have done so far. The appendix contains a brief introduction to continuous time processes that covers what you need to know for this book. Even if you want to end up with a discrete time representation, manipulations are often easier in continuous time. For example, relating interest rates and Sharpe ratios to consumption growth in the last section required a clumsy lognormal approximation; you’ll see the same sort of thing done much more cleanly in this section. The choice of discrete vs. continuous time is one of modeling convenience. The richness of the theory of continuous time processes often allows one to obtain analytical results that would be unavailable in discrete time. On the other hand, in the complexity of most practical situations, one often ends up resorting to numerical simulation of a discretized model anyway. In those cases, it might be clearer to start with a discrete model. But I emphasize this is all a choice of language. One should become familiar enough with discrete as well as continuous time representations of the same ideas to pick the representation that is most convenient for a particular application. First, we need to think about how to model securities, in place of price pt and one-period payoff xt+1 . Let a generic security have price pt at any moment in time, and let it pay dividends at the rate Dt dt. (I will continue to denote functions of time as pt rather than p(t) to maintain continuity with the discrete-time treatment, and I will drop the time subscripts where they are obvious, e.g. dp in place of dpt . In an interval dt, the security pays dividends Dt dt. I use capital D for dividends to distinguish them from the differential operator d. ) The instantaneous total return is dpt Dt + dt. pt pt We model the price of risky assets as diffusions, for example dpt = µ(·)dt + σ(·)dz. pt (I will reserve the notation dz for increments to a standard Brownian motion, e.g. zt+∆ − zt ∼ N (0, ∆). I use the notation (·) to indicate that the drift and diffusions can be functions of state variables. I limit the discussion to diffusion processes – no jumps.) What’s nice about this diffusion model is that the increments dz are normal; the dependence of µ and σ on state 34

S ECTION 1.5

D ISCOUNT FACTORS IN CONTINUOUS TIME

variables means that the finite time distribution of prices f (pt+∆ |It ) need not be normal.

We can think of a riskfree security as one that has a constant price equal to one and pays the riskfree rate as a dividend, p = 1; Dt = rtf ,

(25)

or as a security that pays no dividend but whose price climbs deterministically at a rate dpt = rtf dt. pt

(26)

Next, we need to express the first order conditions in continuous time. The utility function is U ({ct }) = E

Z

∞

e−δt u(ct )dt.

t=0

Suppose the investor can buy a security whose price is pt and that pays a dividend stream Dt . As we did in deriving the present value price relation in discrete time, the first order condition for this problem gives us the infinite period version of the basic pricing equation right away1 , Z ∞ e−δs u0 (ct+s )Dt+s ds (27) pt u0 (ct ) = Et s=0

This equation is an obvious continuous time analogue to pt = Et

∞ X j=0

βt

u0 (ct+j ) Dt+j . u0 (ct )

It turns out that dividing by u0 (ct ) is not a good idea in continuous time, since the ratio u (ct+∆ )/u0 (ct ) isn’t well behaved for small time intervals. Instead, we keep track of the level of marginal utility. Therefore, define the “discount factor” in continuous time as 0

Λt ≡ e−δt u0 (ct ). Then we can write the pricing equation as pt Λt = Et

Z

∞

Λt+s Dt+s ds.

(28)

s=0

1 One unit of the security pays the dividend stream D , i.e. D dt units of the numeraire consumption good in a t t time interval dt. The security costs pt units of the consumption good. The investor can finance the purchase of ξ units of the security by reducing consumption from et to ct = et − ξpt /dt during time interval dt. The loss in utility from doing so is u0 (ct )(et − ct )dt = u0 (ct )ξpt . The gain is the right hand side of (1.27)

35

C HAPTER 1

C ONSUMPTION - BASED MODEL AND OVERVIEW

(Some people like to define Λt = u0 (ct ), in which case you keep the e−δt in the equation, or R f − τs=0 rt+τ dτ in the equation. to scale Λt by the riskfree rate, in which case you get an extra e The latter procedure makes it look like a risk-neutral or present-value formula valuation.) The analogue to the one period pricing equation p = E(mx) is 0 = ΛD dt + Et [d(Λp)] .

(29)

To derive this fundamental equation, take the difference of equation (1.28) at t and t + ∆. (Or, start directly with the first order condition for buying the security at t and selling it at t + ∆.) Z ∆ Λt+s Dt+s ds + Et [Λt+∆ pt+∆ ] pt Λt = Et s=0

For ∆ small the term in the integral can be approximated pt Λt ≈ Λt Dt ∆ + Et [Λt+∆ pt+∆ ] .

(30)

We want to get to d something, so introduce differences by writing pt Λt ≈ Λt Dt ∆ + Et [Λt pt + (Λt+∆ pt+∆ − Λt pt )] .

(31)

Canceling pt Λt , 0 ≈ Λt Dt ∆ + Et (Λt+∆ pt+∆ − Λt pt ). Taking the limit as ∆ → 0, 0 = Λt Dt dt + Et [d(Λt pt )] or, dropping time subscripts, equation (1.29). Equation (1.29) looks different than p = E(mx) because there is no price on the left hand side; we are used to thinking of the one period pricing equation as determining price at t given other things, including price at t + 1. But price at t is really here, of course, as you can see from equation (1.30) or (1.31). It is just easier to express the difference in price over time rather than price today on the left and payoff (including price tomorrow) on the right. With no dividends and constant Λ, 0 = Et (dpt ) = Et (pt+∆ − pt ) says that price should follow a martingale. Thus, Et [d(Λp)] = 0 means that marginal utility-weighted price should follow a martingale, and (1.29) adjusts for dividends. Thus, it’s the same as the equation (1.21), pt u0 (ct ) = Et (mt+1 (pt+1 + dt+1 )) that we derived in discrete time. Since we will write down price processes for dp and discount factor processes for dΛ, and to interpret (1.29) in terms of expected returns, it is often convenient to break up the d(Λt pt ) term using Ito’s lemma: d(Λp) = pdΛ + Λdp + dpdΛ. 36

(32)

S ECTION 1.5

D ISCOUNT FACTORS IN CONTINUOUS TIME

Using the expanded version (1.32) in the basic equation (1.29), and dividing by pΛ to make it pretty, we obtain an equivalent, slightly less compact but slightly more intuitive version, · ¸ D dΛ dp dΛ dp dt + Et + + 0= . (33) p Λ p Λ p (This formula only works when both Λ and p can never be zero. That is often enough the case that this formula is useful. If not, multiply through by Λ and p and keep them in numerators.) Applying the basic pricing equations (1.29) or (1.33) to a riskfree rate, defined as (1.25) or (1.26), we obtain µ ¶ dΛt (34) rtf dt = −Et Λt This equation is the obvious continuous time equivalent to Rtf =

1 . Et (mt+1 )

If a riskfree rate is not traded, we can use (1.34) to define a shadow riskfree rate or zero-beta rate. With this interpretation, we can rearrange equation (1.33) as µ ¶ · ¸ dpt Dt dΛt dpt f dt = rt dt − Et + . Et pt pt Λt pt

(35)

This is the obvious continuous-time analogue to E(R) = Rf − Rf cov(m, R).

(36)

The last term in (1.35) is the covariance of the return with the discount factor or marginal utility. Since means are order dt, there is no difference between covariance and second moment in the last term of (1.35). The interest rate component of the last term of (1.36) naturally vanishes as the time interval gets short. Ito’s lemma makes many transformations simple in continuous time. For example, the nonlinear transformation between consumption and the discount factor led us to some tricky approximations in discrete time. This transformation is easy in continuous time (diffusions are locally normal, so it’s really the same trick). With Λt = e−δt u0 (ct ) we have 1 dΛt = −δe−δt u0 (ct )dt + e−δt u00 (ct )dct + e−δt u000 (ct )dc2t 2 ct u00 (ct ) dct 1 c2t u000 (ct ) dc2t dΛt = −δdt + 0 + Λt u (ct ) ct 2 u0 (ct ) c2t 37

(37)

C HAPTER 1

C ONSUMPTION - BASED MODEL AND OVERVIEW

Denote the local curvature and third derivative of the utility function as ct u00 (ct ) u0 (ct ) c2t u000 (ct ) . u0 (ct )

γt

= −

ηt

=

(For power utility, the former is the power coefficient γ and the latter is ηt = γ(γ + 1).) Using this formula we can quickly redo the relationship between interest rates and consumption growth, equation (1.7), µ ¶ µ ¶ µ 2¶ 1 1 dΛt dct 1 1 dct f = δ + γ t Et − η t Et . rt = − Et dt Λt dt ct 2 dt c2t We can also easily express asset prices in terms of consumption risk rather than discount factor risk, as in equation (1.16). Using (1.37) in (1.35), µ ¶ µ ¶ dpt Dt dct dpt f dt − rt dt = γEt Et + (38) pt pt ct pt Thus, assets whose returns covary more strongly with consumption get higher mean returns, and the constant relating covariance to mean return is the utility curvature coefficient γ. Since correlations are less than one, equation (1.38) implies that Sharpe ratios are related to utility curvature and consumption volatility directly; we don’t need the ugly lognormal facts and an approximation that we needed in (1.20). Using µp ≡ Et (dpt /pt ) ; h i h i σ2p = Et (dpt /pt )2 ; σ 2c = Et (dct /ct )2 , µp +

Dt pt

dt − rtf dt

σp

1.6

≤ γσ c .

Problems

1. (a)

The absolute risk aversion coefficient is u00 (c) . u0 (c) We scale by u0 (c) because expected utility is only defined up to linear transformations – a + bu(c) gives the same predictions as u(c) – and this measure of the second derivative is invariant to linear transformations. Show that the utility 38

S ECTION 1.6

P ROBLEMS

function with constant absolute risk aversion is u(c) = −e−αc . (b)

The coefficient of relative risk aversion in a one-period model (i.e. when consumption equals wealth) is defined as rra =

(c)

For power utility u0 (c) = c−γ , show that the risk aversion coefficient equals the power. The elasticity of intertemporal substitution is defined as ξI ≡ −

2. 3.

c2 /c1 d(c1 /c2 ) . dR/R

Show that with power utility u0 (c) = c−γ , the intertemporal substitution elasticity is equal to 1/γ. Show that the “idiosyncratic risk” line in Figure 1 is horizontal. (a)

(b)

4.

cu00 (c) . u0 (c)

Suppose you have a mean-variance efficient return Rmv and the risk free rate. using the fact that Rmv is perfectly correlated with the discount factor, construct a discount factor m in terms of Rf and Rmv , with no free parameters. (the constants in m = a + bRmv will be functions of things like E(Rmv )) Using this result, and the beta model in terms of m, show that expected returns can be described in a single - beta representation using any mean-variance efficient return (except the riskfree rate). £ ¤ E(Ri ) = Rf + β i,mv E(Rmv ) − Rf .

Can the “Sharpe ratio” between two risky assets exceed the slope of the mean-variance frontier? I.e. if Rmv is on the frontier, is it possible that E(Rmv ) − Rf E(Ri ) − E(Rj ) > ? σ(Ri − Rj ) σ(Rmv )

5.

6.

Show that if consumption growth is lognormal, then ¯ ¯ −γ p ¯ E(Rmv ) − Rf ¯ ¯ ¯ = σ h[(ct+1 /ct ) ]i = eγ 2 σ2 (∆ ln ct+1 ) − 1 ≈ γσ(∆ ln c). ¯ σ(Rmv ) ¯ E (ct+1 /ct )−γ

1

2

(Start with σ 2 (x) = E(x2 ) − E(x)2 and the lognormal property E(ez ) = eEz+ 2 σ (z) .) There are assets with mean return equal to the riskfree rate, but substantial standard deviation of returns. Long term bonds are pretty close examples. Why would anyone 39

C HAPTER 1

7.

C ONSUMPTION - BASED MODEL AND OVERVIEW

hold such an asset? The first order conditions for an infinitely lived consumer who can buy an asset with dividend stream {dt } are pt = Et

∞ X j=1

βj

u0 (ct+j ) dt+j . u0 (ct )

(39)

The first order conditions for buying a security with price pt and payoff xt+1 = dt+1 + pt+1 are ¸ · 0 u (ct+1 ) (pt+1 + dt+1 ) pt = Et β 0 (40) u (ct )

(a) (b)

8.

9.

Derive (1.40) from (1.39) Derive (1.39) from (1.40). You need an extra condition. Show that this extra condition is a first order condition for maximization. To do this, think about what strategy the consumer could follow to improve utility if the condition did not hold. Suppose a consumer has a utility function that includes leisure. (This could also be a second good, or a good produced in another country.) Using the continuous time setup, show that expected returns will now depend on two covariances, the covariance of returns with leisure and the covariance of returns with consumption, so long as leisure enters non-separably, i.e. u(c, l) cannot be written v(c) + w(l). (This is a three line problem, but you need to apply Ito’s lemma to Λ.) From 1 = E(mR) show that the negative of the mean log discount factor must be larger than any mean return, −E(ln m) > E(ln R). How is it possible that E(ln R) is bounded – what about returns of the form R = (1 − α)Rf + αRm for arbitrarily large α? (Hint: start by assuming m and R are lognormal. Then see if you can generalize the results using Jensen’s inequality, E(f (x)) > f(E(x)) for f convex. The return that solves maxR E(ln R) is known as the growth optimal portfolio.)

40

Chapter 2. 2.1

Applying the basic model

Assumptions and applicability

Writing p = E(mx), we do not assume 1. 2. 3. 4. 5.

Markets are complete, or there is a representative investor Asset returns or payoffs are normally distributed (no options), or independent over time. Two period investors, quadratic utility, or separable utility Investors have no human capital or labor income The market has reached equilibrium, or individuals have bought all the securities they want to. All of these assumptions come later, in various special cases, but we haven’t made them yet. We do assume that the investor can consider a small marginal investment or disinvestment. The theory of asset pricing contains lots of assumptions to derive analytically convenient special cases and empirically useful representations. In writing p = E(mx) or pu0 (ct ) = Et [βu0 (ct+1 )xt+1 ] we have not made most of these assumptions. We have not assumed complete markets or a representative investor. These equations apply to each individual investor, for each asset to which he has access, independently of the presence or absence of other investors or other assets. Complete markets/representative agent assumptions are used if one wants to use aggregate consumption data in u0 (ct ), or other specializations and simplifications of the model. We have not said anything about payoff or return distributions. In particular, we have not assumed that returns are normally distributed or that utility is quadratic. The basic pricing equation should hold for any asset, stock, bond, option, real investment opportunity, etc., and any monotone and concave utility function. In particular, it is often thought that mean-variance analysis and beta pricing models require these kind of limiting assumptions or quadratic utility, but that is not the case. A mean-variance efficient return carries all pricing information no matter what the distribution of payoffs, utility function, etc. This is not a “two-period model.” The fundamental pricing equation holds for any two periods of a multi-period model, as we have seen. Really, everything involves conditional moments, so we have not assumed i.i.d. returns over time. I have written things down in terms of a time- and state-separable utility function and I have extensively used the convenient power utility example. Nothing important lies in either 41

C HAPTER 2

A PPLYING THE BASIC MODEL

choice. Just interpret u0 (ct ) as the partial derivative of a general utility function with respect to consumption at time t. State- or time-nonseparable utility (habit persistence, durability) complicates the relation between the discount factor and real variables, but does not change p = E(mx) or any of the basic structure. We will look at several examples below. We do not assume that investors have no non-marketable human capital, or no outside sources of income. The first order conditions for purchase of an asset relative to consumption hold no matter what else is in the budget constraint. By contrast, the portfolio approach to asset pricing as in the CAPM and ICAPM relies heavily on the assumption that the investor has no non-asset income, and we will study these special cases below. For example, leisure in the utility function just means that u0 (c, l) may depend on l as well as c. We don’t even really need the assumption (yet) that the market is “in equilibrium,” that investor has bought all of the asset that he wants to, or even that he can buy the asset at all. We can interpret p = E(mx) as giving us the value, or willingness to pay for, a small amount of a payoff xt+1 that the investor does not yet have. Here’s why: If the investor had a little ξ more of the payoff xt+1 at time t + 1, his utility u(ct ) + βEt u(ct+1 ) would increase by ¸ · 1 βEt [u(ct+1 + ξxt+1 ) − u(ct+1 )] = βEt u0 (ct+1 )xt+1 ξ + u00 (ct+1 ) (xt+1 ξ)2 + ... 2 If ξ is small, only the first term on the right matters. If the investor has to give up a small amount of money vt ξ at time t, that loss lowers his utility by 1 2 u(ct − vt ξ) = u0 (ct )vt ξ + u00 (ct ) (vt ξ) + .... 2 Again, for small ξ, only the first term matters. Therefore, in order to receive the small extra payoff ξxt+1 , the investor is willing to pay the small amount vt ξ where · 0 ¸ u (ct+1 ) vt = Et β 0 xt+1 . u (ct ) If this private valuation is higher than the market value pt , and if the investor can buy some more of the asset, he will. As he buys more, his consumption will change; it will be higher in states where xt+1 is higher, driving down u0 (ct+1 ) in those states, until the value to the investor has declined to equal the market value. Thus, after an investor has reached his optimal portfolio, the market value should obey the basic pricing equation as well, using post-trade or equilibrium consumption. But the formula can also be applied to generate the marginal private valuation, using pre-trade consumption, or to value a potential, not yet traded security. We have calculated the value of a “small” or marginal portfolio change for the investor. For some investment projects, an investor cannot take a small (“diversified”) position. For example, a venture capitalist or entrepreneur must usually take all or nothing P of a project with payoff stream {xt }. Then the value of a project not already taken, E j β j u(ct+j + xt+j ) 42

S ECTION 2.2

G ENERAL E QUILIBRIUM

P might be substantially different from its marginal counterpart, E β j u0 (ct+j )xt+j . Once the project is taken of course, ct+j + xt+j becomes ct+j , so the marginal valuation still applies to the ex-post consumption stream. Analysts often forget this point and apply marginal (diversified) valuation models such as the CAPM to projects that must be bought in discrete chunks. Also, we have abstracted from short sales and bid/ask spreads; this modification changes p = E(mx) from an equality to a set of inequalities.

2.2

General Equilibrium

Asset returns and consumption: which is the chicken and which is the egg? The exogenous return model, the endowment economy model, and the argument that it doesn’t matter for studying p = E(mx). So far, we have not said where the joint statistical properties of the payoff xt+1 and marginal utility mt+1 or consumption ct+1 come from. We have also not said anything about the fundamental exogenous shocks that drive the economy. The basic pricing equation p = E(mx) tells us only what the price should be, given the joint distribution of consumption (marginal utility, discount factor) and the asset payoff. There is nothing that stops us from writing the basic pricing equation as u0 (ct ) = Et [βu0 (ct+1 )xt+1 /pt ] . We can think of this equation as determining today’s consumption given asset prices and payoffs, rather than determining today’s asset price in terms of consumption and payoffs. Thinking about the basic first order condition in this way gives the permanent income model of consumption. Which is the chicken and which is the egg? Which variable is exogenous and which is endogenous? The answer is, neither, and for many purposes, it doesn’t matter. The first order conditions characterize any equilibrium; if you happen to know E(mx), you can use them to determine p; if you happen to know p, you can use them to determine consumption and savings decisions. For most asset pricing applications we are interested in understanding a wide cross-section of assets. Thus, it is interesting to contrast the cross-sectional variation in their prices (expected returns) with cross-sectional variation in their second moments (betas) with a single discount factor. In most applications, the discount factor is a function of aggregate variables (market return, aggregate consumption), so is plausible to hold the properties of the discount factor constant as we study one individual asset after another. Permanent income studies typically dramatically restrict the number of assets under consideration, often to just an interest rate, and study the time-series evolution of aggregate or individual consumption. Nonetheless, it is an obvious next step to complete the solution of our model economy; to 43

C HAPTER 2

A PPLYING THE BASIC MODEL

find c and p in terms of truly exogenous forces. The results will of course depend on what the rest of the economy looks like, in particular the production or intertemporal transformation technology and the set of markets. Figure 2 shows one possibility for a general equilibrium. Suppose that the production technologies are linear: the real, physical rate of return (the rate of intertemporal transformation) is not affected by how much is invested. Now consumption must adjust to these technologically given rates of return. If the rates of return on the intertemporal technologies were to change, the consumption process would have to change as well. This is, implicitly, how the permanent income model works. This is how many finance theories such as the CAPM and ICAPM and the Cox, Ingersoll and Ross (1986) model of the term structure work as well. These models specify the return process, and then solve the consumer’s portfolio and consumption rules.

Ct+1

R Ct Figure 2. Consumption adjusts when the rate of return is determined by a linear technology. Figure 3 shows another extreme possibility for the production technology. This is an “endowment economy.” Nondurable consumption appears (or is produced by labor) every period. There is nothing anyone can do to save, store, invest or otherwise transform consumption goods this period to consumption goods next period. Hence, asset prices must adjust until people are just happy consuming the endowment process. In this case consumption is exogenous and asset prices adjust. Lucas (1978) and Mehra and Prescott (1985) are two very famous applications of this sort of “endowment economy.” Which of these possibilities is correct? Well, neither, of course. The real economy and all serious general equilibrium models look something like figure 4: one can save or transform consumption from one date to the next, but at a decreasing rate. As investment increases, 44

S ECTION 2.2

G ENERAL E QUILIBRIUM

Ct+1

R Ct Figure 3. Asset prices adjust to consumption in an endowment economy. rates of return decline

Ct+1

R Ct Figure 4. General equilibrium. The solid lines represent the indifference curve and production possibility set. The dashed straight line represents the equilibrium rate of return. The dashed box represents an endowment economy that predicts the same consumption-asset return process.

45

C HAPTER 2

A PPLYING THE BASIC MODEL

Does this observation invalidate any modeling we do with the linear technology (CAPM, CIR, permanent income) model, or the endowment economy model? No. Start at the equilibrium in figure 4. Suppose we model this economy as a linear technology, but we happen to choose for the rate of return on the linear technologies exactly the same stochastic process for returns that emerges from the general equilibrium. The resulting joint consumption, asset return process is exactly the same as in the original general equilibrium! Similarly, suppose we model this economy as an endowment economy, but we happen to choose for the endowment process exactly the stochastic process for consumption that emerges from the equilibrium with a concave technology. Again, the joint consumption-asset return process is exactly the same. Therefore, there is nothing wrong in adopting one of the following strategies for empirical work: 1. 2. 3.

Form a statistical model of bond and stock returns, solve the optimal consumptionportfolio decision. Use the equilibrium consumption values in p = E(mx). Form a statistical model of the consumption process, calculate asset prices and returns directly from the basic pricing equation p = E(mx). Form a completely correct general equilibrium model, including the production technology, utility function and specification of the market structure. Derive the equilibrium consumption and asset price process, including p = E(mx) as one of the equilibrium conditions. If the statistical models for consumption and/or asset returns are right, i.e. if they coincide with the equilibrium consumption or return process generated by the true economy, either of the first two approaches will give correct predictions for the joint consumption-asset return process. As we will see, most finance models, developed from the 1950s through the early 1970s, take the return process as given, implicitly assuming linear technologies. The endowment economy approach, introduced by Lucas (1978), is a breakthrough because it turns out to be much easier. It is much easier to evaluate p = E(mx) for fixed m than it is to solve joint consumption-portfolio problems for given asset returns, all to derive the equilibrium consumption process. To solve a consumption-portfolio problem we have to model the investor’s entire environment: we have to specify all the assets to which he has access, what his labor income process looks like (or wage rate process, and include a labor supply decision). Once we model the consumption stream directly, we can look at each asset in isolation, and the actual computation is almost trivial. This breakthrough accounts for the unusual structure of the presentation in this book. It is traditional to start with an extensive study of consumption-portfolio problems. But by modeling consumption directly, we have been able to study pricing directly, and portfolio problems are an interesting side trip which we can defer. Most uses of p = E(mx) do not require us to take any stand on exogeneity or endogeneity, or general equilibrium. This is a condition that must hold for any asset, for any 46

S ECTION 2.3

C ONSUMPTION - BASED MODEL IN PRACTICE

production technology. Having a taste of the extra assumptions required for a general equilibrium model, you can now appreciate why people stop short of full solutions when they can address an application using only the first order conditions, using knowledge of E(mx) to make a prediction about p. It is enormously tempting to slide into an interpretation that E(mx) determines p. We routinely think of betas and factor risk prices – components of E(mx) – as determining expected returns. For example, we routinely say things like “the expected return of a stock increased because the firm took on riskier projects, thereby increasing its β.” But the whole consumption process, discount factor, and factor risk premia change when the production technology changes. Similarly, we are on thin ice if we say anything about the effects of policy interventions, new markets and so on. The equilibrium consumption or asset return process one has modeled statistically may change in response to such changes in structure. For such questions one really needs to start thinking in general equilibrium terms. It may help to remember that there is an army of permanent-income macroeconomists who make precisely the opposite assumption, taking our asset return processes as exogenous and studying (endogenous) consumption and savings decisions.

2.3

Consumption-based model in practice

The consumption-based model is, in principle, a complete answer to all asset pricing questions, but works poorly in practice. This observation motivates other asset pricing models. The model I have sketched so far can, in principle, give a compete answer to all the questions of the theory of valuation. It can be applied to any security—bonds, stocks, options, futures, etc.—or to any uncertain cash flow. All we need is a functional form for utility, numerical values for the parameters, and a statistical model for the conditional distribution of consumption and payoffs. To be specific, consider the standard power utility function u0 (c) = c−γ .

(41)

" µ # ¶−γ ct+1 e 0 = Et β Rt+1 ct

(42)

Then, excess returns should obey

Taking unconditional expectations and applying the covariance decomposition, expected ex47

C HAPTER 2

A PPLYING THE BASIC MODEL

cess returns should follow e E(Rt+1 )

f

= −R cov

"µ

ct+1 ct

¶−γ

,

e Rt+1

#

.

(43)

Given a value for γ, and data on consumption and returns, one can easily estimate the mean and covariance on the right hand side, and check whether actual expected returns are, in fact, in accordance with the formula. Similarly, the present value formula is pt = Et

∞ X

βj

j=1

µ

ct+j ct

¶−γ

(44)

dt+j .

Given data on consumption and dividends or another stream of payoffs, we can estimate the right hand side and check it against prices on the left. Bonds and options do not require separate valuation theories. For example, an N-period default-free nominal discount bond (a U.S. Treasury strip) is a claim to one dollar at time t + N . Its price should be Ã

pt = Et β

N

µ

ct+N ct

¶−γ

! Πt 1 Πt+N

where Π = price level ($/good). A European option is a claim to the payoff max(St+T − K, 0), where St+T = stock price at time t + T, K = strike price. The option price should be "

pt = Et β

T

µ

ct+T ct

¶−γ

#

max(St+T − K, 0)

again, we can use data on consumption, prices and payoffs to check these predictions. Unfortunately, the above specification of the consumption-based model does not work very well. To give a flavor of some of the problems, Figure 5 presents the mean excess returns on the ten size-ranked portfolios of NYSE stocks vs. the predictions – the right hand side of (2.43) – of the consumption-based model. I picked the utility curvature parameter γ = 241 to make the picture look as good as possible (The section on GMM estimation below goes into detail on how to do this. The Figure presents the first-stage GMM estimate.) As you can see, the model isn’t hopeless–there is some correlation between sample average returns and the consumption-based model predictions. But the model does not do very well. The pricing error (actual expected return - predicted expected return) for each portfolio is of the same order of magnitude as the spread in expected returns across the portfolios. 48

S ECTION 2.4

A LTERNATIVE ASSET PRICING MODELS : OVERVIEW

Figure 5. Mean excess returns of 10 CRSP size portfolios vs. predictions of the power utility consumption-based model. The predictions are generated by −Rf cov(m, Ri ) with m = β(ct+1 /ct )−γ . β = 0.98 and γ = 241 are picked by first-stage GMM to minimize the sum of squared pricing errors (deviation from 45◦ line). Source: Cochrane (1996).

2.4

Alternative asset pricing models: Overview

I motivate exploration of different utility functions, general equilibrium models, and linear factor models such as the CAPM, APT and ICAPM as approaches to circumvent the empirical difficulties of the consumption-based model. The poor empirical performance of the consumption-based model motivates a search for alternative asset pricing models – alternative functions m = f (data). All asset pricing models amount to different functions for m. I give here a bare sketch of some of the different approaches; we study each in detail in later chapters. 1) Different utility functions. Perhaps the problem with the consumption-based model is simply the functional form we chose for utility. The natural response is to try different utility functions. Which variables determine marginal utility is a far more important question than the functional form. Perhaps the stock of durable goods influences the marginal utility of nondurable goods; perhaps leisure or yesterday’s consumption affect today’s marginal utility. These possibilities are all instances of nonseparabilities. One can also try to use micro data on individual consumption of stockholders rather than aggregate consumption. Aggregation of heterogenous investors can make variables such as the cross-sectional variance of income 49

C HAPTER 2

A PPLYING THE BASIC MODEL

appear in aggregate marginal utility. 2) General equilibrium models. Perhaps the problem is simply with the consumption data. General equilibrium models deliver equilibrium decision rules linking consumption to other variables, such as income, investment, etc. Substituting the decision rules ct = f(yt , it , . . . ) in the consumption-based model, we can link asset prices to other, hopefully better-measured macroeconomic aggregates. In addition, true general equilibrium models completely describe the economy, including the stochastic process followed by all variables. They can answer questions such as why is the covariance (beta) of an asset payoff x with the discount factor m the value that it is, rather than take this covariance as a primitive. They can in principle answer structural questions, such as how asset prices might be affected by different government policies. Neither kind of question can be answered by just manipulating investor first order conditions. 3) Factor pricing models. Another sensible response to bad consumption data is to model marginal utility in terms of other variables directly. Factor pricing models follow this approach. They just specify that the discount factor is a linear function of a set of proxies, A B + bB ft+1 +... . mt+1 = a + bA ft+1

(45)

where f i are factors and a, bi are parameters. (This is a different sense of the use of the word “factor” than “discount factor.” I didn’t invent the confusing terminology.) By and large, the factors are just selected as plausible proxies for marginal utility; events that describe whether typical investors are happy or unhappy. Among others, the Capital Asset Pricing Model (CAPM) is the model mt+1 = a + bRW t+1 where RW is the rate of return on a claim to total wealth, often proxied by a broad-based portfolio such as the value-weighted NYSE portfolio. The Arbitrage Pricing Theory (APT) uses returns on broad-based portfolios derived from a factor analysis of the return covariance matrix. The Intertemporal Capital Asset Pricing Model (ICAPM) suggests macroeconomic variables such as GNP and inflation and variables that forecast macroeconomic variables or asset returns as factors. Term structure models such as the Cox-Ingersoll-Ross model specify that the discount factor is a function of a few term structure variables, for example the short rate of interest and a few interest rate spreads. Many factor pricing models are derived as general equilibrium models with linear technologies and no labor income; thus they also fall into the general idea of using general equilibrium relations (from, admittedly, very stylized general equilibrium models) to substitute out for consumption. 4) Arbitrage or near-arbitrage pricing. The mere existence of a representation p = E(mx) and the fact that marginal utility is positive m ≥ 0 (these facts are discussed in the next chapter) can often be used to deduce prices of one payoff in terms of the prices of other payoffs. The Black-Scholes option pricing model is the paradigm of this approach: 50

S ECTION 2.5

P ROBLEMS

Since the option payoff can be replicated by a portfolio of stock and bond, any m that prices the stock and bond gives the price for the option. Recently, there have been several suggestions on how to use this idea in more general circumstances by using very weak further restrictions on m, and we will study these suggestions in Chapter 17. We return to a more detailed derivation and discussion of these alternative models of the discount factor m below. First, and with this brief overview in mind, we look at p = E(mx) and what the discount factor m represents in a little more detail.

2.5 1.

Problems

The representative consumer maximizes a CRRA utility function. X j 1−γ β ct+j . Et Consumption is given by an endowment stream.

(a)

2.

Show that with log utility, the price/consumption ratio of the consumption stream is constant, no matter what the distribution of consumption growth. (b) Suppose there is news at time t that future consumption will be higher. For γ < 1, γ = 1,and γ > 1, evaluate the effect of this news on the price. Make sense of your results. (Note: there is a real-world interpretation here. It’s often regarded as a puzzle that the market declines on good economic news. This is attributed to an expectation by the market that the Fed will respond to such news by raising interest rates. Note that γ > 0 in this problem gives a completely real and frictionless interpretation to this phenomenon! I thank Pete Hecht for this nice problem.) The linear quadratic permanent income model is a very useful general equilibrium model that we can solve in closed form. It specifies a production technology rather than fixed endowments, and it easily allows aggregation of disparate consumers. (Hansen 1987 is a wonderful exposition of what one can do with this setup.) The consumer maximizes µ ¶ ∞ X 1 βt − (ct − c∗ )2 E 2 t=0 subject to a linear technology kt+1 = (1 + r)kt + it it = et − ct et is an exogenous endowment or labor income stream. Assume β = 1/(1 + r); the discount rate equals the interest rate or marginal productivity of capital. 51

C HAPTER 2 (a)

A PPLYING THE BASIC MODEL

Show that optimal consumption follows ct = rkt + rβ

∞ X

β j Et et+j

j=0 ∞ X

ct = ct−1 + (Et − Et−1 ) rβ

(b)

β j et+j

(2.46) (2.47)

j=0

i.e., consumption equals permanent income, precisely defined, and consumption follows a random walk whose innovations are equal to innovations in permanent income. Assume that the endowment et follows an AR(1) et = ρet−1 + εt

3.

and specialize (2.46) and (2.47). Calculate and interpret the result for ρ = 1 and ρ = 0. (The result looks like a “consumption function” relating consumption to capital and current income, except that the slope of that function depends on the persistence of income shocks. Transitory shocks will have little effect on consumption, and permanent shocks a larger effect.) (c) Calculate the one period interest rate (it should come out to r of course) and the price of a claim to the consumption stream. e and k are the only state variables, so the price should be a function of e and k. Interpret the time-variation in the price of the consumption stream. (This consumer gets more risk averse as consumption rises to c∗ . c∗ is the bliss point, so at the bliss point there is no average return that can compensate the consumer for greater risk.) Consider again CRRA utility, Et

X

β j c1−γ t+j .

Consumption growth follows a two-state Markov process. The states are ∆ct = ct /ct−1 = h, l, and a 2×2 matrix π governs the set of transition probabilities, i.e. pr(∆ct+1 = h|∆ct = l) = πl→h . (This is the Mehra-Prescott 1986 model, but it will be faster to do it than to look it up. It is a useful and simple endowment economy.) (a)

Find the riskfree rate (price of a certain real payoff of one) in this economy. This price is generated by pbt = Et (mt,t+1 1).

(b)

You are looking for two values, the price in the l state and the price in the h state. Find the price of the consumption stream (the price at t of {ct+1 , ct+2 , ...}). To do this, guess that the price/consumption ratio must be a function of state (h,l), and find 52

S ECTION 2.5

P ROBLEMS

that function. From ¡ ¢ pct = Et mt,t+1 (pct+1 + ct+1 )

(c)

(d)

find a recursive relation for pct /ct , and hence find the two values of pct /ct , one for the h state and one for the l state. Pick β = 0.99 and try γ = 0.5, 5 (Try more if you feel like it). Calibrate the consumption process to have a 1% mean and 1% standard deviation, and consumption growth uncorrelated over time. Calculate prices and returns in each state. Now introduce serial correlation in consumption growth with γ = 5. (You can do this by adding weight to the diagonal entries of the transition matrix π.) What effect does this have on the model?

53

Chapter 3.

Contingent Claims Markets

Our first task is to understand the p = E(mx) representation a little more deeply. In this chapter I introduce a very simple market structure, contingent claims. This leads us to an inner product interpretation of p = E(mx) which allows an intuitive visual representation of most of the theorems. We see that discount factors exist, are positive, and the pricing function is linear, just starting from prices and payoffs in a complete market, without any utility functions. The next chapter shows that these properties can be built up in incomplete markets as well.

3.1

Contingent claims

I describe contingent claims. I interpret the stochastic discount factor m as contingent claims prices divided by probabilities, and p = E(mx) as a bundling of contingent claims. Suppose that one of S possible states of nature can occur tomorrow, i.e. specialize to a finite-dimensional state space. Denote the individual states by s. For example, we might have S = 2 and s = rain or s = shine. A contingent claim is a security that pays one dollar (or one unit of the consumption good) in one state s only tomorrow. pc(s) is the price today of the contingent claim. I write pc to specify that it is the price of a contingent claim and (s) to denote in which state s the claim pays off. In a complete market investors can buy any contingent claim. They don’t necessarily have to be faced with explicit contingent claims; they just need enough other securities to span or synthesize all contingent claims. For example, if the possible states of nature are (rain, shine), one can span or synthesize any contingent claim or portfolio achieved by combining contingent claims by forming portfolios of a security that pays 2 dollars if it rains and one if it shines, or x1 = (2, 1), and a riskfree security whose payoff pattern is x2 = (1, 1). Now, we are on a hunt for discount factors, and the central point is: If there are complete contingent claims, a discount factor exists, and it is equal to the contingent claim price divided by probabilities. Let x(s) denote an asset’s payoff in state of nature s. We can think of the asset as a bundle of contingent claims—x(1) contingent claims to state 1, x(2) claims to state 2, etc. The asset’s price must then equal the value of the contingent claims of which it is a bundle,

p(x) =

X

pc(s)x(s).

s

54

(48)

S ECTION 3.2

R ISK NEUTRAL PROBABILITIES

I denote the price p(x) to emphasize that it is the price of the payoff x. Where the payoff in question is clear, I suppress the (x). I like to think of equation (3.48) as a happy-meal theorem: the price of a happy meal (in a frictionless market) should be the same as the price of one hamburger, one small fries, one small drink and the toy. It is easier to take expectations rather than sum over states. To this end, multiply and divide the bundling equation (3.48) by probabilities, µ ¶ X pc(s) π(s) x(s) p(x) = π(s) s where π(s) is the probability that state s occurs. Then define m as the ratio of contingent claim price to probability m(s) =

pc(s) . π(s)

Now we can write the bundling equation as an expectation, X p= π(s)m(s)x(s) = E(mx). s

Thus, in a complete market, the stochastic discount factor m in p = E(mx) exists, and it is just a set of contingent claims prices, scaled by probabilities. As a result of this interpretation, the combination of discount factor and probability is sometimes called a state-price density. The multiplication and division by probabilities seems very artificial in this finite-state context. In general, we posit states of nature ω that can take continuous (uncountably infinite) values in a space Ω. In this case, the sums become integrals, and we have to use some measure to integrate over Ω. Thus, scaling contingent claims prices by some probability-like object is unavoidable.

3.2

Risk neutral probabilities

I interpret the discount factor m as a transformation to risk-neutral probabilities such that p = E ∗ (x)/Rf . Another common transformation of p = E(mx) results in “risk-neutral” probabilities. Define π∗ (s) ≡ Rf m(s)π(s) = Rf pc(s) 55

C HAPTER 3

C ONTINGENT C LAIMS M ARKETS

where Rf ≡ 1/

X

pc(s) = 1/E(m).

The π∗ (s) are positive, less than or equal to one and sum to one, so they are a legitimate set of probabilities. Then we can rewrite the asset pricing formula as p(x) =

X

pc(s)x(s) =

s

1 X ∗ E ∗ (x) (s)x(s) = . π Rf Rf

I use the notation E ∗ to remind us that the expectation uses the risk neutral probabilities π∗ instead of the real probabilities π. Thus, we can think of asset pricing as if agents are all risk neutral, but with probabilities π∗ in the place of the true probabilities π. The probabilities π∗ gives greater weight to states with higher than average marginal utility m. There is something very deep in this idea: risk aversion is equivalent to paying more attention to unpleasant states, relative to their actual probability of occurrence. People who report high subjective probabilities of unpleasant events like plane crashes may not have irrational expectations, they may simply be reporting the risk neutral probabilities or the product m × π. This product is after all the most important piece of information for many decisions: pay a lot of attention to contingencies that are either highly probable or that are improbable but have disastrous consequences. The transformation from actual to risk-neutral probabilities is given by π∗ (s) =

m(s) π(s). E(m)

We can also think of the discount factor m as the derivative or change of measure from the real probabilities π to the subjective probabilities π ∗ . The risk-neutral probability representation of asset pricing is quite common, especially in derivative pricing where the results are independent of risk adjustments. The risk-neutral representation is particularly popular in continuous time diffusion processes, because we can adjust only the means, leaving the covariances alone. In discrete time, changing the probabilities typically changes first and second moments. Suppose we start with a process for prices and discount factor dp p dΛ Λ

= µp dt + σ p dz = µΛ dt + σΛ dz. 56

S ECTION 3.3

I NVESTORS AGAIN

The discount factor prices the assets, µ ¶ µ ¶ dp dΛ dp D Et + dt − rf dt = −Et = −σp σ Λ dt p p Λ p In the “risk-neutral measure” we just increase the drift of each price process by its covariance with the discount factor, and write a risk-neutral discount factor, dp p dΛ Λ

=

¡ p ¢ µ + σp σΛ dt + σ p dz = µp∗ dt + σ p dz

= µΛ dt.

Under this new set of probabilities, we can just write, µ ¶ dp D Et∗ + dt − rf dt = 0 p p with Et∗ (dp/p) = µp∗ dt.

3.3

Investors again

We look at investor’s first order conditions in a contingent claims market. The marginal rate of substitution equals the discount factor and the contingent claim price ratio. Though the focus of this chapter is on how to do without utility functions, It’s worth looking at the investor’s first order conditions again in the contingent claim context. The investor starts with a pile of initial wealth y and a state-contingent income y(s). He purchases contingent claims to each possible state in the second period. His problem is then X X X βπ(s)u[c(s)] s.t. c + pc(s)c(s) = y + pc(s)y(s). max u(c) + {c,c(s)}

s

s

s

Introducing a Lagrange multiplier λ on the budget constraint, the first order conditions are u0 (c) = λ βπ(s)u0 [c(s)] = λpc(s).

Eliminating the Lagrange multiplier λ, pc(s) = βπ(s) 57

u0 [c(s)] u0 (c)

C HAPTER 3

C ONTINGENT C LAIMS M ARKETS

or m(s) =

pc(s) u0 [c(s)] =β 0 π(s) u (c)

Coupled with p = E(mx), we obtain the consumption-based model again. The investor’s first order conditions say that marginal rates of substitution between states tomorrow equals the relevant price ratio, u0 [c(s1 )] m(s1 ) = 0 . m(s2 ) u [c(s2 )] m(s1 )/m(s2 ) gives the rate at which the investor can give up consumption in state 2 in return for consumption in state 1 through purchase and sales of contingent claims. u0 [c(s1 )]/u0 [c(s2 )] gives the rate at which the investor is willing to make this substitution. At an optimum, the marginal rate of substitution should equal the price ratio, as usual in economics. We learn that the discount factor m is the marginal rate of substitution between date and state contingent commodities. That’s why it, like c(s), is a random variable. Also, scaling contingent claims prices by probabilities gives marginal utility, and so is not so artificial as it may have seemed above. Figure 6 gives the economics behind this approach to asset pricing. We observe the investor’s choice of date or state-contingent consumption. Once we know his utility function, we can calculate the contingent claim prices that must have led to the observed consumption choice, from the derivatives of the utility function. State 2 or date 2

(c1, c2) Indifference curve

State 1, or date 1

Figure 6. Indifference curve and contingent claim prices The relevant probabilities are the investor’s subjective probabilities over the various states. 58

S ECTION 3.4

R ISK SHARING

Asset prices are set, after all, by investor’s demands for assets, and those demands are set by investor’s subjective evaluations of the probabilities of various events. We often assume rational expectations, namely that subjective probabilities are equal to objective frequencies. But this is an additional assumption that we may not always want to make.

3.4

Risk sharing

Risk sharing: In complete markets, consumption moves together. Only aggregate risk matters for security markets. We deduced that the marginal rate of substitution for any individual investor equals the contingent claim price ratio. But the prices are the same for all investors. Therefore, marginal utility growth should be the same for all investors βi

0 j u0 (cit+1 ) j u (ct+1 ) = β u0 (cit ) u0 (cjt )

(49)

where i and j refer to different investors. If investors have the same homothetic utility function (for example, power utility), then consumption itself should move in lockstep, cj cit+1 = t+1 . i ct cjt More generally, shocks to consumption are perfectly correlated across individuals. This is so radical, it’s easy to misread it at first glance. It doesn’t say that expected consumption growth should be equal; it says that consumption growth should be equal expost. If my consumption goes up 10%, yours goes up exactly 10% as well, and so does everyone else’s. In a complete contingent claims market, all investors share all risks, so when any shock hits, it hits us all equally (after insurance payments). It doesn’t say the consumption level is the same – this is risk-sharing, not socialism. The rich have higher levels of consumption, but rich and poor share the shocks equally. This risk sharing is Pareto-optimal. Suppose a social planner wished to maximize everyone’s utility given the available resources. For example, with two investors i and j, he would maximize max λi

X

β t u(cit ) + λj

X

β t u(cjt ) s.t. cit + cjt = cat

where ca is the total amount available and λi and λj are i and j’s relative weights in the 59

C HAPTER 3

C ONTINGENT C LAIMS M ARKETS

planner’s objective. The first order condition to this problem is λi u0 (cit ) = λj u0 (cjt ) and hence the same risk sharing that we see in a complete market, equation (3.49). This simple fact has profound implications. First, it shows you why only aggregate shocks should matter for risk prices. Any idiosyncratic income risk will be equally shared, and so 1/N of it becomes an aggregate shock. Then the stochastic discount factors m that determine asset prices are no longer affected by truly idiosyncratic risks. Much of this sense that only aggregate shocks matter stays with us in incomplete markets as well. Obviously, the real economy does not yet have complete markets or full risk sharing – individual consumptions do not move in lockstep. However, this observation tells us much about the function of securities markets. Security markets – state-contingent claims – bring individual consumptions closer together by allowing people to share some risks. In addition, better risk sharing is much of the force behind financial innovation. Many successful new securities can be understood as devices to more widely share risks.

3.5

State diagram and price function

I introduce the state space diagram and inner product representation for prices, p(x) = E(mx) = m · x. p(x) = E(mx) implies p(x) is a linear function.

Think of the contingent claims price pc and asset payoffs x as vectors in RS , where each element gives the price or payoff to the corresponding state, £ ¤0 pc = pc(1) pc(2) · · · pc(S) , x=

£

x(1) x(2) · · ·

x(S)

¤0

.

Figure 7 is a graph of these vectors in RS . Next, I deduce the geometry of Figure 7.

The contingent claims price vector pc points in to the positive orthant. We saw in section 3.3 that m(s) = u0 [c(s)]/u0 (c). Now, marginal utility should always be positive (people always want more), so the marginal rate of substitution and discount factor are always nonnegative, m > 0 and pc > 0. Don’t forget, m and pc are vectors, or random variables. Thus, m > 0 means the realization of m is positive in every state of nature, or, equivalently every element of the vector m is positive. The set of payoffs with any given price lie on a (hyper)plane perpendicular to the contin60

S ECTION 3.5

S TATE DIAGRAM AND PRICE FUNCTION

State 2 Payoff

Price = 2 Riskfree rate pc

Price = 1 (returns)

State 1 contingent claim State 1 Payoff Price = 0 (excess returns)

Figure 7. Contingent claims prices (pc) and payoffs. gent claim price vector. We reasoned above that the price of the payoff x must be given by its contingent claim value (3.48), X p(x) = pc(s)x(s). (50) s

Interpreting pc and x as vectors, this means that the price is given by the inner product of the contingent claim price and the payoff. If two vectors are orthogonal – if they point out from the origin at right angles to each other – then their inner product is zero. Therefore, the set of all zero price payoffs must lie on a plane orthogonal to the contingent claims price vector, as shown in figure 7. More generally, the inner product of two vectors x and pc equals the product of the magnitude of the projection of x onto pc times the magnitude of pc. Using a dot to denote inner product, X pc(s)x(s) = pc · x = |pc| × |proj(x|pc)| = |pc| × |x| × cos(θ) p(x) = s

where |x| means the length of the vector x and θ is the angle between the vectors pc and x. Since all payoffs on planes (such as the price planes in figure 7) that are perpendicular to pc have the same projection onto pc, they must have the same price. (Only the price = 0 plane is, strictly speaking, orthogonal to pc. Lacking a better term, I’ve called the nonzero 61

C HAPTER 3

C ONTINGENT C LAIMS M ARKETS

price planes “perpendicular” to pc.) When vectors are finite-dimensional, the prime notation is commonly used for inner products, pc0 x. This notation does not extend well to infinitedimensional spaces. The notation hpc|xi is also often used for inner products. Planes of constant price move out linearly, and the origin x = 0 must have a price of zero. If payoff y = 2x, then its price is twice the price of x, X X pc(s)y(s) = pc(s)2x(s) = 2 p(x). p(y) = s

s

Similarly, a payoff of zero must have a price of zero. We can think of p(x) as a pricing function, a map from the state space or payoff space in which x lies (RS in this case) to the real line. We have just deduced from the definition (3.50) that p(x) is a linear function, i.e. that p(ax + by) = ap(x) + bp(y). The constant price lines in Figure 7 are of course exactly what one expects from a linear function from RS to R. (One might draw the price on the z axis coming out of the page. Then the price function would be a plane going through the origin and sloping up with isoprice lines as given in Figure 7.) Figure 7 also includes the payoffs to a contingent claim to the first state. This payoff is one in the first state and zero in other states and thus located on the axis. The plane of price = 1 payoffs is the plane of asset returns; the plane of price = 0 payoffs is the plane of excess returns. A riskfree unit payoff (the payoff to a risk-free pure discount bond) lies on the (1, 1) point in Figure 7; the riskfree return lies on the intersection of the 45o line (same payoff in both states) and the price = 1 plane (the set of all returns). Geometry with m in place of pc. The geometric interpretation of Figure 7 goes through with the discount factor m in the place of pc. We can define an inner product between the random variables x and y by x · y ≡ E(xy), and retain all the mathematical properties of an inner product. For this reason, random variables for which E(xy) = 0 are often called “orthogonal.” This language may be familiar from linear regressions. When we run a regression of y on x, y = b0 x + ε we find the linear combination of x that is “closest” to y, by minimizing the variance or “size” of the residual ε. We do this by forcing the residual to be “orthogonal” to the right hand variable E(xε) = 0. The projection of y on x is defined as the fitted value, proj(y|x) = 62

S ECTION 3.5

S TATE DIAGRAM AND PRICE FUNCTION

b0 x =E(xx0 )−1 E(yx0 )x. This ideal is often illustrated by a residual vector ε that is perpendicular to a plane defined by the right hand variables x. Thus, when the inner product is defined by a second moment, the operation “project y onto x” is a regression. (If x does not include a constant, you don’t add one.) The geometric interpretation of Figure 7 also is valid if we generalize the setup to an infinite-dimensional state space, i.e. if we think of continuously-valued random variables. Instead of vectors, which are functions from RS to R, random variables are (measurable) functions from Ω to R. Nonetheless, we can still think of them as vectors. The equivalent of Rs is now a Hilbert space L2 , which denotes spaces generated by linear combinations of square integrable functions from Ω to the real line, or the space of random variables with finite second moments. We can still define an “inner product” between two such elements by x · y = E(xy), and p(x) = E(mx) can still be interpreted as “m is perpendicular to (hyper)planes of constant price.” Proving theorems in this context is a bit harder. You can’t just say things like “we can take a line perpendicular to any plane,” such things have to be proved. Sometimes, finite-dimensional thinking can lead you to errors, so it’s important to prove things the right way, keeping the finite dimensional pictures in mind for interpretation. Hansen and Richard (1987) is a very good reference for the Hilbert space machinery.

63

Chapter 4.

The discount factor

Now we look more closely at the discount factor. Rather than derive a specific discount factor as with the consumption-based model in the last chapter, I work backwards. A discount factor is just some random variable that generates prices from payoffs, p = E(mx). What does this expression mean? Can one always find such a discount factor? Can we use this convenient representation without implicitly assuming all the structure of the investors, utility functions, complete markets, and so forth? The chapter focuses on two famous theorems. The law of one price states that if two portfolios have the same payoffs (in every state of nature), then they must have the same price. A violation of this law would give rise to an immediate kind of arbitrage profit, as you could sell the expensive version and buy the cheap version of the same portfolio. The first theorem states that there is a discount factor that prices all the payoffs by p = E(mx) if and only if this law of one price holds. In finance, we reserve the term absence of arbitrage for a stronger idea, that if payoff A is always at least as good as payoff B, and sometimes A is better, then the price of A must be greater than the price of B. The second theorem is that there is a positive discount factor that prices all the payoffs by p = E(mx) if and only if there are no arbitrage opportunities, so defined. These theorems are useful to show that we can use stochastic discount factors without implicitly assuming anything about utility functions, aggregation, complete markets and so on. All we need to know about investors in order to represent prices and payoffs via a discount factor is that they won’t leave law of one price violations or arbitrage opportunities on the table. These theorems can be used to describe aspects of a payoff space (such as law of one price, absence of arbitrage) by restrictions on the discount factor (such as it exists and it is positive). Chapter 18 shows how it can be more convenient to impose and check restrictions on a single discount factor than it is to check the corresponding restrictions on all possible portfolios. Chapter 7 discusses these and other implications of the existence theorems. The theorems are credited to Ross (1978), and Harrison and Kreps (1979). My presentation follows Hansen and Richard (1987).

4.1

Law of one price and existence of a discount factor

Definition of law of one price; price is a linear function. p = E(mx) implies law of one price. The law of one price implies that a discount factor exists: There exists a unique x∗ in X such that p = E(x∗ x) for all x ∈ X = space of all available payoffs. 64

S ECTION 4.1

L AW OF ONE PRICE AND EXISTENCE OF A DISCOUNT FACTOR

Furthermore, for any valid discount factor m, x∗ = proj(m | X).

So far we have derived the basic pricing relation p = E(mx) from environments with a lot of structure: either the consumption-based model or complete markets. Suppose we observe a set of prices p and payoffs x, and that markets — either the markets faced by investors or the markets under study in a particular application — are incomplete, meaning they do not span the entire set of contingencies. In what minimal set of circumstances does some discount factor exists which represents the observed prices by p = E(mx)? This section and the following answer this important question. This treatment is a simplified version of Hansen and Richard (1987), which contains rigorous proofs and some technical assumptions. Payoff space The payoff space X is the set (or a subset) of all the payoffs that investors can purchase, or it is a subset of the tradeable payoffs that is used in a particular study. For example, if there are complete contingent claims to S states of nature, then X = RS . But the whole point is that markets are (as in real life) incomplete, so we will generally think of X as a proper subset of complete markets RS .

The payoff space will include some set of primitive assets, but investors can also form new payoffs by forming portfolios. I assume that investors can form any portfolio of traded assets: A1: (Portfolio formation) x1 , x2 ∈ X ⇒ ax1 + bx2 ∈ X for any real a, b.

Of course, X = RS for complete markets satisfies the portfolio formation assumption. If there is a single basic payoff x, then the payoff space must be at least the ray from the origin through x. If there are two basic payoffs in R3 , then the payoff space X must include the plane defined by these two payoffs and the origin. Figure 8 illustrates these possibilities. The payoff space is not the space of returns. The return space is a subset of the payoff space; if a return R is in the payoff space, then you can pay a price $2 to get a payoff 2R, so the payoff 2R with price 2 is also in the payoff space. Also, −R is in the payoff space.

Free portfolio formation is in fact an important and restrictive simplifying assumption. It rules out short sales constraints, bid/ask spreads, leverage limitations and so on. The theory can be modified to incorporate these frictions, but it is a substantial modification. If investors can form portfolios of a vector of basic payoffs x (say, the returns on the NYSE stocks), then the payoff space consists of all portfolios or linear combinations of these original payoffs X = {c0 x} where c is a vector of portfolio weights. We also can allow truly infinite-dimensional payoff spaces. For example, investors might be able to trade nonlinear 65

C HAPTER 4

T HE DISCOUNT FACTOR

State 3 (into page) State 2

State 2

x2

X

x1

x

Single Payoff in R2

State 1

Two Payoffs in R3

State 1

Figure 8. Payoff spaces X generated by one (left) and two (right) basis payoffs. functions of a basis payoff x, such as call options on x with strike price K, which have payoff max [x(s) − K, 0] . The law of one price. A2: (Law of one price, linearity) p(ax1 + bx2 ) = ap(x1 ) + bp(x2 ) It doesn’t matter how one forms the payoff x. The price of a burger, shake and fries must be the same as the price of a happy meal. Graphically, if the iso-price curves were not planes, then one could buy two payoffs on the same iso-price curve, form a portfolio whose payoff is on the straight line connecting the two original payoffs, and sell the portfolio for a higher price than it cost to assemble it. The law of one price basically says that investors can’t make instantaneous profits by repackaging portfolios. If investors can sell securities, this is a very weak characterization of preferences. It says there is at least one investor for whom marketing doesn’t matter, who values a package by its contents. The law is meant to describe a market that has already reached equilibrium. If there are any violations of the law of one price, traders will quickly eliminate them so they can’t survive in equilibrium. A1 and A2 also mean that the 0 payoff must be available, and must have price 0. The Theorem The existence of a discount factor implies the law of one price. This is obvious to the point of triviality: if x = y + z then E(mx) = E[m(y + z)]. The hard, and interesting part of the theorem reverses this logic. We show that the law of one price implies the existence of a discount factor.

66

S ECTION 4.1

L AW OF ONE PRICE AND EXISTENCE OF A DISCOUNT FACTOR

Theorem: Given free portfolio formation A1, and the law of one price A2, there exists a unique payoff x∗ ∈ X such that p(x) = E(x∗ x) for all x ∈ X. x∗ is a discount factor. A1 and A2 imply that the price function on X looks like Figure 7: parallel hyperplanes marching out from the origin. The only difference is that X may be a subspace of the original state space, as shown in Figure 8. The essence of the proof, then, is that any linear function on a space X can be represented by inner products with a vector that lies in X. Proof 1: (Geometric.) We have established that the price is a linear function as shown in Figure 9. (Figure 9 can be interpreted as the plane X of a larger dimensional space as in the right hand panel of Figure 8, laid flat on the page for clarity.) Now we can draw a line from the origin perpendicular to the price planes. Choose a vector x∗ on this line. Since the line is orthogonal to the price zero plane we have 0 = p(x) = E(x∗ x) for price zero payoffs x immediately. The inner product between any payoff x on the price = 1 plane and x∗ is |proj(x|x∗ )| × |x∗ | Thus, every payoff on the price = 1 plane has the same inner product with x∗ . All we have to do is pick x∗ to have the right length, and we obtain p(x) = 1 = E(x∗ x) for every x on the price = 1 plane. Then, of course we have p(x) = E(x∗ x) for payoffs x on the other planes as well. Thus, the linear pricing function implied by the Law ¤ of One Price can be represented by inner products with x∗ .

Price = 2 x*

Price = 1 (returns)

Price = 0 (excess returns)

Figure 9. Existence of a discount factor x∗ . The basic mathematical point is just that any linear function can be represented by an 67

C HAPTER 4

T HE DISCOUNT FACTOR

inner product. The the Riesz representation theorem extends the proof to infinite-dimensional payoff spaces. See Hansen and Richard (1987). Proof 2: (Algebraic.) We can prove the theorem by construction when the payoff space X is generated by portfolios of a N basis payoffs (for example, N stocks). This is a common situation, so the formulas are ¤0 also useful in practice. Organize the basis payoffs into a vector £ x = x1 x2 ... xN and similarly their prices p. The payoff space is then X = {c0 x}. We want a discount factor that is in the payoff space, as the theorem requires. Thus, it must be of the form x∗ = c0 x. Construct c so that x∗ prices the basis assets. We want p =E(x∗ x) = E(xx0 c). Thus we need c = E(xx0 )−1 p. If E(xx0 ) is nonsingular, this c exists and is unique. A2 implies that E(xx0 ) is nonsingular (after pruning redundant rows of x). Thus, x∗ = p0 E(xx0 )−1 x

(51)

is our discount factor. It is a linear combination of x so it is in X. It prices the basis assets x by construction. It prices every x ∈ X : E[x∗ (x0 c)] = E[p0 E(xx0 )−1 xx0 c] = p0 c. By 0 linearity, p(c0 x) = c p. What the theorem does and does not say The theorem says there is a unique x∗ in X. There may be many other discount factors m not in X. In fact, unless markets are complete, there are an infinite number of random variables that satisfy p = E(mx). If p = E(mx) then p = E [(m + ε)x] for any ε orthogonal to x, E(εx) = 0. Not only does this construction generate some additional discount factors, it generates all of them: Any discount factor m (any random variable that satisfies p = E(mx)) can be represented as m = x∗ +ε with E(εx) = 0. Figure 10 gives an example of a one-dimensional X in a two-dimensional state space, in which case there is a whole line of possible discount factors m. If markets are complete, there is nowhere to go orthogonal to the payoff space X, so x∗ is the only possible discount factor. Reversing the argument, x∗ is the projection of any stochastic discount factor m on the space X of payoffs. This is a very important fact: the pricing implications of any discount factor m for a set of payoffs X are the same as those of the projection of m on X. This discount factor is known as the mimicking portfolio for m. Algebraically, p = E(mx) = E [(proj(m|X) + ε)x] = E [proj(m|X) x] Let me repeat and emphasize the logic. Above, we started with investors or a contingent claim market, and derived a discount factor. p = E(mx) implies the linearity of the pricing function and hence the law of one price, a pretty obvious statement in those contexts. Here we work backwards. Markets are incomplete in that contingent claims to lots of states of nature are not available. We found that the law of one price implies a linear pricing function, and a linear pricing function implies that there exists at least one and usually many discount factors. 68

S ECTION 4.2

N O -A RBITRAGE AND POSITIVE DISCOUNT FACTORS

Payoff space X

x*

m = x* + ε space of discount factors

Figure 10. Many discount facotors m can price a given set of assets in incomplete markets. We do allow arbitrary portfolio formation, and that sort of “completeness” is important to the result. If investors cannot form a portfolio ax + by, they cannot force the price of this portfolio to equal the price of its constituents. The law of one price is not innocuous; it is an assumption about preferences albeit a weak one. The point of the theorem is that this is just enough information about preferences to deduce the existence of a discount factor.

4.2

No-Arbitrage and positive discount factors

The definition of arbitrage: positive payoff implies positive price. There is a strictly positive discount factor m such that p = E(mx) if and only if there are no arbitrage opportunities. No arbitrage is another, slightly stronger, implication of marginal utility, that can be reversed to show that there is a positive discount factor. We need to start with the definition of arbitrage: Definition (Absence of arbitrage): A payoff space X and pricing function p(x) leave no arbitrage opportunities if every payoff x that is always non-negative, x ≥ 0 (almost surely), and positive, x > 0, with some positive probability, has positive

69

C HAPTER 4

T HE DISCOUNT FACTOR

price, p(x) > 0. No-arbitrage says that you can’t get for free a portfolio that might pay off positively, but will certainly never cost you anything. This definition is different from the colloquial use of the word “arbitrage.” Most people use “arbitrage” to mean a violation of the law of one price – a riskless way of buying something cheap and selling it for a higher price. “Arbitrages” here might pay off, but then again they might not. The word “arbitrage” is also widely abused. “Risk arbitrage” is a Wall Street oxymoron that means making specific kinds of bets. An equivalent statement is that if one payoff dominates another, then its price must be higher – if x ≥ y, then p(x) ≥ p(y) (Or, a bit more carefully but more long-windedly, if x ≥ y almost surely and x > y with positive probability, then p(x) > p(y). You can’t forget that x and y are random variables.) m > 0 ⇒No-arbitrage The absence of arbitrage opportunities is clearly a consequence of a positive discount factor, and a positive discount factor naturally results from any sort of utility maximization. Recall, m(s) = β

u0 [c(s)] > 0. u0 (c)

It is a sensible characterization of preferences that marginal utility is always positive. Few people are so satiated that they will throw away money. Therefore, the marginal rate of substitution is positive. The marginal rate of substitution is a random variable, so “positive” means “positive in every state of nature” or “in every possible realization.” Now, if contingent claims prices are all positive, a bundle of positive amounts of contingent claims must also have a positive price, even in incomplete markets. A little more formally, Theorem: p = E(mx) and m(s) > 0 imply no-arbitrage. Proof: m > 0; x ≥ 0 and there are some states where x > 0. Thus, in some states mx > 0 and in other states mx = 0. Therefore E(mx) > 0. ¤ No arbitrage ⇒ m > 0 Now we turn the observation around, which is again the hard and interesting part. As the law of one price property guaranteed the existence of a discount factor m, no-arbitrage guarantees the existence of a positive m. The basic idea is pretty simple. No-arbitrage means that the prices of any payoff in the positive orthant (except zero, but including the axes) must be strictly positive. The price = 70

S ECTION 4.2

N O -A RBITRAGE AND POSITIVE DISCOUNT FACTORS

0 plane divides the region of positive prices from the region of negative prices. Thus, if the region of negative prices is not to intersect the positive orthant, the iso-price lines must march up and to the right, and the discount factor m, must point up and to the right. This is how we have graphed it all along, most recently in figure 9. Figure 11 illustrates the case that is ruled out: a whole region of negative price payoffs lies in the positive orthant. For example, the payoff x is strictly positive, but has a negative price. As a result, the (unique, since this market is complete) discount factor m is negative in the y-axis state. p = -1

p=0

x

p = +1

x*, m

Figure 11. Counter-example for no-arbitrage ⇒ m > 0 theorem. The payoff x is positive, but has negative price. The discount factor is not strictly positive The theorem is easy to prove in complete markets. There is only one m, x∗ . If it isn’t positive in some state, then the contingent claim in that state has a positive payoff and a negative price, which violates no arbitrage. More formally, Theorem: In complete markets, no-arbitrage implies that there exists a unique m > 0 such that p = E(mx).

Proof: No-arbitrage implies the law of one price, so there is an x∗ such that p = E(x∗ x), and in a complete market this is the unique discount factor. Suppose that in those states, and zero x∗ ≤ 0 for some states. Then, form a payoff x that is 1 P elsewhere. This payoff is strictly positive, but its price, s:x∗ (s) 0, p = E(mx) ∀ x ∈ X. Proof : Join (−p(x), x) together to form vectors in RS+1 . Call M the set of all (−p(x), x) pairs, M = {(−p(x), x); x ∈ X} M is still a linear space: m1 ∈ M, m2 ∈ M ⇒ am1 + bm2 ∈ M. No-arbitrage means that elements of M can’t have all positive elements. If x is positive, −p(x) must be negative. Thus, M is a hyperplane that only intersects the positive orthant at the point 0. We can then create a linear function F : RS+1 ⇒ R such that RS+1 + F (−p, x) = 0 for (−p, x) ∈ M, and F (−p, x) > 0 for (−p, x) ∈ RS+1 except + the origin. Since we can represent any linear function by a perpendicular vector, there is a vector (1, m) such that F (−p, x) = (1, m) · (−p, x) = −p + m · x or −p + E(mx) using the second moment inner product. Finally, since F (−p, x) is positive for (−p, x) > 0, m must be positive. ¤ In a larger space than RS+1 + , as generated by continuously valued random variables, the separating hyperplane theorem assures us that there is a linear function that separates the two convex sets M and the equivalent of RS+1 + , and the Riesz representation theorem tells us that we can represent F as an inner product with some vector by F (−p, x) = −p + m · x. What the theorem does and does not say The theorem says that a discount factor m > 0 exists, but it does not say that m > 0 is unique. The left hand panel of Figure 12 illustrates the situation. Any m on the line through x∗ perpendicular to X also prices assets. Again, p = E[(m + ε)x] if E(εx) = 0. All of these discount factors that lie in the positive orthant are positive, and thus satisfy the theorem. There are lots of them! In a complete market, m is unique, but not otherwise. 72

S ECTION 4.2

N O -A RBITRAGE AND POSITIVE DISCOUNT FACTORS

The theorem says that a positive m exists, but it also does not say that every discount factor m must be positive. The discount factors in the left hand panel of Figure 12 outside the positive orthant are perfectly valid – they satisfy p = E(mx), and the prices they generate on X are arbitrage free, but they are not positive in every state of nature. In particular, the discount factor x∗ in the payoff space is still perfectly valid — p(x) = E(x∗ x) — but it need not be positive, again as illustrated in the left hand panel of Figure 12. p=2 p=1

m m>0

x*

p=1 p=2 X

x* X

Figure 12. Existence of a discount factor and extensions. The left graph shows that the positive discount factor is not unique, and that discount factors may also exist that are not strictly positive. In particular, x∗ need not be positive. The right hand graph shows that each particular choice of m > 0 induces an arbitrage free extension of the prices on X to all contingent claims. This theorem shows that we can extend the pricing function defined on X to all possible payoffs RS , and not imply any arbitrage opportunities on that larger space of payoffs. It says that there is a pricing function p(x) defined over all of RS , that assigns the same (correct, or observed) prices on X and that displays no arbitrage on all of RS . Graphically, it says that we can draw parallel planes to represent prices on all of RS in such a way that the planes intersect X in the right places, and the price planes march up and to the right so the positive orthant always has positive prices. Any positive m generates such a no-arbitrage extension, as illustrated in the right hand panel of Figure 12. In fact, there are many ways to do this. Each different choice of m > 0 generates a different extension of the pricing function. We can think of strictly positive discount factors as possible contingent claims prices. We can think of the theorem as answering the question: is it possible that an observed and incomplete set of prices and payoffs is generated by some complete markets, contingent claim economy? The answer is, yes, if there is no arbitrage on the observed prices and payoffs. In fact, since there are typically many positive m’s consistent with a {X, p(x)}, there exist 73

C HAPTER 4

T HE DISCOUNT FACTOR

many contingent claims economies consistent with our observations. Finally, the absence of arbitrage is another very weak characterization of preferences. The theorem tells us that this is enough to allow us to use the p = E(mx) formalism with m > 0. As usual, this theorem and proof do not require that the state space is RS . State spaces generated by continuous random variables work just as well.

4.3

An alternative formula, and x∗ in continuous time

In terms of the covariance matrix of payoffs, x∗ = E(x∗ ) + [p−E(x∗ )E(x)]0 Σ−1 (x−E(x)). Just like x∗ in discrete time, ¶0 µ D dΛ∗ f − r = −r dt − µ + Σ−1 dz. Λ∗ p prices assets by construction in continuous time. Being able to compute x∗ is useful in many circumstances. This section gives an alternative formula in discrete time, and the continuous time counterpart. A formula that uses covariance matrices E(xx0 ) in our previous formula (4.51) is a second moment matrix. We typically summarize data in terms of covariance matrices instead. Therefore, a convenient alternative formula is x∗ = E(x∗ ) + [p−E(x∗ )E(x)]0 Σ−1 (x−E(x))

(52)

where ¡ ¢ Σ ≡ E [x−E(x)] [x−E(x)]0

denotes the covariance matrix of the x payoffs. (We could just substitute E(xx0 ) = Σ + E(x)E(x0 ), but the inverse of the sum is not very useful.) We can derive this formula by postulating a discount factor that is a linear function of the shocks to the payoffs, x∗ = E(x∗ ) + (x−E(x))0 b, and then finding b to ensure that x∗ prices the assets x : ¤ £ p = E(x∗ )E(x) + E (x−Ex)x0 b 74

S ECTION 4.3

A N ALTERNATIVE FORMULA , AND X∗ IN CONTINUOUS TIME

so b = Σ−1 [p−E(x∗ )E(x)] . If a riskfree rate is traded, then we know E(x∗ ) = 1/Rf . If a riskfree rate is not traded – if 1 is not in X – then this formula does not necessarily produce a discount factor x∗ that is in X. In many applications, however, all that matters is producing some discount factor, and the arbitrariness of the risk-free or zero beta rate is not a problem. This formula is particularly useful when the payoff space consists solely of excess returns or price-zero payoffs. In that case, x∗ = p0 E(xx0 )−1 x gives x∗ = 0. x∗ = 0 is in fact the only discount factor in X that prices all the assets, but in this case it’s more interesting (and avoids 1/0 difficulties when we want to transform to expected return/beta or other representations) to pick a discount factor not in X by picking a zero-beta rate or price of the riskfree payoff. In the case of excess returns, for arbitrarily chosen Rf , then, (4.52) gives us x∗ =

1 1 − f E(Re )0 Σ−1 (Re −E(Re )); Σ ≡ cov(Re ) f R R

Continuous time The law of one price implies the existence of a discount factor process, and absence of arbitrage a positive discount factor process in continuous time as well as discrete time. At one level, this statement requires no new mathematics. If we reinvest dividends for simplicity, then a discount factor must satisfy pt Λt = Et (Λt+s pt+s ) . Calling pt+s = xt+s , this is precisely the discrete time p = E(mx) that we have studied all along. Thus, the law of one price or absence of arbitrage are equivalent to the existence of a or a positive Λt+s . The same conditions at all horizons s are thus equivalent to the existence of a discount factor process, or a positive discount factor process Λt for all time t. For calculations it is useful to find explicit formulas for a discount factors. Suppose a set of securities pays dividends Dt dt and their prices follow dpt = µt dt + σt dzt pt where p and z are N × 1 vectors, µt and σ t may vary over time, µ(pt , t,other variables), E (dzt dzt0 ) = I and the division on the left hand side is element-by element. (As usual, I’ll drop the t subscripts when not necessary for clarity, but everything can vary over time.) We can form a discount factor that prices these assets from a linear combination of the 75

C HAPTER 4

T HE DISCOUNT FACTOR

shocks that drive the original assets, µ ¶0 D dΛ∗ f f −r = −r dt − µ + Σ−1 σdz. Λ∗ p

(53)

where Σ ≡ σσ0 again is the covariance matrix of returns. You can easily check that this equation solves µ ¶ µ ∗ ¶ dp dΛ dp D Et + dt − rf dt = −Et (54) p p Λ∗ p and Et

µ

dΛ∗ Λ∗

¶

= −rf dt,

or you can show that this is the only diffusion driven by dz, dt with these properties. If there is a risk free rate rtf (also potentially time-varying), then that rate determines rtf . If there is no risk free rate, (4.53) will price the risky assets for any arbitrary (or convenient) choice of rtf . As usual, this discount factor is not unique; Λ∗ plus orthogonal noise will also act as a discount factor: dΛ∗ dΛ = ∗ + dw; E(dw) = 0; E(dzdw) = 0. Λ Λ You can see that (4.53) is exactly analogous to the discrete time formula (4.52). (If you don’t like answers popping out of hats like this, guess a solution of the form dΛ = µΛ dt + σΛ dz. Λ Then find µΛ and σΛ to satisfy (4.54) for the riskfree and risky assets.)

4.4 1. 2.

3.

Problems

Show that the law of one price loop implies that price is a linear function of payoff and vice versa Does the absence of arbitrage imply the law of one price? Does the law of one price imply the absence of arbitrage? Answer directly using portfolio arguments, and indirectly using the corresponding discount factors. If the law of one price or absence of arbitrage hold in population, must they hold in a sample drawn from that population?

76

Chapter 5. Mean-variance frontier and beta representations Much empirical work in asset pricing is couched in terms of expected return - beta representations and mean-variance frontiers. This chapter introduces expected return - beta representations and mean-variance frontiers. I discuss here the beta representation, most commonly applied to factor pricing models. Chapter 6 shows how an expected return/beta model is equivalent to a linear model for the discount factor, i.e. m = b0 f where f are the right hand variables in the time-series regressions that define betas. Chapter 9 then discusses the derivation of popular factor models such as the CAPM, ICAPM and APT, i.e. under what assumptions the discount factor is a linear function of other variables f such as the market return. I summarize the classic Lagrangian approach to the mean-variance frontier. I then introduce a powerful and useful representation of the mean-variance frontier due to Hansen and Richard (1987). This representation uses the state-space geometry familiar from the existence theorems. It is also useful because it is valid and useful in infinite-dimensional payoff spaces, which we shall soon encounter when we add conditioning information, dynamic trading or options.

5.1

Expected return - Beta representations

The expected return-beta expression of a factor pricing model is E(Ri ) = α + β i,a λa + β i,b λb + . . . The model is equivalent to a restriction that the intercept is the same for all assets in time-series regressions. When the factors are returns excess returns, then λa = E(f a ). If the test assets are also excess returns, then the intercept should be zero, α = 0. Much empirical work in finance is cast in terms of expected return - beta representations of linear factor pricing models, of the form E(Ri ) = α + β i,a λa + β i,b λb + . . . , i = 1, 2, ...N.

77

(55)

C HAPTER 5

M EAN - VARIANCE FRONTIER AND BETA REPRESENTATIONS

The β terms are defined as the coefficients in a multiple regression of returns on factors, Rit = ai + β i,a fta + β i,b ftb + . . . + εit ; t = 1, 2, ...T.

(56)

This is often called a time-series regression, since one runs a regression across time for each security i. The “factors” f are proxies for marginal utility growth. I discuss the stories used to select factors at some length in chapter 9. For the moment keep in mind the canonical examples, f = consumption growth, or f = the return on the market portfolio (CAPM). Notice that we run returns Rti on contemporaneous factors ftj . This regression is not about predicting returns from variables seen ahead of time. Its objective is to measure contemporaneous relations or risk exposure; whether returns are typically high in “good times” or “bad times” as measured by the factors. The point of the beta model(5.55) is to explain the variation in average returns across assets. I write i = 1, 2, ...N in (5.55) to emphasize this fact. The model says that assets with higher betas should get higher average returns. Thus the betas in (5.55) are the explanatory (x) variables, which vary asset by asset. The α and λ – common for all assets – are the intercept and slope in this cross-sectional relation. For example, equation (5.55) says that if we plot expected returns versus betas in a one-factor model, we should expect all (E(Ri ), β i,a ) pairs to line up on a straight line with slope λa and intercept α. β i,a is interpreted as the amount of exposure of asset i to factor a risks, and λa is interpreted as the price of such risk-exposure. Read the beta pricing model to say: “for each unit of exposure β to risk factor a, you must provide investors with an expected return premium λa .” Assets must give investors higher average returns (low prices) if they pay off well in times that are already good, and pay off poorly in times that are already bad, as measured by the factors. One way to estimate the free parameters (α, λ) and to test the model (5.55) is to run a cross sectional regression of average returns on betas, E(Ri ) = α + β i,a λa + β i,b λb + . . . + αi , i = 1, 2, ...N.

(57)

Again, the β i are the right hand variables, and the α and λ are the intercept and slope coefficients that we estimate in this cross-sectional regression. The errors αi are pricing errors. The model predicts αi = 0, and they should be statistically insignificant in a test. (I intentionally use the same symbol for the intercept, or mean of the pricing errors, and the individual pricing errors αi .) In the chapters on empirical technique, we will see test statistics based on the sum of squared pricing errors. The fact that the betas are regression coefficients is crucially important. If the betas are also free parameters then there is no content to the equation. More importantly (and this is an easier mistake to make), the betas cannot be asset-specific or firm-specific characteristics, such as the size of the firm, book to market ratio, or (to take an extreme example) the letter of the alphabet of its ticker symbol. It is true that expected returns are associated with or correlated with many such characteristics. Stocks of small companies or of companies with high 78

S ECTION 5.1

E XPECTED RETURN - B ETA REPRESENTATIONS

book/market ratios do have higher average returns. But this correlation must be explained by some beta. The proper betas should drive out any characteristics in cross-sectional regressions. If, for example, expected returns were truly related to size, one could buy many small companies to form a large holding company. It would be a “large” company, and hence pay low average returns to the shareholders, while earning a large average return on its holdings. The managers could enjoy the difference. What ruins this promising idea? . The “large” holding company will still behave like a portfolio of small stocks. Thus, only if asset returns depend on how you behave, not who you are – on betas rather than characteristics – can a market equilibrium survive such simple repackaging schemes. Some common special cases If there is a risk free rate, its betas in (5.55) are all zero,2 so the intercept is equal to the risk free rate, Rf = α. We can impose this condition rather than estimate α in the cross-sectional regression (5.57). If there is no risk-free rate, then α must be estimated in the cross-sectional regression. Since it is the expected return of a portfolio with zero betas on all factors, α is called the (expected) zero-beta rate in this circumstance. We often examine factor pricing models using excess returns directly. (There is an implicit, though not necessarily justified, division of labor between models of interest rates and models of equity risk premia.) Differencing (5.55) between any two returns Rei = Ri − Rj (Rj does not have to be risk free), we obtain E(Rei ) = β i,a λa + β i,b λb + . . . , i = 1, 2, ...N.

(58)

Here, β ia represents the regression coefficient of the excess return Rei on the factors. It is often the case that the factors are also returns or excess returns. For example, the CAPM uses the return on the market portfolio as the single factor. In this case, the model should apply to the factors as well, and this fact allows us to directly measure the λ coefficients. Each factor has beta of one on itself and zero on all the other factors, of course. Therefore, if the factors are excess returns, we have E(f a ) = λa , and so forth. We can then write the factor model as E(Rei ) = β i,a E(f a ) + β i,b E(f b ) + . . . , i = 1, 2, ...N. The cross-sectional beta pricing model (5.55)-(5.58) and the time-series regression definition of the betas in (5.56) look very similar. It seems that one can take expectations of 2 The betas are zero because the risk free rate is known ahead of time. When we consider the effects of conditioning information, i.e. that the interest rate could vary over time, we have to interpret the means and betas as conditional moments. Thus, if you are worried about time-varying risk free rates, betas, and so forth, either assume all variables are i.i.d. (and thus that the risk free rate is constant), or interpret all moments as conditional on time t information.

79

C HAPTER 5

M EAN - VARIANCE FRONTIER AND BETA REPRESENTATIONS

the time-series regression (5.56) and arrive at the beta model (5.55), in which case the latter would be vacuous since one can always run a regression of anything on anything. The difference is subtle but crucial: the time-series regressions (5.56) will in general have a different intercept ai for each return i, while the intercept α is the same for all assets in the beta pricing equation (5.55). The beta pricing equation is a restriction on expected returns, and thus imposes a restriction on intercepts in the time-series regression. In the special case that the factors are themselves excess returns, the restriction is particularly simple: the time-series regression intercepts should all be zero. In this case, we can avoid the cross-sectional regression entirely, since there are no free parameters left.

5.2

Mean-variance frontier: Intuition and Lagrangian characterization

The mean-variance frontier of a given set of assets is the boundary of the set of means and variances of the returns on all portfolios of the given assets. One can find or define this boundary by minimizing return variance for a given mean return. Many asset pricing propositions and test statistics have interpretations in terms of the mean-variance frontier. Figure 13 displays a typical mean-variance frontier. As displayed in Figure 13, it is common to distinguish the mean-variance frontier of all risky assets, graphed as the hyperbolic region, and the mean-variance frontier of all assets, i.e. including a risk free rate if there is one, which is the larger wedge-shaped region. Some authors reserve the terminology “meanvariance frontier” for the upper portion, calling the whole thing the minimum variance frontier. The risky asset frontier is a hyperbola, which means it lies between two asymptotes, shown as dotted lines. The risk free rate is typically drawn below the intersection of the asymptotes and the vertical axis, or the point of minimum variance on the risky frontier. If it were above this point, investors with a mean-variance objective would try to short the risky assets, which cannot represent an equilibrium. In general, portfolios of two assets or portfolios fill out a hyperbolic curve through the two assets. The curve is sharper the less correlated are the two assets, because the portfolio variance benefits from increasing diversification. Portfolios of a risky asset and risk free rate give rise to straight lines in mean-standard deviation space. In Chapter 1, we derived a similar wedge-shaped region as the set of means and variances of all assets that are priced by a given discount factor. This chapter is about incomplete markets, so we think of a mean-variance frontier generated by a given set of assets, typically less than complete. When does the mean-variance frontier exist? I.e., when is the set of portfolio means and variances less than the whole {E, σ} space? We basically have to rule out a special case: two returns are perfectly correlated but yield different means. In that case one could short one, long the other, and achieve infinite expected returns with no risk. More formally, eliminate purely redundant securities from consideration, then 80

S ECTION 5.2

M EAN - VARIANCE FRONTIER : I NTUITION AND L AGRANGIAN CHARACTERIZATION

Theorem: So long as the variance-covariance matrix of returns is non-singular, there is a mean-variance frontier. To prove this theorem, just follow the construction below. This theorem should sound very familiar: Two perfectly correlated returns with different mean are a violation of the law of one price. Thus, the law of one price implies that there is a mean variance frontier as well as a discount factor. E(R)

Mean-variance frontier Risky asset frontier

Tangency portfolio of risky assets Original assets Rf

σ(R)

Figure 13. Mean-variance frontier

5.2.1

Lagrangian approach to mean-variance frontier

The standard definition and the computation of the mean-variance frontier follows a brute force approach. Problem: Start with a vector of asset returns R. Denote by E the vector £ of mean returns, 0¤ E ≡ E(R), and denote by Σ the variance-covariance matrix Σ = E (R − E)(R − E) . A portfolio is defined by its weights w on the initial securities. The portfolio return is w0 R where the weights sum to one, w0 1 =1. The problem “choose a portfolio to minimize variance for a given mean” is then min{w} w0 Σw s.t. w0 E = µ; w0 1 = 1. Solution: Let A = E 0 Σ−1 E; B = E 0 Σ−1 1; C = 10 Σ−1 1. 81

(59)

C HAPTER 5

M EAN - VARIANCE FRONTIER AND BETA REPRESENTATIONS

Then, for a given mean portfolio return µ, the minimum variance portfolio has variance var (Rp ) = and is formed by portfolio weights w = Σ−1

Cµ2 − 2Bµ + A AC − B 2

(60)

E (Cµ − B) + 1 (A − Bµ) . (AC − B 2 )

Equation (5.60) shows that the variance is a quadratic function of the mean. The square root of a parabola is a hyperbola, which is why we draw hyperbolic regions in mean-standard deviation space. The minimum-variance portfolio is interesting in its own right. It appears as a special case in many theorems and it appears in several test statistics. We can find it by minimizing (5.60) over µ, giving µmin var = B/C. The weights of the minimum variance portfolio are thus w = Σ−1 1/(10 Σ−1 1). We can get to any point on the mean-variance frontier by starting with two returns on the frontier and forming portfolios. The frontier is spanned by any two frontier returns. To see this fact, notice that w is a linear function of µ. Thus, if you take the portfolios corresponding to any two distinct mean returns µ1 and µ2 , the weights on a third portfolio with mean µ3 = λµ1 + (1 − λ)µ2 are given by w3 = λw1 + (1 − λ)w2 . Derivation: To derive the solution, introduce Lagrange multipliers 2λ and 2δ on the constraints. The first order conditions to (5.59) are then Σw−λE − δ1 = 0 w = Σ−1 (λE + δ1).

(61)

We find the Lagrange multipliers from the constraints, E 0 w = E 0 Σ−1 (λE + δ1) = µ 10 w = 10 Σ−1 (λE + δ1) = 1 or

·

E 0 Σ−1 E 10 Σ−1 E ·

A B

E 0 Σ−1 1 10 Σ−1 1

¸·

λ δ

¸

=

¸·

¸

·

µ 1

¸

B C

λ δ

82

=

·

µ 1

¸

S ECTION 5.3

A N ORTHOGONAL CHARACTERIZATION OF THE MEAN - VARIANCE FRONTIER

Hence, λ=

Cµ−B AC−B 2

δ=

A−Bµ AC − B 2

Plugging in to (5.61), we get the portfolio weights and variance.

5.3

An orthogonal characterization of the mean-variance frontier

Every return can be expressed as Ri = R∗ + wi Re∗ + ni . The mean-variance frontier is Rmv = R∗ + wRe∗ Re∗ is defined as Re∗ = proj(1|Re ). It represents mean excess returns, E(Re ) = E(Re∗ Re ) ∀R ∈ Re e

The Lagrangian approach to the mean-variance frontier is straightforward but cumbersome. Our further manipulations will be easier if we follow an alternative approach due to Hansen and Richard (1987). Technically, Hansen and Richard’s approach is also valid when we can’t generate the payoff space by portfolios of a finite set of basis payoffs c0 x. This happens, for example, when we think about conditioning information in Chapter 8. Also, it is the natural geometric way to think about the mean-variance frontier given that we have started to think of payoffs, discount factors and other random variables as vectors in the space of payoffs. Rather than write portfolios as combinations of basis assets, and pose and solve a minimization problem, we first describe any return by a three-way orthogonal decomposition. The mean-variance frontier then pops out easily without any algebra. 5.3.1

Definitions of R∗ , Re∗

I start by defining two special returns. R∗ is the return corresponding to the payoff x∗ that can act as the discount factor. The price of x∗ , is, like any other price, p(x∗ ) = E(x∗ x∗ ). Thus, The definition of R∗ is R∗ ≡

x∗ x∗ = p(x∗ ) E(x∗2 ) 83

(62)

C HAPTER 5

M EAN - VARIANCE FRONTIER AND BETA REPRESENTATIONS

The definition of Re∗ is Re∗ ≡ proj(1 | Re )

(63)

Re ≡ space of excess returns = {x ∈ X s.t. p(x) = 0} Why Re∗ ? We are heading towards a mean-variance frontier, so it is natural to seek a special return that changes means. Re∗ is an excess return that represents means on Re with an inner product in the same way that x∗ is a payoff in X that represents prices with an inner product. As p(x) = E(mx) = E[proj(m|X)x] = E(x∗ x), so E(Re ) = E(1 × Re ) = E [proj(1 | Re ) × Re ] = E(Re∗ Re ). If R∗ and Re∗ are still a bit mysterious at this point, they will make more sense as we use them, and discover their many interesting properties. Now we can state a beautiful orthogonal decomposition. Theorem: Every return Ri can be expressed as Ri = R∗ + wi Re∗ + ni where wi is a number, and ni is an excess return with the property E(ni ) = 0. The three components are orthogonal, E(R∗ Re∗ ) = E(R∗ ni ) = E(Re∗ ni ) = 0. This theorem quickly implies the characterization of the mean variance frontier which we are after: Theorem: Rmv is on the mean-variance frontier if and only if Rmv = R∗ + wRe∗

(64)

for some real number w. As you vary the number w, you sweep out the mean-variance frontier. E(Re∗ ) 6= 0, so adding more w changes the mean and variance of Rmv . You can interpret (5.64) as a “two84

S ECTION 5.3

A N ORTHOGONAL CHARACTERIZATION OF THE MEAN - VARIANCE FRONTIER

fund” theorem for the mean-variance frontier. It expresses every frontier return as a portfolio of R∗ and Re∗ , with varying weights on the latter. As usual, first I’ll argue why the theorems are sensible, then I’ll offer a simple algebraic proof. Hansen and Richard (1987) give a much more careful algebraic proof. 5.3.2

Graphical construction

Figure 14 illustrates the decomposition. Start at the origin (0). Recall that the x∗ vector is perpendicular to planes of constant price; thus the R∗ vector lies perpendicular to the plane of returns as shown. Go to R∗ . Re∗ is the excess return that is closest to the vector 1; it lies at right angles to planes (in R ) of constant mean return, shown in the E = 1, E = 2 lines, just as the return R∗ lies at right angles to planes of constant price. Since Re∗ is an excess return, it is orthogonal to R∗ . Proceed an amount wi in the direction of Re∗ , getting as close to Ri as possible. e

Now move, again in an orthogonal direction, by an amount ni to get to the return Ri . We have thus expressed Ri = R∗ +wi Re∗ +ni in a way that all three components are orthogonal. Returns with n = 0, R∗ + wRe∗ , are the mean-variance frontier. Here’s why. Since E(R2 ) = σ 2 (R) + E(R)2 , we can define the mean-variance frontier by minimizing second moment for a given mean. The length of each vector in Figure 14 is its second moment, so we want the shortest vector that is on the return plane for a given mean. The shortest vectors in the return plane with given mean are on the R∗ + wRe∗ line. The graph also shows how Re∗ represents means in the space of excess returns. Expectation is the inner product with 1. Planes of constant expected value in Figure 14 are perpendicular to the 1 vector, just as planes of constant price are perpendicular to the x∗ or R∗ vectors. I don’t show the full extent of the constant expected payoff planes for clarity; I do show lines of constant expected excess return in Re , which are the intersection of constant expected payoff planes with the Re plane. Therefore, just as we found an x∗ in X to represent prices in X by projecting m onto X, we find Re∗ in Re by projecting of 1 onto Re . Yes, a regression with one on the left hand side. Planes perpendicular to Re∗ in Re are payoffs with constant mean, just as planes perpendicular to x∗ in X are payoffs with the same price. 5.3.3

Algebraic argument

Now, an algebraic proof of the decomposition and characterization of mean variance frontier. The algebra just represents statements about what is at right angles to what with second moments. Proof: Straight from their definitions, (5.62) and (5.63) we know that Re∗ is an

85

C HAPTER 5

M EAN - VARIANCE FRONTIER AND BETA REPRESENTATIONS

R=space of returns (p=1) R*+wiRe* R*

ni Ri=R*+wiRe*+ni

1

Re*

0

E=1

E=0

Re = space of excess returns (p=0) Figure 14. Orthogonal decomposition and mean-variance frontier. excess return (price zero), and hence that R∗ and Re∗ are orthogonal, E(R∗ Re∗ ) =

E(x∗ Re∗ ) = 0. E(x∗2 )

We define ni so that the decomposition adds up to Ri as claimed, and we define wi to make sure that ni is orthogonal to the other two components. Then we prove that E(ni ) = 0. Pick any wi and then define ni ≡ Ri − R∗ − wi Re∗ . ni is an excess return so already orthogonal to R∗ , E(R∗ ni ) = 0. To show E(ni ) = 0 and ni orthogonal to Re∗ , we exploit the fact that since ni is an excess return, E(ni ) = E(Re∗ ni ). Therefore, Re∗ is orthogonal to ni if and only if we pick wi so that E(ni ) = 0. We 86

S ECTION 5.4

S PANNING THE MEAN - VARIANCE FRONTIER

don’t have to explicitly calculate wi for the proof.3 Once we have constructed the decomposition, the frontier drops out. Since E(ni ) = 0 and the three components are orthogonal, E(Ri ) = E(R∗ ) + wi E(Re∗ )

σ2 (Ri ) = σ2 (R∗ + wi Re∗ ) + σ2 (ni ). Thus, for each desired value of the mean return, there is a unique wi . Returns with ¥ ni = 0 minimize variance for each mean. 5.3.4

Decomposition in mean-variance space

Figure 15 illustrates the decomposition in mean-variance space rather than in state-space. First, let’s locate R∗ . R∗ is the minimum second moment return. One can see this fact from the geometry of Figure 14: R∗ is the return closest to the origin, and thus the return with the smallest “length” which is second moment. As with OLS regression, minimizing the length of R∗ and creating an R∗ orthogonal to all excess returns is the same thing. One can also verify this property algebraically. Since any return can be expressed as R = R∗ + wRe∗ + n, E(R2 ) = E(R∗2 ) + w2 E(Re∗2 ) + E(n2 ). n = 0 and w = 0 thus give the minimum second moment return. In mean-standard deviation space, lines of constant second moment are circles. Thus, the minimum second-moment return R∗ is on the smallest circle that intersects the set of all assets, which lie in the mean-variance frontier in the right hand panel of Figure 19. Notice that R∗ is on the lower, or “inefficient” segment of the mean-variance frontier. It is initially surprising that this is the location of the most interesting return on the frontier! R∗ is not the “market portfolio” or “wealth portfolio,” which typically lie on the upper portion of the frontier. Adding more Re∗ moves one along the frontier. Adding n does not change mean but does change variance, so it is an idiosyncratic return that just moves an asset off the frontier as graphed. α is the “zero-beta rate” corresponding to R∗ . It is the expected return of any return that is uncorrelated with R∗ . I demonstrate these properties in section 6.5. 3

Its value wi =

E(Ri ) − E(R∗ ) E(Re∗ )

is not particularly enlightening.

87

C HAPTER 5

M EAN - VARIANCE FRONTIER AND BETA REPRESENTATIONS

E(R)

Ri

R* + wiRe* ni

R*

σ(R) Figure 15. Orthogonal decomposition of a return Ri in mean-standard deviation space.

5.4

Spanning the mean-variance frontier

The characterization of the mean-variance frontier in terms of R∗ and Re∗ is most natural in our setup. However, you can equivalently span the mean-variance frontier with any two portfolios that are on the frontier – any two distinct linear combinations of R∗ and Re∗ . In particular, take any return Rα = R∗ + γRe∗ , γ 6= 0.

(65)

Using this return in place of Re∗ , Rα − R∗ γ

Re∗ =

you can express the mean variance frontier in terms of R∗ and Rα : R∗ + wRe∗

= R∗ + y (Rα − R∗ ) = (1 − y)R∗ + yRα

where I have defined a new weight y = w/γ. 88

(5.66)

S ECTION 5.5

A COMPILATION OF PROPERTIES OF R∗ , Re∗ AND x∗

The most common alternative approach is to use a risk free rate or a risky rate that somehow behaves like the risk free rate in place of Re∗ to span the frontier. When there is a risk free rate, it is on the frontier with representation Rf = R∗ + Rf Re∗ I derive this expression in equation (5.72) below. Therefore, we can use (5.66) with Ra = Rf . When there is no risk free rate, several risky returns that retain some properties of the risk free rate are often used. In section 5.3 below I present a “zero beta” return, which is uncorrelated with R∗ , a “constant-mimicking portfolio” return, which is the return on the traded payoff ˆ = proj(1|X)/p[proj(1|X)] and the minimum variance return. Each of closest to unity, R these returns is on the mean-variance frontier, with form 5.65, though different values of α. Therefore, we can span the mean-variance frontier with R∗ and any of these risk-free rate proxies.

5.5

A compilation of properties of R∗ , Re∗ and x∗

The special returns R∗ , Re∗ that generate the mean variance frontier have lots of interesting and useful properties. Some I derived above, some I will derive and discuss below in more detail, and some will be useful tricks later on. Most properties and derivations are extremely obscure if you don’t look at the picture! 1) 1 . E(x∗2 )

E(R∗2 ) =

(67)

To derive this fact, multiply both sides of (5.62) by R∗ , take expectations, and remember R∗ is a return so 1 = E(x∗ R∗ ). 2) We can reverse the definition and recover x∗ from R∗ via x∗ =

R∗ . E(R∗2 )

(68)

To derive this formula, start with the definition R∗ = x∗ /E(x∗2 ) and substitute from (5.67) for E(x∗2 ) 3) R∗ can be used to represent prices just like x∗ . This is not surprising, since they both point in the same direction, orthogonal to planes of constant price. Most obviously, from 5.68 p(x) = E(x∗ x) =

E (R∗ x) ∀x ∈ X E(R∗2 )

For returns, we can nicely express this result as E(R∗2 ) = E(R∗ R) ∀R ∈ R. 89

(69)

C HAPTER 5

M EAN - VARIANCE FRONTIER AND BETA REPRESENTATIONS

This fact can also serve as an alternative defining property of R∗ . 4) Re∗ represents means on Re via an inner product in the same way that x∗ represents prices on X via an inner product. Re∗ is orthogonal to planes of constant mean in Re as x∗ is orthogonal to planes of constant price. Algebraically, in analogy to p(x) = E(x∗ x) we have E(Re ) = E(Re∗ Re ) ∀Re ∈ Re .

(70)

This fact can serve as an alternative defining property of Re∗ . 5) Re∗ and R∗ are orthogonal, E(R∗ Re∗ ) = 0. More generally, R∗ is orthogonal to any excess return. 6) The mean variance frontier is given by Rmv = R∗ + wRe∗ . ¤ £ To prove this, E(R2 ) = E (R∗ + wRe∗ + n)2 = E(R∗2 ) + w2 E(Re2 ) + E(n2 ), and E(n) = 0, so set n to zero. The conditional mean-variance frontier allows w in the conditioning information set. The unconditional mean variance frontier requires w to equal a constant. 7) R∗ is the minimum second moment return. Graphically, R∗ is the return closest to the origin. To see this, using the decomposition in #6, and set w2 and n to zero to minimize second moment. 8) Re∗ has the same first and second moment, E(Re∗ ) = E(Re∗2 ). Just apply fact (5.70) to Re∗ itself. Therefore var(Re∗ ) = E(Re∗2 ) − E(Re∗ )2 = E(Re∗ ) [1 − E(Re∗ )] . 9) If there is a riskfree rate, then Re∗ can also be defined as the residual in the projection of 1 on R∗ : Re∗ = 1 − proj(1|R∗ ) = 1 −

E(R∗ ) ∗ 1 R = 1 − f R∗ ∗2 E(R ) R

(71)

You’d never have thought of this without looking at Figure 14! Since R∗ and Re are orthogonal and together span X, 1 = proj(1|Re ) + proj(1|R∗ ). You can also verify this statement analytically. Check that Re∗ so defined is an excess return in X – its price is zero–, and E(Re∗ Re ) = E(Re ); E(R∗ Re∗ ) = 0. 90

S ECTION 5.5

A COMPILATION OF PROPERTIES OF R∗ , Re∗ AND x∗

As a result, Rf has the decomposition Rf = R∗ + Rf Re∗ .

(72)

Since Rf > 1 typically, this means that R∗ +Re∗ is located on the lower portion of the meanvariance frontier in mean-variance space, just a bit to the right of Rf . If the risk free rate were one, then the unit vector would lie in the return space, and Rf = R∗ + Re∗ . Typically, the space of returns is a little bit above the unit vector. As you stretch the unit vector by the amount Rf to arrive at the return Rf , so you stretch the amount Re∗ that you add to R∗ to get to Rf . If there is no riskfree rate, then we can use proj(1|X) = proj(proj (1|X) |Re ) + proj(proj (1|X) |R∗ ) = proj(1|Re ) + proj(1|R∗ ) to deduce an analogue to equation (5.71), Re∗ = proj(1|X) − proj(1|R∗ ) = proj(1|X) −

E(R∗ ) ∗ R E(R∗2 )

(73)

10) If a riskfree rate is traded, we can construct Rf from R∗ via Rf =

E(R∗2 ) 1 = . ∗ E(x ) E(R∗ )

(74)

If not, this gives a “zero beta rate” interpretation of the right hand expression. 11) Since we have a formula x∗ = p0 E(xx0 )−1 x for constructing x∗ from basis assets (see section 4.1), we can construct R∗ in this case from R∗ =

p0 E(xx0 )−1 x x∗ = . p(x∗ ) p0 E(xx0 )−1 p

(p(x∗ ) = E(x∗ x∗ ) leading to the denominator.) 12) We can construct Re∗ from a set of basis assets as well. Following the definition to project one on the space of excess returns, Re∗ = E(Re )0 E(Re Re0 )−1 Re where Re is the basis set of excess returns. (You can always use Re = R−R∗ if you want). This construction obviously mirrors the way we constructed x∗ in section 4.1, and you can see the similarity in the result, with E in place of p, since Re∗ represents means rather than prices. . 91

C HAPTER 5

M EAN - VARIANCE FRONTIER AND BETA REPRESENTATIONS

If there is a riskfree rate, we can also use (5.71), Re∗ = 1 −

1 ∗ 1 p0 E(xx0 )−1 x . R = 1 − Rf Rf p0 E(xx0 )−1 p

(75)

If there is no riskfree rate, we can use (5.73) to construct Re∗ . The central ingredient is proj(1|X) = E(x)0 E(xx0 )−1 x.

5.6

Mean-variance frontiers for m: the Hansen-Jagannathan bounds

The mean-variance frontier of all discount factors that price a given set of assets is related to the mean-variance frontier of asset returns by |E(Re )| σ(m) ≥ . E(m) σ(Re ) and hence σ(m) E(Re ) = max e {all m that price x∈X} E(m) {all excess returns R in X} σ(Re ) min

The discount factors on the frontier can be characterized analogously to the mean-variance frontier of asset returns, m = x∗ + we∗ 0 e∗ ≡ 1 − proj(1|X) = proj(1|E) = 1 − E(x) E(xx0 )−1 x E = {m − x∗ } . We derived in Chapter 1 a relation between the Sharpe ratio of an excess return and the volatility of discount factors necessary to price that return, |E(Re )| σ(m) ≥ . E(m) σ(Re ) Quickly, 0 = E(mRe ) = E(m)E(Re ) + ρm,Re σ(m)σ(Re ), 92

S ECTION 5.6

M EAN - VARIANCE FRONTIERS FOR m: THE H ANSEN -JAGANNATHAN BOUNDS

and |ρ| ≤ 1. If we had a riskfree rate, then we know in addition E(m) = 1/Rf . Hansen and Jagannathan (1991) had the brilliant insight to read this equation as a restriction on the set of discount factors that can price a given set of returns, as well as a restriction on the set of returns we will see given a specific discount factor. This calculation teaches us that we need very volatile discount factors with a mean near one to understand stock returns. This and more general related calculations turn out to be a central tool in understanding and surmounting the equity premium puzzle, surveyed in Chapter 21. We would like to derive a bound that uses a large number of assets, and that is valid if there is no riskfree rate. What is the set of {E(m), σ(m)} consistent with a given set of asset prices and payoffs? What is the mean-variance frontier for discount factors? Obviously, the higher the Sharpe ratio, the tighter the bound. This suggests a way to construct the frontier we’re after. For any hypothetical risk-free rate, find the highest Sharpe ratio. That is, of course the tangency portfolio. Then the slope to the tangency portfolio gives the ratio σ(m)/E(m). Figure 16 illustrates. E(R)

σ(m) = Ε(Re)/σ(Re) 1/E(m)

E(Re)/σ(Re)

σ(R)

E(m)

Figure 16. Graphical construction of the Hansen-Jagannathan bound. As we sweep through values of E(m), the slope to the tangency becomes lower, and the Hansen-Jagannathan bound declines. At the mean return corresponding to the minimum variance point, the HJ bound attains its minimum. Continuing, the Sharpe ratio rises again and so does the bound. If there were a riskfree rate, then we know E(m), the return frontier is a V shape, and the HJ bound is purely a bound on variance. This discussion implies a beautiful duality between discount factor volatility and Sharpe ratios. σ(m) E(Re ) = max e . {all m that price x∈X} E(m) {all excess returns R in X} σ(Re ) min

93

(76)

C HAPTER 5

M EAN - VARIANCE FRONTIER AND BETA REPRESENTATIONS

We need formulas for an explicit calculation. In equation (), we found a representation for the set of discount factors that price a given set of asset returns – that satisfy p = E(mx) : m = E(m) + [p − E(m)E(x)] Σ−1 [x − E(x)] + ε

(77)

where Σ ≡ cov(x, x0 ) and E(ε) = 0, E(εx) = 0. You can think of this as a regression or projection of any discount factor on the space of payoffs, plus an error. Since σ2 (ε) > 0, this representation leads immediately to an explicit expression for the Hansen-Jagannathan bound, 0 σ 2 (m) ≥ (p − E(m)E(x)) Σ−1 (p − E(m)E(x)) .

(78)

As all asset returns must lie in a cup-shaped σ(R)} space, all discount ª © region in {E(R), factors must lie in a parabolic region in E(m), σ 2 (m) space, as illustrated in the right hand panel of Figure 16.

We would like an expression for the discount factors on the bound, as we wanted an expression for the returns on the mean variance frontier instead of just a formula for the means and variances. As we found a three way decomposition of all returns, two of which generated the mean-variance frontier of returns, so we can find a three way decomposition of discount factors, two of which generate the mean-variance frontier of discount factors (5.78). I illustrate the construction in Figure 17. Any m must line in the plane marked M, perpendicular to X through x∗ . Any m must be of the form m = x∗ + we∗ + n. Here, I have just broken up the residual ε in the familiar representation m = x∗ + ε into two components. e∗ is defined as the residual from the projection of 1 onto X or, equivalently the projection of 1 on the space E of “excess m’s,” random variables of the form m − x∗ . e∗ ≡ 1 − proj(1|X) = proj(1|E). e∗ generates means of m just as Re∗ did for returns: E(m − x∗ ) = E[1 × (m − x∗ )] = E[proj(1|E)(m − x∗ )]. Finally n, defined as the leftovers, has mean zero since it’s orthogonal to 1 and is orthogonal to X. As with returns, then, the mean-variance frontier of m0 s is given by m∗ = x∗ + we∗ .

(79)

If the unit payoff is in the payoff space, then we know E(m), and the frontier and bound are just m = x∗ , σ 2 (m) ≥ σ2 (x∗ ). This is exactly like the case of risk-neutrality for return mean-variance frontiers. 94

S ECTION 5.6

M EAN - VARIANCE FRONTIERS FOR m: THE H ANSEN -JAGANNATHAN BOUNDS

X = payoff space M = space of discount factors x*+we* x*

n m = x*+we*+n

1

proj(1| X)

e*

0

E(.)=1

E(.)=0 E = space of m-x*

Figure 17. Decomposition of any discount factor m = x∗ + we + n.

95

C HAPTER 5

M EAN - VARIANCE FRONTIER AND BETA REPRESENTATIONS

The construction (5.79) can be used to derive the formula (5.78) for the Hansen-Jagannathan bound for the finite-dimensional cases discussed above. It’s more general, since it can be used in infinite-dimensional payoff spaces as well. Along with the corresponding return formula Rmv = R∗ + wRe∗ , we see in Chapter 8 that it extends more easily to the calculation of conditional vs. unconditional mean-variance frontiers (Gallant, Hansen and Tauchen 1995). It will make construction (5.79) come alive if we find equations for its components. We find x∗ as before, it is the portfolio c0 x in X that prices x: x∗ = p0 E(xx0 )−1 x. Similarly, let’s find e∗ . The projection of 1 on X is proj(1|X) = E(x)0 E(xx0 )−1 x. (After a while you get used to the idea of running regressions with 1 on the left hand side and random variables on the right hand side!) Thus, e∗ = 1 − E(x)0 E(xx0 )−1 x. Again, you can construct time-series of x∗ and e∗ from these definitions.

or

Finally, we now can construct the variance-minimizing discount factors £ ¤ m∗ = x∗ + we∗ = p0 E(xx0 )−1 x + w 1 − E(x)0 E(xx0 )−1 x 0 m∗ = w + [p − wE(x)] E(xx0 )−1 x

(80)

As w varies, we trace out discount factors m∗ on the frontier with varying means and variances. It’s easiest to find mean and second moment: E(m∗ ) = w + [p − wE(x)]0 E(xx0 )−1 E(x) 0 E(m∗2 ) = [p − wE(x)] E(xx0 )−1 [p − wE(x)] ;

variance follows from σ 2 (m) = E(m2 ) − E(m)2 . With a little algebra one can also show that these formulas are equivalent to equation (5.78). As you can see, Hansen-Jagannathan frontiers are equivalent to mean-variance frontiers. For example, an obvious exercise is to see how much the addition of assets raises the HansenJagannathan bound. This is exactly the same as asking how much those assets expand the mean-variance frontier. It was, in fact, this link between Hansen-Jagannathan bounds and mean-variance frontiers rather than the logic I described that inspired Knez and Chen (1996) and DeSantis (1994) to test for mean-variance efficiency using, essentially, HansenJagannathan bounds. 96

S ECTION 5.7

P ROBLEMS

Hansen-Jagannathan bounds have the potential to do more than mean-variance frontiers. Hansen and Jagannathan show how to solve the problem min σ2 (m) s.t. p = E(mx), m > 0. This is the “Hansen-Jagannathan bound with positivity,” and is strictly tighter than the HansenJagannathan bound. It allows you to impose no-arbitrage conditions. In stock applications, this extra bound ended up not being that informative. However, in the option application of this idea of Chapter (18), positivity is really important. That chapter shows how to solve for a bound with positivity. Hansen, Heaton and Luttmer (1995) develop a distribution theory for the bounds. Luttmer (1996) develops bounds with market frictions such as short-sales constraints and bid-ask spreads, to account for ludicrously high apparent Sharpe ratios and bounds in short term bond data. Cochrane and Hansen (1992) survey a variety of bounds, including bounds that incorporate information that discount factors are poorly correlated with stock returns (the HJ bounds use the extreme ρ = 1), bounds on conditional moments that illustrate how many models imply excessive interest rate variation, bounds with short-sales constraints and market frictions, etc. Chapter 21 discusses what the results of Hansen Jagannathan bound calculations and what they mean for discount factors that can price stock and bond return data.

5.7 1. 2. 3. 4.

5. 6. 7.

Problems

Prove that Re∗ lies at right angles to planes (in Re ) of constant mean return, as shown in figure 14. Should we typically draw x∗ above, below or on the plane of returns? Must x∗ always lie in this position? Can you construct Re∗ from knowledge of m, x∗ , or R∗ ? What happens to R∗ , Re∗ and the mean-variance frontier if investors are risk neutral? (a) If a riskfree rate is traded. (b) If no riskfree rate is traded? (Hint: make a drawing or think about the case that payoffs are generated by an N dimensional vector of basis assets x) x∗ = proj(m|X). Is R∗ = proj(m|R)? We showed that all m are of the form x∗ + ε. What about R−1 R? Show that if there is a risk-free rate—if the unit payoff is in the payoff space X—then Re∗ = (Rf − R∗ )/Rf .

97

Chapter 6. Relation between discount factors, betas, and mean-variance frontiers In this chapter, I draw the connection between discount factors, mean-variance frontiers, and beta representations. In the first chapter, I showed how mean-variance and a beta representation follow from p = E(mx) and (in the mean-variance case) complete markets. Here, I discuss the connections in both directions and in incomplete markets, drawing on the representations studied in the last chapter. The central theme of the chapter is that all three representations are equivalent. Figure 18 summarizes the ways one can go from one representation to another. A discount factor, a reference variable for betas – the thing you put on the right hand side in the regressions that give betas – and a return on the mean-variance frontier all carry the same information, and given any one of them, you can find the others. More specifically, 1. 2. 3. 4. 5.

p = E(mx) ⇒ β: Given m such that p = E(mx), m, x∗ , R∗ , or R∗ + wRe∗ all can serve as reference variables for betas. p = E(mx) ⇒ mean-variance frontier. You can construct R∗ from x∗ = proj(m|X), R∗ = x∗ /E(x∗2 ), and then R∗ , R∗ + wRe∗ are on the mean-variance frontier. Mean-variance frontier ⇒ p = E(mx). If Rmv is on the mean-variance frontier, then m = a + bRmv linear in that return is a discount factor; it satisfies p = E(mx). β ⇒ p = E(mx). If we have an expected return/beta model with factors f , then m = b0 f linear in the factors satisfies p = E(mx) (and vice-versa). If a return is on the mean-variance frontier, then there is an expected return/beta model with that return as reference variable. The following subsections discuss the mechanics of going from one representation to the other in detail. The last section of the chapter collects some special cases when there is no riskfree rate. The next chapter discusses some of the implications of these equivalence theorems, and why they are important. Roll (197x) pointed out the connection between mean-variance frontiers and beta pricing. Ross (1978) and Dybvig and Ingersoll (1982) pointed out the connection between linear discount factors and beta pricing. Hansen and Richard (1987) pointed out the connection between a discount factor and the mean-variance frontier.

6.1

From discount factors to beta representations

m, x∗ , and R∗ can all be the single factor in a single beta representation.

98

S ECTION 6.1

F ROM DISCOUNT FACTORS TO BETA REPRESENTATIONS

LOOP Þ m exists

p = E(mx) f = m, x*, R* m = b’f

m = a + bRmv R* is on m.v.f.

E(Ri) = α + βi’λ f = Rmv proj(f|R) on m.v.f.

Rmv on m.v.f.

E(RR’) nonsingular Þ Rmv exists

Figure 18. Relation between three views of asset pricing. 6.1.1

Beta representation using m

p = E(mx) implies E(Ri ) = α + β i,m λm . Start with 1 = E(mRi ) = E(m)E(Ri ) + cov(m, Ri ). Thus, E(Ri ) =

cov(m, Ri ) 1 − . E(m) E(m)

Multiply and divide by var(m), define α ≡ 1/E(m) to get µ ¶µ ¶ cov(m, Ri ) var(m) i E(R ) = α + − = α + β i,m λm . var(m) E(m) As advertised, we have a single beta representation. For example, we can equivalently state the consumption-based model as: mean asset returns should be linear in the regression betas of asset returns on (ct+1 /ct )−γ . Furthermore, the slope of this cross-sectional relationship λm is not a free parameter, though it is usually treated as such in empirical evaluation of factor pricing models. λm should equal the ratio of 99

C HAPTER 6

R ELATION BETWEEN DISCOUNT FACTORS , BETAS , AND MEAN - VARIANCE FRONTIERS

variance to mean of (ct+1 /ct )−γ . The factor risk premium λm for marginal utility growth is negative. Positive expected returns are associated with positive correlation with consumption growth, and hence negative correlation with marginal utility growth and m. Thus, we expect λm < 0. 6.1.2

β representation using x∗ and R∗

It is often useful to express a pricing model in a way that the factor is a payoff rather than a real factor such as consumption growth. In applications, we can then avoid measurement difficulties of real data. We have already seen the idea of “factor mimicking portfolios” formed by projection: project m on to X, and the resulting payoff x∗ also serves as a discount factor. Unsurprisingly, x∗ can also serve as the factor in an expected return-beta representation. It’s even more useful if the reference payoff is a return. Unsurprisingly, the return R∗ = x∗ /E(x∗2 ) can also serve as the factor in a beta pricing model. When the factor is also a return, the model is particularly simple, since the factor risk premium is also the expected excess return. Theorem. 1 = E(mRi ) implies an expected return - beta model with x∗ = proj(m|X) or R∗ ≡ x∗ /E(x∗2 ) as factors, e.g. E(Ri ) = α + β i,x∗ λx∗ and E(Ri ) = α + β i,R∗ [E(R∗ ) − α]. Proof: Recall that p = E(mx) implies p = E [proj(m | X) x], or p = E(x∗ x). Then 1 = E(mRi ) = E(x∗ Ri ) = E(x∗ )E(Ri ) + cov(x∗ , Ri ). Solving for the expected return, E(Ri ) =

cov(x∗ , Ri ) 1 cov(x∗ , Ri ) var(x∗ ) 1 − = − ∗ ∗ ∗ E(x ) E(x ) E(x ) var(x∗ ) E(x∗ )

which we can write as the desired single-beta model, E(Ri ) = α + β i,x∗ λx∗ . Notice that the zero-beta rate 1/E(x∗ ) appears when there is no riskfree rate. To derive a single beta representation with R∗ , recall the definition, R∗ =

x∗ E(x∗2 )

Substituting R∗ for x∗ , equation (6.81) implies that we can in fact construct a return R∗ from m that acts as the single factor in a beta model, µ ¶µ ¶ E(R∗2 ) cov(R∗ , Ri ) E(R∗2 ) cov(R∗ , Ri ) var(R∗ ) i − = + E(R ) = − E(R∗ ) E(R∗ ) E(R∗ ) var(R∗ ) E(R∗ ) 100

(81)

S ECTION 6.2

F ROM MEAN - VARIANCE FRONTIER TO A DISCOUNT FACTOR AND BETA REPRESENTATION

or, defining Greek letters in the obvious way, E(Ri ) = α + β Ri ,R∗ λR∗

(82)

Since the factor R∗ is also a return, its expected excess return over the zero beta rate gives the factor risk premium λR∗ . Applying equation (6.82) to R∗ itself, E(R∗ ) = α −

var(R∗ ) . E(R∗ )

(83)

So we can write the beta model in an even more traditional form E(Ri ) = α + β Ri ,R∗ [E(R∗ ) − α].

(84)

¥ Recall that R∗ is the minimum second moment frontier, on the lower portion of the meanvariance frontier. This is why R∗ has an unusual negative expected excess return or factor risk premium, λR∗ = −var(R∗ )/E(R∗ ) < 0. α is the zero-beta rate on R∗ shown in Figure15. Special cases A footnote to these constructions is that E(m), E(x∗ ), or E(R∗ ) cannot be zero, or you couldn’t divide by them. This is a pathological case: E(m) = 0 implies a zero price for the riskfree asset, and an infinite riskfree rate. If a riskfree rate is traded, we can simply observe that it is not infinite and verify the fact. Also, in a complete market, E(m) cannot be zero since, by absence of arbitrage, m > 0. We will see similar special cases in the remaining theorems: the manipulations only work for discount factor choices that do not imply zero or infinite riskfree rates. I discuss the issue in section 6.6 The manipulation from expected return-covariance to expected return-beta breaks down if var(m), var(x∗ ) or var(R∗ ) are zero. This is the case of pure risk neutrality. In this case, the covariances go to zero faster than the variances, so all betas go to zero and all expected returns become the same as the risk free rate.

6.2

From mean-variance frontier to a discount factor and beta representation

Rmv is on mean-variance frontier ⇒ m = a + bRmv ; E(Ri ) − α = β i [E(Rmv ) − α] We have seen that p = E(mx) implies a single−β model with a mean-variance efficient reference return, namely R∗ . The converse is also true: for (almost) any return on the meanvariance frontier, we can define a discount factor m that prices assets as a linear function of 101

C HAPTER 6

R ELATION BETWEEN DISCOUNT FACTORS , BETAS , AND MEAN - VARIANCE FRONTIERS

the mean-variance efficient return. Also, expected returns mechanically follow a single−β representation using the mean-variance efficient return as reference. I start with the discount factor. Theorem: There is a discount factor of the form m = a + bRmv if and only if Rmv is on the mean-variance frontier, and Rmv is not the riskfree rate. (If there is no riskfree rate, if Rmv is not the constant-mimicking portfolio return.) Graphical argument The basic idea is very simple, and Figure 19 shows the geometry for the complete markets case. The discount factor m = x∗ is proportional to R∗ . The mean-variance frontier is R∗ + wRe∗ . Pick a vector Rmv on the mean-variance frontier as shown in Figure 19. Then stretch it (bRmv ) and then subtract some of the 1 vector (a). Since Re∗ is generated by the unit vector, we can get rid of the Re∗ component and get back to the discount factor x∗ if we pick the right a and b. If the original return vector were not on the mean-variance frontier, then any linear combination a + bRmv with b 6= 0 would point in some of the n direction, which R∗ and x∗ do not. If b = 0, though, just stretching up and down the 1 vector will not get us to x∗ . Thus, we can only get a discount factor of the form a + bRmv if Rmv is on the frontier. You may remember that x∗ is not the only discount factor – all discount factors are of the form m = x∗ + ε with E(εx) = 0. Perhaps a + bR gives one of these discount factors, when R is not on the mean-variance frontier? This doesn’t work, however; n is still in the payoff space X while, by definition, ε is orthogonal to this space. If the mean-variance efficient return Rmv that we start with happens to lie right on the intersection of the stretched unit vector and the frontier, then stretching the Rmv vector and adding some unit vector are the same thing, so we again can’t get back to x∗ by stretching and adding some unit vector. The stretched unit payoff is the riskfree rate, so the theorem rules out the riskfree rate. When there is no riskfree rate, we have to rule out the “constant-mimicking portfolio return.” I treat this case in section 6.1. Algebraic proof Now, an algebraic proof that captures the same ideas. Proof. For an arbitrary R, try the discount factor model m = a + bR = a + b(R∗ + wRe∗ + n). We show that this discount factor prices an arbitrary payoff if and only if n = 0, and except for the w choice that makes R the riskfree rate, or the constant-mimicking portfolio return if there is no riskfree rate. We can determine a and b by forcing m to price any two assets. I find a and b to 102

(85)

S ECTION 6.2

F ROM MEAN - VARIANCE FRONTIER TO A DISCOUNT FACTOR AND BETA REPRESENTATION

bRmv er E - σ fronti

Rmv

Rf

R* 1 x* = a + bRmv Re* 0

Figure 19. There is a discount factor m = a + bRmv if and only if Rmv is on the mean-variance frontier and not the risk free rate. make the model price R∗ and Re∗ . 1 = E(mR∗ ) = aE(R∗ ) + bE(R∗2 ) 0 = E(mRe∗ ) = aE(Re∗ ) + bwE(Re∗2 ) = (a + bw) E(Re∗ ). Solving for a and b, a =

w wE(R∗ ) −

E(R∗2 ) 1 . b = − wE(R∗ ) − E(R∗2 )

Thus, if it is to price R∗ and Re∗ , the discount factor must be m=

w − (R∗ + wRe∗ + n) . wE(R∗ ) − E(R∗2 )

Now, let’s see if m prices an arbitrary payoff xi . Any xi ∈ X can also be decom103

(86)

C HAPTER 6

R ELATION BETWEEN DISCOUNT FACTORS , BETAS , AND MEAN - VARIANCE FRONTIERS

posed as xi = y i R∗ + wi Re∗ + ni . (See Figure 14 if this isn’t obvious.) The price of xi is yi , since both Re∗ and ni are zero-price (excess return) payoffs. Therefore, we want E(mxi ) = y i . Does it? µ ¶ (w − R∗ − wRe∗ − n)(y i R∗ + wi Re∗ + ni ) i E(mx ) = E wE(R∗ ) − E(R∗2 ) Using the orthogonality of R∗ , Re∗ n; E(n) = 0 and E(Re∗2 ) = E(Re∗ ) to simplify the product, E(mxi ) =

wyi E(R∗ ) − y i E(R∗2 ) − E(nni ) E(nni ) i = y . − wE(R∗ ) − E(R∗2 ) wE(R∗ ) − E(R∗2 )

To get p(xi ) = y i = E(mxi ), we need E(nni ) = 0. The only way to guarantee this condition for every payoff xi ∈ X is to insist that n = 0. Obviously, this construction can’t work if the denominator of (6.86) is zero, i.e. if w = E(R∗2 )/E(R∗ ) = 1/E(x∗ ). If there is a riskfree rate, then Rf = 1/E(x∗ ), so we are ruling out the case Rmv = R∗ + Rf Re∗ , which is the risk free rate. If ˆ = R∗ + E(R∗2 )/E(R∗ )Re∗ as a “constant there is no riskfree rate, I interpret R mimicking portfolio return” in section 5.3, and I give a graphical interpretation of this special case in section 6.1 ¥ We can generalize the theorem somewhat. Nothing is special about returns; any payoff of the form yR∗ +wRe∗ or yx∗ +wRe∗ can be used to price assets; such payoffs have minimum variance among all payoffs with given mean and price. Of course, we proved existence not uniqueness: m = a + bRmv + ², E(²x) = 0 also price assets as always. To get from the mean-variance frontier to a beta pricing model, we can just chain this theorem and the theorem of the last section together. There is a slight subtlety about special cases when there is no riskfree rate, but since it is not important for the basic points I relegate the direct connection and the special cases to section 6.2.

6.3

Factor models and discount factors

Beta-pricing models are equivalent to linear models for the discount factor m. E(Ri ) = α + λ0 β i ⇔ m = a + b0 f

104

S ECTION 6.3

FACTOR MODELS AND DISCOUNT FACTORS

We have shown that p = E(mx) implies a single beta representation using m, x∗ or R as factors. Let’s ask the converse question: suppose we have an expected return - beta model such as CAPM, APT, ICAPM, etc. What discount factor model does this imply? I show that an expected return - beta model is equivalent to a model for the discount factor that is a linear function of the factors in the beta model. This is an important and central result. It gives the connection between the discount factor formulation emphasized in this book and the expected return/beta, factor model formulation common in empirical work. ∗

You can write a linear factor model most compactly as m = b0 f , letting one of the factors be a constant. However, since we want a connection to the beta representation based on covariances rather than second moments, it is easiest to fold means of the factors in to the constant, and write m = a + b0 f with E(f) = 0 and hence E(m) = a. The connection is easiest to see in the special case that all the test assets are excess returns. Then 0 = E(mRe ) does not identify the mean of m, and we can normalize a arbitrarily. I find it convenient to normalize to E(m) = 1, or m = 1 + b0 [f − E (f )]. Then, Theorem: Given the model m = 1 + b0 [f − E (f )] ; 0 = E(mRe )

(87)

one can find λ such that E(Re ) = β 0 λ

(88)

where β are the multiple regression coefficients of excess returns Re on the factors. Conversely, given λ in (6.88), we can find b such that (6.87) holds. Proof: From (6.87) 0 = E(mRe ) = E(Re ) + b0 cov(f, Re ) E(Re ) = −b0 cov(f, Re ). From covariance to beta is quick, E(Re ) = −b0 var(f )var(f )−1 cov(f, Re ) = λ0 β Thus, λ and b are related by λ = −var(f )b. ¥ When the test assets are returns, the same idea works just as well, but gets a little more drowned in algebra since we have to keep track of the constant in m and the zero-beta rate in the beta model. 105

C HAPTER 6

R ELATION BETWEEN DISCOUNT FACTORS , BETAS , AND MEAN - VARIANCE FRONTIERS

Theorem: Given the model m = a + b0 f, 1 = E(mRi ),

(89)

one can find α and λ such that E(Ri ) = α + λ0 β i ,

(90)

where β i are the multiple regression coefficients of Ri on f with a constant. Conversely, given α and λ in a factor model of the form (6.90), one can find a, b such that (6.89) holds. Proof: We just have to construct the relation between (α, λ) and (a, b) and show that it works. Start with m = a + b0 f , 1 = E(mR), and hence E(R) =

cov(m, R) 1 E(Rf 0 )b 1 − = − E(m) E(m) a a

(91)

β i is the vector of the appropriate regression coefficients, ¡ ¢−1 β i ≡E ff 0 E(fRi ), so to get β in the formula, continue with E(R) =

1 E(Rf 0 )E(f f 0 )−1 E(f f 0 )b 1 E(f f 0 )b − = − β0 a a a a

Now, define α and λ to make it work, 1 1 = E (m) a 1 λ ≡ − E(ff 0 )b = − αE [mf ] a

α ≡

(6.92)

Using (6.92) we can just as easily go backwards from the expected return-beta representation to m = a + b0 f . As always, we have to worry about a special case of zero or infinite riskfree rates. We rule out E(m) = E(a + b0 f ) = 0 to keep (6.91) from exploding, and we rule ¥ out α = 0 and E(f f 0 ) singular to go from α, β, λ in (6.92) back to m. Given either model there is a model of the other form. They are not unique. We can add to m any random variable orthogonal to returns, and we can add spurious risk factors with zero β and/or λ , leaving pricing implications unchanged. We can also express the multiple beta model as a single beta model with m = a + b0 f as the single factor, or use its corresponding R∗ . Equation (6.92) shows that the factor risk premium λ can be interpreted as the price of the factor; A test of λ 6= 0 is often called a test of whether the “factor is priced.” More precisely, 106

S ECTION 6.3

FACTOR MODELS AND DISCOUNT FACTORS

λ captures the price E(mf ) of the (de-meaned) factors brought forward at the risk free rate. ˜ If we start with underlying factors f˜ such that the demeaned factors are f = f˜−E(f), h i ˜ = −α λ ≡ −α p f˜−E(f)

"

˜ ˜ E(f) p(f)− α

#

λ represents the price of the factors less their risk-neutral valuation, i.e. the factor risk premium. If the factors are not traded, λ is the model’s predicted price rather than a market price. Low prices are high risk premia, resulting in the negative sign. If the factors are returns with price one, then the factor risk premium is the expected return of the factor, less α, λ = E(f ) − α.

Note that the “factors” need not be returns (though they may be); they need not be orthogonal, and they need not be serially uncorrelated or conditionally or unconditionally meanzero. Such properties may occur as natural special cases, or as part of the economic derivation of specific factor models, but they are not required for the existence of a factor pricing representation. For example, if the riskfree rate is constant then Et (mt+1 ) is constant and at least the sum b0f should be uncorrelated over time. But if the riskfree rate is not constant, then Et (mt+1 ) = Et (b0f t+1 ) should vary over time. Factor-mimicking portfolios

It is often convenient to use factor-mimicking payoffs f ∗ = proj(f|X) factor-mimicking returns f∗ =

proj(f|X) p [proj(f |X)]

or factor-mimicking excess returns f ∗ = proj(f|Re ) in place of true factors. These payoffs carry the same pricing information as the original factors, and can serve as reference variables in expected return-beta representations When the factors are not already returns or excess returns, it is convenient to express a beta pricing model in terms of its factor mimicking portfolios rather than the factors themselves. Recall that x∗ = proj(m|X) carries all of m0 s pricing implications on X; p(x) = E(mx) = E(x∗ x). The factor-mimicking portfolios are just the same idea using the individual factors. 107

C HAPTER 6

R ELATION BETWEEN DISCOUNT FACTORS , BETAS , AND MEAN - VARIANCE FRONTIERS

Define the payoffs f ∗ by f ∗ = proj(f|X) Then, m = b0 f ∗ carries the same pricing implications on X as does m = b0 f : p = E(mx) = E(b0 f x) = E [b0 (projf |X) x] = E [b0 f ∗ x] .

(93)

(I include the constant as one of the factors.) The factor-mimicking portfolios also form a beta representation. Just go from (6.93) back to an expected return- beta representation E(Ri ) = α∗ + β ∗0 λ∗ ,

(94)

and find λ∗ , α∗ using (6.92). The β ∗ are the regression coefficients of the returns Ri on the factor-mimicking portfolios, not on the factors, as they should be. It is more traditional to use the returns or excess returns on the factor-mimicking portfolios rather than payoffs as I have done so far. To generate returns, divide the payoff by its price, f∗ =

proj(f |X) . p [proj(f|X)]

The resulting b will be scaled down by the price of the factor-mimicking payoff, and the model is the same. Note you project on the space of payoffs, not of returns. Returns R are not a space, since they don’t contain zero. If the test assets are all excess returns, you can even more easily project the factors on the set of excess returns, which are a space since they do include zero. If we define f ∗ = proj(f|Re ) then of course the excess returns f ∗ carry the same pricing implications as the factors f for a set of excess returns; m = b0 f ∗ satisfies 0 = E(mRei ) and E(Rei ) = β i,f ∗ λ = β i,f ∗ E(f ∗ )

6.4

Discount factors and beta models to mean - variance frontier

From m, we can construct R∗ which is on the mean variance frontier If a beta pricing model holds, then the return R∗ on the mean-variance frontier is a linear combination of the factor-mimicking portfolio returns. 108

S ECTION 6.5

T HREE RISKFREE RATE ANALOGUES

Any frontier return is a combination of R∗ and one other return, a risk free rate or a risk free rate proxy. Thus, any frontier return is a linear function of the factor-mimicking returns plus a risk free rate proxy. It’s easy to show that given m that we can find a return on the mean-variance frontier. Given m construct x∗ = proj(m|X) and R∗ = x∗ /E(x∗2 ). R∗ is the minimum second moment return, and hence on the mean-variance frontier. Similarly, if you have a set of factors f for a beta model, then a linear combination of the factor-mimicking portfolios is on the mean-variance frontier. A beta model is the same as m = b0 f . Since m is linear in f , x∗ is linear in f ∗ = proj(f |X), so R∗ is linear in the factor mimicking payoffs f ∗ or their returns f ∗ /p(f ∗ ). Section 5.4 showed how we can span the mean-variance frontier with R∗ and a risk free rate, if there is one, or the zero-beta, minimum variance, or constant-mimicking portfolio ˆ = proj(1|X)/p[proj(1|X)] if there is no risk free rate. The latter is particularly return R nice in the case of a linear factor model, since we may consider the constant as a factor, so the frontier is entirely generated by factor-mimicking portfolio returns.

6.5

Three riskfree rate analogues

I introduce three counterparts to the risk free rate that show up in asset pricing formulas when there is no risk free rate. The three returns are the zero-beta return, the minimumvariance return and the constant-mimicking portfolio return. Three different generalizations of the riskfree rate are useful when a risk free rate or unit payoff is not in the set of payoffs. These are the zero-beta return, the minimum-variance return and the constant-mimicking portfolio return. I introduce the returns in this section, and I use them in the next section to state some special cases involving the mean-variance frontier. Each of these returns maintains one property of the risk free rate in a market in which there is no risk free rate. The zero-beta return is a mean-variance efficient return that is uncorrelated with another given mean-variance efficient return. The minimum-variance return is just that. The constant-mimicking portfolio return is the return on the payoff “closest” to the unit payoff. Each of these returns one has a representation in the standard form R∗ + wRe∗ with slightly different w. In addition, the expected returns of these risky assets are used in some asset pricing representations. For example, the zero beta rate is often used to refer to the expected value of the zero beta return. Each of these riskfree rate analogues is mean-variance efficient. Thus, I characterize each one by finding its weight w in a representation of the form R∗ + wRe∗ . We derived such a

109

C HAPTER 6

R ELATION BETWEEN DISCOUNT FACTORS , BETAS , AND MEAN - VARIANCE FRONTIERS

representation above for the riskfree rate as equation (5.72), Rf = R∗ + Rf Re∗ .

(95)

In the last subsection, I show how each riskfree rate analogue reduces to the riskfree rate when there is one. 6.5.1

Zero-beta return for R∗

The zero beta return for R∗ , denoted Rα , is the mean-variance efficient return uncorrelated with R∗ . Its expected return is the zero beta rate α = E(Ra ). This zero beta return has representation Ra = R∗ +

var(R∗ ) Re∗ , E(R∗ )E(Re∗ )

and the corresponding zero beta rate is α = E(Rα ) =

1 E(R∗2 ) = . E(R∗ ) E(x∗ )

The zero beta rate is found graphically in mean-standard deviation space by extending the tangency at R∗ to the vertical axis. It is also the inverse of the price that x∗ and R∗ assign to the unit payoff. The riskfree rate Rf is of course uncorrelated with R∗ . Risky returns uncorrelated with R earn the same average return as the risk free rate if there is one, so they might take the place of Rf when the latter does not exist. For any return Rα that is uncorrelated with R∗ we have E(R∗ Rα ) = E(R∗ )E(Rα ), so ∗

α = E(Rα ) =

1 E(R∗2 ) = . E(R∗ ) E(x∗ )

The first equality introduces a popular notation α for this rate. I call α the zero beta rate, and Ra the zero beta return. There is no riskfree rate, so there is no security that just pays α. As you can see from the formula, the zero-beta rate is the inverse of the price that R∗ and x assign to the unit payoff, which is another natural generalization of the riskfree rate. It is called the zero beta rate because cov(R∗ , Rα ) = 0 implies that the regression beta of Rα on R∗ is zero. More precisely, one might call it the zero beta rate on R∗ , since one can calculate zero-beta rates for returns other than R∗ and they are not the same as the zero-beta rate for R∗ In particular, the zero-beta rate on the “market portfolio” will generally be different from the zero beta rate on R∗ . ∗

110

S ECTION 6.5

T HREE RISKFREE RATE ANALOGUES

E(R) α= E(R*2 )/ E(R* ) = 1/E(x*)

Rα R*

σ(R) Figure 20. Zero-beta rate α and zero-beta return Ra for R∗ .

111

C HAPTER 6

R ELATION BETWEEN DISCOUNT FACTORS , BETAS , AND MEAN - VARIANCE FRONTIERS

I draw α in Figure 20 as the intersection of the tangency and the vertical axis. This is a property of any return on the mean variance frontier: The expected return on an asset uncorrelated with the mean-variance efficient asset (a zero-beta asset) lies at the point so constructed. To check this geometry, use similar triangles: p The length ofp R∗ in Figure 20 is p ∗ ∗2 ∗2 E(R ), and its vertical extent is E(R ). Therefore, α/ E(R ) = E(R∗2 )/E(R∗ ), or α = E(R∗2 )/E(R∗ ). Since R∗ is on the lower portion of the mean-variance frontier, this zero beta rate α is above the minimum variance return. Note that in general α 6= 1/E(m). Projecting m on X preserves asset pricing implications on X but not for payoffs not in X. Thus if a risk free rate is not traded, x∗ and m may differ in their predictions for the riskfree rate as for other nontraded assets. The zero beta return is the rate of return on the mean-variance frontier with mean equal to the zero beta rate, as shown in Figure 20. We want to characterize this return in R∗ + wRe∗ form. To do this, we want to find w such that E(Ra ) =

E(R∗2 ) = E(R∗ ) + wE(Re∗ ). E(R∗ )

Solving, the answer is w=

var(R∗ ) E(R∗2 ) − E(R∗ )2 = . ∗ e∗ E(R )E(R ) E(R∗ )E(Re∗ )

Thus, the zero beta return is Ra = R∗ +

var(R∗ ) Re∗ , E(R∗ )E(Re∗ )

expression (6.103). Note that the weight is not E(Ra ) = E(R∗2 )/E(R∗ ). When there is no risk free rate, the weight and the mean return are different. 6.5.2

Minimum variance return

The minimum variance return has the representation Rmin. var. = R∗ +

E(R∗ ) Re∗ . 1 − E(Re∗ )

The riskfree rate obviously is the minimum variance return when it exists. When there is no risk free rate, the minimum variance return is Rmin. var. = R∗ +

E(R∗ ) Re∗ . 1 − E(Re∗ )

112

(96)

S ECTION 6.5

T HREE RISKFREE RATE ANALOGUES

Taking expectations, E(Rmin. var. ) = E(R∗ ) +

E(R∗ ) E(R∗ ) E(Re∗ ) = . e∗ 1 − E(R ) 1 − E(Re∗ )

The minimum variance return retains the nice property of the risk free rate, that its weight on Re∗ is the same as its mean, Rmin. var. = R∗ + E(Rmin. var. )Re∗ just as Rf = R∗ + Rf Re∗ . When there is no risk free rate, the zero-beta and minimum variance returns are not the same. You can see this fact clearly in Figure 20. We can derive expression (6.96) for the minimum variance return by brute force: choose w in R∗ + wRe∗ to minimize variance. min var(R∗ + wRe∗ ) = E[(R∗ + wRe∗ )2 ] − E(R∗ + wRe∗ )2 = w

= E(R∗2 ) + w2 E(Re∗ ) − E(R∗ )2 − 2wE(R∗ )E(Re∗ ) − w2 E(Re∗ )2 . The first order condition is 0 = wE(Re∗ )[1 − E(Re∗ )] − E(R∗ )E(Re∗ ) w=

6.5.3

E(R∗ ) . 1 − E(Re∗ )

Constant-mimicking portfolio return

The constant-mimicking portfolio return is defined as the return on the projection of the unit vector on the payoff space, ˆ = proj(1|X) . R p [proj(1|X)] It has the representation ∗2 ˆ = R∗ + E(R ) Re∗ . R E(R∗ )

When there is a risk free rate, it is the rate of return on a unit payoff, Rf = 1/p(1). When there is no risk free rate, we might define the rate of return on the mimicking portfolio for a 113

C HAPTER 6

R ELATION BETWEEN DISCOUNT FACTORS , BETAS , AND MEAN - VARIANCE FRONTIERS

unit payoff, ˆ = proj(1|X) . R p [proj(1|X)] I call this object the constant-mimicking portfolio return. The mean-variance representation of the constant-mimicking portfolio return is ∗2 ˆ = R∗ + αRe∗ = R∗ + E(R ) Re∗ . R E(R∗ )

(97)

Note that the weight α equal to the zero beta rate creates the constant-mimicking return, not the zero beta return. To show (6.97), start with property (5.73), Re∗ = proj(1|X) −

E(R∗ ) ∗ R . E(R∗2 )

(98)

Take the price of both sides. Since the price of Re∗ is zero and the price of R∗ is one, we establish p [proj(1|X)] =

E(R∗ ) . E(R∗2 )

(99a)

Solving (6.98) for proj(1|X), dividing by (6.99a) we obtain the right hand side of (6.97). 6.5.4

Risk free rate

The risk free rate has the mean-variance representation Rf = R∗ + Rf Re∗ . The zero-beta, minimum variance and constant-mimicking portfolio returns reduce to this formula when there is a risk free rate. Again, we derived in equation (5.72) that the riskfree rate has the representation, Rf = R∗ + Rf Re∗ .

(100)

Obviously, we should expect that the zero-beta return, minimum-variance return, and constantmimicking portfolio return reduce to the riskfree rate when there is one. These other rates are ∗2 ˆ = R∗ + E(R ) Re∗ constant-mimicking: R E(R∗ )

114

(101)

S ECTION 6.6

M EAN - VARIANCE SPECIAL CASES WITH NO RISKFREE RATE

minimum-variance: Rmin. var. = R∗ +

zero-beta: Rα = R∗ +

E(R∗ ) Re∗ 1 − E(Re∗ )

var(R∗ ) Re∗ . E(R∗ )E(Re∗ )

(102)

(103)

To establish that these are all the same when there is a riskfree rate, we need to show that Rf =

E(R∗ ) var(R∗ ) E(R∗2 ) = = E(R∗ ) 1 − E(Re∗ ) E(R∗ )E(Re∗ )

(104)

We derived the first equality above as equation (5.74). To derive the second equality, take expectations of (6.95), Rf = E(R∗ ) + Rf E(Re∗ )

(105)

and solve for Rf . To derive the third equality, use the first equality from (6.104) in (6.105), E(R∗2 ) = E(R∗ ) + Rf E(Re∗ ). E(R∗ ) Solving for Rf , Rf =

6.6

var(R∗ ) E(R∗2 ) − E(R∗ )2 = . E(R∗ )E(Re∗ ) E(R∗ )E(Re∗ )

Mean-variance special cases with no riskfree rate

We can find a discount factor from any mean-variance efficient return except the constantmimicking return. We can find a beta representation from any mean-variance efficient return except the minimum-variance return. I collect in this section the special cases for the equivalence theorems of this chapter. The special cases all revolve around the problem that the expected discount factor, price of a unit payoff or riskfree rate must not be zero or infinity. This is typically an issue of theoretical rather than practical importance. In a complete, arbitrage free market, m > 0 so we know E(m) > 0. If a riskfree rate is traded you can observe ∞ > E(m) = 1/Rf > 0. However, in an incomplete market in which no riskfree rate is traded, there are many discount factors with the same asset pricing implications, and you might have happened to choose one with E(m) = 0 in your manipulations. By and large, this is easy to avoid: choose another of the 115

C HAPTER 6

R ELATION BETWEEN DISCOUNT FACTORS , BETAS , AND MEAN - VARIANCE FRONTIERS

many discount factors with the same pricing implications that does not have E(m) = 0. More generally, when you choose a particular discount factor you are choosing an extension of the current set of prices and payoffs; you are viewing the current prices and payoffs as a subset of a particular contingent-claim economy. Make sure you pick a sensible one. Therefore, we could simply state the special cases as “when a riskfree rate is not traded, make sure you use discount factors with 0 < E(m) < ∞.” However, it is potentially useful and it certainly is traditional to specify the special return on the mean-variance frontier that leads to the infinite or zero implied riskfree rate, and to rule it out directly. This section works out what those returns are and shows why they must be avoided. 6.6.1

The special case for mean variance frontier to discount factor

When there is no riskfree rate, we can find a discount factor that is a linear function of any mean-variance efficient return except the constant-mimicking portfolio return. In section 6.2, we saw that we can form a discount factor a + bRmv from any mean∗2 ) e∗ variance efficient return Rmv except one particular return, of the form R∗ + E(R E(R∗ ) R . This return led to an infinite m. We now recognize this return as the risk-free rate, when there is one, or the constant-mimicking portfolio return, if there is no riskfree rate. Figure 21 shows the geometry of this case. To use no more than three dimensions I had to reduce the return and excess return spaces to lines. The payoff space X is the plane joining the return and excess return sets as shown. The set of all discount factors is m = x∗ + ε, E(εx) = 0, the line through x∗ orthogonal to the payoff space X in the figure. I draw the unit payoff (the dot marked “1” in Figure 21) closer to the viewer than the plane X, and I draw a vector through the unit payoff coming out of the page. Take any return on the mean-variance frontier, Rmv . (Since the return space only has two dimensions, all returns are on the frontier.) For a given Rmv , the space a + bRmv is the plane spanned by Rmv and the unit payoff. This plane lies sideways in the figure. As the figure shows, there is a vector a + bRmv in this plane that lies on the line of discount factors. Next, the special case. This construction would go awry if the plane spanning the unit payoff and the return Rmv were parallel to the plane containing the discount factor. Thus, ˆ in the Figure. This is a return the construction would not work for the return marked R corresponding to a payoff that is the projection of the unit payoff on to X, so that the residual will be orthogonal to X, as is the line of discount factors. With Figure 21 in front of us, we can also see why the constant-mimicking portfolio return is not the same thing as the minimum-variance return. Variance is the size or second moment

116

S ECTION 6.6

M EAN - VARIANCE SPECIAL CASES WITH NO RISKFREE RATE

Discount factors Rmv

^ Constant-mimicking return R

R*

R

a+b Rmv x* 1

X Re 0 Figure 21. One can construct a discount factor m = a + bRmv from any ˆ mean-variance-efficient return except the constant-mimicking return R. of the residual in a projection (regression) on 1. ¤ £ ¤ £ var(x) = E (x − E(x))2 = E (x − proj(x|1))2 = ||x − proj(x|1)||2

Thus, the minimum variance return is the return closest to extensions of the unit vector. It is formed by projecting returns on the unit vector. The constant-mimicking portfolio return is the return on the payoff closest to 1 It is formed by projecting the unit vector on the set of payoffs. 6.6.2

The special case for mean-variance frontier to a beta model

We can use any return on the mean-variance frontier as the reference return for a single beta representation, except the minimum-variance return. We already know mean variance frontiers ⇔ discount factor and discount factor ⇔ single beta representation, so at a superficial level we can string the two theorems together to go 117

C HAPTER 6

R ELATION BETWEEN DISCOUNT FACTORS , BETAS , AND MEAN - VARIANCE FRONTIERS

from a mean-variance efficient return to a beta representation. However it is more elegant to go directly, and the special cases are also a bit simpler this way. Theorem: There is a single beta representation with a return Rmv as factor, E(Ri ) = αRmv + β i,Rmv [E(Rmv ) − α] , if and only if Rmv is mean-variance efficient and not the minimum variance return. This famous theorem is given by Roll (1976) and Hansen and Richard (1987). We rule out minimum variance to rule out the special case E(m) = 0. Graphically, the zero-beta rate is formed from the tangency to the mean-variance frontier as in Figure 20. I use the notation αRmv to emphasize that we use the zero-beta rate corresponding to the particular meanvariance return Rmv that we use as the reference return. If we used the minimum-variance return, that would lead to an infinite zero-beta rate. Proof: The mean-variance frontier is Rmv = R∗ + wRe∗ . Any return is Ri = R∗ + wi Re∗ + ni . Thus, E(Ri ) = E(R∗ ) + wi E(Re∗ )

(106)

Now, £ ¤ cov(Ri , Rmv ) = cov (R∗ + wRe∗ ) , (R∗ + wi Re∗ )

= var(R∗ ) + wwi var(Re∗ ) − (w + wi )E(R∗ )E(Re∗ ) = var(R∗ ) − wE(R∗ )E(Re∗ ) + wi [w var(Re∗ ) − E(R∗ )E(Re∗ )]

Thus, cov(Ri , Rmv ) and E(Ri ) are both linear functions of wi . We can solve cov(Ri , Rmv ) for wi , plug into the expression for E(Ri ) and we’re done. To do this, of course, we must be able to solve cov(Ri , Rmv ) for wi . This requires w 6=

E(R∗ )E(Re∗ ) E(R∗ ) E(R∗ )E(Re∗ ) = = var(Re∗ ) E(Re∗2 ) − E(Re∗ )2 1 − E(Re∗ )

which is the condition for the minimum variance return.

6.7 1.

2.

(107) ¥

Problems

In the argument that Rmv on the mean variance frontier, Rmv = R∗ + wRe∗ , implies a discount factor m = a + bRmv , do we have to rule out the case of risk neutrality? (Hint: What is Re∗ when the economy is risk-neutral?) If you use factor mimicking portfolios as in (6.93), you know that the predictions for expected returns are the same as they are if you use the factors themselves . Are the α∗ , 118

S ECTION 6.7

3. 4.

5.

6.

P ROBLEMS

λ∗ , and β ∗ for the factor mimicking portfolio representation the same as the original α, λ, and β of the factor pricing model? Suppose the CAPM is true, m = a − bRm prices a set of assets, and there is a risk-free rate Rf . Find R∗ in terms of the moments of Rm , Rf . If you express the mean-variance frontier as a linear combination of factor-mimicking portfolios from a factor model, do the relative weights of the various factor portfolios in the mean-variance efficient return change as you sweep out the frontier, or do they stay the same? (Start with the riskfree rate case) For an arbitrary mean-variance efficient return of the form R∗ + wRe∗ , find its zero-beta return and zero-beta rate. Show that your rate reduces to the riskfree rate when there is one. When the economy is risk neutral, and if there is no risk-free rate, show that the zero-beta, minimum-variance, and constant-mimicking portfolio returns are again all equivalent, though not equal to the risk-free rate. (In this case, the mean-variance frontier is just the minimum-variance point.)

119

Chapter 7. Implications of existence and equivalence theorems Existence of a discount factor means p = E(mx) is innocuous, and all content flows from the discount factor model. The theorems apply to sample moments too; the dangers of fishing up ex-post or sample mean-variance efficient portfolios. Sources of discipline in factor fishing expeditions. The joint hypothesis problem. How efficiency tests are the same as tests of economic discount factor models. Factors vs. their mimicking portfolios. Testing the number of factors. Plotting contingent claims on the axis vs. mean and variance. The theorems on the existence of a discount factor, and the equivalence between the p = E(mx), expected return - beta, and mean-variance views of asset pricing have important implications for how we approach and evaluate empirical work. The equivalence theorems are obviously important, especially to the theme of this book, to show that the choice of discount factor language versus expected return-beta language or mean-variance frontier is entirely one of convenience. Nothing in the more traditional statements is lost. p = E(mx) is innocuous Before Roll (1976), expected return – beta representations had been derived in the context of special and explicit economic models, especially the CAPM. In empirical work, the success of any expected return - beta model seemed like a vindication of the whole structure. The fact that, for example, one might use the NYSE value-weighted index portfolio in place of the return on total wealth predicted by the CAPM seemed like a minor issue of empirical implementation. When Roll showed that mean-variance efficiency implies a single beta representation, all that changed. Some single beta representation always exists, since there is some meanvariance efficient return. The asset pricing model only serves to predict that a particular return (say, the “market return”) will be mean-variance efficient. Thus, if one wants to “test the CAPM” it becomes much more important to be choosy about the reference portfolio, to guard against stumbling on something that happens to be mean-variance efficient and hence prices assets by construction. 120

This insight led naturally to the use of broader wealth indices (Stambaugh 1982) in the reference portfolio to provide a more grounded test of the CAPM. However, this approach has not caught on. Stocks are priced with stock factors, bonds with bond factors, and so on. More recently, stocks sorted on size, book/market, and past performance characteristics are priced by portfolios sorted on those characteristics. Part of the reason for this is that the betas are small; stocks and bonds are not highly correlated so risk premia from one source of betas have small impacts on another set of average returns. Larger measures of wealth including human capital and real estate do not come with high frequency price data, so adding them to a wealth portfolio has little effect on betas. The good news in this existence theorem is that you can always start by writing an expected return-beta model, knowing that you have imposed almost no structure in doing so. The bad news is that you haven’t gotten very far. All the economic, statistical and predictive content comes in picking the factors. The theorem that, from the law of one price, there exists some discount factor m such that p = E(mx) is just an updated restatement of Roll’s theorem. The content is all in m = f (data) not in p = E(mx). Again, an asset pricing framework that initially seemed to require a lot of completely unbelievable structure–the representative consumer consumptionbased model in complete frictionless markets–turns out to require (almost) no structure at all. Again, the good news is that you can always start by writing p = E(mx), and need not suffer criticism about hidden contingent claim or representative consumer assumptions in so doing. The bad news is that you haven’t gotten very far by writing p = E(mx) as all the economic, statistical and predictive content comes in picking the discount factor model m = f (data). Ex-ante and ex-post. I have been deliberately vague about the probabilities underlying expectations and other moments in the theorems. The fact is, the theorems hold for any set of probabilities4 . Thus, the existence and equivalence theorems work equally well ex-ante as ex-post: E(mx), β, E(R) and so forth can refer to agent’s subjective probability distributions, objective population probabilities, or to the moments realized in a given sample. Thus, if the law of one price holds in a sample, one may form an x∗ from sample moments that satisfies p(x) = E(x∗ x), exactly, in that sample, where p(x) refers to observed prices and E(x∗ x) refers to the sample average. Equivalently, if the sample covariance matrix of a set of returns is nonsingular, there exists an ex-post mean-variance efficient portfolio for which sample average returns line up exactly with sample regression betas. This observation points to a great danger in the widespread exercise of searching for and statistically evaluating ad-hoc asset pricing models. Such models are guaranteed empirical success in a sample if one places little enough structure on what is included in the discount factor function. The only reason the model doesn’t work perfectly is the restrictions the researcher has imposed on the number or identity of the factors included in m, or the parameters of the function relating the factors to m. Since these restrictions are the entire content of the 4

Precisely, any set of probabilities that agree on impossible (zero-probability) events.

121

C HAPTER 7

I MPLICATIONS OF EXISTENCE AND EQUIVALENCE THEOREMS

model, they had better be interesting, carefully described and well motivated! Obviously, this is typically not the case or I wouldn’t be making such a fuss about it. Most empirical asset pricing research posits an ad-hoc pond of factors, fishes around a bit in that set, and reports statistical measures that show “success,” in that the model is not statistically rejected in pricing an ad-hoc set of portfolios. The set of discount factors is usually not large enough to give the zero pricing errors we know are possible, yet the boundaries are not clearly defined. Discipline What is wrong, you might ask, with finding an ex-post efficient portfolio or x∗ that prices assets by construction? Perhaps the lesson we should learn from the existence theorems is to forget about economics, the CAPM, marginal utility and all that, and simply price assets with ex-post mean variance efficient portfolios that we know set pricing errors to zero! The mistake is that a portfolio that is ex-post efficient in one sample, and hence prices all assets in that sample, is unlikely to be mean-variance efficient, ex-ante or ex-post, in the next sample, and hence is likely to do a poor job of pricing assets in the future. Similarly, the portfolio x∗ = p0 E(xx0 )−1 x (using the sample second moment matrix) that is a discount factor by construction in one sample is unlikely to be a discount factor in the next sample; the required portfolio weights p0 E(xx0 )−1 change, often drastically, from sample to sample. For example, suppose the CAPM is true, the market portfolio is ex-ante mean-variance efficient, and sets pricing errors to zero if you use true or subjective probabilities. Nonetheless, the market portfolio is unlikely to be ex-post mean-variance efficient in any given sample. In any sample, there will be lucky winners and unlucky losers. An ex-post mean variance efficient portfolio will be a Monday-morning quarterback; it will tell you to put large weights on assets that happened to be lucky in a given sample, but are no more likely than indicated by their betas to generate high returns in the future. “Oh, if I had only bought Microsoft in 1982...” is not a useful guide to forming a mean-variance efficient portfolio today. (In fact, mean-reversion in the market and book/market effects in individual stocks suggest that if anything, assets with unusually good returns in the past are likely to do poorly in the future!) The only solution is to impose some kind of discipline in order to avoid dredging up spuriously good in-sample pricing. The situation is the same as in traditional regression analysis. Regressions are used to forecast or to explain a variable y by other variables x in a regression y = x0 β + ε. By blindly including right hand variables, one can produce models with arbitrarily good statistical measures of fit. But this kind of model is typically unstable out of sample or otherwise useless for explanation or forecasting. One has to carefully and thoughtfully limit the search for right hand variables x to produce good models. What makes for an interesting set of restrictions? Econometricians wrestling with y = x0 β + ε have been thinking about this question for about 50 years, and the best answers are 1) use economic theory to carefully specify the right hand side and 2) use a battery of 122

cross-sample and out-of-sample stability checks. Alas, this advice is hard to follow. Economic theory is usually either silent on what variables to put on the right hand side of a regression, or allows a huge range of variables. The same is true in finance. “What are the fundamental risk factors?” is still an unanswered question. At the same time one can appeal to the APT and ICAPM to justify the inclusion of just about any desirable factor (Fama 1991 calls these theories a “fishing license.”) Thus, you will grow old waiting for theorists to provide useful answers to this kind of question. Following the purely statistical advice, the battery of cross-sample and out-of-sample tests often reveals the model is unstable, and needs to be changed. Once it is changed, there is no more out-of-sample left to check it. Furthermore, even if one researcher is pure enough to follow the methodology of classical statistics, and wait 50 years for another fresh sample to be available before contemplating another model, his competitors and journal editors are unlikely to be so patient. In practice, then, out of sample validation is not a strong guard against fishing. Nonetheless, these are the only standards we have to guard against fishing. In my opinion, the best hope for finding pricing factors that are robust out of sample and across different markets, is to try to understand the fundamental macroeconomic sources of risk. By this I mean, tying asset prices to macroeconomic events, in the way the ill-fated consumption based model does via mt+1 = βu0 (ct+1 )/u0 (ct ). The difficulties of the consumption-based model have made this approach lose favor in recent years. However, the alternative approach is also running into trouble that the number and identity of empirically-determined risk factors does not seem stable. Every time a new anomaly or data set pops up, a new set of ad-hoc factors gets created to explain them! In any case, one should always ask of a factor model, “what is the compelling economic story that restricts the range of factors used?” and / or “what statistical restraints are used” to keep from discovering ex-post mean variance efficient portfolios, or to ensure that the results will be robust across samples. The existence theorems tell us that the answers to these questions are the only content of the exercise. If the purpose of the model is not just to predict asset prices but also to explain them, this puts an additional burden on economic motivation of the risk factors. There is a natural resistance to such discipline built in to our current statistical methodology for evaluating models (and papers). When the last author fished around and produced an ad-hoc factor pricing model that generates 1% average pricing errors, it is awfully hard to persuade readers, referees, journal editors, and clients that your economically motivated factor pricing model is better despite 2% average pricing errors. Your model may really be better and will therefore continue to do well out of sample when the fished model falls by the wayside of financial fashion, but it is hard to get past statistical measures of in-sample fit. One hungers for a formal measurement of the number of hurdles imposed on a factor fishing ¯ 2 . Absent a numerical correction, we expedition, like the degrees of freedom correction in R have to use judgment to scale back apparent statistical successes by the amount of economic and statistical fishing that produced them. 123

C HAPTER 7

I MPLICATIONS OF EXISTENCE AND EQUIVALENCE THEOREMS

Mimicking portfolios The theorem x∗ = proj(m|X) also has interesting implications for empirical work. The pricing implications of any model can be equivalently represented by its factor-mimicking portfolio. If there is any measurement error in a set of economic variables driving m, the factor-mimicking portfolios for the true m will price assets better than an estimate of m that uses the measured macroeconomic variables. Thus, it is probably not a good idea to evaluate economically interesting models with statistical horse races against models that use portfolio returns as factors. Economically interesting models, even if true and perfectly measured, will just equal the performance of their own factor-mimicking portfolios, even in large samples. They will always lose in sample against ad-hoc factor models that find nearly ex-post efficient portfolios. This said, there is an important place for models that use returns as factors. After we have found the underlying true macro factors, practitioners will be well advised to look at the factor-mimicking portfolio on a day-by-day basis. Good data on the factor-mimicking portfolios will be available on a minute-by-minute basis. For many purposes, one does not have to understand the economic content of a model. But this fact does not tell us to circumvent the process of understanding the true macroeconomic factors by simply fishing for factor-mimicking portfolios. The experience of practitioners who use factor models seems to bear out this advice. Large commercial factor models resulting from extensive statistical analysis (otherwise known as fishing) perform poorly out of sample, as revealed by the fact that the factors and loadings (β) change all the time. Also models specified with economic fundamentals will always seem to do poorly in a given sample against ad-hoc variables (especially if one fishes an ex-post mean-variance efficient portfolio out of the latter!). But what other source of discipline do we have? Irrationality and Joint Hypothesis Finance contains a long history of fighting about “rationality” vs. “irrationality” and “efficiency” vs. “inefficiency” of asset markets. The results of many empirical asset pricing papers are sold as evidence that markets are “inefficient” or that investors are “irrational.” For example, the crash of October 1987, and various puzzles such as the small-firm, book/market, seasonal effects or long-term predictability have all been sold this way. However, none of these puzzles documents an arbitrage opportunity5 . Therefore, we know that there is a “rational model”–a stochastic discount factor, an efficient portfolio to use in a single-beta representation—that rationalizes them all. And we can confidently predict this situation to continue; real arbitrage opportunities do not last long! Fama (1970) contains a famous statement of the same point. Fama emphasized that any test of “efficiency” is a joint test of efficiency and a “model of market equilibrium.” Translated, an asset pricing model, or a model of m. 5 The closed-end fund puzzle comes closest since it documents an apparent violation of the law of one price. However, you can’t costlessly short closed end funds, and we have ignored short sales constraints so far.

124

But surely markets can be “irrational” or “inefficient” without requiring arbitrage opportunities? Yes, they can, if (and only if) the discount factors that generate asset prices are disconnected from marginal rates of substitution or transformation in the real economy. But now we are right back to specifying and testing economic models of the discount factor! At best, an asset pricing puzzle might be so severe that we can show that the required discount factors are completely “unreasonable” (by some standard) measures of real marginal rates of substitution and/or transformation, but we still have to say something about what a reasonable marginal rate looks like. In sum, the existence theorems mean that there are no quick proofs of “rationality” or “irrationality.” The only game in town for the purpose of explaining asset prices is thinking about economic models of the discount factor. The number of factors. Many asset pricing tests focus on the number of factors required to price a cross-section of assets. The equivalence theorems imply that this is a silly question. A linear factor model m = b0 f or its equivalent expected return / beta model E(Ri ) = α + β 0if λf are not unique representations. In particular, given any multiple-factor or multiple-beta representation we can easily find a single-beta representation. The single factor m = b0 f will price assets just as well as the original factors f, as will x∗ = proj(b0 f | X) or the corresponding R∗ . All three options give rise to single-beta models with exactly the same pricing ability as the multiple factor model. We can also easily find equivalent representations with different numbers (greater than one) of factors. For example, write µ ¶ b3 m = a + b1 f1 + b2 f2 + b3 f3 = a + b1 f1 + b2 f2 + f3 = a + b1 f1 + b2 fˆ2 b2 to reduce a “three factor” model to a “two factor” model. In the ICAPM language, consumption itself could serve as a single state variable, in place of the S state variables presumed to drive it. There are times when one is interested in a multiple factor representation. Sometimes the factors have an economic interpretation that is lost on taking a linear combination. But the pure number of pricing factors is not a meaningful question. Discount factors vs. mean, variance and beta. The point of the previous chapter was to show how the discount factor, mean-variance, and expected return- beta models are all equivalent representations of asset pricing. It seems a good moment to contrast them as well; to understand why the mean-variance and beta language developed first, and to think about why the discount factor language seems to be taking over. Asset pricing started by putting mean and variance of returns on the axes, rather than payoff in state 1 payoff in state 2, etc. as we do now. The early asset pricing theorists posed the question just right: they wanted to treat assets in the apples-and-oranges, indifference 125

C HAPTER 7

I MPLICATIONS OF EXISTENCE AND EQUIVALENCE THEOREMS

curve and budget set framework of macroeconomics. The problem was, what labels to put on the axis? Clearly, “IBM stock” and “GM stock” is not a good idea; investors do not value securities per se, but value some aspects of the stream of random cash flows that those securities give rise to. Their brilliant insight was to put the mean and variance of the portfolio return on the axis; to treat these as “hedonics” by which investors valued their portfolios. Investors plausibly want more mean and less variance. They gave investors “utility functions” defined over this mean and variance, just as standard utility functions are defined over apples and oranges. The mean-variance frontier is the “budget set.” With this focus on portfolio mean and variance, the next step was to realize that each security’s mean return measures its contribution to the portfolio mean, and that regression betas on the overall portfolio give each security’s contribution to the portfolio variance. The mean-return vs. beta description for each security followed naturally. In a deep sense, the transition from mean-variance frontiers and beta models to discount factors represents the realization that putting consumption in state 1 and consumption in state 2 on the axes — specifying preferences and budget constraints over state-contingent consumption — is a much more natural mapping of standard microeconomics into finance than putting mean, variance, etc. on the axes. If for no other reason, the contingent claim budget constraints are linear, while the mean-variance frontier is not. Thus, I think, the focus on means and variance, the mean-variance frontier and expected return/beta models is all due to an accident of history, that the early asset pricing theorists happened to put mean and variance on the axes rather than state contingent consumption. Well, here we are, why prefer one language over another? The discount factor language has an advantage for its simplicity, generality, mathematical convenience, and elegance. These virtues are to some extent in the eye of the beholder, but to this beholder, it is inspiring to be able to start every asset pricing calculation with one equation, p = E(mx). This equation covers all assets, including bonds, options, and real investment opportunities, while the expected return/beta formulation is not useful or very cumbersome in the latter applications. Thus, it has seemed that there are several different asset pricing theories: expected return/beta for stocks, yield-curve models for bonds, arbitrage models for options. In fact all three are just cases of p = E(mx). As a particular example, arbitrage, in the precise sense of positive payoffs with negative prices, has not entered the equivalence discussion at all. I don’t know of any way to cleanly graft absence of arbitrage on to expected return/beta models. You have to tack it on after the fact – “by the way, make sure that every portfolio with positive payoffs has a positive price.” It is trivially easy to graft it on to a discount factor model: just add m > 0. The discount factor and state space language also makes it easier to think about different horizonsP and the present value statement of models. p = E(mx) generalizes quickly to pt = Et j mt,t+j xt+j , while returns have to be chained together to think about multiperiod models. Papers are still written arguing about geometric vs. arithmetic average returns for multiperiod discounting. 126

The choice of language is not about normality or return distributions. There is a lot of confusion about where return distribution assumptions show up in finance. I have made no distributional assumptions in any of the discussion so far. Second moments as in betas and the variance of the mean-variance frontier show up because p = E(mx) involves a second moment. One does not need to assume normality to talk about the mean-variance frontier. Returns on the mean-variance frontier price other assets even when returns are not normally distributed.

127

Chapter 8.

Conditioning information

The asset pricing theory I have sketched so far really describes prices at time t in terms of conditional moments. The investor’s first order conditions are pt u0 (ct ) = βEt [u0 (ct+1 )xt+1 ] where Et means expectation conditional on the investor’s time t information. Sensibly, the price at time t should be higher if there is information at time t that the discounted payoff is likely to be higher than usual at time t + 1. The basic asset pricing equation should be pt = Et (mt+1 xt+1 ). (Conditional expectation can also be written pt = E [mt+1 xt+1 |It ] when it is important to specify the information set It .). If payoffs and discount factors were independent and identically distributed (i.i.d.) over time, then conditional expectations would be the same as unconditional expectations and we would not have to worry about the distinction between the two concepts. But stock price/dividend ratios, bond and option prices all change over time, which must reflect changing conditional moments of something on the right hand side. One approach is to specify and estimate explicit statistical models of conditional distributions of asset payoffs and discount factor variables (e.g. consumption growth). This approach is sometimes used, and is useful in some applications, but it is usually cumbersome. As we make the conditional mean, variance, covariance, and other parameters of the distribution of (say) N returns depend flexibly on M information variables, the number of required parameters can quickly exceed the number of observations. More importantly, this explicit approach typically requires us to assume that investors use the same model of conditioning information that we do. We obviously don’t even observe all the conditioning information used by economic agents, and we can’t include even a fraction of observed conditioning information in our models. The basic feature and beauty of asset prices (like all prices) is that they summarize an enormous amount of information that only individuals see. The events that make the price of IBM stock change by a dollar, like the events that make the price of tomatoes change by 10 cents, are inherently unobservable to economists or would-be social planners (Hayek 1945). Whenever possible, our treatment of conditioning information should allow agents to see more than we do. If we don’t want to model conditional distributions explicitly, and if we want to avoid assuming that investors only see the variables that we include in an empirical investigation, we eventually have to think about unconditional moments, or at least moments conditioned on less information than agents see. Unconditional implications are also interesting in and of themselves. For example, we may be interested in finding out why the unconditional mean 128

S ECTION 8.1

S CALED PAYOFFS

returns on some stock portfolios are higher than others, even if every agent fundamentally seeks high conditional mean returns. Most statistical estimation essentially amounts to characterizing unconditional means, as we will see in the chapter on GMM. Thus, rather than model conditional distributions, this chapter focuses on what implications for unconditional moments we can derive from the conditional theory.

8.1

Scaled payoffs

pt = Et (mt+1 xt+1 ) ⇒ E(pt zt ) = E(mt+1 xt+1 zt ) One can incorporate conditioning information by adding scaled payoffs and doing everything unconditionally. I interpret scaled returns as payoffs to managed portfolios.

8.1.1

Conditioning down

The unconditional implications of any pricing model are pretty easy to state. From pt = Et (mt+1 xt+1 ) we can take unconditional expectations to obtain6 E(pt ) = E(mt+1 xt+1 ).

(108)

Thus, if we just interpret p to stand for E(pt ), everything we have done above applies to unconditional moments. In the same way, we can also condition down from agents’ fine information sets to coarser sets that we observe, pt

= E(mt+1 xt+1 | Ω) ⇒ E(pt |I ⊂ Ω) = E(mt+1 xt+1 | I ⊂ Ω) ⇒ pt = E(mt+1 xt+1 | It ⊂ Ωt ) if pt ∈ It .

In making the above statements I used the law of iterated expectations, which is important enough to highlight it. This law states that if you take an expected value using less information of an expected value that is formed on more information, you get back the expected value using less information. Your best forecast today of your best forecast tomorrow is the same We need a small technical assumption that the unconditional moment³or moment ´ conditioned ³ on ´ a coarser information set exists. For example, if X and Y are normal (0, 1), then E X |Y = 0 but E X is infinite. Y Y 6

129

C HAPTER 8

C ONDITIONING INFORMATION

as your best forecast today. In various useful guises, E(Et (x)) = E(x), Et−1 (Et (xt+1 )) = Et−1 (xt+1 ) E [E(x|Ω) | I ⊂ Ω] = E [x|I] 8.1.2

Instruments and managed portfolios

We can do more than just condition down. Suppose we multiply the payoff and price by an instrument zt observed at time t. Then, zt pt = Et (mt+1 xt+1 zt ) and, taking unconditional expectations, E(pt zt ) = E(mt+1 xt+1 zt ).

(109)

This is an additional implication of the conditional model, not captured by just conditioning down as in (8.108). This trick originates from the GMM method of estimating asset pricing models, discussed below. The word instruments for the z variables comes from the instrumental variables estimation heritage of GMM. To think about equation (8.109), group (xt+1 zt ). Call this product a payoff x = xt+1 zt , with price p = E(pt zt ). Then 8.109 reads p = E(mx) once again. Rather than thinking about (8.109) as a instrumental variables estimate of a conditional model, we can think of it as a price and a payoff, and apply all the asset pricing theory directly. This interpretation is not as artificial as it sounds. zt xt+1 are the payoffs to managed portfolios. An investor who observes zt can, rather than “buy and hold,” invest in an asset according to the value of zt . For example, if a high value of zt forecasts that asset returns are likely to be high the next period, the investor might buy more of the asset when zt is high and vice-versa. If the investor follows a linear rule, he puts zt pt dollars into the asset each period and receives zt xt+1 dollars the next period. This all sounds new and different, but practically every test uses managed portfolios. For example, the size, beta, industry, book/market and so forth portfolios of stocks are all managed portfolios, since their composition changes every year in response to conditioning information – the size, beta, etc. of the individual stocks. This idea is also closely related to the deep idea of dynamic spanning. Markets that are apparently very incomplete can in 130

S ECTION 8.2

S UFFICIENCY OF ADDING SCALED RETURNS

reality provide many more state-contingencies through dynamic (conditioned on information) trading strategies. Equation (8.109) offers a very simple view of how to incorporate the extra information in conditioning information: Add managed portfolio payoffs, and proceed with unconditional moments as if conditioning information didn’t exist! Linearity is not important. If the investor wanted to place, say, 2 + 3z 2 dollars in the asset, we could capture this desire with an instrument z2 = 2 + 3z 2 . Nonlinear (measurable) transformations of time−t random variables are again random variables. We can thus incorporate conditioning information while still looking at unconditional moments instead of conditional moments, without any of the statistical machinery of explicit models with time-varying moments. The only subtleties are 1) The set of asset payoffs expands dramatically, since we can consider all managed portfolios as well as basic assets, potentially multiplying every asset return by every information variable. 2) Expected prices of managed portfolios show up for p instead of just p = 0 and p = 1 if we started with basic asset returns and excess returns.

8.2

Sufficiency of adding scaled returns

Checking the expected price of all managed portfolios is, in principle, sufficient to check all the implications of conditioning information. E(zt ) = E(mt+1 Rt+1 zt ) ∀zt ∈ It ⇒ 1 = E(mt+1 Rt+1 |It ) E(pt ) = E(mt+1 xt+1 ) ∀ xt+1 ∈ X t+1 ⇒ pt = E(mt+1 xt+1 |It ) We have shown that we can derive some extra implications from the presence of conditioning information by adding scaled returns. But does this exhaust the implications of conditioning information? Are we missing something important by relying on this trick? The answer is, in principle no. I rely on the following mathematical fact: The conditional expectation of a variable yt+1 given an information set It , E(yt+1 | It ) is equal to a regression forecast of yt+1 using every variable zt ∈ It . Now, “every random variable” means every variable and every nonlinear (measurable) transformation of every variable, so there are a lot of variables in this regression! (The word projection and proj(yt+1 |zt ) is used to distinguish the best forecast of yt+1 using only linear combinations of zt from the conditional expectation.) Applying this fact to our case, let yt+1 = mt+1 Rt+1 − 1. Then E [(mt+1 Rt+1 − 1) zt ] = 0 for every zt ∈ It implies 1 = E(mt+1 Rt+1 | It ). Thus, no implications are lost in principle by looking at scaled returns. 131

C HAPTER 8

C ONDITIONING INFORMATION

Another way of looking at the same idea is that Rt+1 zt+1 is the return on a payoff available at time t + 1. Thus, the space of all payoffs X t+1 should be understood to include the time-t + 1 payoff you can generate with a basis set of assets Rt+1 and all dynamic strategies that use information in the set It . With that definition of the space X t+1 we can write the sufficiency of scaled returns with the more general second equality above. “All linear and nonlinear transformations of all variables observed at time t” sounds like a lot of instruments, and it is. But there is a practical limit to the number of instruments zt one needs to scale by, since only variables that forecast returns or m (or their higher moments and co-moments) add any information. Since adding instruments is the same thing as including potential managed portfolios, thoughtfully choosing a few instruments is the same thing as the thoughtful choice of a few assets or portfolios that one makes in any test of an asset pricing model. Even when evaluating completely unconditional asset pricing models, one always forms portfolios and omits many possible assets from analysis. Few studies, in fact, go beyond checking whether a model correctly prices 10-25 stock portfolios and a few bond portfolios. Implicitly, one feels that the chosen payoffs do a pretty good job of spanning the set of available risk-loadings (mean returns) and hence that adding additional assets will not affect the results. Nonetheless, since data are easily available on all 2000 or so NYSE stocks, plus AMEX and NASDAQ stocks, to say nothing of government and corporate bonds, returns of mutual funds, foreign exchange, foreign equities, real investment opportunities, etc., the use of a few portfolios means that a tremendous number of potential asset payoffs are left out in an ad-hoc manner. In a similar manner, if one had a small set of instruments that capture all the predictability of discounted returns mt+1 Rt+1 , then there would be no need to add more instruments. Thus, we carefully but arbitrarily select a few instruments that we think do a good job of characterizing the conditional distribution of returns. Exclusion of potential instruments is exactly the same thing as exclusion of assets. It is no better founded, but the fact that it is a common sin may lead one to worry less about it. There is nothing special about unscaled returns, and no economic reason to place them above scaled returns. A mutual fund might come into being that follows the managed portfolio strategy and then its unscaled returns would be the same as an original scaled return. Models that cannot price scaled returns are no more interesting than models that can only price (say) stocks with first letter A through L. (There may be econometric reasons to trust results for nonscaled returns a bit more, but we haven’t gotten to statistical issues yet.) Of course, the other way to incorporate conditioning information is by constructing explicit parametric models of conditional distributions. With this procedure one can in fact check all of a model’s implications about conditional moments. However, the parametric model may be incorrect, or may not reflect some variable used by investors. Including instruments may not be as efficient, but it is still consistent if the parametric model is incorrect. The wrong parametric model of conditional distributions may lead to inconsistent estimates. In addition, one avoids estimating nuisance parameters of the parametric distribution model.

132

S ECTION 8.3

8.3

C ONDITIONAL AND UNCONDITIONAL MODELS

Conditional and unconditional models

A conditional factor model does not imply a fixed-weight or unconditional factor model: mt+1 = b0t ft+1 , pt = Et (mt+1 xt+1 ) does not imply that ∃b s.t. mt+1 = b0 ft+1 , E(pt ) = E(mt+1 xt+1 ). Et (Rt+1 ) = β 0t λt does not imply E(Rt+1 ) = β 0 λ. Conditional mean-variance efficiency does not imply unconditional mean-variance efficiency. The converse statements are true, if managed portfolios are included. For explicit discount factor models—models whose parameters are constant over time— the fact that one looks at a conditional vs. unconditional implications makes no difference to the statement of the model. pt = Et (mt+1 xt+1 ) ⇒ E(pt ) = E(mt+1 xt+1 ) and that’s it. Examples include the consumption-based model with power utility, mt+1 = W . β(ct+1 /ct )−γ , and the log utility CAPM, mt+1 = 1/Rt+1 However, linear factor models include parameters that may vary over time and as functions of conditioning information. In these cases the transition from conditional to unconditional moments is much more subtle. We cannot easily condition down the model at the same time as the prices and payoffs. 8.3.1

Conditional vs. unconditional factor models in discount factor language

As an example, consider the CAPM m = a − bRW where RW is the return on the market or wealth portfolio. We can find a and b from the condition that this model correctly price any two returns, for example RW itself and a riskfree rate: W ½ ) a = R1f + bEt (Rt+1 ) 1 = Et (mt+1 RW t t+1 f ⇒ . (110) Et (RW )−R f t 1 = Et (mt+1 )Rt b = f t+1 R σ2 (RW ) t

t

t+1

As you can see, b > 0 and a > 0: to make a payoff proportional to the minimum secondmoment return (on the inefficient part of the mean-variance frontier) we need a portfolio long the risk free rate and short the market RW . 133

C HAPTER 8

C ONDITIONING INFORMATION

2 W More importantly for our current purposes, a and b vary over time, as Et (RW t+1 ), σ t (Rt+1 ), f and Rt vary over time. If it is to price assets conditionally, the CAPM must be a linear factor model with time-varying weights, of the form

mt+1 = at − bt RW t+1 . This fact means that we can no longer transparently condition down. The statement that £ ¤ 1 = Et (at + bt RW t+1 )Rt+1

does not imply that we can find constants a and b so that ¤ £ 1 = E (a + bRW t+1 )Rt+1 .

Just try it. Taking unconditional expectations, ¤ £ ¤ £ W W )Rt+1 = E at Rt+1 + bt Rt+1 Rt+1 1 = E (at + bt Rt+1 W = E(at )E(Rt+1 ) + E(bt )E(RW t+1 Rt+1 ) + cov(at , Rt+1 ) + cov(bt , Rt+1 Rt+1 )

Thus, the unconditional model 1=E

£¡

¢ ¤ W Rt+1 E(at ) + E(bt )Rt+1

only holds if the covariance terms above happen to be zero. Since at and bt are formed from conditional moments of returns, the covariances will not, in general be zero. On the other hand, suppose it is true that at and bt are constant over time. Then £ ¤ 1 = Et (a + bRW t+1 )Rt+1

does imply

¤ £ W )Rt+1 , 1 = E (a + bRt+1

just like any other constant-parameter factor pricing model. Furthermore, the latter unconditional model implies the former conditional model, if the latter holds for all managed portfolios. 8.3.2

Conditional vs. unconditional in an expected return / beta model

To put the same observation in beta-pricing language, Et (Ri ) = Rtf + β t λt 134

(111)

S ECTION 8.3

C ONDITIONAL AND UNCONDITIONAL MODELS

does not imply that E(Ri ) = α + βλ

(112)

The reason is that β t and β represent conditional and unconditional regression coefficients respectively. Again, if returns and factors are i.i.d., the unconditional model can go through. In that case, cov(·) = covt (·), var(·) = vart (·), so the unconditional regression beta is the same as the conditional regression beta, β = β t . Then, we can take expectations of (8.111) to get (8.112), with λ = E(λt ). But to condition down in this way, the covariance and variance must each be constant over time. It is not enough that their ratio, or conditional betas are constant. If covt and vart change over time, then the unconditional regression beta, β = cov/var is not equal to the average conditional regression beta, E(β t ) or E(covt /vart ). Some models specify that covt and vart vary over time, but covt /vart is a constant. This specification still does not imply that the unconditional regression beta β ≡ cov/var is equal to the constant covt /vart . Similarly, it is not enough that λ be constant, since E(β t ) 6= β. The betas must be regression coefficients, not just numbers. If the betas do not vary over time, the λt may still vary and λ = E(λt ). 8.3.3

A precise statement

Let’s formalize these observations somewhat. Let X denote the space of all portfolios of the primitive assets, including managed portfolios in which the weights may depend on conditioning information, i.e. scaled returns. A conditional factor pricing model is a model mt+1 = at + b0t ft+1 that satisfies pt = Et+1 (mt+1 xt+1 ) for all xt+1 ∈ X. An unconditional factor pricing model is model mt+1 = a + b0 ft+1 satisfies E(pt ) = E(mt+1 xt+1 ) for all xt+1 ∈ X. It might be more appropriately called a fixed-weight factor pricing model. Given these definitions it’s almost trivial that the unconditional model is just a special case of the conditional model, one that happens to have fixed weights. Thus, a conditional factor model does not imply an unconditional factor model (because the weights may vary) but an unconditional factor model does imply a conditional factor model. There is one important subtlety. The payoff space X is common, and contains all managed portfolios in both cases. The payoff space for the unconditional factor pricing model is not just fixed combinations of a set of basis assets. For example, we might simply check that the static (constant a, b) CAPM captures the unconditional mean returns of a set of assets. If this model does not also price those assets scaled by instruments, then it is not a conditional model, or, as I argued above, really a valid factor pricing model at all. Of course, everything applies for the relation between a conditional factor pricing model 135

C HAPTER 8

C ONDITIONING INFORMATION

using a fine information set (like investors’ information sets) and conditional factor pricing models using coarser information sets (like ours). If you think a set of factors prices assets with respect to investors’ information, that does not mean the same set of factors prices assets with respect to our, coarser, information sets. 8.3.4

Mean-variance frontiers

Define the conditional mean-variance frontier as the set of returns that minimize vart (Rt+1 ) given Et (Rt+1 ). (This definition includes the lower segment as usual.) Define the unconditional mean-variance frontier as the set of returns including managed portfolio returns that minimize var(Rt+1 ) given E(Rt+1 ). These two frontiers are related by: If a return is on the unconditional mean-variance frontier, it is on the conditional mean-variance frontier. However, If a return is on the conditional mean-variance frontier, it need not be on the unconditional mean-variance frontier. These statements are exactly the opposite of what you first expect from the language. The law of iterated expectations E(Et (x)) = E(x) leads you to expect that “conditional” should imply “unconditional.” But we are studying the conditional vs. unconditional mean-variance frontier, not raw conditional and unconditional expectations, and it turns out that exactly the opposite words apply. Of course “unconditional” can also mean “conditional on a coarser information set.” Again, keep in mind that the unconditional mean variance frontier includes returns on managed portfolios. This definition is eminently reasonable. If you’re trying to minimize variance for given mean, why tie your hands to fixed weight portfolios? Equivalently, why not allow yourself to include in your portfolio the returns of mutual funds whose advisers promise the ability to adjust portfolios based on conditioning information? You could form a mean-variance frontier of fixed-weight portfolios of a basis set of assets, and this is what many people often mean by “unconditional mean-variance frontier.” The return on the true unconditional mean-variance frontier will, in general, include some managed portfolio returns, and so will lie outside this mean-variance frontier of fixed-weight portfolios. Conversely, a return on the fixed-weight portfolio MVF is, in general, not on the unconditional or conditional mean-variance frontier. All we know is that the fixed-weight frontier lies inside the other two. It may touch, but it need not. This is not to say the fixed-weight unconditional frontier is uninteresting. For example, returns on this frontier will price fixed-weight portfolios of the basis assets. The point is that this frontier has no connection to the other two frontiers. In particular, a conditionally mean-variance efficient return (conditional CAPM) need not unconditionally price the fixed weight portfolios. 136

S ECTION 8.3

C ONDITIONAL AND UNCONDITIONAL MODELS

I offer several ways to see this important statement. Using the connection to factor models We have seen that the conditional CAPM mt+1 = at − bt RW t+1 does not imply an unconW . We have seen that the existence of such a conditional ditional CAPM mt+1 = a − bRt+1 W lies on the conditional meanfactor model is equivalent to the statement that the return Rt+1 variance frontier, and the existence of an unconditional factor model mt+1 = a − bRW t+1 is equivalent to the statement that RW is on the unconditional mean-variance frontier. Then, from the “trivial” fact that an unconditional factor model is a special case of a conditional one, we know that RW on the unconditional frontier implies RW on the conditional frontier but not vice-versa. Using the orthogonal decomposition We can see the relation between conditional and unconditional mean-variance frontiers using the orthogonal decomposition characterization of mean-variance efficiency given above. This beautiful proof is the main point of Hansen and Richard (1987). By the law of iterated expectations, x∗ and R∗ generate expected prices and Re∗ generates unconditional means as well as conditional means: E [p = Et (x∗ x)] ⇒ E(p) = E(x∗ x) ¤ £ E Et (R∗2 ) = Et (R∗ R) ⇒ E(R∗2 ) = E(R∗ R)

E [Et (Re∗ Re ) = Et (Re )] ⇒ E(Re∗ Re ) = E(Re ) This fact is subtle and important. For example, starting with x∗ = p0t Et (xt+1 x0t+1 )−1 xt+1 , you might think we need a different x∗ , R∗ , Re∗ to represent expected prices and unconditional means, using unconditional probabilities to define inner products. The three lines above show that this is not the case. The same old x∗ , R∗ , Re∗ represent conditional as well as unconditional prices and means. Recall that a return is mean-variance efficient if and only if it is of the form Rmv = R∗ + wRe∗ . Thus, Rmv is conditionally mean-variance efficient if w is any number in the time t information set. mv ∗ = Rt+1 + wt Re∗ conditional frontier: Rt+1 t+1 ,

and Rmv is unconditionally mean-variance efficient if w is any constant. mv ∗ unconditional frontier: Rt+1 = Rt+1 + wRe∗ t+1 .

137

C HAPTER 8

C ONDITIONING INFORMATION

Constants are in the t information set; time t random variables are not necessarily constant. Thus unconditional efficiency (including managed portfolios) implies conditional efficiency but not vice versa. As with the factor models, once you see the decomposition, it is a trivial argument about whether a weight is constant or time-varying. Brute force and examples. If you’re still puzzled, an additional argument by brute force may be helpful. If a return is on the unconditional MVF it must be on the conditional MVF at each date. If not, you could improve the unconditional mean-variance trade-off by moving to the conditional MVF at each date. Minimizing unconditional variance given mean is the same as minimizing unconditional second moment given mean, min E(R2 ) s.t. E(R) = µ Writing the unconditional moment in terms of conditional moments, the problem is £ ¤ min E Et (R2 ) s.t. E [Et (R)] = µ

Now, suppose you could lower Et (R2 ) at one date t without affecting Et (R) at that date. This change would lower the objective, without changing the constraint. Thus, you should have done it: you should have picked returns on the conditional mean variance frontiers. It almost seems that reversing the argument we can show that conditional efficiency implies unconditional efficiency, but it doesn’t. Just because you have minimized Et (R2 ) for given value of Et (R) at each date t does not imply that you have minimized E(R2 ) for a given value of E(R). In showing that unconditional efficiency implies conditional efficiency we held fixed Et (R) at each date at µ, and showed it is a good idea to minimize σ t (R). In trying to go backwards, the problem is that a given value of E(R) does not specify what Et (R) should be at each date. We can increase Et (R) in one conditioning information set and decrease it in another, leaving the return on the conditional MVF. Figure 22 presents an example. Return B is conditionally mean-variance efficient. It also has zero unconditional variance, so it is the unconditionally mean-variance efficient return at the expected return shown. Return A is on the conditional mean-variance frontiers, and has the same unconditional expected return as B. But return A has some unconditional variance, and so is inside the unconditional mean-variance frontier. As a second example,the riskfree rate is only on the unconditional mean-variance frontier if it is a constant. Remember the expression (6.95) for the risk free rate, Rf = R∗ + Rf Re∗ . The unconditional mean-variance frontier is R∗ + wRe∗ with w a constant. Thus, the riskfree rate is only unconditionally mean-variance efficient if it is a constant. 138

S ECTION 8.4

S CALED FACTORS : A PARTIAL SOLUTION

Et(R)

Info. set 1

A Info. set 2 B A

σt(R) Figure 22. Return A is on the conditional mean-variance frontiers but not on the unconditional mean variance frontier. 8.3.5

Implications: Hansen-Richard Critique.

Many models, such as the CAPM, imply a conditional linear factor model mt+1 = at + b0t ft+1 . These theorems show that such a model does not imply an unconditional model. Equivalently, if the model predicts that the market portfolio is conditionally mean-variance efficient, this does not imply that the market is unconditionally mean-variance efficient. We often test the CAPM by seeing if it explains the average returns of some portfolios or (equivalently) if the market is on the unconditional mean-variance frontier. The CAPM may quite well be true (conditionally) and fail these tests; many assets may do better in terms of unconditional mean vs. unconditional variance. The situation is even worse than these comments seem, and is not repaired by simple inclusion of some conditioning information. Models such as the CAPM imply a conditional linear factor model with respect to investors’ information sets. However, the best we can hope to do is to test implications conditioned down on variables that we can observe and include in a test. Thus, a conditional linear factor model is not testable! I like to call this observation the “Hansen-Richard critique” by analogy to the “Roll Critique.” Roll pointed out, among other things, that the wealth portfolio might not be observable, making tests of the CAPM impossible. Hansen and Richard point out that the conditioning information of agents might not be observable, and that one cannot omit it in testing a conditional model. Thus, even if the wealth portfolio were observable, the fact that we cannot observe agents’ information sets dooms tests of the CAPM.

139

C HAPTER 8

8.4

C ONDITIONING INFORMATION

Scaled factors: a partial solution

You can expand the set of factors to test conditional factor pricing models factors = ft+1 ⊗ zt The problem is that the parameters of the factor pricing model mt+1 = at + bt ft+1 may vary over time. A partial solution is to model the dependence of parameters at and bt on variables in the time−t information set; let at = a(zt ), bt = b(zt ) where zt is a vector of variables observed at time t (including a constant). In particular, why not try linear models at = a0 z t , bt = b0 z t Linearity is not restrictive: zt2 is just another instrument. The only criticism one can make is that some instrument zjt is important for capturing the variation in at and bt , and was omitted. For instruments on which we have data, we can meet this objection by trying zjt and seeing whether it does, in fact, enter significantly. However, for instruments zt that are observed by agents but not by us, this criticism remains valid. Linear discount factor models lead to a nice interpretation as scaled factors, in the same way that linearly managed portfolios are scaled returns. With a single factor and instrument, write mt = a(zt ) + b(zt )ft+1

(113)

= a0 + a1 zt + (b0 + b1 zt )ft+1 = a0 + a1 zt + b0 ft+1 + b1 (zt ft+1 ) .

(114)

Thus, in place of the one-factor model with time-varying coefficients (8.113), we have a three-factor model (zt , ft+1 , zt ft+1 ) with fixed coefficients, (8.114). Since the coefficients are now fixed, we can use the scaled-factor model with unconditional moments. pt = Et [(a0 + a1 zt + b0 ft+1 + b1 (zt ft+1 )) xt+1 ] ⇒ E(pt ) = E [(a0 + a1 zt + b0 ft+1 + b1 (zt ft+1 )) xt+1 ] For example, in standard derivations of CAPM, the market (wealth portfolio) return is conditionally mean-variance efficient; investors want to hold portfolios on the conditional 140

S ECTION 8.5

S UMMARY

mean-variance frontier; conditionally expected returns follow a conditional single-beta representation, or the discount factor m follows a conditional linear factor model W mt+1 = at − bt Rt+1

as we saw above. But none of these statements mean that we can use the CAPM unconditionally. Rather than throw up our hands, we can add some scaled factors. Thus, if, say, the dividend/price ratio and term premium do a pretty good job of summarizing variation in conditional moments, the conditional CAPM implies an unconditional, five-factor (plus constant) model. The factors are a constant, the market return, the dividend/price ratio, the term premium, and the market return times the dividend-price ratio and the term premium. The unconditional pricing implications of such a five-factor model could, of course, be summarized by a single−β representation. (See the caustic comments in the section on implications and equivalence.) The reference portfolio would not be the market portfolio, of course, but a mimicking portfolio of the five factors. However, the single mimicking portfolio would not be easily interpretable in terms of a single factor conditional model and two instruments. In this case, it might be more interesting to look at a multiple −β or multiplefactor representation. If we have many factors f and many instruments z, we should in principle multiply every factor by every instrument, m = b1 f1 + b2 f1 z1 + b3 f1 z2 + ... + bN+1 f2 + bN+2 f2 z1 + bN+3 f2 z2 + ... This operation can be compactly summarized with the Kronecker product notation, a ⊗ b, which means “multiply every element in vector a by every element in vector b, or mt+1 = b0 (ft+1 ⊗ zt ).

8.5

Summary

When you first think about it, conditioning information sounds scary – how do we account for time-varying expected returns, betas, factor risk premia, variances, covariances, etc. However, the methods outlined in this chapter allow a very simple and beautiful solution to the problems raised by conditioning information. To express the conditional implications of a given model, all you have to do is include some scaled or managed portfolio returns, and then pretend you never heard about conditioning information. Some factor models are conditional models, and have coefficients that are functions of investors’ information sets. In general, there is no way to test such models, but if you are willing to assume that the relevant conditioning information is well summarized by a few variables, then you can just add new factors, equal to the old factors scaled by the conditioning 141

C HAPTER 8

C ONDITIONING INFORMATION

variables, and again forget that you ever heard about conditioning information. You may want to remember conditioning information as a diagnostic and in economic interpretation of the results. It may be interesting to take estimates of a many factor model, mt = a0 + a1 zt + b0 ft+1 + b1 zt ft+1 , and see what they say about the implied conditional model, mt = (a0 + a1 zt ) + (b0 + b1 zt )ft+1 . You may want to make plots of conditional bs, betas, factor risk premia, expected returns,etc. But you don’t have to worry about it in estimation and testing.

8.6 1. 2.

3.

Problems

If there is a risk free asset, is it on the a) conditional b) unconditional c) both mean-variance frontier? If there is a conditionally riskfree asset – a claim to 1 is traded at each date, does this mean that there is an unconditionally risk free asset? (Define the latter first!) How about vice versa? Suppose you took the unconditional population moments E(R), E(RR0 ) of assets returns and constructed the mean-variance frontier. Does this frontier correspond to the conditional or the unconditional MV frontier, or neither? What is the key assumption underlying your answer?

142

Chapter 9.

Factor pricing models

In Chapter 2, I noted that the consumption-based model, while a complete answer to most asset pricing questions in principle, does not (yet) work well in practice. This observation motivates efforts to tie the discount factor m to other data. Linear factor pricing models are the most popular models of this sort in finance. They dominate discrete time empirical work. Factor pricing models replace the consumption-based expression for marginal utility growth with a linear model of the form mt+1 = a + b0 f t+1 a and b are free parameters. This specification is equivalent to a multiple-beta model E(Rt+1 ) = α + β 0 λ where β are multiple regression coefficients of returns R on the factors f . Here, α and λ are the free parameters. The big question is, what should one use for factors ft+1 ? Factor pricing models look for variables that are good proxies for aggregate marginal utility growth, i.e., variables for which β

u0 (ct+1 ) ≈ a + b0 f t+1 u0 (ct )

(115)

is a sensible and economically interpretable approximation. More directly and interpretably, the essence of asset pricing is that there are special states of the world in which investors are especially concerned that their portfolios not do badly. They are willing to trade off some overall performance – average return – to make sure that portfolios do not do badly in these particular states of nature. The factors are variables that indicate that these “bad states” have occurred. The factors that result from this search are and should be intuitively sensible. In any sensible economic model, as well as in the data, consumption is related to returns on broadbased portfolios, to interest rates, to growth in GNP, investment, or other macroeconomic variables, and to returns on production processes. All of these variables measure “wealth” or the state of the economy. Consumption is and should be high in “good times” and low in “bad times.” Furthermore, consumption and marginal utility respond to news: if a change in some variable today signals high income in the future, then consumption rises now, by permanent income logic. This fact opens the door to forecasting variables: any variable that forecasts asset returns (“changes in the investment opportunity set”) or macroeconomic variables is a candidate factor. Variables such as the term premium, dividend/price ratio, stock returns, etc. can be defended as pricing factors on this logic. Though they themselves are not measures of aggregate good or bad times, they forecast such times. 143

C HAPTER 9

FACTOR PRICING MODELS

Should factors be independent over time? The answer is, sort of. If there is a constant real interest rate, then marginal utility growth should be unpredictable. (“Consumption is a random walk” in the quadratic utility permanent income model.) To see this, just look at the first order condition with a constant interest rate, u0 (ct ) = βRf Et [u0 (ct+1 )] or in a more time-series notation, 1 u0 (ct+1 ) = + εt+1 ; Et (εt+1 ) = 0. 0 u (ct ) βRf The real risk free rate is not constant, but it does not vary a lot, especially compared to asset returns. Measured consumption growth is not exactly unpredictable but it is the least predictable macroeconomic time series, especially if one accounts properly for temporal aggregation (consumption data are quarterly averages). Thus, factors that proxy for marginal utility growth, though they don’t have to be totally unpredictable, should not be highly predictable. If one chooses highly predictable factors, the model will counterfactually predict large interest rate variation. In practice, this consideration means that one should choose the right units: Use GNP growth rather than level, portfolio returns rather than prices or price/dividend ratios, etc. However, unless one wants to impose an exactly constant risk free rate, one does not have to filter or prewhiten factors to make them exactly unpredictable. This view of factors as intuitively motivated proxies for marginal utility growth is sufficient to carry the reader through current empirical tests of factor models. The extra constraints of a formal exposition of theory in this part have not yet constrained the factor-fishing expedition. The precise derivations all proceed in the way I have motivated factor models: One writes down a general equilibrium model, in particular a specification of the production technology by which real investment today results in real output tomorrow. This general equilibrium produces relations that express the determinants of consumption from exogenous variables, and relations linking consumption and other endogenous variables; equations of the form ct = g(ft ). One then uses this kind of equation to substitute out for consumption in the basic first order conditions. The formal derivations accomplish two things: they determine one particular list of factors that can proxy for marginal utility growth, and they prove that the relation should be linear. Some assumptions can often be substituted for others in the quest for these two features of a factor pricing model. This is a point worth remembering: all factor models are derived as specializations of the consumption-based model. Many authors of factor model papers disparage the consumptionbased model, forgetting that their factor model is the consumption-based model plus extra assumptions that allow one to proxy for marginal utility growth from some other variables. 144

S ECTION 9.1

C APITAL A SSET P RICING MODEL (CAPM)

My presentation follows Constantinides’ (1989) derivation of traditional models as instances of the consumption-based model in this regard. Above, I argued that clear economic foundation was important for factor models, since it is the only guard against fishing. Alas, we discover here that the current state of factor pricing models is not a particularly good guard against fishing. One can call for better theories or derivations, more carefully aimed at limiting the list of potential factors and describing the fundamental macroeconomic sources of risk, and thus providing more discipline for empirical work. The best minds in finance have been working on this problem for 40 years though, so a ready solution is not immediately in sight. On the other hand, we will see that even current theory can provide much more discipline than is commonly imposed in empirical work. For example, the derivations of the CAPM and ICAPM do leave predictions for the risk free rate and for factor risk premia that are often ignored. The ICAPM gives tighter restrictions on state variables than are commonly checked: “State variables” do have to forecast something! We also see how special and unrealistic are the general equilibrium setups necessary to derive popular specifications such as CAPM and ICAPM. This observation motivates a more serious look at real general equilibrium models below.

9.1

Capital Asset Pricing Model (CAPM)

The CAPM is the model m = a + bRw ; Rw = wealth portfolio return. I derive it from the consumption based model by 1) Two period quadratic utility; 2) Two periods, exponential utility and normal returns; 3) Infinite horizon, quadratic utility and i.i.d. returns; 4) Log utility and normally distributed returns. The CAPM is the first, most famous and (so far) most widely used model in asset pricing. It ties the discount factor m to the return on the “wealth portfolio.” The function is linear, W . mt+1 = a + bRt+1

a and b are free parameters. One can find theoretical values for the parameters a and b by requiring the discount factor m to price any two assets, such as the wealth portfolio return and risk-free rate, 1 = E(mRW ) and 1 = E(m)Rf . (As an example, we did this in equation (8.110) above.) In empirical applications, we can also pick a and b to “best” price larger cross-sections of assets. We do not have good data on, or even a good empirical definition for, the return on total wealth. It is conventional to proxy RW by the return on a broad-based stock portfolio such as the value- or equally-weighted NYSE, S&P500, etc. The CAPM is of course most frequently stated in equivalent expected return / beta language, E(Ri ) = α + β i,RW [E(Rw ) − α] . 145

C HAPTER 9

FACTOR PRICING MODELS

This section briefly describes some classic derivations of the CAPM. Again, we need to find assumptions that defend which factors proxy for marginal utility (RW here), and assumptions to defend the linearity between m and the factor. I present several derivations of the same model. Many of these derivations use classic modeling assumptions which are important in their own sake. This is also an interesting place in which to see that various sets of assumptions can often be used to get to the same place. The CAPM is often criticized for one or another assumption. By seeing several derivations, we can see how one assumption can be traded for another. For example, the CAPM does not in fact require normal distributions, if one is willing to swallow quadratic utility instead. 9.1.1

Two-period quadratic utility

Two period investors with no labor income and quadratic utility imply the CAPM. Investors have quadratic preferences and only live two periods, 1 1 U(ct , ct+1 ) = − (ct − c∗ )2 − βE[(ct+1 − c∗ )2 ]. 2 2

(116)

Their marginal rate of substitution is thus mt+1 = β

(ct+1 − c∗ ) u0 (ct+1 ) =β . 0 u (ct ) (ct − c∗ )

The quadratic utility assumption means marginal utility is linear in consumption. Thus, the first target of the derivation, linearity. Investors are born with wealth Wt in the first period and earn no labor income. They can invest in lots of assets with prices pit and payoffs xit+1 , or, to keep the notation simple, i . They choose how much to consume at the two dates, ct and ct+1 , and the returns Rt+1 portfolio weights αi for their investment portfolio. Thus, the budget constraint is (117)

ct+1 = Wt+1 Wt+1 = RW t+1 (Wt − ct ) RW =

N X

αi Ri ;

i=1

N X i=1

RW is the rate of return on total wealth. 146

αi = 1.

S ECTION 9.1

C APITAL A SSET P RICING MODEL (CAPM)

The two-period assumption means that investors consume everything in the second period, by constraint (9.117). This fact allows us to substitute wealth and the return on wealth for consumption, achieving the second goal of the derivation, naming the factor that proxies for consumption or marginal utility: mt+1 = β i.e.

W Rt+1 (Wt − ct ) − c∗ −βc∗ β(Wt − ct ) W = + Rt+1 ct − c∗ ct − c∗ ct − c∗

mt+1 = at + bt RW t+1 . 9.1.2

Exponential utility, normal distributions

u(c) = −e−αc and a normally distributed set of returns also produces the CAPM. The combination of exponential utility and normal distributions is another set of assumptions that deliver the CAPM in a one or two period model. This structure has a particularly convenient analytical form. Since it gives rise to linear demand curves, it is very widely used in models that complicate the trading structure, by introducing incomplete markets or asymmetric information. I present a model with consumption only in the last period. (You can do the quadratic utility model of the last section this way as well.) Utility is ¤ £ E [u(c)] = E −e−αc . α is known as the coefficient of absolute risk aversion. If consumption is normally distributed, we have Eu(c) = −e−αE(c)+

α2 2

σ 2 (c)

.

Suppose this investor has initial wealth W which can be split between a riskfree asset paying Rf and a set of risky assets paying return R. Let y denote the amount of this wealth W (amount, not fraction) invested in each security. Then, the budget constraint is c = y f Rf + y 0 R W = y f + y0 1 Plugging the first constraint into the utility function we obtain Eu(c) = −e−α[y

f

2

Rf +y 0 E(R)]+ α2 y 0 Σy

147

.

(118)

C HAPTER 9

FACTOR PRICING MODELS

As with quadratic utility, the two-period model is what allows us to set consumption to wealth and then substitute the return on the wealth portfolio for consumption growth in the discount factor. Maximizing (9.118) with respect to y, y f , we obtain the first order condition describing the optimal amount to be invested in the risky asset, y = Σ−1

E(R) − Rf α

Sensibly, the investor invests more in risky assets if their expected return is higher, less if his risk aversion coefficient is higher, and less if the assets are riskier. Notice that total wealth does not appear in this expression. With this setup, the amount invested in risky assets is independent of the level of wealth. This is why we say that this investor has an aversion to absolute rather than relative (to wealth) risk aversion. Note also that these “demands” for the risky assets are linear in expected returns, which is a very convenient property. Inverting the first order conditions, we obtain E(R) − Rf = αΣy = α cov(R, Rm ).

(119)

The investor’s total risky portfolio is y 0 R. Hence, Σy gives the covariance of each return with y0 R, and also with the investor’s overall portfolio y f Rf + y0 R. If all investors are identical, then the market portfolio is the same as the individual’s portfolio so Σy also gives the correlation of each return with Rm = yf Rf + y 0 R. (If investors differ in risk aversion α, the same thing goes through but with an aggregate risk aversion coefficient.) Thus, we have the CAPM. This version is especially interesting because it ties the market price of risk to the risk aversion coefficient. Applying (9.119) to the market return itself, we have E(Rm ) − Rf = α. σ2 (Rm ) 9.1.3

Quadratic value function, dynamic programming.

We can let investors live forever in the quadratic utility CAPM so long as we assume that the environment is independent over time. Then the value function is quadratic, taking the place of the quadratic second-period utility function. This case is a nice first introduction to dynamic programming. The two-period structure given above is unpalatable, since (most) investors do in fact live longer than two periods. It is natural to try to make the same basic ideas work with less 148

S ECTION 9.1

C APITAL A SSET P RICING MODEL (CAPM)

restrictive and more palatable assumptions. We can derive the CAPM in a multi-period context by replacing the second-period quadratic utility function with a quadratic value function. However, the quadratic value function requires the additional assumption that returns are i.i.d. (no “shifts in the investment opportunity set”). This observation, due to Fama (1970), is also a nice introduction to dynamic programming, which is a powerful way to handle multiperiod problems by expressing them as two period problems. Finally, I think this derivation makes the CAPM more realistic, transparent and intuitively compelling. Buying stocks amounts to taking bets over wealth; really the fundamental assumption driving the CAPM is that marginal utility of wealth is linear in wealth and does not depend on other state variables. Let’s start in a simple ad-hoc manner by just writing down a “utility function” defined over this period’s consumption and next period’s wealth, U = u(ct ) + βEt V (Wt+1 ). This is a reasonable objective for an investor, and does not require us to make the very artificial assumption that he will die tomorrow. If an investor with this “utility function” can buy an asset at price pt with payoff xt+1 , his first order condition (buy a little more, then x contributes to wealth next period) is pt u0 (ct ) = βEt [V 0 (Wt+1 )xt+1 ] . Thus, the discount factor uses next period’s marginal value of wealth in place of the more familiar marginal utility of consumption mt+1 = β

V 0 (Wt+1 ) u0 (ct )

(The envelope condition states that, at the optimum, a penny saved has the same value as a penny consumed u0 (ct ) = V 0 (Wt ). We could use this condition to express the denominator in terms of wealth also.) Now, suppose the value function were quadratic, η V (Wt+1 ) = − (Wt+1 − W ∗ )2 . 2 Then, we would have mt+1

W Rt+1 (Wt − ct ) − W ∗ Wt+1 − W ∗ = −βη = −βη u0 (ct ) u0 (ct ) · ¸ · ¸ ∗ βηW βη(Wt − ct ) W = , + − Rt+1 u0 (ct ) u0 (ct )

149

C HAPTER 9

FACTOR PRICING MODELS

or, once again, mt+1 = at + bt RW t+1 , the CAPM! Let’s be clear about the assumptions and what they do. 1) The value function only depends on wealth. If other variables entered the value function, then ∂V /∂W would depend on those other variables, and so would m. This assumption bought us the first objective of any derivation: the identity of the factors. The ICAPM, below, allows other variables in the value function, and obtains more factors. (Actually, other variables could enter so long as they don’t affect the marginal value of wealth. The weather is an example: You like me might be happier on sunny days, but you do not value additional wealth more on sunny than on rainy days. Hence, covariance with weather does not affect how you value stocks.) 2) The value function is quadratic. We wanted the marginal value function V 0 (W ) be linear, to buy us the second objective, showing m is linear in the factor. Quadratic utility and value functions deliver a globally linear marginal value function V 0 (W ). By the usual Taylor series logic, linearity of V 0 (W ) is probably not a bad assumption for small perturbations, and not a good one for large perturbations. Why is the value function quadratic? You might think we are done. But economists are unhappy about a utility function that has wealth in it. Few of us are like Disney’s Uncle Scrooge, who got pure enjoyment out of a daily swim in the coins in his vault. Wealth is valuable because it gives us access to more consumption. Utility functions should always be written over consumption. One of the few real rules in economics that keep our theories from being vacuous is that ad-hoc “utility functions” over other objects like wealth (or means and variances of portfolio returns, or “status” or “political power”) should be defended as arising from a more fundamental desire for consumption. More practically, being careful about the derivation makes clear that the superficially plausible assumption that the value function is only a function of wealth derives from the much less plausible, in fact certainly false, assumption that interest rates are constant, the distribution of returns is i.i.d., and that the investor has no risky labor income. So, let us see what it takes to defend the quadratic value function in terms of some utility function. Suppose investors last forever, and have the standard sort of utility function ∞ 1 X j β u(ct+j ). U = − Et 2 j=0

Again, investors start with wealth W0 which earns a random return RW and they have no other source of income. In addition, suppose that interest rates are constant, and stock returns 150

S ECTION 9.1

C APITAL A SSET P RICING MODEL (CAPM)

are i.i.d. over time. Define the value function as the maximized value of the utility function in this environment. Thus, define V (W ) as7 V (Wt ) ≡ max{ct ,ct+1 ,ct+2 ...αt ,αt+1 ,...} Et s.t. Wt+1

∞ X

β j u(ct+j )

(9.120)

j=0

W 0 0 = RW t+1 (Wt − ct ); Rt = αt Rt ; αt 1 = 1

¤0 £ (I used vector notation to simplify the statement of the portfolio problem; R ≡ R1 R2 ... RN , etc.) The value function is the total level of utility the investor can achieve, given how much wealth he has, and any other variables constraining him. This is where the assumptions of no labor income, a constant interest rate and i.i.d. returns come in. Without these assumptions, the value function as defined above might depend on these other characteristics of the investor’s environment. For example, if there were some variable, say, “D/P” that indicated returns would be high or low for a while, then the investor would be happier, and have a high value, when D/P is high, for a given level of wealth. Thus, we would have to write V (Wt , D/Pt ) Value functions allow you to express an infinite period problem as a two period problem. Break up the maximization into the first period and all the remaining periods, as follows ∞ X Et+1 β j u(ct+1+j ) s. t. .. max V (Wt ) = max{ct ,αt } u(ct ) + βEt {ct+1 ,ct+2 ..,αt+1 ,αt+2 ....} j=0

or

V (Wt ) = max{ct ,αt } {u(ct ) + βEt V (Wt+1 )} s.t. ...

(121)

Thus, we have defended the existence of a value function. Writing down a two period “utility function” over this period’s consumption and next period’s wealth is not as crazy as it might seem. The value function is also an attractive view of how people actually make decisions. You don’t think “If I buy a sandwich today, I won’t be able to go out to dinner one night 20 years from now” – trading off goods directly as expressed by the utility function. You think “I can’t afford a new car” meaning that the decline in the value of wealth is not worth the increase in the marginal utility of consumption. Thus, the maximization in (9.121) describes your psychological approach to utility maximization. 7 There is also a transversality condition or a lower limit on wealth in the budget constraints. This keeps the consumer from consuming a bit more and rolling over more and more debt, and it means we can write the budget constraint in present value form.

151

C HAPTER 9

FACTOR PRICING MODELS

The remaining question is, can the value function be quadratic? What utility function assumption leads to a quadratic value function? Here is the fun fact: A quadratic utility function leads to a quadratic value function in this environment. This is not a law of nature; it is not true that for any u(c), V (W ) has the same functional form. But it is true here and a few other special cases. The “in this environment” clause is not innocuous. The value function – the achieved level of expected utility – is a result of the utility function and the constraints. How could we show this fact? One way would be to try to calculate the value function by brute force from its definition, equation (9.120). This approach is not fun, and it does not exploit the beauty of dynamic programming, which is the reduction of an infinite period problem to a two period problem. Instead solve (9.121) as a functional equation. Guess that the value function V (Wt+1 ) is quadratic, with some unknown parameters. Then use the recursive definition of V (Wt ) in (9.121), and solve a two period problem–find the optimal consumption choice, plug it into (9.121) and calculate the value function V (Wt ). If the guess was right, you obtain a quadratic function for V (Wt ), and determine any free parameters. Let’s do it. Specify 1 u(ct ) = − (ct − c∗ )2 . 2 Guess γ V (Wt+1 ) = − (Wt+1 − W ∗ )2 2 with γ and W ∗ parameters to be determined later. Then the problem (9.121) is (I don’t write the portfolio choice α part for simplicity; it doesn’t change anything) · ¸ γ 1 V (Wt ) = max − (ct − c∗ )2 − β E(Wt+1 − W ∗ )2 s. t. Wt+1 = RW t+1 (Wt − ct ). {ct } 2 2 (Et is now E since I assumed i.i.d.) Substituting the constraint into the objective, · ¸ ¤ γ £ W 1 ∗ 2 ∗ 2 V (Wt ) = max − (ct − c ) − β E Rt+1 (Wt − ct ) − W . 2 2 {ct }

The first order condition with respect to ct , using cˆ to denote the optimal value, is ¤ W ª ©£ ˆt ) − W ∗ Rt+1 cˆt − c∗ = βγE RW t+1 (Wt − c

Solving for cˆt ,

cˆt = c∗ + βγE

¤ª ©£ W 2 W2 W Rt+1 Wt − cˆt Rt+1 − W ∗ Rt+1 152

(122)

S ECTION 9.1

C APITAL A SSET P RICING MODEL (CAPM)

£ ¤ 2 ∗ W2 ∗ W cˆt 1 + βγE(RW t+1 ) = c + βγE(Rt+1 )Wt − βγW E(Rt+1 ) cˆt =

∗ W2 c∗ − βγE(RW t+1 )W + βγE(Rt+1 )Wt 2 1 + βγE(RW t+1 )

(123)

This is a linear function of Wt . Writing (9.122) in terms of the optimal value of c, we get ¤2 1 γ £ W V (Wt ) = − (ˆ ct − c∗ )2 − β E Rt+1 (Wt − cˆt ) − W ∗ 2 2

(124)

This is a quadratic function of Wt and cˆ. A quadratic function of a linear function is a quadratic function, so the value function is a quadratic function of Wt . If you want to spend a pleasant few hours doing algebra, plug (9.123) into (9.124), check that the result really is quadratic in Wt , and determine the coefficients γ, W ∗ in terms of fundamental parameters β, c∗ , E(RW ), E(RW 2 ) (or σ2 (RW )). The expressions for γ, W ∗ do not give much insight, so I don’t do the algebra here. 9.1.4

Log utility

Log utility rather than quadratic utility also implies a CAPM. Log utility implies that consumption is proportional to wealth, allowing us to substitute the wealth return for consumption data. The point of the CAPM is to avoid the use of consumption data, and so to use wealth or the rate of return on wealth instead. Log utility is another special case that allows this substitution. Log utility is much more plausible than quadratic utility. Suppose that the investor has log utility u(c) = ln(c). Define the wealth portfolio as a claim to all future consumption. Then, with log utility, the price of the wealth portfolio is proportional to consumption itself. pW t = Et

∞ X j=1

βj

∞ X β u0 (ct+j ) ct c ct = E βj ct+j = t+j t u0 (ct ) c 1 − β t+j j=1

The return on the wealth portfolio is proportional to consumption growth, W Rt+1

pW + ct+1 = t+1 W = pt

β 1−β + 1 ct+1 β ct 1−β

153

=

1 ct+1 1 u0 (ct ) . = β ct β u0 (ct+1 )

C HAPTER 9

FACTOR PRICING MODELS

Thus, the log utility discount factor equals the inverse of the wealth portfolio return, mt+1 =

1 . RW t+1

(125)

Equation (9.125) could be used by itself: it attains the goal of replacing consumption data by some other variable. (Brown and Gibbons 1982 test a CAPM in this form.) Note that log utility is the only assumption so far. We do not assume constant interest rates, i.i.d. returns or the absence of labor income. Log utility has a special property that “income effects offset substitution effects,” or in an asset pricing context that “discount rate effects offset cashflow effects.” News of higher consumption = dividend should make the claim to consumption more valuable. However, through u0 (c) it also raises the discount rate, lowering the value of the claim to consumption. For log utility, these two effects exactly offset. 9.1.5

Linearizing any model: Taylor approximations and normal distributions.

Any nonlinear model m = f(z) can be turned into a linear model m = a + bz in discrete time by assuming normal returns. It is traditional in the CAPM literature to try to derive a linear relation between m and the wealth portfolio return. We could always do this by a Taylor approximation, mt+1 ∼ = at + bt RW t+1 . We can make this approximation exact in a special case, that the factors and all asset returns are normally distributed. (We can also take the continuous time limit, which is really the same thing. However, this discrete-time trick is common and useful.) First, I quote without proof the central mathematical trick as a lemma Lemma 1 (Stein’s lemma) If f, R are bivariate normal, g(f ) is differentiable and E | g0 (f) |< ∞, then cov [g(f), R] = E[g 0 (f )] cov(f, R).

(126)

Now we can use the lemma to state the theorem. Theorem 2 If m = g(f ), if f and a set of the payoffs priced by m are normally distributed returns, and if |E[g0 (f )]| < ∞, then there is a linear model m = a + bf that prices the normally distributed returns. 154

S ECTION 9.1

C APITAL A SSET P RICING MODEL (CAPM)

Proof: First, the definition of covariance means that the pricing equation can be rewritten as a restriction between mean returns and the covariance of returns with m: 1 = E(mR) ⇔ 1 = E(m)E(R) + cov(m, R).

(127)

Now, given m = g(f), f and R jointly normal, apply Stein’s lemma (9.126) and (9.127), 1 = E[g(f)]E(R) + E[g0 (f)]cov(f, R) 1 = E[g(f)]E(R) + cov(E[g 0 (f)]f, R) Exploiting the ⇐ part of (9.127), we know that an m with mean E(g(f )) and that depends on f via E(g 0 (f ))f will price assets, m = E[g(f )] + E[g0 (f )][f − E(f )]. ¥ Using this trick, and recalling that we have not assumed i.i.d. so all these moments are conditional, the log utility CAPM implies the linear model "µ µ ¶ ¶2 # ¤ £ W 1 1 mt+1 = Et (128) Rt+1 − Et (RW − Et t+1 ) W Rt+1 RW t+1 if RW t+1 and all asset returns to be priced are normally distributed. From here it is a short step to an expected return-beta representation using the wealth portfolio return as the factor.

In the same way, we can trade the quadratic utility function for normal distributions in the dynamic programming derivation of the CAPM. Starting from £ ¤ V 0 RW V 0 (Wt+1 ) t+1 (Wt − ct ) =β mt+1 = β u0 (ct ) u0 (ct ) we can derive an expression that links m linearly to RW t+1 by assuming normality. Using the same trick, the consumption-based model can be written in linear fashion, i.e. expected returns can be expressed as a linear function of betas on consumption growth rather than betas on consumption growth raised to a power. However, for large risk aversion coefficients (more than about 10 in postwar consumption data) or other transformations, the inaccuracies due to the normal or lognormal approximation can be very significant in discrete data. The normal distribution assumption seems rather restrictive, and it is. However, the most popular class of continuous-time models specify instantaneously normal distributions even for things like options that have very non-normal distributions for discrete time intervals. 155

C HAPTER 9

FACTOR PRICING MODELS

Therefore, one can think of the Stein’s lemma tricks as a way to get to continuous time approximations without doing it in continuous time. I demonstrate the explicit continuous time approach with the ICAPM, in the next section. 9.1.6

Portfolio intuition

The classic derivation of the CAPM contains some useful intuition. The classic derivation starts with a mean-variance objective for portfolio wealth, max Eu(W ). Beta drives average returns because beta measures how much adding a bit of the asset to a diversified portfolio increases the volatility of the portfolio. The central insight that started it all is that investors care about portfolio returns, not about the behavior of specific assets. Once the characteristics of portfolios replaced demand curves for individual stocks, modern finance was born.

9.2

Intertemporal Capital Asset Pricing Model (ICAPM)

Any “state variable” zt can be a factor. The ICAPM is a linear factor model with wealth and state variables that forecast changes in the distribution of future returns or income. The ICAPM generates linear discount factor models mt+1 = a + b0 ft+1 in which the factors are “state variables” for the investor’s consumption-portfolio decision. The “state variables” are the variables that determine how well the investor can do in his maximization. Current wealth is obviously a state variable. Additional state variables describe the conditional distribution of income and asset returns the agent will face in the future or “shifts in the investment opportunity set.” In multiple good or international models, relative price changes are also state variables. Optimal consumption is a function of the state variables, ct = g(zt ). We can use this fact once again to substitute out consumption, and write mt+1 = β

u0 [g(zt+1 )] . u0 [g(zt )]

From here, it is a simple linearization to deduce that the state variables zt+1 will be factors. Alternatively, the value function depends on the state variables V (Wt+1 , zt+1 ), 156

S ECTION 9.2

I NTERTEMPORAL C APITAL A SSET P RICING M ODEL (ICAPM)

so we can write mt+1 = β

VW (Wt+1 , zt+1 ) VW (Wt , zt )

(The marginal value of a dollar must be the same in any use, so I made the denominator pretty by writing u0 (ct ) = VW (Wt , zt ). This fact is known as the envelope condition.) This completes the first step, naming the proxies. To obtain a linear relation, we can take a Taylor approximation, assume normality and use Stein’s lemma, or, most conveniently, move to continuous time (which is really just a more convenient way of making the normal approximation.) We saw above that we can write the basic pricing equation in continuous time as µ ¶ dΛ dp dp − rf dt = −E . E p Λ p (for simplicity of the formulas, I’m folding any dividends into the price process). The discount factor is marginal utility, which is the same as the marginal value of wealth, dVW (Wt , zt ) du0 (ct ) dΛt = = 0 Λt u (ct ) VW Our objective is to express the model in terms of factors z rather than marginal utility or value, and Ito’s lemma makes this easy VW z W VW W dW 1 dVW + = dz + (second derivative terms) VW VW W VW 2 (We don’t have to grind out the second derivative terms if we are going to take rf dt = Et (dΛ/Λ) , though this approach removes a potentially interesting and testable implication of the model). The elasticity of marginal value with respect to wealth is often called the coefficient of relative risk aversion, rra ≡ −

W VW W . VW

Substituting, we obtain the ICAPM, which relates expected returns to the covariance of returns with wealth, and also with the other state variables, µ µ ¶ ¶ dW dp dp dp VW z − rf dt = rra E E dz E − . p W p VW p From here, it is fairly straightforward to express the ICAPM in terms of betas rather than covariances, or as a linear discount factor model. Most empirical work occurs in discrete time; we often simply approximate the continuous time result as E(R) − Rf ≈ rra cov(R, ∆W ) + λz cov(R, ∆z). 157

C HAPTER 9

FACTOR PRICING MODELS

One often substitutes covariance with the wealth portfolio for covariance with wealth, and one uses factor-mimicking portfolios for the other factors dz as well. The factor-mimicking portfolios are interesting for portfolio advice as well, as they give the purest way of hedging against or profiting from state variable risk exposure. This short derivation does not do justice to the beauty of Merton’s portfolio theory and ICAPM. What remains is to actually state the consumer’s problem and prove that the value function depends on W and z, the state variables for future investment opportunities, and that the optimal portfolio holds the market and hedge portfolios for the investment opportunity variables.

9.3

Comments on the CAPM and ICAPM

Conditional vs. unconditional models. Do they price options? Why bother linearizing? The wealth portfolio. Ex-post returns. The implicit consumption-based model. What are the ICAPM state variables? CAPM and ICAPM as general equilibrium models Is the CAPM conditional or unconditional? Is the CAPM a conditional or an unconditional factor model? I.e., are the parameters a and b in m = a − bRW constants, or do they change at each time period, as conditioning information changes? We saw above that a conditional CAPM does not imply an unconditional CAPM, so additional steps must be taken to say anything about observed average returns. The two period quadratic utility based derivation results in a conditional CAPM, since the parameters at and bt depend on consumption which changes over time. Also we know that a and b must vary over time if the conditional moments of RW , Rf vary over time. This twoperiod investor chooses a portfolio on the conditional mean variance frontier, which is not on the unconditional frontier. The multiperiod quadratic utility CAPM only holds if returns are i.i.d. so it only holds if there is no difference between conditional and unconditional models. The log utility CAPM expressed with the inverse market return is a beautiful model, since it holds both conditionally and unconditionally. There are no free parameters that can change

158

S ECTION 9.3

C OMMENTS ON THE CAPM AND ICAPM

with conditioning information: 1 = Et

µ

1 RW t+1

Rt+1

¶

⇔1=E

µ

1 RW t+1

¶

Rt+1 .

In fact there are no free parameters at all! Furthermore, the model makes no distributional assumptions, so it can apply to any asset, including options. Finally it requires no specification of the investment opportunity set, or (macro language) no specification of technology. Linearizing the log utility CAPM comes at enormous price. The expectations in the linearized log utility CAPM (9.128) are conditional. Thus, the apparent simplification of linearity destroys the nice unconditional feature of the log utility CAPM. Should the CAPM price options? As I have derived them, the quadratic utility CAPM and the nonlinear log utility CAPM should apply to all payoffs: stocks, bonds, options, contingent claims, etc. However, if we assume normal return distributions to obtain a linear CAPM from log utility, we can no longer hope to price options, since option returns are non-normally distributed (that’s the point of options!) Even the normal distribution for regular returns is a questionable assumption. You may hear the statement “the CAPM is not designed to price derivative securities”; the statement refers to the log utility plus normal-distribution derivation of the linear CAPM. Why linearize? Why bother linearizing a model? Why take the log utility model m = 1/RW which W that loses the clean conditioningshould price any asset, and turn it into mt+1 = at + bt Rt+1 down property and cannot price non-normally distributed payoffs? These tricks were developed before the p = E(mx) expression of asset pricing models, when (linear) expected return-beta models were the only thing around. You need a linear model of m to get an expected return - beta model. More importantly, the tricks were developed when it was hard to estimate nonlinear models. It’s clear how to estimate a β and a λ by regressions, but estimating nonlinear models used to be a big headache. Now, GMM has made it easy to estimate and evaluate nonlinear models. Thus, in my opinion, linearization is mostly intellectual baggage. The desire for linear representations and this normality trick is one of the central reasons why many asset pricing models are written in continuous time. In most continuous time models, everything is locally normal. Unfortunately for empiricists, this approach adds timeaggregation and another layer of unobservable conditioning information into the predictions of the model. For this reason, most empirical work is still based on discrete-time models. However, the local normal distributions in continuous time, even for option returns, is a good reminder that normal approximations probably aren’t that bad, so long as the time interval is kept reasonably short. What about the wealth portfolio? The log utility derivation makes clear just how expansive is the concept of the wealth portfolio. To own a (share of) the consumption stream, you have to own not only all stocks, 159

C HAPTER 9

FACTOR PRICING MODELS

but all bonds, real estate, privately held capital, publicly held capital (roads, parks, etc.), and human capital – a nice word for “people.” Clearly, the CAPM is a poor defense of common proxies such as the value-weighted NYSE portfolio. And keep in mind that since it is easy to find ex-post mean-variance efficient portfolios of any subset of assets (like stocks) out there, taking the theory seriously is our only guard against fishing. Implicit consumption-based models Many users of alternative models clearly are motivated by a belief that the consumptionbased model doesn’t work, no matter how well measured consumption might be. This view is not totally unreasonable; as above, perhaps transactions costs de-link consumption and asset returns at high frequencies, and some diagnostic evidence suggests that the consumption behavior necessary to save the consumption model is too wild to be believed. However, the derivations make clear that the CAPM and ICAPM are not alternatives to the consumption-based model, they are special cases of that model. In each case mt+1 = βu0 (ct+1 )/u0 (ct ) still operates. We just added assumptions that allowed us to substitute other variables in place of ct . One cannot adopt the CAPM on the belief that the consumption based model is wrong. If you think the consumption-based model is wrong, the economic justification for the alternative factor models evaporates. The only plausible excuse for factor models is a belief that consumption data are unsatisfactory. However, while asset return data are well measured, it is not obvious that the S&P500 or other portfolio returns are terrific measures of the return to total wealth. “Macro factors” used by Chen, Roll and Ross (1986) and others are distant proxies for the quantities they want to measure, and macro factors based on other NIPA aggregates (investment, output, etc.) suffer from the same measurement problems as aggregate consumption. In large part, the “better performance” of the CAPM and ICAPM relative to consumptionbased models comes from throwing away content. Again mt+1 = δu0 (ct+1 )/u0 (ct ) is there in any CAPM or ICAPM. The CAPM and ICAPM make predictions concerning consumption data that are wildly implausible, not only of admittedly poorly measured aggregate consumption data but any imaginable perfectly measured individual consumption data as well. For example, equation (9.129) says that the standard deviation of the wealth portfolio return equals the standard deviation of consumption growth. The latter is about 1% per year. All the miserable failures of the log-utility consumption-based model apply equally to the log utility CAPM. Finally, most models take the market price of risk as a free parameter. Of course it isn’t; it is related to risk aversion and consumption volatility and is very hard to justify as such. Ex-post returns The log utility model also allows us for the first time to look at what moves returns ex-post as well as ex-ante. Recall that, in the log utility model, we have RW t+1 =

1 ct+1 . β ct

160

(129)

S ECTION 9.3

C OMMENTS ON THE CAPM AND ICAPM

Thus, the wealth portfolio return is high, ex-post, when consumption is high. This holds at every frequency: If stocks go up between 12:00 and 1:00, it must be because (on average) we all decided to have a big lunch. This seems silly. Aggregate consumption and asset returns are likely to be de-linked at high frequencies, but how high (quarterly?) and by what mechanism are important questions to be answered. In any case, this is another implication of the log utility CAPM that is just thrown out. In sum, the poor performance of the consumption-based model is an important nut to chew on, not just a blind alley or failed attempt that we can safely disregard and go on about our business. Identity of state variables The ICAPM does not tell us the identity of the state variables zt , and many authors use the ICAPM as an obligatory citation to theory on the way to using factors composed of ad-hoc portfolios, leading Fama (1991) to characterize the ICAPM as a “fishing license.” The ICAPM really isn’t quite such an expansive license. One could do a lot to insist that the factor-mimicking portfolios actually are the projections of some identifiable state variables on to the space of returns, and one could do a lot to make sure the candidate state variables really are plausible state variables for an explicitly stated optimization problem. For example, one could check that investment-opportunity set state variables actually do forecast something. The fishing license comes as much from habits of applying the theory as from the theory itself. General equilibrium models The CAPM and other models are really general equilibrium models. Looking at the derivation through general-equilibrium glasses, we have specified a set of linear technologies with returns Ri that do not depend on the amount invested. Some derivations make further assumptions, such as an initial capital stock, and no labor or labor income. The CAPM is obviously very artificial. Its central place really comes from its long string of empirical successes rather than its theoretical purity. The theory was extended and multiple factors anticipated long before they became empirically popular. Portfolio intuition I have derived all the models as instances of the consumption-based model. The more traditional portfolio intuition for multifactor models is also useful. The intuition (and historical development) comes from looking past consumption to its determinants in sources of income or news. The CAPM simplifies matters by assuming that the average investor only cares the performance of his investment portfolio. Most of us have jobs, so events like recessions hurt the majority of investors. People with jobs will prefer stocks that don’t fall in recessions, even if their market betas, mean returns, and standard deviations are the same as stocks that do fall in recessions. Demanding such stocks, they drive down the corresponding expected returns. Thus, we expect expected returns to depend on additional betas that capture labor market 161

C HAPTER 9

FACTOR PRICING MODELS

conditions. The traditional ICAPM intuition works the same way. Even jobless investors have long horizons. Thus, they will prefer stocks that do well when news comes that future returns are lower. Demanding more of such stocks, they depress expected returns. Thus, expected returns come to depend on covariation with news of future returns, not just covariation with the current market return. The ICAPM remained on the theoretical shelf for 20 years mostly because it took that long to accumulate empirical evidence that returns are, in fact, predictable. It is vitally important that the extra factors affect the average investor. If an event makes investor A worse off and investor B better off, then investor A buys assets that do well when the event happens, and investor B sells them. They transfer the risk of the event, but the price or expected return of the asset is unaffected. For a factor to affect prices or expected returns, the average investor must be affected by it, so investors collectively bid up or down the price and expected return of assets that covary with the event rather than just transfer the risk without affecting equilibrium prices. As you can see, this traditional intuition is encompassed by consumption. Bad labor market outcomes or bad news about future returns are bad news that raise the marginal utility of wealth, which equals the marginal utility of consumption.

9.4

Arbitrage Pricing Theory (APT)

The APT: If a set of asset returns are generated by a linear factor model Ri = E(Ri ) +

N X

β ij f˜j + εi

j=1

E(εi ) = E(εi f˜j ) = 0. Then (with additional assumptions) there is a discount factor m linear in the factors m = a + b0 f that prices the returns. The APT starts from a statistical characterization. There is a big common component to stock returns: when the market goes up, most individual stocks also go up. Beyond the market, groups of stocks move together such as computer stocks, utilities, small stocks, value stocks and so forth. Finally, each stock’s return has some completely idiosyncratic movement. This is a characterization of realized returns, outcomes or payoffs. The point of the APT is to start with this statistical characterization of outcomes, and derive something about expected returns or prices. The intuition behind the APT is that the completely idiosyncratic movements in asset 162

S ECTION 9.4

A RBITRAGE P RICING T HEORY (APT)

returns should not carry any risk prices, since investors can diversify them away by holding portfolios. Therefore, risk prices or expected returns on a security should be related to the security’s covariance with the common components or “factors” only. The job of this section is then 1) to describe a mathematical model of the tendency for stocks to move together, and thus to define the “factors” and residual idiosyncratic components, and 2) to think carefully about what it takes for the idiosyncratic components to have zero (or small) risk prices, so that only the common components matter to asset pricing. There are two lines of attack for the second item. 1) If there were no residual, then we could price securities from the factors by arbitrage (really, by the law of one price, but the current distinction between law of one price and arbitrage came after the APT was named.) Perhaps we can extend this logic and show that if the residuals are small, they must have small risk prices. 2) If investors all hold well-diversified portfolios, then only variations in the factors drive consumption and hence marginal utility. Much of the original appeal and marketing of the APT came from the first line of attack, the idea that we could derive pricing implications without the economic structure required of the CAPM, ICAPM, or any other model derived as a specialization of the consumptionbased model. In this section, I will first try to see how far we can in fact get with purely law of one price arguments. I will conclude that the answer is, “not very far,” and that the most satisfactory argument for the APT is in fact just another specialization of the consumptionbased model. 9.4.1

Factor structure in covariance matrices

I define and examine the factor decomposition xi = αi + β 0i f + εi ; E(εi ) = 0, E(fεi ) = 0 The factor decomposition is equivalent to a restriction on the payoff covariance matrix. The APT models the tendency of asset payoffs (returns) to move together via a statistical factor decomposition xi = αi +

M X

β ij fj + εi = αi + β 0i f + εi .

(130)

j=1

The fj are the factors, the β ij are the betas or factor loadings and the εi are residuals. letter without subscripts to denote a vector, for example f = £As usual, I use the ¤same 0 f1 f2 ... fK . A discount factor m, pricing factors f in m = b0 f and this factor decomposition (or factor structure) for returns are totally unrelated uses of the word “factor.” 163

C HAPTER 9

FACTOR PRICING MODELS

I didn’t invent the terminology! The APT is conventionally written with xi = returns, but it ends up being much less confusing to use prices and payoffs. It is a convenient and conventional simplification to fold the factor means into the first, constant, factor and write the factor decomposition with zero-mean factors f˜ ≡ f − E(f). xi = E(xi ) +

M X

β ij f˜j + εi .

(131)

j=1

Remember that E(xi ) is still just a statistical characterization, not a prediction of a model. We can construct the factor decomposition as a regression equation. Define the β ij as regression coefficients, and then the εi are uncorrelated with the factors by construction, E(εi f˜j ) = 0. The content — the assumption that keeps (9.131) from describing any arbitrary set of returns — is an assumption that the εi are uncorrelated with each other. E(εi εj ) = 0. (More general versions of the model allow some limited correlation across the residuals but the basic story is the same.) The factor structure is thus a restriction on the covariance matrix of payoffs. For example, if there is only one factor, then ½ 2 σεi if i = j i j i j 2 ˜ ˜ cov(x , x ) = E[(β i f + ε )(β j f + ε )] = β i β j σ (f) + . 0 if i 6= j Thus, with N = number of securities, the N (N − 1)/2 elements of a variance-covariance matrix are described by N betas, and N + 1 variances. A vector version of the same thing is 2 0 σ1 0 2 cov(x, x0 ) = ββ 0 σ 2 (f ) + 0 σ 2 0 . .. . 0 0

With multiple (orthogonalized) factors, we obtain

cov(x, x0 ) = β 1 β 01 σ2 (f1 ) + β 2 β 02 σ2 (f2 ) + . . . + (diagonal matrix) In all these cases, we describe the covariance matrix a singular matrix ββ 0 (or a sum of a few such singular matrices) plus a diagonal matrix. If we know the factors we want to use ahead of time, say the market (value-weighted portfolio) and industry portfolios, or size and book to market portfolios, we can estimate a factor structure by running regressions. Often, however, we don’t know the identities of the factor portfolios ahead of time. In this case we have to use one of several statistical 164

S ECTION 9.4

A RBITRAGE P RICING T HEORY (APT)

techniques under the broad heading of factor analysis (that’s where the word “factor” came from in this context) to estimate the factor model. One can estimate a factor structure quickly by simply taking an eigenvalue decomposition of the covariance matrix, and then setting small eigenvalues to zero. 9.4.2

Exact factor pricing

With no error term, xi = E(xi )1 + β 0i f˜. implies p(xi ) = E(xi )p(1) + β 0i p(f˜) and thus m = a + b0 f ; p(xi ) = E(mxi ) E(Ri ) = Rf + β 0i λ. using only the law of one price. Suppose that there are no idiosyncratic terms εi . This is called an exact factor model. Now look again at the factor decomposition, xi = E(xi )1 + β 0i f˜.

(132)

It started as a statistical decomposition. But it also says that the payoff xi can be synthesized as a portfolio of the factors and a constant (risk-free payoff). Thus, the price of xi can only depend on the prices of the factors f, ˜ p(xi ) = E(xi )p(1) + β 0i p(f).

(133)

The law of one price assumption lets you take prices of right and left sides. If the factors are returns, their prices are 1. If the factors are not returns, their prices are free parameters which can be picked to make the model fit as well as possible. Since there are fewer factors than payoffs, this procedure is not vacuous. (Recall that the prices of the factors are related to the λ in expected return beta representations. λ is determined by the expected return of a return factor, and is a free parameter for non-return factor models.) We are really done, but the APT is usually stated as “there is a discount factor linear in f that prices returns Ri ,” or “there is an expected return-beta representation with f as 165

C HAPTER 9

FACTOR PRICING MODELS

factors.” Therefore, we should take a minute to show that the rather obvious relationship (9.133) between prices is equivalent to discount factor and expected return statements. Assuming only the law of one price, we know there is a discount factor m linear in factors that prices the factors. We usually call it x∗ , but call it f ∗ here to remind us that it £ ¤0 prices the factors. Denote fˆ = 1 f˜ the factors including the constant. As with x∗ , f ∗ = p(fˆ)0 E(fˆfˆ0 )−1 fˆ = a + b0 f satisfies p(fˆ) = E(f ∗ fˆ) and p(1) = E(f ∗ ). If the discount factor prices the factors, it must price any portfolio of the factors; hence f ∗ prices all payoffs xi that follow the factor structure (9.132). We could now go from m linear in the factors to an expected return-beta model using the above theorems that connect the two representations. But there is a more direct and elegant connection. Start with (9.133), specialized to returns xi = Ri and of course p(Ri ) = 1. Use p(1) = 1/Rf and solve for expected return as h i ˜ = Rf + β 0i λ. E(Ri ) = Rf + β 0i −Rf p(f)

The last equality defines λ. Expected returns are linear in the betas, and the constants (λ) are related to the prices of the factors. In fact, this is the same definition of λ that we arrived at above connecting m = b0 f to expected return-beta models. 9.4.3

Approximate APT using the law of one price

Attempts to extend the exact factor model to an approximate factor pricing model when errors are “small,” or markets are “large,” still only using law of one price. For fixed m, the APT gets better and better as R2 or the number of assets increases. However, for any fixed R2 or size of market, the APT can be arbitrarily bad. These observations mean that we must go beyond the law of one price to derive factor pricing models. Actual returns do not display an exact factor structure. There is some idiosyncratic or residual risk; we cannot exactly replicate the return of a given stock with a portfolio of a few large factor portfolios. However, the idiosyncratic risks are often small. For example, factor model regressions of the form (9.130) often have very high R2 , especially when portfolios rather than individual securities are on the left hand side. And the residual risks are still idiosyncratic: Even if they are a large part of an individual security’s variance, they should be a small contributor to the variance of well diversified portfolios. Thus, there is reason to hope that the APT holds approximately, especially for reasonably large portfolios. Surely, if the residuals are “small” and/or “idiosyncratic,” the price of an asset can’t be “too different” from the price predicted from its factor content? 166

S ECTION 9.4

A RBITRAGE P RICING T HEORY (APT)

To think about these issues, start again from a factor structure, but this time put in a residual, xi = E(xi )1 + β 0i f˜ + εi Again take prices of both sides, p(xi ) = E(xi )p(1) + β 0i p(f˜) + E(mεi ) Now, what can we say about the price of the residual p(εi ) = E(mεi )? Figure 23 illustrates the situation. Portfolios of the factors span a payoff space, the line from the origin through β 0i f in the Figure. The payoff we want to price, xi is not in that space, since the residual εi is not zero. A discount factor f ∗ that is in the f payoff space prices the factors. The set of all discount factors that price the factors is the line m perpendicular to f ∗ . The residual εi is orthogonal to the factor space, since it is a regression residual, and to f ∗ in particular, E(f ∗ εi ) = 0. This means that f ∗ assigns zero price to the residual. But the other discount factors on the m line are not orthogonal to εi , so generate non-zero price for the residual εi . As we sweep along the line of discount factors m that price the f, in fact, we generate every price from −∞ to ∞ for the residual. Thus, the law of one price does not nail down the price of the residual εi and hence the price or expected return of xi .

β’if

All m m>0

εi xi

f*

m: σ2(m) < A

m

Figure 23. Approximate arbitrage pricing. 167

C HAPTER 9

FACTOR PRICING MODELS

Limiting arguments We would like to show that the price of xi has to be “close to” the price ofβ 0i f . One notion of “close to” is that in some appropriate limit the price of xi converges to the price of β 0i f. “Limit” means, of course, that you can get arbitrarily good accuracy by going far enough in the direction of the limit (for every ε > 0 there is a δ....). Thus, establishing a limit result is a way to argue for an approximation. Here is one theorem that seems to imply that the APT should be a good approximation for portfolios that have high R2 on the factors. I state the argument for the case that there is a constant factor, so the constant is in the f space and E(εi ) = 0. The same ideas work in the less usual case that there is no constant factor, using second moments in place of variance. Theorem: Fix a discount factor m that prices the factors. Then, as var(εi ) → 0, p(xi ) → p(β 0i f ). This is easiest to see by just looking at the graph. E(εi ) = 0 so var(εi ) = E(εi2 ) = ||ε || . Thus, as the size of the εi vector in Figure 23 gets smaller, xi gets closer and closer to β 0i f. For any fixed m, the induced pricing function (lines perpendicular to the chosen m) is continuous. Thus, as xi gets closer and closer to β 0i f , its price gets closer and closer to β 0i f. i 2

The factor model is defined as a regression, so var(xi ) = var(β 0i f ) + var(εi ) Thus, the variance of the residual is related to the regression R2 . var(εi ) = 1 − R2 var(xi ) The theorem says that as R2 → 1, the price of the residual goes to zero.

We were hoping for some connection between the fact that the risks are idiosyncratic and factor pricing. Even if the idiosyncratic risks are a large part of the payoff at hand, they are a small part of a well-diversified portfolio. The next theorem shows that portfolios with high R2 don’t have to happen by chance; well-diversified portfolios will always have this characteristic. Theorem: As the number of primitive assets increases, the R2 of well-diversified portfolios increases to 1. Proof: Start with an equally weighted portfolio xp =

N 1 X i x. N i=1

Going back to the factor decomposition (9.130) for each individual asset xi , the 168

S ECTION 9.4

A RBITRAGE P RICING T HEORY (APT)

factor decomposition of xp is xp =

N N N N ¢ 0 1 X 1 X¡ 1 X 0 1 X i ai + β 0i f + εi = ai + β if + ε = ap + β p f + εp . N i=1 N i=1 N i=1 N i=1

The last equality defines notation αp , β p , εp . But p

var(ε ) = var

Ã

N 1 X i ε N i=1

!

So long as the variance of εi are bounded, and given the factor assumption E(εi εj ) = 0, lim var(εp ) = 0.

N→∞

Obviously, the same idea goes through so long as the portfolio spreads some weight on all the new assets, i.e. so long as it is “well-diversified.” ¥ These two theorems can be interpreted to say that the APT holds approximately (in the usual limiting sense) for portfolios that either naturally have high R2 , or well-diversified portfolios in large enough markets. We have only used the law of one price. Law of one price arguments fail Now, let me pour some cold water on these results. I fixed m and then let other things take limits. The flip side is that for any nonzero residual εi , no matter how small, we can pick a discount factor m that prices the factors and assigns any price to xi ! As often in mathematics, the order of “for all” and “there exists” matters a lot. Theorem: For any nonzero residual εi there is a discount factor that prices the factors f (consistent with the law of one price) and that assigns any desired price in (−∞, ∞) to the payoff xi . So long as ||εi || > 0, as we sweep the choice of m along the dashed line, the inner product of m with εi and hence xi varies from −∞ to ∞. Thus, for a given size R2 < 1, or a given finite market, the law of one price says absolutely nothing about the prices of payoffs that do not exactly follow the factor structure. The law of one price says that two ways of constructing the same portfolio must give the same price. If the residual is not exactly zero, there is no way of replicating the payoff xi from the factors and no way to infer anything about the price of xi from the price of the factors. I think the contrast between this theorem and those of the last subsection accounts for most of the huge theoretical controversy over the APT. If you fix m and take limits of N or 169

C HAPTER 9

FACTOR PRICING MODELS

ε, the APT gets arbitrarily good. But if you fix N or ε, as one does in any application, the APT can get arbitrarily bad as you search over possible m. The lesson I learn is that the effort to extend prices from an original set of securities (f in this case) to new payoffs that are not exactly spanned by the original set of securities, using only the law of one price, is fundamentally doomed. To extend a pricing function, you need to add some restrictions beyond the law of one price. 9.4.4

Beyond the law of one price: arbitrage and Sharpe ratios

We can find a well-behaved approximate APT if we impose the law of one price and a restriction on the volatility of discount factors, or, equivalently, a bound on the Sharpe ratio achievable by portfolios of the factors and test assets. The approximate APT based on the law of one price fell apart because we could always choose a discount factor sufficiently “far out” to generate an arbitrarily large price for an arbitrarily small residual. But those discount factors are surely “unreasonable.” Surely, we can rule them out, reestablishing an approximate APT, without jumping all the way to fully specified discount factor models such as the CAPM or consumption-based model A natural first idea is to impose the no-arbitrage restriction that m must be positive. Graphically, we are now restricted to the solid m line in Figure 23. Since that line only extends a finite amount, restricting us to strictly positive m0 s gives rise to finite upper and lower arbitrage bounds on the price of εi and hence xi . (The word arbitrage bounds comes from option pricing, and we will see these ideas again in that context. If this idea worked, it would restore the APT to “arbitrage pricing” rather than “law of one-pricing.”) Alas, in applications of the APT (as often in option pricing), the arbitrage bounds are too wide to be of much use. The positive discount factor restriction is equivalent to saying “if portfolio A gives a higher payoff than portfolio B in every state of nature, then the price of A must be higher than the price of B.” Since stock returns and factors are continuously distributed, not two-state distributions as I have graphed for figure 23, there typically are no strictly dominating portfolios, so adding m > 0 does not help. A second restriction does let us derive an approximate APT that is useful in finite markets with R2 < 1. We can restrict the variance and hence the size (||m|| = E(m2 ) = σ 2 (m) + E(m)2 = σ 2 (m) + 1/Rf2 ) of the discount factor. Figure 23 includes a plot of the discount factors with limited variance, size, or length in the geometry of that Figure. The restricted range of discount factors produces a restricted range of prices for xi . The restricted range of discount factors gives us upper and lower price bounds for the price of xi in terms of the

170

S ECTION 9.5

APT VS . ICAPM

factor prices. Precisely, the upper and lower bounds solve the problem min ( or max) p(xi ) = E(mxi ) s.t. E(mf ) = p(f), m ≥ 0, σ2 (m) ≤ A. {m}

{m}

Limiting the variance of the discount factor is of course the same as limiting the maximum Sharpe ratio (mean / standard deviation of excess return) available from portfolios of the factors and xi . Recall that σ(m) E (Re ) ≤ . σ(Re ) E(m) Though a bound on Sharpe ratios or discount factor volatility is not a totally preferencefree concept, it clearly imposes a great deal less structure than the CAPM or ICAPM which are essentially full general equilibrium models. Ross (1976) included this suggestion in his original APT paper, though it seems to have disappeared from the literature since then in the failed effort to derive an APT from the law of one price alone. Ross pointed out that deviations from factor pricing could provide very high Sharpe ratio opportunities, which seem implausible though not violations of the law of one price. Saá-Requejo and I (2000) dub this idea “good-deal” pricing, as an extension of “arbitrage pricing.” Limiting σ(m) rules out “good deals” as well as pure arbitrage opportunities. Having imposed a limit on discount factor volatility or Sharpe ratio A, then the APT limit does work, and does not depend on the order of “for all” and “there exists.” Theorem: As εi → 0 and R2 → 1, the price p(xi ) assigned by any discount factor m that satisfies E(mf) = p(f ), m ≥ 0, σ 2 (m) ≤ A approaches p(β 0i f ).

9.5

APT vs. ICAPM

A factor structure in the covariance of returns or high R2 in regressions of returns on factors can imply factor pricing (APT) but factors can price returns without describing their covariance matrix (ICAPM). Differing inspiration for factors. The disappearance of absolute pricing. The APT and ICAPM stories are often confused. Factor structure can imply factor pricing (APT), but factor pricing does not require a factor structure. In the ICAPM there is no presumption that factors f in a pricing model m = b0 f describe the covariance matrix of returns. The factors don’t have to be orthogonal or i.i.d. either. High R2 in time-series regressions of the returns on the factors may imply factor pricing (APT), but again are not 171

C HAPTER 9

FACTOR PRICING MODELS

necessary (ICAPM). The regressions of returns on factors can have low R2 in the ICAPM. Factors such as industry may describe large parts of returns’ variances but not contribute to the explanation of average returns. The biggest difference between APT and ICAPM for empirical work is in the inspiration for factors. The APT suggests that one start with a statistical analysis of the covariance matrix of returns and find portfolios that characterize common movement. The ICAPM suggests that one start by thinking about state variables that describe the conditional distribution of future asset returns and non-asset income. More generally, the idea of proxying for marginal utility growth suggests macroeconomic indicators, and indicators of shocks to non-asset income in particular. The difference between the derivations of factor pricing models, and in particular an approximate law-of-one-price basis vs. a proxy for marginal utility basis seems not to have had much impact on practice. In practice, we just test models m = b0 f and rarely worry about derivations. The best evidence for this view is the introductions of famous papers. Chen, Roll and Ross (1986) describe one of the earliest popular multifactor models, using industrial production and inflation as some of the main factors. They do not even present a factor decomposition of test asset returns, or the time-series regressions. A reader might well categorize the paper as a macroeconomic factor model or perhaps an ICAPM. Fama and French (1993) describe the currently most popular multifactor model, and their introduction describes it as an ICAPM in which the factors are state variables. But the factors are sorted on size and book/market just like the test assets, the time-series R2 are all above 90%, and much of the explanation involves “common movement” in test assets captured by the factors. A a reader might well categorize the model as much closer to an APT. In the first chapter, I made a distinction between relative pricing and absolute pricing. In the former, we price one security given the prices of others, while in the latter, we price each security by reference to fundamental sources of risk. The factor pricing stories are interesting in that they start with a nice absolute pricing model, the consumption-based model, and throw out enough information to end up with relative models. The CAPM prices Ri given the market, but throws out the consumption-based model’s description of where the market return came from.

9.6 1.

2. 3.

Problems

Suppose the investor only has a one-period horizon. He invests wealth W at date zero, and only consumes with expected utility Eu(c) = Eu(W ) in period one. Derive the quadratic utility CAPM in this case. (This is an even simpler derivation. The Lagrange multiplier on initial wealth W now becomes the denominator of m in place of u0 (c0 )). Express the log utility CAPM in continuous time to derive a discount factor linear in wealth. Figure 23 suggests that m > 0 is enough to establish a well-behaved approximate APT. 172

S ECTION 9.6

4.

P ROBLEMS

The text claims this is not true. Which is right? Can you use any excess return for the market factor in the CAPM, or must it be the market less the riskfree rate?

173

PART II Estimating and evaluating asset pricing models

174

S ECTION 9.6

P ROBLEMS

Our first task in bringing an asset pricing model to data is to estimate the free parameters; the β and γ in m = β(ct+1 /ct )−γ , or the b in m = b0 f. Then we want to evaluate the model. Is it a good model or not? Is another model better? Statistical analysis helps us to evaluate a model by providing a distribution theory for numbers such as parameter estimates that we create from the data. A distribution theory pursues the following idea: Suppose that we generate artificial data over and over again from a statistical model. For example, we could specify that the market return is an i.i.d. normal + εit . After random variable, and a set of stock returns is generated by Rtei = αi + β i Rem t 2 i picking values for the mean and variance of the market return and the αi , β i , σ (ε ), we could ask a computer to simulate many artificial data sets. We can repeat our statistical procedure in each of these artificial data sets, and graph the distribution of any statistic which we have estimated from the real data, i.e. the frequency that it takes on any particular value in our artificial data sets. In particular, we are interested in a distribution theory for the estimated parameters, to give us some sense of how much the data really has to say about their values; and for the pricing errors, which helps us to judge whether pricing errors are just bad luck of one particular historical accident or if they indicate a failure of the model. We also will want to generate distributions for statistics that compare one model to another, or provide other interesting evidence, to judge how much sample luck affects those calculations. All of the statistical methods I discuss in this part achieve these ends. They give methods for estimating free parameters; they provide a distribution theory for those parameters, and they provide distributions for statistics that we can use to evaluate models, most often a ˆ. quadratic form of pricing errors in the form α ˆ 0 V −1 α I start by focusing on the GMM approach. The GMM approach is a natural fit for a discount factor formulation of asset pricing theories, since we just use sample moments in the place of population moments. As you will see, there is no singular “GMM estimate and test.” GMM is a large canvas and a big set of paints and brushes; a flexible tool for doing all kinds of sensible (and, unless you’re careful, not-so-sensible) things to the data. Then I consider traditional regression tests (naturally paired with expected return-beta statements of factor models) and their maximum likelihood formalization. I emphasize the fundamental similarities between these three methods, as I emphasized the similarity between p = E(mx),

175

C HAPTER 9

FACTOR PRICING MODELS

expected return-beta models, and mean-variance frontiers. A concluding chapter highlights some of the differences between the methods, as I contrasted p = E(mx) and beta or meanvariance representations of the models.

176

Chapter 10. GMM in explicit discount factor models The basic idea in the GMM approach is very straightforward. The asset pricing model predicts E(pt ) = E [m(datat+1 , parameters) xt+1 ] .

(134)

The most natural way to check this prediction is to examine sample averages, i.e. to calculate T T 1X 1X pt and [m(datat+1 , parameters) xt+1 ] . T t=1 T t=1

(135)

GMM estimates the parameters by making the sample averages as close to each other as possible. It seems natural, before evaluating a model, to pick parameters that give it its best chance. GMM then works out a distribution theory for the estimates. This distribution theory is a generalization of the simplest exercise in statistics: the distribution of the sample mean. Then, it suggests that we evaluate the model by looking at how close the sample averages of price and discounted payoff are to each other, or equivalently by looking at the pricing errors. It gives a statistical test of the hypothesis that the underlying population means are in fact zero.

10.1

The Recipe

Definitions ut+1 (b) ≡ mt+1 (b)xt+1 − pt gT (b) ≡ ET [ut (b)] ∞ X S ≡ E [ut (b) ut−j (b)0 ] j=−∞

GMM estimate ˆb2 = argminb gT (b)0 Sˆ−1 gT (b). Standard errors 1 ∂gT (b) var(ˆb2 ) = (d0 S −1 d)−1 ; d ≡ T ∂b 177

C HAPTER 10

GMM IN EXPLICIT DISCOUNT FACTOR MODELS

Test of the model (“overidentifying restrictions”) £ ¤ T JT = T min gT (b)0 S −1 gT (b) ∼ χ2 (#moments − #parameters). It’s easiest to start our discussion of GMM in the context of an explicit discount factor model, such as the consumption-based model. I treat the special structure of linear factor models later. I start with the basic classic recipe as given by Hansen and Singleton (1982). Discount factor models involve some unknown parameters as well as data, so I write mt+1 (b) when it’s important to remind ourselves of this dependence. For example, if mt+1 = β(ct+1 /ct )−γ , then b ≡ [β γ]0 . I write ˆb to denote an estimate when it is important to distinguish estimated from other values. Any asset pricing model implies E(pt ) = E [mt+1 (b)xt+1 ] .

(136)

It’s easiest to write this equation in the form E(·) = 0 E [mt+1 (b)xt+1 − pt ] = 0.

(137)

x and p are typically vectors; we typically check whether a model for m can price a number of assets simultaneously. Equations (10.137) are often called the moment conditions. It’s convenient to define the errors ut (b) as the object whose mean should be zero, ut+1 (b) = mt+1 (b)xt+1 − pt Given values for the parameters b, we could construct a time series on ut and look at its mean. Define gT (b) as the sample mean of the ut errors, when the parameter vector is b in a sample of size T : T 1X ut (b) = ET [ut (b)] = ET [mt+1 (b)xt+1 − pt ] . gT (b) ≡ T t=1

The second equality introduces the handy notation ET for sample means, T 1X (·). ET (·) = T t=1

ˆ and gˆ. However, Hansen’s T subscript (It might make more sense to denote these estimates E notation is so widespread that doing so would cause more confusion than it solves.) 178

S ECTION 10.1

T HE R ECIPE

The first stage estimate of b minimizes a quadratic form of the sample mean of the errors, ˆb1 = argmin ˆ gT (ˆb)0 W gT (ˆb) {b} for some arbitrary matrix W (often, W = I). This estimate is consistent and asymptotically normal. You can and often should stop here, as I explain below. Using ˆb1 , form an estimate Sˆ of ∞ X

S≡

E [ut (b) ut−j (b)0 ] .

(138)

j=−∞

(Below I discuss various interpretations of and ways to construct this estimate.) Form a second stage estimate ˆb2 using the matrix Sˆ in the quadratic form, ˆb2 = argmin gT (b)0 Sˆ−1 gT (b). b ˆb2 is a consistent, asymptotically normal, and asymptotically efficient estimate of the parameter vector b. “Efficient” means that it has the smallest variance-covariance matrix among all estimators that set different linear combinations of gT (b) to zero. The variance-covariance matrix of ˆb2 is 1 var(ˆb2 ) = (d0 S −1 d)−1 T where d≡

∂gT (b) ∂b

or, more explicitly, d = ET

µ

¶¯ ¯ ∂ [(mt+1 (b)xt+1 − pt )] ¯¯ ∂b b=ˆ b

(More precisely, d should be written as the object to which ∂gT /∂b converges, and then ∂gT /∂b is an estimate of that object used to form a consistent estimate of the asymptotic variance-covariance matrix.) This variance-covariance matrix can be used to test whether a parameter or group of parameters are equal to zero, via ˆb q i ∼ N (0, 1) var(ˆb)ii 179

C HAPTER 10

GMM IN EXPLICIT DISCOUNT FACTOR MODELS

and h i−1 ˆbj ∼ χ2 (#included b0 s) ˆbj var(ˆb)jj

where bj =subvector, var(b)jj =submatrix.

Finally, the test of overidentifying restrictions is a test of the overall fit of the model. It states that T times the minimized value of the second-stage objective is distributed χ2 with degrees of freedom equal to the number of moments less the number of estimated parameters. £ ¤ T JT = T min gT (b)0 S −1 gT (b) ∼ χ2 (#moments − #parameters). {b}

10.2

Interpreting the GMM procedure

gT (b) is a pricing error. It is proportional to α. GMM picks parameters to minimize a weighted sum of squared pricing errors. The second-stage picks the linear combination of pricing errors that are best measured, by having smallest sampling variation. First and second stage are like OLS and GLS regressions. The standard error formula is a simple application of the delta method. The JT test evaluates the model by looking at the sum of squared pricing errors.

Pricing errors The moment conditions are gT (b) = ET [mt+1 (b)xt+1 ] − ET [pt ] . Thus, each moment is the difference between actual (ET (p)) and predicted (ET (mx)) price, or pricing error. What could be more natural than to pick parameters so that the model’s predicted prices are as close as possible to the actual prices, and then to evaluate the model by how large these pricing errors are? In the language of expected returns, the moments gT (b) are proportional to the difference between actual and predicted returns; Jensen’s alphas, or the vertical distance between the points and the line in Figure 5. To see this fact, recall that 0 = E(mRe ) can be translated to a predicted expected return, E(Re ) = −

cov(m, Re ) . E(m)

180

S ECTION 10.2

I NTERPRETING THE GMM PROCEDURE

Therefore, we can write the pricing error as µ µ ¶¶ cov(m, Re ) e g(b) = E(mR ) = E(m) E(R ) − − E(m) 1 (actual mean return - predicted mean return.) g(b) = Rf e

If we express the model in expected return-beta language, E(Rei ) = αi + β 0i λ then the GMM objective is proportional to the Jensen’s alpha measure of mis-pricing, g(b) =

1 αi . Rf

First-stage estimates If we could, we’d pick b to make every element of gT (b) = 0 — to have the model price assets perfectly in sample. However, there are usually more moment conditions (returns times instruments) than there are parameters. There should be, because theories with as many free parameters as facts (moments) are vacuous. Thus, we choose b to make gT (b) as small as possible, by minimizing a quadratic form, min gT (b)0 W gT (b). {b}

(139)

W is a weighting matrix that tells us how much attention to pay to each moment, or how to trade off doing well in pricing one asset or linear combination of assets vs. doing well in pricing another. In the common case W = I, GMM treats all assets symmetrically, and the objective is to minimize the sum of squared pricing errors. The sample pricing error gT (b) may be a nonlinear function of b. Thus, you may have to use a numerical search to find the value of b that minimizes the objective in (10.139). However, since the objective is locally quadratic, the search is usually straightforward. Second-stage estimates: Why S −1 ? What weighting matrix should you use? The weighting matrix directs GMM to emphasize some moments or linear combinations of moments at the expense of others. You might start with W = I, i.e., try to price all assets equally well. A W that is not the identity matrix can be used to offset differences in units between the moments. You also might also start with different elements on the diagonal of W if you think some assets are more interesting, more informative, or better measured than others. The second-stage estimate picks a weighting matrix based on statistical considerations. 181

C HAPTER 10

GMM IN EXPLICIT DISCOUNT FACTOR MODELS

Some asset returns may have much more variance than other assets. For those assets, the sample mean gT = ET (mt Rt − 1) will be a much less accurate measurement of the population mean E(mR − 1), since the sample mean will vary more from sample to sample. Hence, it seems like a good idea to pay less attention to pricing errors from assets with high variance of mt Rt − 1. One could implement this idea by using a W matrix composed of inverse variances of ET (mt Rt − 1) on the diagonal. More generally, since asset returns are correlated, one might think of using the covariance matrix of ET (mt Rt −1). This weighting matrix pays most attention to linear combinations of moments about which the data set at hand has the most information. This idea is exactly the same as heteroskedasticity and cross-correlation corrections that lead you from OLS to GLS in linear regressions. The covariance matrix of gT = ET (ut+1 ) is the variance of a sample mean. Exploiting the assumption that E(ut ) = 0, and that ut is stationary so E(u1 u2 ) = E(ut ut+1 ) depends only on the time interval between the two us, we have

var(gT ) = var =

Ã

T 1 X ut+1 T t=1

!

¡ ¢ ¤ 1 £ T E(ut u0t ) + (T − 1) E(ut u0t−1 ) + E(ut u0t+1 )) + ... 2 T

As T → ∞, (T − j)/T → 1, so

var(gT ) →

∞ 1 X 1 E(ut u0t−j ) = S. T j=−∞ T

The last equality denotes S, known for other reasons as the spectral density matrix at frequency zero of ut . (Precisely, S so defined is the variance-covariance matrix of the gT for fixed b. The actual variance-covariance matrix of gT must take into account the fact that we chose b to set a linear combination of the gT to zero in each sample. I give that formula below. The point here is heuristic.) This fact suggests that a good weighting matrix might be the inverse of S. In fact, Hansen (1982) shows formally that the choice W = S −1 , S ≡

∞ X

E(ut u0t−j )

j=−∞

is the statistically optimal weighing matrix, meaning that it produces estimates with lowest asymptotic variance. √ You may be more used to the formula σ(u)/ T for the standard deviation of a sample mean. This formula is a special case that holds when the u0t s are uncorrelated over time. If 182

S ECTION 10.2

I NTERPRETING THE GMM PROCEDURE

Et (ut u0t−j ) = 0, j 6= 0, then the previous equation reduces to var

Ã

T 1X ut+1 T t=1

!

=

var(u) 1 E(uu0 ) = . T T

This is probably the first statistical formula you ever saw – the variance of the sample mean. In GMM, it is the last statistical formula you’ll ever see as well. GMM amounts to just generalizing the simple ideas behind the distribution of the sample mean to parameter estimation and general statistical contexts. The first and second stage estimates should remind you of standard linear regression models. You start with an OLS regression. If the errors are not i.i.d., the OLS estimates are consistent, but not efficient. If you want efficient estimates, you can use the OLS estimates to obtain a series of residuals, estimate a variance-covariance matrix of residuals, and then do GLS. GLS is also consistent and more efficient, meaning that the sampling variation in the estimated parameters is lower. Standard errors The formula for the standard error of the estimate, 1 var(ˆb2 ) = (d0 S −1 d)−1 T

(140)

can be understood most simply as an instance of the “delta method” that the asymptotic variance of f (x) is f 0 (x)2 var(x). Suppose there is only one parameter and one moment. S/T is the variance matrix of the moment gT . d−1 is [∂gT /∂b]−1 = ∂b/∂gT . Then the delta method formula gives 1 ∂b ∂b var(ˆb2 ) = var(gT ) . T ∂gT ∂gT The actual formula (10.140) just generalizes this idea to vectors. 10.2.1

JT Test

Once you’ve estimated the parameters that make a model “fit best,” the natural question is, how well does it fit? It’s natural to look at the pricing errors and see if they are “big.” The JT test asks whether they are “big” by statistical standards – if the model is true, how often should we see a (weighted) sum of squared pricing errors this big? If not often, the model is “rejected.” The test is h i T JT = T gT (ˆb)0 S −1 gT (ˆb) ∼ χ2 (#moments − #parameters).

Since S is the variance-covariance matrix of gT , this statistic is the minimized pricing errors 183

C HAPTER 10

GMM IN EXPLICIT DISCOUNT FACTOR MODELS

divided by their variance-covariance matrix. Sample means converge to a normal distribution, so sample means squared divided by variance converges to the square of a normal, or χ2 . The reduction in degrees of freedom corrects for the fact that S is really the covariance matrix of gT for fixed b. We set a linear combination of the gT to zero in each sample, so the actual covariance matrix of gT is singular, with rank #moments - #parameters. More details below.

10.3

Applying GMM

Notation. Forecast errors and instruments. Stationarity and choice of units.

Notation; instruments and returns Most of the effort involved with GMM is simply mapping a given problem into the very general notation. The equation E [mt+1 (b)xt+1 − pt ] = 0 can capture a lot. We often test asset pricing models using returns, in which case the moment conditions are E [mt+1 (b)Rt+1 − 1] = 0. It is common to add instruments as well. Mechanically, you can multiply both sides of 1 = Et [mt+1 (b)Rt+1 ] by any variable zt observed at time t before taking unconditional expectations, resulting in E(zt ) = E [mt+1 (b)Rt+1 zt ] . Expressing the result in E(·) = 0 form, 0 = E {[mt+1 (b)Rt+1 − 1] zt } .

(141)

We can do this for a whole vector of returns and instruments, multiplying each return by each instrument. For example, if we start with two returns R = [Ra Rb ]0 and one instrument z, 184

S ECTION 10.3

A PPLYING GMM

equation (10.141) looks like mt+1 (b) Rat+1 mt+1 (b) Rbt+1 E mt+1 (b) Rat+1 zt mt+1 (b) Rbt+1 zt

1 1 − = zt zt

0 0 . 0 0

Using the Kronecker product ⊗ meaning “multiply every element by every other element” we can denote the same relation compactly by E {[mt+1 (b) Rt+1 − 1] ⊗ zt } = 0,

(142)

or, emphasizing the managed-portfolio interpretation and p = E(mx) notation, E [mt+1 (b)(Rt+1 ⊗ zt ) − (1 ⊗ zt )] = 0. Forecast errors and instruments The asset pricing model says that, although expected returns can vary across time and assets, expected discounted returns should always be the same, 1. The error ut+1 = mt+1 Rt+1 − 1 is the ex-post discounted return. ut+1 = mt+1 Rt+1 − 1 represents a forecast error. Like any forecast error, ut+1 should be conditionally and unconditionally mean zero. In an econometric context, z is an instrument because it is uncorrelated with the error ut+1 . E(zt ut+1 ) is the numerator of a regression coefficient of ut+1 on zt ; thus adding instruments basically checks that the ex-post discounted return is unforecastable by linear regressions. If an asset’s return is higher than predicted when zt is unusually high, but not on average, scaling by zt will pick up this feature of the data. Then, the moment condition checks that the discount rate is unusually low at such times, or that the conditional covariance of the discount rate and asset return moves sufficiently to justify the high conditionally expected return. As I explained in Section 8.1, the addition of instruments is equivalent to adding the returns of managed portfolios to the analysis, and is in principle able to capture all of the model’s predictions. Stationarity and distributions The GMM distribution theory does require some statistical assumption. Hansen (1982) and Ogaki (1993) cover them in depth. The most important assumption is that m, p, and x must be stationary random variables. (“Stationary” of often misused to mean constant, or i.i.d.. The statistical definition of stationarity is that the joint distribution of xt , xt−j depends only on j and not on t.) Sample averages must converge to population means as the sample size grows, and stationarity implies this result. Assuring stationarity usually amounts to a choice of sensible units. For example, though 185

C HAPTER 10

GMM IN EXPLICIT DISCOUNT FACTOR MODELS

we could express the pricing of a stock as pt = Et [mt+1 (dt+1 + pt+1 )] it would not be wise to do so. For stocks, p and d rise over time and so are typically not stationary; their unconditional means are not defined. It is better to divide by pt and express the model as · ¸ dt+1 + pt+1 1 = Et mt+1 = Et (mt+1 Rt+1 ) pt The stock return is plausibly stationary. Dividing by dividends is an alternative and I think underutilized way to achieve stationarity (at least for portfolios, since many individual stocks do not pay regular dividends): · µ ¶ ¸ pt+1 dt+1 pt = Et mt+1 1 + . dt dt+1 dt ´ ³ dt+1 pt Now we map 1 + pdt+1 dt into xt+1 and dt into pt . This formulation allows us to focus t+1 on prices rather than one-period returns. Bonds are a claim to a dollar, so bond prices and yields do not grow over time. Hence, it might be all right to examine pbt = E(mt+1 1) with no transformations. Stationarity is not always a clear-cut question in practice. As variables become “less stationary,” as they experience longer swings in a sample, the asymptotic distribution can becomes a less reliable guide to a finite-sample distribution. For example, the level of nominal interest rates is surely a stationary variable in a fundamental sense: it was 6% in ancient Babylon, about 6% in 14th century Italy, and about 6% again today. Yet it takes very long swings away from this unconditional mean, moving slowly up or down for even 20 years at a time. Therefore, in an estimate and test that uses the level of interest rates, the asymptotic distribution theory might be a bad approximation to the correct finite sample distribution theory. This is true even if the number of data points is large. 10,000 data points measured every minute are a “smaller” data set than 100 data points measured every year. In such a case, it is particularly important to develop a finite-sample distribution by simulation or bootstrap, which is easy to do given today’s computing power. It is also important to choose test assets in a way that is stationary. For example, individual stocks change character over time, increasing or decreasing size, exposure to risk factors, leverage, and even nature of the business. For this reason, it is common to sort stocks into portfolios based on characteristics such as betas, size, book/market ratios, industry and so forth. The statistical characteristics of the portfolio returns may be much more constant than 186

S ECTION 10.3

A PPLYING GMM

the characteristics of individual securities, which float in and out of the various portfolios. (One can alternatively include the characteristics as instruments.) Many econometric techniques require assumptions about distributions. As you can see, the variance formulas used in GMM do not include the usual assumptions that variables are i.i.d., normally distributed, homoskedastic, etc. You can put such assumptions in if you want to – we’ll see how below, and adding such assumptions simplifies the formulas and can improve the small-sample performance when the assumptions are justified – but you don’t have to add these assumptions.

187

Chapter 11. GMM: general formulas and applications Lots of calculations beyond formal parameter estimation and overall model testing are useful in the process of evaluating a model and comparing it to other models. But you want to understand sampling variation in such calculations, and mapping the questions into the GMM framework allows you to do this easily. In addition, alternative estimation and evaluation procedures may be more intuitive or robust to model misspecification than the two (or multi) stage procedure described above. In this chapter I lay out the general GMM framework, and I discuss four applications and variations on the basic GMM method. 1) I show how to derive standard errors of nonlinear functions of sample moments, such as correlation coefficients. 2) I apply GMM to OLS regressions, easily deriving standard error formulas that correct for autocorrelation and conditional heteroskedasticity. 3) I show how to use prespecified weighting matrices W in asset pricing tests in order to overcome the tendency of efficient GMM to focus on spuriously lowvariance portfolios 4) As a good parable for prespecified linear combination of moments a, I show how to mimic “calibration” and “evaluation” phases of real business cycle models. 5) I show how to use the distribution theory for the gT beyond just forming the JT test in order to evaluate the importance of individual pricing errors. The next chapter continues, and collects GMM variations useful for evaluating linear factor models and related mean-variance frontier questions. Many of these calculations amount to creative choices of the aT matrix that selects which linear combination of moments are set to zero, and reading off the resulting formulas for variance covariance matrix of the estimated coefficients, equation (11.146) and variance covariance matrix of the moments gT , equation (11.147).

11.1

General GMM formulas

The general GMM estimate aT gT (ˆb) = 0 Distribution of ˆb : T cov(ˆb) = (ad)−1 aSa0 (ad)−10 Distribution of gT (ˆb) : h i ¡ ¢ ¡ ¢0 T cov gT (ˆb) = I − d(ad)−1 a S I − d(ad)−1 a 188

S ECTION 11.1

G ENERAL GMM FORMULAS

The “optimal” estimate uses a = d0 S −1 . In this case, T cov(ˆb) = (d0 S −1 d)−1 h i T cov gT (ˆb) = S − d(d0 S −1 d)−1 d0

and

T JT = T gT (ˆb)0 S −1 gT (ˆb) → χ2 (#moments − #parameters). An analogue to the likelihood ratio test, T JT (restricted) − T JT (unrestricted) ∼ χ2Number of restrictions GMM procedures can be used to implement a host of estimation and testing exercises. Just about anything you might want to estimate can be written as a special case of GMM. To do so, you just have to remember (or look up) a few very general formulas, and then map them into your case. Express a model as E[f (xt , b)] = 0 Everything is a vector: f can represent a vector of L sample moments, xt can be M data series, b can be N parameters. f(xt , b) is a slightly more explicit statement of the errors ut (b) in the last chapter Definition of the GMM estimate. We estimate parameters ˆb to set some linear combination of sample means of f to zero, ˆb : set aT gT (ˆb) = 0

(143)

where gT (b) ≡

T 1X f(xt , b) T t=1

and aT is a matrix that defines which linear combination of gT (b) will be set to zero. This defines the GMM estimate. If there are as many moments as parameters, you will set each moment to zero; when there are fewer parameters than moments, (11.143) just captures the natural idea that you will set some moments, or some linear combination of moments to zero in order to estimate the parameters. The minimization of the last chapter is a special case. If you estimate b by 189

C HAPTER 11

GMM: GENERAL FORMULAS AND APPLICATIONS

min gT0 (b)0 W gT (b), the first order conditions are ∂gT0 W gT (b) = 0, ∂b which is of the form (11.143) with aT = ∂gT0 /∂bW . The general GMM procedure allows you to pick arbitrary linear combinations of the moments to set to zero in parameter estimation. Standard errors of the estimate. Hansen (1982), Theorem 3.1 tells us that the asymptotic distribution of the GMM estimate is √ ¤ £ T (ˆb − b) → N 0, (ad)−1 aSa0 (ad)−10

where

d≡E

·

(144)

¸ ∂f ∂gT (b) (x , b) = t ∂b0 ∂b0

(i.e., d is defined as the population moment in the first equality, which we estimate in sample by the second equality), where a ≡ plim aT , and where S≡ Don’t forget the

√

∞ X

E [f(xt , b), f (xt−j b)0 ] .

(145)

j=−∞

T in (11.144)! In practical terms, this means to use 1 var(ˆb) = (ad)−1 aSa0 (ad)−10 T

(146)

as the covariance matrix for standard errors and tests. As in the last chapter, you can understand this formula as an application of the delta method. Distribution of the moments. Hansen’s Lemma 4.1 gives the sampling distribution of the moments gT (b) : h ¡ √ ¢ ¡ ¢0 i T gT (ˆb) → N 0, I − d(ad)−1 a S I − d(ad)−1 a .

(147)

As we have seen, S would be the asymptotic variance-covariance matrix of sample means, if we did not estimate any parameters, which sets some linear combinations of the gT to zero. The I − d(ad)−1 a terms account for the fact that in each sample some linear combinations of gT are set to zero. Thus, this variance-covariance matrix is singular. 190

S ECTION 11.1

G ENERAL GMM FORMULAS

χ2 tests. A sum of squared standard normals is distributed χ2 . Therefore, it is natural to use the distribution theory for gT to see if the gT are jointly “too big.” Equation (11.147) suggests that we form the statistic h¡ ¢ ¡ ¢0 i−1 T gT (ˆb)0 I − d(ad)−1 a S I − d(ad)−1 a gT (ˆb) (148)

and that it should have a χ2 distribution. It does, but with a hitch: The variance-covariance matrix is singular, soP you have to pseudo-invert it. For example, you can perform an eigenvalue decomposition = QΛQ0 and then invert only the non-zero eigenvalues. Also, the χ2 distribution has degrees of freedom given by the number non-zero linear combinations of gT , the number of moments less number of estimated parameters. You can similarly use (11.147) to construct tests of individual moments (“are the small stocks mispriced?”) or groups of moments. Efficient estimates

The theory so far allows us to estimate parameters by setting any linear combination of moments to zero. Hansen shows that one particular choice is statistically optimal, a = d0 S −1 .

(149)

This choice is the first order condition to min{b} gT (b)0 S −1 gT (b) that we studied in the last Chapter. With this weighting matrix, the standard error formula (11.146) reduces to √ ¤ £ (150) T (ˆb − b) → N 0, (d0 S −1 d)−1 .

This is Hansen’s Theorem 3.2. The sense in which (11.149) is “efficient” is that the sampling variation of the parameters for arbitrary a matrix, (11.146), equals the sampling variation of the “efficient” estimate in (11.150) plus a positive semidefinite matrix. With the optimal weights (11.149), the variance of the moments (11.147) simplifies to cov(gT ) =

¢ 1¡ S − d(d0 S −1 d)−1 d0 . T

(151)

We can use this matrix in a test of the form (11.148). However, Hansen’s Lemma 4.2 tells us that there is an equivalent and simpler way to construct this test, T gT (ˆb)0 S −1 gT (ˆb) → χ2 (#moments − #parameters).

(152)

This result is nice since we get to use the already-calculated and non-singular S −1 . To derive (11.152) from (11.147), factor S = CC 0 and then find the asymptotic covariance matrix of C −1 gT (ˆb) using (11.147). The result is i h√ var T C −1 gT (ˆb) = I − C −1 d(d0 S −1 d)−1 d0 C −10 . 191

C HAPTER 11

GMM: GENERAL FORMULAS AND APPLICATIONS

This is an idempotent matrix of rank #moments-#parameters, so (11.152) follows. Alternatively, note that S −1 is a pseudo-inverse of the second stage cov(gT ). (A pseudoinverse times cov(gT ) should result in an idempotent matrix of the same rank as cov(gT ).) ¡ ¢ S −1 cov(gT ) = S −1 S − d(d0 S −1 d)−1 d0 = I − S −1 d(d0 S −1 d)−1 d0

Then, check that the result is idempotent. ¡ ¢¡ ¢ I − S −1 d(d0 S −1 d)−1 d0 I − S −1 d(d0 S −1 d)−1 d0 = I − S −1 d(d0 S −1 d)−1 d0 .

This derivation not only verifies that JT has the same distribution as gT0 cov(gT )−1 gT , but that they are numerically the same in every sample. I emphasize that (11.150) and (11.152) only apply to the “optimal” choice of weights, (11.149). If you use another set of weights, as in a first-stage estimate, you must use the general formulas (11.146) and (11.147). Model comparisons You often want to compare one model to another. If one model can be expressed as a special or “restricted” case of the other or “unrestricted” model we can perform a statistical comparison that looks very much like a likelihood ratio test. If we use the same S matrix – usually that of the unrestricted model – the restricted JT must rise. But if the restricted model is really true, it shouldn’t rise “much.” How much? T JT (restricted) − T JT (unrestricted) ∼ χ2 (#of restrictions) This is a “χ2 difference” test, due to Newey and West (1987a), who call it the “D-test.”

11.2

Testing moments

How to test one or a group of pricing errors. 1) Use the formula for var(gT ) 2) A χ2 difference test. You may want to see how well a model does on particular moments or particular pricing errors. For example, the celebrated “small firm effect” states that an unconditional CAPM (m = a+ bRW , no scaled factors) does badly in pricing the returns on a portfolio that always holds the smallest 1/10th or 1/20th of firms in the NYSE. You might want to see whether a new model prices the small returns well. The standard error of pricing errors also allows you to add error bars to a plot of predicted vs. actual mean returns such as Figure 5 or other diagnostics based on pricing errors. We have already seen that individual elements of gT measure the pricing errors or expected return errors. Thus, the sampling variation of gT given by (11.147) provides exactly 192

S ECTION 11.3

S TANDARD ERRORS OF ANYTHING BY DELTA METHOD

the standard error we are looking for. You can use the sampling distribution of gT , to evaluate the significance of individual pricing errors, to construct a t-test (for a single gT , such as small firms) or χ2 test (for groups of gT , such as small firms ⊗ instruments). As usual this is the Wald test. Alternatively, you can use the χ2 difference approach. Start with a general model that includes all the moments, and form an estimate of the spectral density matrix S. Now set to zero the moments you want to test, and denote gsT (b) the vector of moments, including the zeros (s for “smaller”). Choose bs to minimize gsT (bs )0 S −1 gsT (bs ) using the same weighting matrix S. The criterion will be lower than the original criterion gT (b)0 S −1 gT (b), since there are the same number of parameters and fewer moments. But, if the moments we want to test truly are zero, the criterion shouldn’t be that much lower. The χ2 difference test applies, T gT (ˆb)0 S −1 gT (ˆb) − T gsT (ˆbs )S −1 gsT (ˆbs ) ∼ χ2 (#eliminated moments). Of course, don’t fall into the obvious trap of picking the largest of 10 pricing errors and noting it’s more than two standard deviations from zero. The distribution of the largest of 10 pricing errors is much wider than the distribution of a single one. To use this distribution, you have to pick which pricing error you’re going to test before you look at the data.

11.3

Standard errors of anything by delta method

One quick application illustrates the usefulness of the GMM formulas. Often, we want to estimate a quantity that is a nonlinear function of sample means, b = φ [E(xt )] = φ(µ). In this case, the formula (11.144) reduces to · ¸ · ¸0 X ∞ 1 dφ dφ cov(xt , x0t−j ) var(bT ) = . T dµ j=−∞ dµ

(153)

The formula is very intuitive. The variance of the sample mean is the covariance term inside. The derivatives just linearize the function φ near the true b. For example, a correlation coefficient can be written as a function of sample means as

Thus, take

E(xt yt ) − E(xt )E(yt ) p corr(xt , yt ) = p E(x2t ) − E(xt )2 E(yt2 ) − E(yt )2 µ=

£

E(xt ) E(x2t ) E(yt ) E(yt2 ) E(xt yt )

¤0

.

A problem at the end of the chapter asks you to take derivatives and derive the standard error of the correlation coefficient. One can derive standard errors for impulse-response functions, 193

C HAPTER 11

GMM: GENERAL FORMULAS AND APPLICATIONS

variance decompositions, and many other statistics in this way.

11.4

Using GMM for regressions

By mapping OLS regressions in to the GMM framework, we derive formulas for OLS standard errors that correct for autocorrelation and conditional heteroskedasticity of the errors. The general formula is ∞ X 1 ˆ = E(xt x0 )−1 E(ut xt x0t−j ut−j ) E(xt x0t )−1 . var(β) t T j=−∞ and it simplifies in special cases.

Mapping any statistical procedure into GMM makes it easy to develop an asymptotic distribution that corrects for statistical problems such as non-normality, serial correlation and conditional heteroskedasticity. To illustrate, as well as to develop the very useful formulas, I map OLS regressions into GMM. Correcting OLS standard errors for econometric problems is not the same thing as GLS. When errors do not obey the OLS assumptions, OLS is consistent, and often more robust than GLS, but its standard errors need to be corrected. OLS picks parameters β to minimize the variance of the residual: £ ¤ min ET (yt − β 0 xt )2 . {β}

ˆ from the first order condition, which states that the residual is orthogonal to the We find β right hand variable: h i ˆ =0 ˆ = ET xt (yt − x0 β) gT (β) (154) t This condition is exactly identified–the number of moments equals the number of parameters. Thus, we set the sample moments exactly to zero and there is no weighting matrix (a = I). We can solve for the estimate analytically, ˆ = [ET (xt x0t )]−1 ET (xt yt ). β This is the familiar OLS formula. The rest of the ingredients to equation (11.144) are d = E(xt x0t ) 194

S ECTION 11.4

U SING GMM FOR REGRESSIONS

f (xt , β) = xt (yt − x0t β) = xt et where et is the regression residual. Equation (11.144) gives a formula for OLS standard errors, ∞ X 1 ˆ = E(xt x0 )−1 var(β) E(ut xt x0t−j ut−j ) E(xt x0t )−1 . (155) t T j=−∞ This formula reduces to some interesting special cases. Serially uncorrelated, homoskedastic errors These are the usual OLS assumptions, and it’s good the usual formulas emerge. Formally, the OLS assumptions are E(et | xt , xt−1 ...et−1 , et−2 ...) = 0

(156)

E(e2t | xt , xt−1 ...et , et−1 ...) = constant = σ2e .

(157)

To use these assumptions, I use the fact that E(ab) = E(E(a|b)b). The first assumption means that only the j = 0 term enters the sum ∞ X

E(et xt x0t−j et−j ) = E(e2t xt x0t ).

j=−∞

The second assumption means that E(e2t xt x0t ) = E(e2t )E(xt x0t ) = σ2e E(xt x0t ). Hence equation (11.155) reduces to our old friend, 1 2 −1 σe E(xt x0t )−1 = σ 2e (X 0 X) . T £ The last notation is typical of econometrics texts, in which X = x1 resents the data matrix. ˆ = var(β)

x2

... xT

¤0

rep-

Heteroskedastic errors If we delete the conditional homoskedasticity assumption (11.157), we can’t pull the u out 195

C HAPTER 11

GMM: GENERAL FORMULAS AND APPLICATIONS

of the expectation, so the standard errors are ˆ = var(β)

1 E(xt x0t )−1 E(u2t xt x0t )E(xt x0t )−1 . T

These are known as “Heteroskedasticity consistent standard errors” or “White standard errors” after White (1980). Hansen-Hodrick errors Hansen and Hodrick (1982) run forecasting regressions of (say) six month returns, using monthly data. We can write this situation in regression notation as yt+k = β 0 xt + εt+k t = 1, 2, ...T. Fama and French (1988) also use regressions of overlapping long horizon returns on variables such as dividend/price ratio and term premium. Such regressions are an important part of the evidence for predictability in asset returns. Under the null that one-period returns are unforecastable, we will still see correlation in the εt due to overlapping data. Unforecastable returns imply E(εt εt−j ) = 0 for |j| ≥ k but not for |j| < k. Therefore, we can only rule out terms in S lower than k. Since we might as well correct for potential heteroskedasticity while we’re at it, the standard errors are k X 1 E(ut xt x0t−j ut−j ) E(xt x0t )−1 . var(bT ) = E(xt x0t )−1 T j=−k

11.5

Prespecified weighting matrices and moment conditions

Prespecified rather than “optimal” weighting matrices can emphasize economically interesting results, they can avoid the trap of blowing up standard errors rather than improving pricing errors, they can lead to estimates that are more robust to small model misspecifications. This is analogous to the fact that OLS is often preferable to GLS in a regression context. The GMM formulas for a fixed weighting matrix W are 1 var(ˆb) = (d0 W d)−1 d0 W SW d(d0 W d)−1 T var(gT ) =

1 (I − d(d0 W d)−1 d0 W )S(I − W d(d0 W d)−1 d0 ). T 196

S ECTION 11.5

P RESPECIFIED WEIGHTING MATRICES AND MOMENT CONDITIONS

In the basic approach outlined in Chapter 10, our final estimates were based on the “efficient” S −1 weighting matrix. This objective maximizes the asymptotic statistical information in the sample about a model, given the choice of moments gT . However, you may want to use a prespecified weighting matrix W 6= S −1 instead, or at least as a diagnostic accompanying more formal statistical tests. A prespecified weighting matrix lets you, rather than the S matrix, specify which moments or linear combination of moments GMM will value in the minimization min{b} gT (b)0 W gT (b). A higher value of Wii forces GMM to pay more attention to getting the ith moment right in the parameter estimation. For example, you might feel that some assets suffer from measurement error, are small and illiquid and hence should be deemphasized, or you may want to keep GMM from looking at portfolios with strong long and short position. I give some additional motivations below. You can also go one step further and impose which linear combinations aT of moment conditions will be set to zero in estimation rather than use the choice resulting from a minimization, aT = d0 S −1 or aT = d0 W . The fixed W estimate still trades off the accuracy of individual moments according £ to the¤0 sensitivity of each moment with respect to the parameter. For example, if gT = gT1 gT2 , W = I, but ∂gT /∂b = [1 10], so that the second moment is 10 times more sensitive to the parameter value than the first moment, then GMM with fixed weighting matrix sets 1 × gT1 + 10 × gT2 = 0. The second moment condition will be 10 times closer to zero than the first. If you really want GMM to pay equal attention to the two moments, then you can fix the aT matrix directly, for example aT = [1 1] or aT = [1 − 1].

Using a prespecified weighting matrix or using a prespecified set of moments is not the same thing as ignoring correlation of the errors ut in the distribution theory. The S matrix will still show up in all the standard errors and test statistics. 11.5.1

How to use prespecified weighting matrices

Once you have decided to use a prespecified weighting matrix W or a prespecified set of moments aT gT (b) = 0, the general distribution theory outlined in section 11.1 quickly gives standard errors of the estimates and moments, and therefore a χ2 statistic that can be used to test whether all the moments are jointly zero. Section 11.1 gives the formulas for the case that aT is prespecified. If we use weighting matrix W , the first order conditions to min{b} gT0 (b)W gT (b) are ∂gT (b)0 W gT (b) = d0 W gT (b) = 0, ∂b so we map into the general case with aT = d0 W. Plugging this value into (11.146), the 197

C HAPTER 11

GMM: GENERAL FORMULAS AND APPLICATIONS

variance-covariance matrix of the estimated coefficients is 1 var(ˆb) = (d0 W d)−1 d0 W SW d(d0 W d)−1 . T

(158)

(You can check that this formula reduces to 1/T (d0 S −1 d)−1 with W = S −1 .) Plugging a = d0 W into equation (11.147), we find the variance-covariance matrix of the moments gT var(gT ) =

1 (I − d(d0 W d)−1 d0 W )S(I − W d(d0 W d)−1 d0 ) T

(159)

As in the general formula, the terms to the left and right of S account for the fact that some linear combinations of moments are set to zero in each sample. Equation (11.159) can be the basis of χ2 tests for the overidentifying restrictions. If we interpret ()−1 to be a generalized inverse, then gT0 var(gT )−1 gT ∼ χ2 (#moments − #parameters). As in the general case, you have to pseudo-invert the singular var(gT ), for example by inverting only the non-zero eigenvalues. The major danger in using prespecified weighting matrices or moments aT is that the choice of moments, units, and (of course) the prespecified aT or W must be made carefully. For example, if you multiply the second moment by 10 times its original value, the S matrix will undo this transformation and weight them in their original proportions. The identity weighting matrix will not undo such transformations, so the units should be picked right initially. 11.5.2

Motivations for prespecified weighting matrices

Robustness, as with OLS vs. GLS. When errors are autocorrelated or heteroskedastic, every econometrics textbook shows you how to “improve” on OLS by making appropriate GLS corrections. If you correctly model the error covariance matrix and if the regression is perfectly specified, the GLS procedure can improve efficiency, i.e. give estimates with lower asymptotic standard errors. However, GLS is less robust. If you model the error covariance matrix incorrectly, the GLS estimates can be much worse than OLS. Also, the GLS transformations can zero in on slightly misspecified areas of the model, producing garbage. GLS is “best,” but OLS is “pretty darn good.” One often has enough data that wringing every last ounce of statistical precision (low standard errors) from the data is less important than producing estimates that do not depend on questionable statistical assumptions, and that transparently focus on the interesting features of the data. In these cases, it is often a good idea to use OLS estimates. The OLS standard error formulas are wrong, though, so you must correct the standard errors of the 198

S ECTION 11.5

P RESPECIFIED WEIGHTING MATRICES AND MOMENT CONDITIONS

OLS estimates for these features of the error covariance matrices, using the formulas we developed in section 11.4. GMM works the same way. First-stage or otherwise fixed weighting matrix estimates may give up something in asymptotic efficiency, but they are still consistent, and they can be more robust to statistical and economic problems. You still want to use the S matrix in computing standard errors, though, as you want to correct OLS standard errors, and the GMM formulas show you how to do this. Even if in the end you want to produce “efficient” estimates and tests, it is a good idea to calculate standard errors and model fit tests for the first-stage estimates. Ideally, the parameter estimates should not change by much, and the second stage standard errors should be tighter. If the “efficient” parameter estimates do change a great deal, it is a good idea to diagnose why this is so. It must come down to the “efficient” parameter estimates strongly weighting moments or linear combinations of moments that were not important in the first stage, and that the former linear combination of moments disagrees strongly with the latter about which parameters fit well. Then, you can decide whether the difference in results is truly due to efficiency gain, or whether it signals a model misspecification. Chapter 16 argues more at length for judicious use of “inefficient” methods such as OLS to guard against inevitable model misspecifications. Near-singular S. The spectral density matrix is often nearly singular, since asset returns are highly correlated with each other, and since we often include many assets relative to the number of data points. As a result, second stage GMM (and, as we will see below, maximum likelihood or any other efficient technique) tries to minimize differences and differences of differences of asset returns in order to extract statistically orthogonal components with lowest variance. One may feel that this feature leads GMM to place a lot of weight on poorly estimated, economically uninteresting, or otherwise non-robust aspects of the data. In particular, portfolios of the form 100R1 − 99R2 assume that investors can in fact purchase such heavily leveraged portfolios. Short-sale costs often rule out such portfolios or significantly alter their returns, so one may not want to emphasize pricing them correctly in the estimation and evaluation. For example, suppose that S is given by S=

·

1 ρ ρ 1

¸

.

so S −1 =

1 1 − ρ2

·

1 −ρ −ρ 1

¸

.

We can factor S −1 into a “square root” by the Choleski decomposition. This produces a 199

C HAPTER 11

GMM: GENERAL FORMULAS AND APPLICATIONS

triangular matrix C such that C 0 C = S −1 . You can check that the matrix " # √1 √−ρ 2 2 1−ρ 1−ρ C= 0 1

(160)

works. Then, the GMM criterion min gT0 S −1 gT is equivalent to min(gT0 C 0 )(CgT ). CgT gives the linear combination of moments that efficient GMM is trying to minimize. Looking at (11.160), as ρ → 1, the (2,2) element stays at 1, but the (1,1) and (1,2) elements get very large and of opposite signs. For example, if ρ = 0.95, then · ¸ 3.20 −3.04 C= . 0 1 In this example, GMM pays a little attention to the second moment, but places three times as much weight on the difference between the first and second moments. Larger matrices produce even more extreme weights. At a minimum, it is a good idea to look at S −1 and its Choleski decomposition to see what moments GMM is prizing. The same point has a classic interpretation, and is a well-known danger with classic regression-based tests. Efficient GMM wants to focus on well-measured moments. In asset pricing applications, the errors are typically close to uncorrelated over time, so GMM is looking for portfolios with small values of var(mt+1 Ret+1 ). Roughly speaking, those will be asset with small return variance. Thus, GMM will pay most attention to correctly pricing the sample minimum-variance portfolio, and GMM’s evaluation of the model by JT test will focus on its ability to price this portfolio. Now, consider what happens in a sample, as illustrated in Figure 24. The sample meanvariance frontier is typically a good deal wider than the true, or ex-ante mean-variance frontier. In particular, the sample minimum-variance portfolio may have little to do with the true minimum-variance portfolio. Like any portfolio on the sample frontier, its composition largely reflects luck – that’s why we have asset pricing models in the first place rather than just price assets with portfolios on the sample frontier. The sample minimum variance return is also likely to be composed of strong long-short positions. In sum, you may want to force GMM not to pay quite so much attention to correctly pricing the sample minimum variance portfolio, and you may want to give less importance to a statistical measure of model evaluation that almost entirely prizes GMM’s ability to price that portfolio. Economically interesting moments. 200

S ECTION 11.5

P RESPECIFIED WEIGHTING MATRICES AND MOMENT CONDITIONS

Sample minimum-variance portfolio E(R)

Sample, ex-post frontier

True, ex-ante frontier

σ(R)

Figure 24. True or ex ante and sample or ex-post mean-variance frontier. The sample often shows a spurious minimum-variance portfolio.

201

C HAPTER 11

GMM: GENERAL FORMULAS AND APPLICATIONS

The optimal weighting matrix makes GMM pay close attention to linear combinations of moments with small sampling error in both estimation and evaluation. One may want to force the estimation and evaluation to pay attention to economically interesting moments instead. The initial portfolios are usually formed on an economically interesting characteristic such as size, beta, book/market or industry. One typically wants in the end to see how well the model prices these initial portfolios, not how well the model prices potentially strange portfolios of those portfolios. If a model fails, one may want to characterize that failure as “the model doesn’t price small stocks” not “the model doesn’t price a portfolio of 900× small firm returns −600× large firm returns −299× medium firm returns.”

Level playing field.

The S matrix changes as the model and as its parameters change. (See the definition, (10.138) or (11.145).) As the S matrix changes, which assets the GMM estimate tries hard to price well changes as well. For example, the S matrix from one model may value strongly pricing the T bill well, while that of another model may value pricing a stock excess return well. Comparing the results of such estimations is like comparing apples and oranges. By fixing the weighting matrix, you can force GMM to pay attention to the various assets in the same proportion while you vary the model. The fact that S matrices change with the model leads to another subtle trap. One model my may “improve” a JT = gT0 S −1 gT statistic because it blows up the estimates of S, rather than making any progress on lowering the pricing errors gT . No one would formally use a comparison of JT tests across models to compare them, of course. But it has proved nearly irresistible for authors to claim success for a new model over previous ones by noting improved JT statistics, despite different weighting matrices, different moments, and sometimes much larger pricing errors. For example, if you take a model mt and create a new model by simply adding noise, unrelated to asset returns (in sample), m0t = mt + εt , then the moment condition gT = ETh(m0t Rte ) = ET ((mit + εt ) Rte ) is unchanged. However, the spectral den2 sity matrix S = E (mt + εt ) Rte Rte0 can rise dramatically. This can reduce the JT leading to a false sense of “improvement.” Conversely, if the sample contains a nearly riskfree portfolio of the test assets, or a portfolio with apparently small variance of mt+1 Ret+1 , then the JT test essentially evaluates the model by how will it can price this one portfolio. This can lead to a false rejection – even a very small gT will produce a large gT0 S −1 gT if there is an eigenvalue of S that is (spuriously) too small. If you use a common weighting matrix W for all models, and evaluate the models by gT0 W gT , then you can avoid this trap. Beware that the individual χ2 statistics are based on gT0 var(gT )−1 gT , and var(gT ) contains S, even with a prespecified weighting matrix W . You should look at the pricing errors, or at some statistic such as the sum of absolute or squared pricing errors to see if they are bigger or smaller, leaving the distribution aside. The question “are the pricing errors small?” is as interesting as the question “if we drew artificial data over and over again from a null statistical model, how often would we estimate a ratio 202

S ECTION 11.5

P RESPECIFIED WEIGHTING MATRICES AND MOMENT CONDITIONS

of pricing errors to their estimated variance gT0 S −1 gT this big or larger?” 11.5.3

Some prespecified weighting matrices

Two examples of economically interesting weighting matrices are the second-moment matrix of returns, advocated by Hansen and Jagannathan (1997) and the simple identity matrix, which is used implicitly in much empirical asset pricing. Second moment matrix. Hansen and Jagannathan (1997) advocate the use of the second moment matrix of payoffs W = E(xx0 )−1 in place of S. They motivate this weighting matrix as an interesting distance measure between a model for m, say y, and the space of true m’s. Precisely, the minimum distance (second moment) between a candidate discount factor y and the space of true discount factors is the same as the minimum value of the GMM criterion with W = E(xx0 )−1 as weighting matrix.

X proj(y| X)

m x*

y Nearest m

Figure 25. Distance between y and nearest m = distance between proj(y|X) and x∗ . To see why this is true, refer to Figure 25. The distance between y and the nearest valid m is the same as the distance between proj(y | X) and x∗ . As usual, consider the case that X is generated from a vector of payoffs x with price p. From the OLS formula, proj(y | X) = E(yx0 )E(xx0 )−1 x. 203

C HAPTER 11

GMM: GENERAL FORMULAS AND APPLICATIONS

x∗ is the portfolio of x that prices x by construction, x∗ = p0 E(xx0 )−1 x. Then, the distance between y and the nearest valid m is: ky − nearest mk = kproj(y|X) − x∗ k ° ° = °E(yx0 )E(xx0 )−1 x − p0 E(xx0 )−1 x° ° ° = °(E(yx0 ) − p0 ) E(xx0 )−1 x° = [E(yx) − p]0 E(xx0 )−1 [E(yx) − p] = gT0 E(xx0 )−1 gT

You might want to choose parameters of the model to minimize this “economic” measure of model fit, or this economically motivated linear combination of pricing errors, rather than the statistical measure of fit S −1 . You might also use the minimized value of this criterion to compare two models. In that way, you are sure the better model is better because it improves on the pricing errors rather than just blowing up the weighting matrix. Identity matrix. Using the identity matrix weights the initial choice of assets or portfolios equally in estimation and evaluation. This choice has a particular advantage with large systems in which S is nearly singular, as it avoids most of the problems associated with inverting a near-singular S matrix. Many empirical asset pricing studies use OLS cross-sectional regressions, which are the same thing as a first stage GMM estimate with an identity weighting matrix. Comparing the second moment and identity matrices. The second moment matrix gives an objective that is invariant to the initial choice of assets or portfolios. If we form a portfolio Ax of the initial payoffs x, with nonsingular A (i.e. a transformation that doesn’t throw away information) then [E(yAx) − Ap]0 E(Axx0 A0 )−1 [E(yAx) − Ap] = [E(yx) − p]0 E(xx0 )−1 [E(yx) − p]. The optimal weighting matrix S shares this property. It is not true of the identity or other fixed matrices. In those cases, the results will depend on the initial choice of portfolios. Kandel and Stambaugh (1995) have suggested that the results of several important asset pricing model tests are highly sensitive to the choice of portfolio; i.e. that authors inadvertently selected a set of portfolios on which the CAPM does unusually badly in a particular sample. Insisting that weighting matrices have this kind of invariance to portfolio selection might be a good device to ward against this problem. On the other hand, if you want to focus on the model’s predictions for economically interesting portfolios, then it wouldn’t make much sense for the weighting matrix to undo the specification of economically interesting portfolios! For example, many studies want 204

S ECTION 11.6

E STIMATING ON ONE GROUP OF MOMENTS , TESTING ON ANOTHER .

to focus on the ability of a model to describe expected returns that seem to depend on a characteristic such as size, book/market, industry, momentum, etc. Also, the second moment matrix is often even more nearly singular than the spectral density matrix, since E(xx0 ) = cov(x)+E(x)E(x)0 . Therefore, it often emphasizes portfolios with even more extreme short and long positions, and is no help on overcoming the near singularity of the S matrix.

11.6

Estimating on one group of moments, testing on another.

You may want to force the system to use one set of moments for estimation and another for testing. The real business cycle literature in macroeconomics does this extensively, typically using “first moments” for estimation (“calibration”) and “second moments” (i.e. first moments of squares) for evaluation. A statistically minded macroeconomist might like to know whether the departures of model from data “second moments” are large compared to sampling variation, and would like to include sampling uncertainty about the parameter estimates in this evaluation. You might want to choose parameters using one set of asset returns (stocks; domestic assets; size portfolios, first 9 size deciles, well-measured assets) and then see how the model does “out of sample” on another set of assets (bonds; foreign assets; book/market portfolios, small firm portfolio, questionably measured assets, mutual funds). However, you want the distribution theory for evaluation on the second set of moments to incorporate sampling uncertainty about the parameters in their estimation on the first set of moments. You can do all this very simply by using an appropriate weighting matrix or a prespecified moment matrix aT . For example, if the first N moments will be used to estimate N parameters, and the remaining M moments will be used to test the model “out of sample,” use aT = [IN 0N×M ] . If there are more moments N than parameters in the “estimation” block, you can construct a weighting matrix W which is an identity matrix in the N × N estimation block and zero elsewhere. Then aT = ∂gT0 /∂bW will simply contain the first N columns of ∂gT0 /∂b followed by zeros. The test moments will not be used in estimation. You could even use the inverse of the upper N × N block of S (not the upper block of the inverse of S!) to make the estimation a bit more efficient.

11.7

Estimating the spectral density matrix

Hints on estimating the spectral density or long run covariance matrix. 1) Use a sensible first stage estimate 2) Remove means 3) Downweight higher order correlations 4) Consider parametric structures for autocorrelation and heteroskedasticity 5) Use the null to limit the number of correlations or to impose other structure on S? 6) Size problems; consider a factor or other parametric cross-sectional structure for S. 7) Iteration and simultaneous b, S estimation. 205

C HAPTER 11

GMM: GENERAL FORMULAS AND APPLICATIONS

The optimal weighting matrix S depends on population moments, and depends on the parameters b. Work back through the definitions, ∞ X

E(ut u0t−j );

S

=

ut

≡ (mt (b)xt − pt−1 )

j=−∞

How do we estimate this matrix? The big picture is simple: following the usual philosophy, estimate population moments by their sample counterparts. Thus, use the first stage b estimates and the data to construct sample versions of the definition of S. This produces a consistent estimate of the true spectral density matrix, which is all the asymptotic distribution theory requires. The details are important however, and this section gives some hints. Also, you may want a different, and less restrictive, estimate of S for use in standard errors than you do when you are estimating S for use in a weighting matrix. 1) Use a sensible first stage W, or transform the data. In the asymptotic theory, you can use consistent first stage b estimates formed by any nontrivial weighting matrix. In practice, of course, you should use a sensible weighting matrix so that the first stage estimates are not ridiculously inefficient. W = I is often a good choice. Sometimes, some moments will have different units than other moments. For example, the dividend/price ratio is a number like 0.04. Therefore, the moment formed by Rt+1 × d/pt will be about 0.04 as big as large as the moment formed by Rt+1 × 1. If you use W = I, GMM will pay much less attention to the Rt+1 × d/pt moment. It is wise, then, to either use an initial weighting matrix that overweights the Rt+1 × d/pt moment, or to transform the data so the two moments are about the same mean and variance. For example, you could use Rt+1 × (1 + d/pt ). It is also useful to start with moments that are not horrendously correlated with each other, or to remove such correlation with a clever W . For example, you might consider Ra and Rb − Ra rather than Ra and Rb . You can accomplish this directly, or by starting with

W =

·

1 −1 0 1

¸·

1 0 −1 1

¸

=

·

2 −1 −1 1

¸

.

2) Remove means. Under the null, E(ut ) = 0, so it does not matter to the asymptotic distribution theory 206

S ECTION 11.7

E STIMATING THE SPECTRAL DENSITY MATRIX

whether you estimate the covariance matrix by removing means, using T T 1X 1X 0 [(ut − u ¯)(ut − u ¯) ] ; u ¯≡ ut T t=1 T t=1

or whether you estimate the second moment matrix by not removing means. However, Hansen and Singleton (1982) advocate removing the means in sample, and this is generally a good idea. It is already a major obstacle to second-stage estimation that estimated S matrices (and even simple variance-covariance matrices) are often nearly singular, providing an unreliable weighting matrix when inverted. Since second moment matrices E(uu0 ) = cov(u, u0 ) + E(u)E(u0 ) add a singular matrix E(u)E(u0 ) they are often even worse. 3) Downweight higher order correlations. You obviously cannot use a direct sample counterpart to the spectral density matrix. In a sample of size 100, there is no way to estimate E(ut u0t+101 ).Your estimate of E(ut u0t+99 ) is based on one data point, u1 u0100 . Hence, it will be a pretty unreliable estimate. For this reason, the estimator using all possible autocorrelations in a given sample is inconsistent. (Consistency means that as the sample grows, the probability distribution of the estimator converges to the true value. Inconsistent estimates typically have very large sample variation.) Furthermore, even S estimates that use few autocorrelations are not always positive definite in sample. This is embarrassing when one tries to invert the estimated spectral density matrix, which you have to do if you use it as a weighting matrix. Therefore, it is a good idea to construct consistent estimates that are automatically positive definite in every sample. One such estimate is the Bartlett estimate, used in this application by Newey and West (1987b). It is ¶ k µ T X k − |j| 1X (ut u0t−j ). (161) Sˆ = k T t=1 j=−k

As you can see, only autocorrelations up to kth (k < T ) order are included, and higher order autocorrelations are downweighted. (It’s important to use 1/T not 1/(T − k); this is a further downweighting.) The Newey-West estimator is basically the variance of kth sums, which is why it is positive definite in sample: k X ut−j = kE(ut u0t ) + (k − 1)[E(ut u0t−1 ) + E(ut−1 u0t )] + · · · V ar j=1

+[E(ut u0t−k ) + E(ut−k u0t )] = k

k X k − |j| E(ut u0t−k ). k

j=−k

Andrews (1991) gives some additional weighting schemes for spectral density estimates. 207

C HAPTER 11

GMM: GENERAL FORMULAS AND APPLICATIONS

This calculation also gives some intuition the S matrix. We’re looking for the variance Pfor T across samples of the sample mean var( T1 t=1 ut ). We only have one sample mean to look at, so we estimate the variance the sample mean by looking at the variance in a single ´ ³ of Pk sample of shorter sums, var k1 j=1 uj . The S matrix is sometimes called the long-run covariance matrix for this reason. In fact, one could estimate S directly as a variance of kth sums and obtain almost the same estimator, that would also be positive definite in any sample, vt

=

k X

ut−j ; v¯ =

j=1

Sˆ =

1 1 kT −k

T X 1 vt T −k t=k+1

T X

t=k+1

(vt − v¯) (vt − v¯)0 .

This estimator has been used when measurement of S is directly interesting (Cochrane 1998, Lo and MacKinlay 1988). A variety of other weighting schemes have been advocated. What value of k, or how wide a window if of another shape, should you use? Here again, you have to use some judgment. Too short values of k, together with a ut that is significantly autocorrelated, and you don’t correct for correlation that might be there in the errors. Too long a value of k, together with a series that does not have much autocorrelation, and the performance of the estimate and test deteriorates. If k = T /2 for example, you are really using only two data points to estimate the variance of the mean. The optimum value then depends on how much persistence or low-frequency movement there is in a particular application, vs. accuracy of the estimate. There is an extensive statistical literature about optimal window width, or size of k. Alas, this literature mostly characterizes the rate at which k should increase with sample size. You must promise to increase k as sample size increases, but not as quickly as the sample size increases – limT →∞ k = ∞, limT →∞ k/T = 0 – in order to obtain consistent estimates. In practice, promises about what you’d do with more data are pretty meaningless, and usually broken once more data arrives. 4) Consider parametric structures for autocorrelation and heteroskedasticity. “Nonparametric” corrections such as (11.161) often don’t perform very well in typical samples. The problem is that “nonparametric” techniques are really very highly parametric; you have to estimate many correlations in the data. Therefore, the nonparametric estimate varies a good deal from sample to sample, while the asymptotic distribution theory ignores sampling variation in covariance matrix estimates. The asymptotic distribution can therefore be a poor approximation to the finite-sample distribution of statistics like the JT . The S −1 weighting matrix will also be unreliable. One answer is to use a Monte-Carlo or bootstrap to estimate the finite-sample distribution parameters and test statistics rather than to rely on asymptotic theory. Alternatively, you can impose a parametric structure on the S matrix. Just because the 208

S ECTION 11.7

E STIMATING THE SPECTRAL DENSITY MATRIX

formulas are expressed in terms of a sum of covariances does not mean you have to estimate them that way; GMM is not inherently tied to “nonparametric” covariance matrix estimates. For example, if you model a scalar u as an AR(1) with parameter ρ, then you can estimate two numbers ρ and σ 2u rather than a whole list of autocorrelations, and calculate S=

∞ X

E(ut ut−j ) = σ2u

j=−∞

∞ X

ρ|j| = σ 2u

j=−∞

1+ρ 1−ρ

If this structure is not a bad approximation, imposing it can result in more reliable estimates and test statistics since one has to estimate many fewer coefficients. You could transform the data in such a way that there is less correlation to correct for in the first place. (This is a very useful formula, by the way. You are probably used to calculating the standard error of the mean as σ(x) σ(¯ x) = √ . T This formula assumes that the x are uncorrelated over time. If an AR(1) is not a bad model for their correlation, you can quickly adjust for correlation by using r σ(x) 1 + ρ √ σ(¯ x) = 1−ρ T instead.) This sort of parametric correction is very familiar from OLS regression analysis. The textbooks commonly advocate the AR(1) model for serial correlation as well as parametric models for heteroskedasticity corrections. There is no reason not to follow a similar approach for GMM statistics. 5) Use the null to limit correlations? In the typical asset pricing setup, the null hypothesis specifies that Et (ut+1 ) = Et (mt+1 Rt+1 − 1) = 0, as well as E(ut+1 ) = 0. This implies that all the autocorrelation terms of S drop out; E(ut u0t−j ) = 0 for j 6= 0. The lagged u could be an instrument z; the discounted return should be unforecastable, using past discounted returns as well as any other variable. In this situation, one could exploit the null to only include one term, and estimate T 1X ut u0t . Sˆ = T t=1

Similarly, if one runs a regression forecasting returns from some variable zt , Rt+1 = a + bzt + εt+1 , the null hypothesis that returns are not forecastable by any variable at time t means that the 209

C HAPTER 11

GMM: GENERAL FORMULAS AND APPLICATIONS

errors should not be autocorrelated. One can then simplify the standard errors in the OLS regression formulas given in section 11.4, eliminating all the leads and lags. In other situations, the null hypothesis can suggest a functional form for E(ut u0t−j ) or that some but not all are zero. For example, as we saw in section 11.4, regressions of long horizon returns on overlapping data lead to a correlated error term, even under the null hypothesis of no return forecastability. We can impose this null, ruling out terms past the overlap, as suggested by Hansen and Hodrick, k X 1 E(et xt x0t−j et−j ) E(xt x0t )−1 . var(bT ) = E(xt x0t )−1 T

(162)

j=−k

However, the null might not be correct, and the errors might be correlated. If so, you might make a mistake by leaving them out. If the null is correct, the extra terms will converge to zero and you will only have lost a few (finite-sample) degrees of freedom needlessly estimating them. If the null is not correct, you have an inconsistent estimate. With this in mind, you might want to include at least a few extra autocorrelations, even when the null says they don’t belong. Furthermore, there is no guarantee that the unweighted sum in (11.162) is positive definite in sample. If the sum in the middle is not positive definite, you could add a weighting to the sum, possibly increasing the number of lags so that the lags near k are not unusually underweighted. Again, estimating extra lags that should be zero under the null only loses a little bit of power. Monte Carlo evidence (Hodrick 1992) suggests that imposing the null hypothesis to simplify the spectral density matrix helps to get the finite-sample size of test statistics right – the probability of rejection given the null is true. One should not be surprised that if the null is true, imposing as much of it as possible makes estimates and tests work better. On the other hand, adding extra correlations can help with the power of test statistics – the probability of rejection given that an alternative is true – since they converge to the correct spectral density matrix. This trade-off requires some thought. For measurement rather than pure testing, using a spectral density matrix that can accommodate alternatives may be the right choice. For example, in the return forecasting regressions, one is really focused on measuring return forecastability rather than just formally testing the hypothesis that it is zero. On the other hand, the small-sample performance of the nonparametric estimators with many lags is not very good. If you are testing an asset pricing model that predicts u should not be autocorrelated, and there is a lot of correlation – if this issue makes a big difference – then this is an indication that something is wrong with the model; that including u as one of your instruments z would result in a rejection or at least substantially change the results. If the u are close to uncorrelated, then it really doesn’t matter if you add a few extra terms or not. 210

S ECTION 11.7

E STIMATING THE SPECTRAL DENSITY MATRIX

6) Size problems; consider a factor or other parametric cross-sectional structure. If you try to estimate a covariance matrix that is larger than the number of data points (say 2000 NYSE stocks and 800 monthly observations), the estimate of S, like any other covariance matrix, is singular by construction. This fact leads to obvious problems when you try to invert S! More generally, when the number of moments is more than around 1/10 the number of data points, S estimates tend to become unstable and near-singular. Used as a weighting matrix, such an S matrix tells you to pay lots of attention to strange and probably spurious linear combinations of the moments, as I emphasized in section 11.5. For this reason, most second-stage GMM estimations are limited to a few assets and a few instruments. A good, but as yet untried alternative might be to impose a factor structure or other wellbehaved structure on the covariance matrix. The near-universal practice of grouping assets into portfolios before analysis already implies an assumption that the true S of the underlying assets has a factor structure. Grouping in portfolios means that the individual assets have no information not contained in the portfolio, so that a weighting matrix S −1 would treat all assets in the portfolio identically. It might be better to estimate an S imposing a factor structure on all the primitive assets. Another response to the difficulty of estimating S is to stop at first stage estimates, and only use S for standard errors. One might also use a highly structured estimate of S as weighting matrix, while using a less constrained estimate for the standard errors. This problem is of course not unique to GMM. Any estimation technique requires us to calculate a covariance matrix. Many traditional estimates simply assume that ut errors are cross-sectionally independent. This false assumption leads to understatements of the standard errors far worse than the small sample performance of any GMM estimate. Our econometric techniques all are designed for large time series and small cross-sections. Our data has a large cross section and short time series. A large unsolved problem in finance is the development of appropriate large-N small-T tools for evaluating asset pricing models. 7) Alternatives to the two-stage procedure: iteration and one-step. Hansen and Singleton (1982) describe the above two-step procedure, and it has become popular for that reason. Two alternative procedures may perform better in practice, i.e. may result in asymptotically equivalent estimates with better small-sample properties. They can also be simpler to implement, and require less manual adjustment or care in specifying the setup (moments, weighting matrices) which is often just as important. a) Iterate. The second stage estimate ˆb2 will not imply the same spectral density as the first stage. It might seem appropriate that the estimate of b and of the spectral density should be consistent, i.e. to find a fixed point of ˆb = min{b} [gT (b)0 S(ˆb)−1 gT (b)]. One way to search for such a fixed point is to iterate: find b2 from ˆb2 = min gT (b)0 S −1 (b1 )gT (b) {b}

(163)

where b1 is a first stage estimate, held fixed in the minimization over b2 . Then use ˆb2 to find 211

C HAPTER 11

GMM: GENERAL FORMULAS AND APPLICATIONS

S(ˆb2 ), find ˆb3 = min[gT (b)0 S(ˆb2 )−1 gT (b)], {b}

and so on. There is no fixed point theorem that such iterations will converge, but they often do, especially with a little massaging. (I once used S [(bj + bj−1 )/2] in the beginning part of an iteration to keep it from oscillating between two values of b). Ferson and Foerster (1994) find that iteration gives better small sample performance than two-stage GMM in Monte Carlo experiments. This procedure is also likely to produce estimates that do not depend on the initial weighting matrix. b) Pick b and S simultaneously. It is not true that S must be held fixed as one searches for b. Instead, one can use a new S(b) for each value of b. Explicitly, one can estimate b by min [gT (b)0 S −1 (b)gT (b)] {b}

(164)

The estimates produced by this simultaneous search will not be numerically the same in a finite sample as the two-step or iterated estimates. The first order conditions to (11.163) are ¶0 µ ∂gT (b) S −1 (b1 )gT (b) = 0 (165) ∂b while the first order conditions in (11.164) add a term involving the derivatives of S(b) with respect to b. However, the latter terms vanish asymptotically, so the asymptotic distribution theory is not affected. Hansen, Heaton and Yaron (1996) conduct some Monte Carlo experiments and find that this estimate may have small-sample advantages in certain problems. A problem is that the one-step minimization may find regions of the parameter space that blow up the spectral density matrix S(b) rather than lower the pricing errors gT . Often, one choice will be much more convenient than another. For linear models, one can find the minimizing value of b from the first order conditions (11.165) analytically. This fact eliminates the need to search so even an iterated estimate is much faster. For nonlinear 0 models, each step involves a numerical search over gT (b) SgT (b). Rather than perform this 0 search many times, it may be much quicker to minimize once over gT (b) S(b)gT (b). On the other hand, the latter is not a locally quadratic form, so the search may run into greater numerical difficulties.

11.8 1. 2.

Problems

Use the delta method version of the GMM formulas to derive the sampling variance of an autocorrelation coefficient. Write a formula for the standard error of OLS regression coefficients that corrects for autocorrelation but not heteroskedasticity 212

S ECTION 11.8 3. 4.

P ROBLEMS

Write a formula for the standard error of OLS regression coefficients if E(et et−j ) = ρj σ2 . If the GMM errors come from an asset pricing model, ut = mt Rt − 1, can you ignore lags in the spectral density matrix? What if you know that returns are predictable? What if the error is formed from an instrument/managed portfolio ut zt−1 ?

213

Chapter 12. Regression-based tests of linear factor models This and the next three chapters study the question, how should we estimate and evaluate linear factor models; models of the form p = E(mx), m = b0 f or equivalently E(Re ) = βλ? These models are by far the most common in empirical asset pricing, and there is a large literature on econometric techniques to estimate and evaluate them. Each technique focuses on the same questions: how to estimate parameters, how to calculate standard errors of the estimated parameters, how to calculate standard errors of the pricing errors, and how to test ˆ. the model, usually with a test statistic of the form α ˆ 0 V −1 α I start with simple and longstanding time-series and cross-sectional regression tests. Then, I pursue GMM approach to the model expressed in p = E(mx), m = b0 f form. The following chapter summarizes the principle of maximum likelihood estimation and derives maximum likelihood estimates and tests. Finally, a chapter compares the different approaches. As always, the theme is the underlying unity. All of the techniques come down to one of two basic ideas: time-series regression or cross-sectional regression. The GMM, p = E(mx) approach turns out to be almost identical to cross-sectional regressions. Maximum likelihood (with appropriate statistical assumptions) justifies the time-series and cross-sectional regression approaches. The formulas for parameter estimates, standard errors, and test statistics are all strikingly similar.

12.1

Time-series regressions

When the factor is also a return, we can evaluate the model E(Rei ) = β i E(f) by running OLS time series regressions i Rei t = αi + β i ft + εt ; t = 1, 2, ...T

for each asset. The OLS distribution formulas (with corrected standard errors) provide standard errors of α and β. With errors that are i.i.d. over time, homoskedastic and independent of the factors, the asymptotic joint distribution of the intercepts gives the model test statistic, "

T 1+

µ

ET (f) σ ˆ (f )

¶2 #−1 214

ˆ −1 α α ˆ 0Σ ˆ ∼χ2N

S ECTION 12.1

T IME - SERIES REGRESSIONS

The Gibbons-Ross-Shanken test is a multivariate, finite sample counterpart to this statistic, when the errors are also normally distributed, ´−1 T −N −K ³ ˆ −1 ET (f ) ˆ −1 α 1 + ET (f)0 Ω α ˆ 0Σ ˆ ∼FN,T −N−K . N

I show how to construct the same test statistics with heteroskedastic and autocorrelated errors via GMM. I start with the simplest case. We have a factor pricing model with a single factor. The factor is an excess return (for example, the CAPM, with Rem = Rm − Rf ), and the test assets are all excess returns. We express the model in expected return - beta form. The betas are defined by regression coefficients Rtei = αi + β i ft + εit

(166)

and the model states that expected returns are linear in the betas: E(Rei ) = β i E(f ).

(167)

Since the factor is also an excess return, the model applies to the factor as well, so E(f ) = 1 × λ.

Comparing the model (12.167) and the expectation of the time series regression (12.166) we see that the model has one and only one implication for the data: all the regression intercepts αi should be zero. The regression intercepts are equal to the pricing errors.

Given this fact, Black Jensen and Scholes (1972) suggested a natural strategy for estimation and evaluation: Run time-series regressions (12.166) for each test asset. The estimate of the factor risk premium is just the sample mean of the factor, ˆ = ET (f ). λ Then, use standard OLS formulas for a distribution theory of the parameters. In particular you can use t-tests to check whether the pricing errors α are in fact zero. These distributions are usually presented for the case that the regression errors in (12.166) are uncorrelated and homoskedastic, but the formulas in section 11.4 show easily how to calculate standard errors for arbitrary error covariance structures. We also want to know whether all the pricing errors are jointly equal to zero. This requires us to go beyond standard formulas for the regression (12.166) taken alone, as we want to know the joint distribution of α estimates from separate regressions running side by side but with errors correlated across assets (E(εit εjt ) 6= 0). (We can think of 12.166 as a panel regression, and then it’s a test whether the firm dummies are jointly zero.) The classic form of these tests assume no autocorrelation or heteroskedasticity, but allow the errors to be correlated across assets. Dividing the α ˆ regression coefficients by their variance-covariance 215

C HAPTER 12

R EGRESSION - BASED TESTS OF LINEAR FACTOR MODELS

matrix leads to a χ2 test, "

T 1+

µ

ET (f) σ ˆ (f )

¶2 #−1

ˆ −1 α α ˆ 0Σ ˆ ∼χ2N

(168)

where ET (f ) denotes sample mean, σ ˆ 2 (f ) denotes sample variance, α ˆ is a vector of the estimated intercepts, ¤0 £ ˆ 2 ... α ˆN ˆ1 α α ˆ= α ˆ is the residual covariance matrix, i.e. the sample estimate of E(εt ε0t ) = Σ, where Σ ¤0 £ . εt = ε1t ε2t · · · εN t

As usual when testing hypotheses about regression coefficients, this test is valid asymptotically. The asymptotic distribution theory assumes that σ2 (f) (i.e. X 0 X) and Σ have converged to their probability limits; therefore it is asymptotically valid even though the factor is stochastic and Σ is estimated, but it ignores those sources of variation in a finite sample. It does not require that the errors are normal, relying on the central limit theorem so that α ˆ is normal. I derive (12.168) below. Also as usual in a regression context, we can derive a finite-sample F distribution for the hypothesis that a set of parameters are jointly zero, for fixed values of the right hand variable ft ,. " µ ¶2 #−1 T −N −1 ET (f) ˆ −1 α α ˆ 0Σ ˆ ∼FN,T −N−1 (169) 1+ N σ ˆ (f ) This is the Gibbons Ross and Shanken (1989) or “GRS” test statistic. The F distribution ˆ which is not included in (12.168). This distribution recognizes sampling variation in Σ, requires that the errors ε are normal as well as uncorrelated and homoskedastic. With normal ˆ is an independent Wishart (the multivariate version of a χ2 ), errors, the α ˆ are normal and Σ so the ratio is F . This distribution is exact in a finite sample. Tests (12.168) and (12.169) have a very intuitive form. The basic part of the test is a ˆ −1 α ˆ . If there were no βf in the model, then the α ˆ quadratic form in the pricing errors, α ˆ 0Σ would simply be the sample mean of the regression errors εt . Assuming i.i.d. εt , the variance ˆ would be a sum of their sample mean is just 1/T Σ. Thus, if we knew Σ then T α ˆ 0 Σ−1 α of squared sample means divided by their variance-covariance matrix, which would have an asymptotic χ2N distribution, or a finite sample χ2N distribution if the εt are normal. But we have to estimate Σ, which is why the finite-sample distribution is F rather than χ2 . We also estimate the β, and the second term in (12.168) and (12.169) accounts for that fact. Recall that a single beta representation exists if and only if the reference return is on the mean-variance frontier. Thus, the test can also be interpreted as a test whether f is ex216

S ECTION 12.1

T IME - SERIES REGRESSIONS

ante mean-variance efficient – whether it is on the mean-variance frontier using population moments – after accounting for sampling error. Even if f is on the true or ex-ante meanvariance frontier, other returns will outperform it in sample due to luck, so the return f will usually be inside the ex-post mean-variance frontier – i.e. the frontier drawn using sample moments. Still, it should not be too far inside the sample frontier. Gibbons Ross and Shanken show that the test statistic can be expressed in terms of how far inside the ex-post frontier the return f is, ´2 ³ ´2 ³ µq ET (f ) − σ ˆ (f ) T − N − 1 σq (170) ´2 . ³ N ET (f ) 1 + σˆ (f ) ³

µq σq

´2

is the Sharpe ratio of the ex-post tangency portfolio (maximum ex-post Sharpe ratio) formed from the test assets plus the factor f . If there are many factors that are excess returns, the same ideas work, with some cost of algebraic complexity. The regression equation is Rei = αi + β 0i ft + εit . The asset pricing model E(Rei ) = β 0i E(f) again predicts that the intercepts should be zero. We can estimate α and β with OLS timeˆ −1 α ˆ has the distriseries regressions. Assuming normal i.i.d. errors, the quadratic form α ˆ0Σ bution, ´−1 T −N −K ³ ˆ −1 ET (f) ˆ −1 α 1 + ET (f )0 Ω α ˆ0Σ ˆ ∼FN,T −N−K (171) N where

N K

= =

ˆ = Ω

Number of assets Number of factors T 1X [ft − ET (f)] [ft − ET (f)]0 T t=1

The main difference is that the Sharpe ratio of the single factor is replaced by the natural 0 ˆ −1 ET (f). generalization ET (f) Ω 12.1.1

Derivation of the χ2 statistic and distributions with general errors.

I derive (12.168) as an instance of GMM. This approach allows us to generate straightforwardly the required corrections for autocorrelated and heteroskedastic disturbances. (MacKin217

C HAPTER 12

R EGRESSION - BASED TESTS OF LINEAR FACTOR MODELS

lay and Richardson (1991) advocate GMM approaches to regression tests in this way.) It also serves to remind us that GMM and p = E(mx) are not necessarily paired; one can do a GMM estimate of an expected return - beta model too. The mechanics are only slightly different than what we did to generate distributions for OLS regression coefficients in section 11.4, since we keep track of N OLS regressions simultaneously. Write the equations for all N assets together in vector form, Ret = α + βft + εt . We use the usual OLS moments to estimate the coefficients, · ¸ µ· ¸¶ ET (Rte − α − βft ) εt = ET =0 gT (b) = ft εt ET [(Rte − α − βft ) ft ] These moments exactly identify the parameters α, β, so the a matrix in agT (ˆb) = 0 is the identity matrix. Solving, the GMM estimates are of course the OLS estimates, ˆ T (ft ) α ˆ = ET (Rte ) − βE e e e ˆ = ET [(Rt − ET (Rt )) ft ] = covT (Rt , ft ) . β ET [(ft − ET (ft )) ft ] varT (ft ) The d matrix in the general GMM formula is · ¸ · ¸ ∂gT (b) IN 1 E(ft ) IN E(ft ) = − d≡ = − ⊗ IN IN E(ft ) IN E(ft2 ) E(ft ) E(ft2 ) ∂b0 where IN is an N × N identity matrix. The S matrix is

¸ ∞ · X E(εt ε0t−j ) E(εt ε0t−j ft−j ) S= . E(ft εt ε0t−j ) E(ft εt ε0t−j ft−j ) j=−∞

Using the GMM variance formula (11.146) with a = I we have µ· ¸¶ α ˆ 1 var = d−1 Sd−10 . ˆ β T

(172)

At this point, we’re done. The upper left hand corner of var(α β) gives us var(ˆ α) and the −1 test we’re looking for is α ˆ 0 var(ˆ α) α ˆ ∼ χ2N .

The standard formulas make this expression prettier by assuming that the errors are uncorrelated over time and not heteroskedastic to simplify the S matrix, as we derived the standard OLS formulas in section 11.4. If we assume that f and ε are independent as well as orthogonal, E(fεε0 ) = E(f )E(εε0 ) and E(f 2 εε0 ) = E(f 2 )E(εε0 ). If we assume that the errors are independent over time as well, we lose all the lead and lag terms. Then, the S matrix 218

S ECTION 12.2

C ROSS - SECTIONAL REGRESSIONS

simplifies to S=

·

E(εt ε0t ) E(εt ε0t )E(ft ) 0 E(ft )E(εt εt ) E(εt ε0t )E(ft2 )

¸

=

·

1 E(ft ) E(ft ) E(ft2 )

¸

⊗Σ

(173)

Now we can plug into (12.172). Using (A⊗B)−1 = A−1 ⊗B −1 and (A⊗B)(C ⊗D) = AC ⊗ BD, we obtain var

µ·

α ˆ ˆ β

¸¶

1 = T

Ã·

1 E(ft ) E(ft ) E(ft2 )

¸−1

!

⊗Σ .

Evaluating the inverse, var

µ·

α ˆ ˆ β

¸¶

1 1 = T var(f)

·

E(ft2 ) −E(ft ) 1 −E(ft )

¸

⊗Σ

We’re interested in the top left corner. Using E(f 2 ) = E(f )2 + var(f), 1 var (ˆ α) = T

µ ¶ E(f )2 1+ Σ. var(f )

This is the traditional formula (12.168), but there is now no real reason to assume that the errors are i.i.d. or independent of the factors. By simply calculating 12.172, we can easily construct standard errors and test statistics that do not require these assumptions.

12.2

Cross-sectional regressions

We can fit E(Rei ) = β 0i λ + αi by running a cross-sectional regression of average returns on the betas. This technique can be used whether the factor is a return or not. I discuss OLS and GLS cross-sectional regressions, I find formulas for the standard errors of λ, and a χ2 test whether the α are jointly zero. I derive the distributions as an instance of GMM, and I show how to implement the same approach for autocorrelated and heteroskedastic errors. I show that the GLS cross-sectional regression is the same as the time-series regression when the factor is also an excess return, and is included in the set of test assets.

219

C HAPTER 12

R EGRESSION - BASED TESTS OF LINEAR FACTOR MODELS

E ( Rei )

αi

Assets i

Slope λ

βi

Figure 26. Cross-sectional regression Start again with the K factor model, written as E(Rei ) = β 0i λ; i = 1, 2, ...N

The central economic question is why average returns vary across assets; expected returns of an asset should be high if that asset has high betas or risk exposure to factors that carry high risk premia. Figure 26 graphs the case of a single factor such as the CAPM. Each dot represents one asset i. The model says that average returns should be proportional to betas, so plot the sample average returns against the betas. Even if the model is true, this plot will not work out perfectly in each sample, so there will be some spread as shown. Given these facts, a natural idea is to run a cross-sectional regression to fit a line through the scatterplot of Figure 26. First find estimates of the betas from a time series regression, 0 i Rei t = ai + β i ft + εt , t = 1, 2, ...T for each i.

(174)

Then estimate the factor risk premia λ from a regression across assets of average returns on the betas, ET (Rei ) = β 0i λ + αi , i = 1, 2....N.

220

(175)

S ECTION 12.2

CROSS - SECTIONAL REGRESSIONS

As in the figure, β are the right hand variables, λ are the regression coefficients, and the cross-sectional regression residuals αi are the pricing errors. This is also known as a twopass regression estimate, because one estimates first time-series and then cross-sectional regressions. You can run the cross-sectional regression with or without a constant. The theory says that the constant or zero-beta excess return should be zero. You can impose this restriction or estimate a constant and see if it turns out to be small. The usual tradeoff between efficiency (impose the null as much as possible to get efficient estimates) and robustness applies. 12.2.1

OLS cross-sectional regression

It will simplify notation to consider a single factor; the case of multiple factors looks the same with vectors in place of scalars. I denote from 1 to N with¤ missing sub or ¤ vectors £ 1 2 £ 0 N 0 superscripts, i.e. εt = εt εt · · · εt , β = β 1 β 2 · · · β N , and similarly for Ret and α. For simplicity take the case of no intercept in the cross-sectional regression. With this notation OLS cross-sectional estimates are ¡ ¢ ˆ = β 0 β −1 β 0 ET (Re ) λ (12.176) e ˆ α ˆ = ET (R ) − λβ. Next, we need a distribution theory for the estimated parameters. The most natural place to start is with the standard OLS distribution formulas. I start with the traditional assumption that the true errors are i.i.d. over time, and independent of the factors. This will give us some easily interpretable formulas, and we will see most of these terms remain when we do the distribution theory right later on. In an OLS regression Y = Xβ + u and E(uu0 ) = Ω, the standard error of the β estimate is (X 0 X)−1 X 0 ΩX(X 0 X)−1 . The residual covariance matrix is (I − X(X 0 X)−1 X 0 )Ω(I − X(X 0 X)−1 X 0 )0 Denote Σ = E (εt ε0t ). Since the αi are just time series averages of the true εit shocks (the average of the sample residuals is always zero), the errors in the cross-sectional regression have covariance matrix E (αα0 ) = T1 Σ. Thus the conventional OLS formulas for the covariance matrix of OLS estimates and residual with correlated errors give ³ ´ ¡ ¢−1 1 ¡ 0 ¢−1 0 ˆ = ββ σ2 λ β Σβ β 0 β (12.177) T ¡ ¢−1 0 ´ ³ ¡ ¢−1 0 ´ 1³ I − β β0β cov(ˆ α) = β Σ I − β β0β β (12.178) T We could test whether all pricing errors are zero with the statistic α ˆ 0 cov(ˆ α)−1 α ˆ ∼χ2N−1 .

221

(179)

C HAPTER 12

R EGRESSION - BASED TESTS OF LINEAR FACTOR MODELS

The distribution is χ2N−1 not χ2N because the covariance matrix is singular. The singularity and the extra terms in (12.178) result from the fact that the λ coefficient was estimated along the way, and means that we have to use a generalized inverse. (If there are K factors, we obviously end up with χ2N−K .) A test of the residuals is unusual in OLS regressions. We do not usually test whether the residuals are “too large,” since we have no information other than the residuals themselves about how large they should be. In this case, however, the first stage time-series regression gives us some independent information about the size of cov(αα0 ), information that we could not get from looking at the cross-sectional residual α itself. 12.2.2

GLS cross-sectional regression

Since the residuals in the cross-sectional regression (12.175) are correlated with each other, standard textbook advice is to run a GLS cross-sectional regression rather than OLS, using E(αα0 ) = T1 Σ as the error covariance matrix: ¡ 0 −1 ¢−1 0 −1 βΣ β β Σ ET (Re ) ˆ α ˆ = ET (Re ) − λβ. ˆ = λ

The standard regression formulas give the variance of these estimates as ³ ´ 1 ¡ 0 −1 ¢−1 ˆ = βΣ β σ2 λ T ³ ¡ ¢−1 0 ´ 1 Σ − β β 0 Σ−1 β cov(ˆ α) = β T

(12.180)

(12.181) (12.182)

The comments of section 11.5 warning that OLS is sometimes much more robust than GLS apply in this case. The GLS regression should improve efficiency, i.e. give more precise estimates. However, Σ may be hard to estimate and to invert, especially if the cross-section N is large. One may well choose the robustness of OLS over the asymptotic statistical advantages of GLS. A GLS regression can be understood as a transformation of the space of returns, to focus attention on the statistically most informative portfolios. Finding (say, by Choleski decomposition) a matrix C such that CC 0 = Σ−1 , the GLS regression is the same as an OLS regression of CET (Re ) on Cβ , i.e. of testing the model on the portfolios CRe . The statistically most informative portfolios are those with the lowest residual variance Σ. But this asymptotic statistical theory assumes that the covariance matrix has converged to its true value. In most samples, the ex-post or sample mean-variance frontier still seems to indicate lots of luck, and this is especially true if the cross section is large, anything more than 1/10 of the time series. The portfolios CRe are likely to contain many extreme long-short positions. Again, we could test the hypothesis that all the α are equal to zero with (12.179). Though the appearance of the statistic is the same, the covariance matrix is smaller, reflecting the 222

S ECTION 12.2

CROSS - SECTIONAL REGRESSIONS

greater power of the GLS test. As with the JT test, (11.152) we can develop an equivalent test that does not require a generalized inverse; Tα ˆ 0 Σ−1 α ˆ ∼χ2N−1 .

(183)

To derive (12.183), I proceed exactly as in the derivation of the JT test (11.152). Define, say 0 −1 by Choleski √ decomposition, a matrix C such that CC = Σ . Now, find the covariance ˆ: matrix of T C 0 α ³ √ ¡ ¢−1 0 ¡ ¢−1 0 ´ −1 cov( T Cα) = C 0 (CC 0 ) − β β 0 CC 0 β β C = I − δ δ0 δ δ

where

δ = C 0 β. √ √ In sum, α ˆ is asymptotically normal so T C 0 α ˆ is asymptotically normal, cov( T C 0 α ˆ ) is an ˆ 0 CC 0 α ˆ = Tα ˆ 0 Σ−1 α ˆ is χ2N−1 . idempotent matrix with rank N − 1; therefore T α 12.2.3 Correction for the fact that β are estimated, and GMM formulas that don’t need i.i.d. errors.

In applying standard OLS formulas to a cross-sectional regression, we assume that the right hand variables β are fixed. The β in the cross-sectional regression are not fixed, of course, but are estimated in the time series regression. This turns out to matter, even as T → ∞.

In this section, I derive the correct asymptotic standard errors. With the simplifying assumption that the errors ε are i.i.d. over time and independent of the factors, the result is i h ´ ¡ ¢ ³ ˆ OLS ) = 1 (β 0 β)−1 β 0 Σβ β 0 β −1 1 + λ0 Σ−1 λ + Σf σ 2 (λ (12.184) f T i h¡ ´ ¢ ³ ˆ GLS ) = 1 β 0 Σ−1 β −1 1 + λ0 Σ−1 λ + Σf σ 2 (λ f T

where Σf is the variance-covariance matrix of the factors. This correction is due to Shanken (1992). Comparing these³ standard errors ´ to (12.177) and (12.181), we see that there is a 0 −1 multiplicative correction 1 + λ Σf λ and an additive correction Σf . The asymptotic variance-covariance matrix of the pricing errors is ´ ¢³ ¡ ¢−1 0 ´ ¡ 1³ IN − β β 0 β cov(ˆ αOLS ) = β Σ IN − β(β 0 β)−1 β 0 1 + λ0 Σ−1 (12.185) λ f T ´ ¡ 0 −1 ¢−1 0 ´ ³ 1³ Σ−β β Σ β 1 + λ0 Σ−1 cov(ˆ αGLS ) = β (12.186) f λ T

Comparing these results to (12.178) and (12.182) we see the same multiplicative correction applies. 223

C HAPTER 12

R EGRESSION - BASED TESTS OF LINEAR FACTOR MODELS

We can form the asymptotic χ2 test of the pricing errors by dividing pricing errors by their ˆ cov(ˆ α)−1 α ˆ . Following (12.183), we can simplify this result for variance-covariance matrix, α the GLS pricing errors resulting in ³ ´ T 1 + λ0 Σ−1 λ α ˆ 0GLS Σ−1 α ˆ GLS ∼ χ2N−K . (187) f Are the corrections important relative to the simple OLS formulas given above? In the CAPM λ = E(Rem ) so λ2 /σ2 (Rem ) ≈ (0.08/0.16)2 = 0.25 in annual data. In annual data, then, the multiplicative term is too large to ignore. However, the mean and variance both scale with horizon, so the Sharpe ration scales with the square root of horizon. Therefore, for a monthly interval λ2 /σ 2 (Rem ) ≈ 0.25/12 ≈ 0.02 which is quite small and ignoring the multiplicative term makes little difference. ˆ can be very important. Consider a one factor The additive term in the standard error of λ model, suppose all the β are 1.0, all the residuals are uncorrelated so Σ is diagonal, suppose all assets have the same residual covariance σ2 (ε), and ignore the multiplicative term. Now we can write either covariance matrix in (12.184) as ¸ · 1 1 2 2 ˆ 2 σ (ε) + σ (f) σ (λ) = T N Even with N = 1, most factor models have fairly high R2 , so σ2 (ε) < σ2 (f ). Typical CAPM values of R2 = 1 − σ2 (ε)/σ2 (f ) for large portfolios are 0.6-0.7; and multifactor models such as the Fama French 3 factor model have R2 often over 0.9. Typical numbers of assets N = 10 to 50 make the first term vanish compared to the second term. More generally, suppose the factor were in fact a return. Then the factor risk premium is λ = E(f), and we’d use Σf /T as the standard error of λ. This is the “correction” term in (12.184), so we expect it to be, in fact, the most important term. Note that Σf /T is the standard error of the mean of f . Thus, in the case that the return is a factor, so E(f ) = λ, this is the only term you would use. This example suggests that Σf is not just an important correction, it is likely to be the ˆ. dominant consideration in the sampling error of the λ Comparing (12.187) to the GRS tests for a time-series regression, (12.168), (12.169), (12.171) we see the same statistic. The only difference is that by estimating λ from the cross-section rather than imposing λ = E(f ), the cross-sectional regression loses degrees of freedom equal to the number of factors. Though these formulas are standard classics, I emphasize that we don’t have to make the severe assumptions on the error terms that are used to derive them. As with the time-series ˆ and α ˆ , and only at the last moment case, I derive a general formula for the distribution of λ make classic error term assumptions to make the spectral density matrix pretty. Derivation and formulas that don’t require i.i.d. errors.

224

S ECTION 12.2

CROSS - SECTIONAL REGRESSIONS

The easy and elegant way to account for the effects of “generated regressors” such as the β in the cross-sectional regression is to map the whole thing into GMM. Then, we treat the moments that generate the regressors β at the same time as the moments that generate the cross-sectional regression coefficient λ, and the covariance matrix S between the two sets of moments captures the effects of generating the regressors on the standard error of the cross-sectional regression coefficients. Comparing this straightforward derivation with the ˆ and noting difficulty of Shanken’s (1992) paper that originally derived the corrections for λ, that Shanken did not go on to find the formulas (12.185) that allow a test of the pricing errors is a nice argument for the simplicity and power of the GMM framework. To keep the algebra manageable, I treat the case of a single factor. The moments are 0 E(Rte − a − βft ) gT (b) = E [(Rte − a − βft )ft ] = 0 0 E (Re − βλ)

(188)

The top two moment conditions exactly identify a and β as the time-series OLS estimates. (Note a not α. The time-series intercept is not necessarily equal to the pricing error in a cross-sectional regression.) The bottom moment condition is the asset pricing model. It is in general overidentified in a sample, since there is only one extra parameter (λ) and N extra moment conditions. If we use a weighting vector β 0 on this condition, we obtain the OLS cross-sectional estimate of λ. If we use a weighting vector β 0 Σ−1 , we obtain the GLS crosssectional estimate of λ. To accommodate both cases, use a weighting vector γ 0 , and then substitute γ 0 = β 0 , γ 0 = β 0 Σ−1 , etc. at the end. ˆ come straight from the general GMM standard error formula The standard errors for λ ˆ are not parameters, but are the last N moments. Their covariance matrix is (11.146). The α thus given by the GMM formula (11.147) for the sample variation of the gT . All we have to do is map the problem into the GMM notation. The parameter vector is b0 =

£

a0

β0

λ

¤

The a matrix chooses which moment conditions are set to zero in estimation, a=

·

I2N 0

0 γ0

¸

.

The d matrix is the sensitivity of the moment conditions to the parameters, −IN −IN E(f ) 0 ∂gT d= = −IN E(f ) −IN E(f 2 ) 0 ∂b0 −β 0 −λIN

225

C HAPTER 12

R EGRESSION - BASED TESTS OF LINEAR FACTOR MODELS

The S matrix is the long-run covariance matrix of the moments.

S

0 e Rt−j − a − βft−j Ret − a − βft e − a − βft−j )ft−j = E (Ret − a − βft )ft (Rt−j e e j=−∞ Rt − βλ Rt−j − βλ 0 ∞ εt εt−j X εt ft εt−j ft−j = E j=−∞ β(ft − Ef ) + εt β(ft−j − Ef) + εt−j ∞ X

In the second expression, I have used the regression model and the restriction under the null that E (Rte ) = βλ. In calculations, of course, you could simply estimate the first expression. We are done. We have the ingredients to calculate the GMM standard error formula (11.146) and formula for the covariance of moments (11.147). We can recover the classic formulas (12.184), (12.185), (12.186) by adding the assumption that the errors are i.i.d. and independent of the factors, and that the factors are uncorrelated over time as well. The assumption that the errors and factors are uncorrelated over time means we can ignore the lead and lag terms. Thus, the top left corner is E(εt ε0t ) = Σ. The assumption that the errors are independent from the factors ft simplifies the terms in which εt and ft are multiplied: E(εt (ε0t ft )) = E(f )Σ for example. The result is Σ E(f )Σ Σ E(f )Σ S = E(f )Σ E(f 2 )Σ 0 2 Σ E(f )Σ ββ σ (f ) + Σ

Multiplying a, d, S together as specified by the GMM formula for the covariance matrix of parameters (11.146) we obtain the covariance matrix of all the parameters, and its (3,3) ˆ . Multiplying the terms together as specified by (11.147), we element gives the variance of λ obtain the sampling distribution of the α ˆ , (12.185). The formulas (12.184) reported above are derived the same way with a vector of factors ft rather than a scalar; the second moment condition in (12.188) then reads E [(Rte − a − βf t ) ⊗ ft ]. The matrix multiplication is not particularly enlightening. Once again, there is really no need to make the assumption that the errors are i.i.d. and especially that they are conditionally homoskedastic – that the factor f and errors ε are independent. It is quite easy to estimate an S matrix that does not impose these conditions and calculate standard errors. They will not have the pretty analytic form given above, but they will more closely report the true sampling uncertainty of the estimate. Furthermore, if one is really interested in efficiency, the GLS cross-sectional estimate should use the spectral density matrix as weighting matrix rather than Σ−1 . 226

S ECTION 12.2 12.2.4

CROSS - SECTIONAL REGRESSIONS

Time series vs. cross-section

How are the time-series and cross-sectional approaches different? Most importantly, you can run the cross-sectional regression when the factor is not a return. The time-series test requires factors that are also returns, so that you can estimate ˆ = ET (f ). The asset pricing model does predict a restriction on the factor risk premia by λ intercepts in the time-series regression. Why not just test these? If you impose the restriction E(Rei ) = β 0i λ, you can write the time-series regression (12.174) as 0 0 i Rei t = β i λ + β i (ft − E(f)) + εt , t = 1, 2, ...T for each i.

Comparing this with (12.174), you see that the intercept restriction is ai = β 0i (λ − E(f )) .

This restriction makes sense. The model says that mean returns should be proportional to betas, and the intercept in the time-series regression controls the mean return. You can also see how λ = E(f ) results in a zero intercept. Finally, however, you see that without an estimate of λ, you can’t check this intercept restriction. If the factor is not a return, you will be forced to do something like a cross-sectional regression. When the factor is a return, so that we can compare the two methods, they are not necessarily the same. The time-series regression estimates the factor risk premium as the sample mean of the factor. Hence, the factor receives a zero pricing error. Also, the predicted zerobeta excess return is also zero. Thus, the time-series regression describes the cross-section of expected returns by drawing a line as in Figure 26 that runs through the origin and through the factor, ignoring all of the other points. The OLS cross-sectional regression picks the slope and intercept, if you include one, to best fit all the points; to minimize the sum of squares of all the pricing errors. If the factor is a return, the GLS cross-sectional regression, including the factor as a test asset, is identical to the time-series regression. The time-series regression for the factor is, of course, ft = 0 + 1ft + 0

so it has a zero intercept, beta equal to one, and zero residual in every sample. The residual variance covariance matrix of the returns, including the factor, is µ· e ¸ ¸ ¶ · R − a − βf Σ 0 E [·]0 = 0 0 f − 0 − 1f Since the factor has zero residual variance, a GLS regression puts all its weight on that asset. ˆ = ET (f ) just as for the time-series regression. The pricing errors are the same, Therefore, λ as is their distribution and the χ2 test. (You gain a degree of freedom by adding the factor to the cross sectional regression, so the test is a χ2N .) 227

C HAPTER 12

R EGRESSION - BASED TESTS OF LINEAR FACTOR MODELS

Why does the “efficient” technique ignore the pricing errors of all of the other assets in estimating the factor risk premium, and focus only on the mean return? The answer is simple, though subtle. In the regression model Ret = a + βft + εt ,

the average return of each asset in a sample is equal to beta times the average return of the factor in the sample, plus the average residual in the sample. An average return carries no additional information about the mean of the factor. A signal plus noise carries no additional information beyond that in the same signal. Thus, an “efficient” cross-sectional regression wisely ignores all the information in the other asset returns and uses only the information in the factor return to estimate the factor risk premium.

12.3

Fama-MacBeth Procedure

I introduce the Fama-MacBeth procedure for running cross sectional regression and calculating standard errors that correct for cross-sectional correlation in a panel. I show that, when the right hand variables do not vary over time, Fama-MacBeth is numerically equivalent to pooled time-series, cross-section OLS with standard errors corrected for cross-sectional correlation, and also to a single cross-sectional regression on time-series averages with standard errors corrected for cross-sectional correlation. Fama-MacBeth standard errors do not include corrections for the fact that the betas are also estimated. Fama and MacBeth (1973) suggest an alternative procedure for running cross-sectional regressions, and for producing standard errors and test statistics. This is a historically important procedure, it is computationally simple to implement, and is still widely used, so it is important to understand it and relate it to other procedures. First, you find beta estimates with a time-series regression. Fama and MacBeth use rolling 5 year regressions, but one can also use the technique with full-sample betas, and I will consider that simpler case. Second, instead of estimating a single cross-sectional regression with the sample averages, we now run a cross-sectional regression at each time period, i.e. Rtei = β 0i λt + αit i = 1, 2, ...N for each t.

I write the case of a single factor for simplicity, but it’s easy to extend the model to multiple factors. Then, Fama and MacBeth suggest that we estimate λ and αi as the average of the cross sectional regression estimates, T T X 1X ˆ= 1 ˆt; α λ λ ˆi = α ˆ it . T t=1 T t=1

228

S ECTION 12.3

FAMA -MAC B ETH P ROCEDURE

Most importantly, they suggest that we use the standard deviations of the cross-sectional regression estimates to generate the sampling errors for these estimates, ˆ = σ2 (λ)

T T ´2 1 X ³ˆ 1 X 2 ˆ λ − λ ; σ (ˆ α ) = (ˆ αit − α ˆ i )2 . t i T 2 t=1 T 2 t=1

It’s 1/T 2 because we’re finding standard errors of sample means, σ2 /T This is an intuitively appealing procedure once you stop to think about it. Sampling error is, after all, about how a statistic would vary from one sample to the next if we repeated the observations. We can’t do that with only one sample, but why not cut the sample in half, and deduce how a statistic would vary from one full sample to the next from how it varies from the first half of the sample to the next half? Proceeding, why not cut the sample in fourths, eights and so on? The Fama-MacBeth procedure carries this idea to is logical conclusion, ˆ t over time to deduce its sampling variation. using the variation in the statistic λ We are used to deducing the sampling variance of the sample mean of a series xt by x) = σ2 (x)/T = looking at the variation of xt through time in the sample, using σ 2 (¯ P 2 1 ¯) . The Fama-MacBeth technique just applies this idea to the slope and prict (xt − x T2 ing error estimates. The formula assumes that the time series is not autocorrelated, but one ˆ t that are correlated over time by using a long run could easily extend the idea to estimates λ variance matrix, i.e. estimate . ˆ = σ 2 (λ)

∞ 1 X ˆt, λ ˆ t−j ) covT (λ T j=−∞

One should of course use some sort of weighting matrix or a parametric description of the ˆ , as explained in section 11.7. Asset return data are usually not highly autocorrelations of λ correlated, but accounting for such correlation could have a big effect on the application of the Fama-MacBeth technique to corporate finance data or other regressions in which the cross-sectional estimates are highly correlated over time. It is natural to use this sampling theory to test whether all the pricing errors are jointly zero as we have before. Denote by α the vector of pricing errors across assets. We could estimate the covariance matrix of the sample pricing errors by α ˆ =

cov(ˆ α) =

T 1X α ˆt T t=1

T 1 X (ˆ αt − α ˆ ) (ˆ αt − α ˆ )0 T 2 t=1

(or a general version that accounts for correlation over time) and then use the test α ˆ 0 cov(ˆ α)−1 α ˆ ∼ χ2N−1 .

229

C HAPTER 12 12.3.1

R EGRESSION - BASED TESTS OF LINEAR FACTOR MODELS

Fama MacBeth in depth

The GRS procedure and the formulas given above for a single cross-sectional regression are familiar from any course in regression. We will see them justified by maximum likelihood below. The Fama MacBeth procedure seems unlike anything you’ve seen in any econometrics course, and it is obviously a useful and simple technique that can be widely used in panel data in economics and corporate finance as well as asset pricing. Is it truly different? Is there something different about asset pricing data that requires a fundamentally new technique not taught in standard regression courses? Or is it similar to standard techniques? To answer these questions it is worth looking in a little more detail at what it accomplishes and why. It’s easier to do this in a more standard setup, with left hand variable y and right hand variable x. Consider a regression yit = β 0 xit + εit i = 1, 2, ...N; t = 1, 2, ...T.

The data in this regression has a cross-sectional element as well as a time-series element. In corporate finance, for example, one might be interested in the relationship between investment and financial variables, and the data set has many firms (N ) as well as time series observations for each firm (T ). In and expected return-beta asset pricing model, the xit standing for the β i and β stands for λ. An obvious thing to do in this context is simply to stack the i and t observations together and estimate β by OLS. I will call this the pooled time-series cross-section estimate. However, the error terms are not likely to be uncorrelated with each other. In particular, the error terms are likely to be cross-sectionally correlated at a given time. If one stock’s return is unusually high this month, another stock’s return is also likely to be high; if one firm invests an unusually great amount this year, another firm is also likely to do so. When errors are correlated, OLS is still consistent, but the OLS distribution theory is wrong, and typically suggests standard errors that are much too small. In the extreme case that the N errors are perfectly correlated at each time period, there really only one observation for each time period, so one really has T rather than NT observations. Therefore, a real pooled time-series cross-section estimate must include corrected standard errors. People often ignore this fact and report OLS standard errors. Another thing we could do is first take time series averages and then run a pure crosssectional regression of ET (yit ) = β 0 ET (xit ) + ui i = 1, 2, ...N

This procedure would lose any information due to variation of the xit over time, but at least it might be easier to figure out a variance-covariance matrix for ui and correct the standard errors for residual correlation. (You could also average cross-sectionally and than run a single time-series regression. We’ll get to that option later.) In either case, the standard error corrections are just applications of the standard formula 230

S ECTION 12.3

FAMA -MAC B ETH P ROCEDURE

for OLS regressions with correlated error terms. Finally, we could run the Fama-MacBeth procedure: run a cross-sectional regression at each point in time; average the cross-sectional βˆ t estimates to get an estimate βˆ , and use the time-series standard deviation of βˆ t to estimate the standard error of βˆ . It turns out that the Fama MacBeth procedure is just another way of calculating the standard errors, corrected for cross-sectional correlation: Proposition: If the xit variables do not vary over time, and if the errors are cross-sectionally correlated but not correlated over time, then the Fama-MacBeth estimate, the pure crosssectional OLS estimate and the pooled time-series cross-sectional OLS estimates are identical. Also, the Fama-MacBeth standard errors are identical to the cross-sectional regression or stacked OLS standard errors, corrected for residual correlation. None of these relations hold if the x vary through time.

Since they are identical procedures, whether one calculates estimates and standard errors in one way or the other is a matter of taste. I emphasize one procedure that is incorrect: pooled time series and cross section OLS with no correction of the standard errors. The errors are so highly cross-sectionally correlated in most finance applications that the standard errors so computed are often off by a factor of 10. The assumption that the errors are not correlated over time is probably not so bad for asset pricing applications, since returns are close to independent. However, when pooled time-series cross-section regressions are used in corporate finance applications, errors are likely to be as severely correlated over time as across firms, if not more so. The “other factors” (ε) that cause, say, company i to invest more at time t than predicted by a set of right hand variables is surely correlated with the other factors that cause company j to invest more. But such factors are especially likely to cause company i to invest more tomorrow as well. In this case, any standard errors must also correct for serial correlation in the errors; the GMM based formulas in section 11.4 can do this easily. The Fama-MacBeth standard errors also do not correct for the fact that βˆ are generated regressors. If one is going to use them, it is a good idea to at least calculate the Shanken correction factors outlined above, and check that the corrections are not large. Proof: We just have to write out the three approaches and compare them. Having assumed that the x variables do not vary over time, the regression is yit = x0i β + εit .

We can stack up the cross-sections i = 1...N and write the regression as yt = xβ + εt . x is now a matrix with the x0i as rows. The error assumptions mean E(εt ε0t ) = Σ.

231

C HAPTER 12

R EGRESSION - BASED TESTS OF LINEAR FACTOR MODELS

Pooled OLS: To run pooled OLS, we stack the time series and cross sections by writing ε1 x y1 ε2 x y2 Y = . ; X = . ; ² = . .. .. .. yT

x

εT

and then

Y = Xβ + ²

with

E(²²0 ) = Ω =

Σ

..

. Σ

The estimate and its standard error are then

ˆ OLS = (X 0 X)−1 X 0 Y β −1 −1 0 ˆ X 0 ΩX (X 0 X) cov(β OLS ) = (X X)

Writing this out from the definitions of the stacked matrices, with X 0 X =T x0 x, ˆ β OLS

−1

= (x0 x) x0 ET (yt ) 1 0 −1 0 −1 ˆ (x x) (x Σx) (x0 x) . cov(β OLS ) = T

We can estimate this sampling variance with ¡ ¢ ˆ OLS ˆ = ET ˆεtˆε0t ; ˆεt ≡ yt − xβ Σ

Pure cross-section: The pure cross-sectional estimator runs one cross-sectional regression of the time-series averages. So, take those averages, ET (yt ) = xβ + ET (εt )

where x = ET (x ) since x is constant. Having assumed i.i.d. errors over time, the error covariance matrix is E (ET (εt ) ET (ε0t )) =

1 Σ. T

The cross sectional estimate and corrected standard errors are then ˆ XS β

−1

= (x0 x) x0 ET (yt ) ˆ ) = 1 (x0 x)−1 x0 Σx−1 (x0 x)−1 σ 2 (β XS T

232

S ECTION 12.3

FAMA -MAC B ETH P ROCEDURE

Thus, the cross-sectional and pooled OLS estimates and standard errors are exactly the same, in each sample. Fama-MacBeth: The Fama–MacBeth estimator is formed by first running the crosssectional regression at each moment in time, ˆ = (x0 x)−1 x0 yt . β t

Then the estimate is the average of the cross-sectional regression estimates, ³ ´ ˆ = (x0 x)−1 x0 ET (yt ) . ˆ F M = ET β β t

Thus, the Fama-MacBeth estimator is also the same as the OLS estimator, in each sample. The Fama-MacBeth standard error is based on the time-series standard deviation of the βˆ t . Using covT to denote sample covariance, ´ ³ ´ ³ ˆ = 1 (x0 x)−1 x0 covT (yt ) x (x0 x)−1 . ˆ F M = 1 covT β cov β t T T with

yt = xβ F M + ˆεt

we have ˆ covT (yt ) = ET (ˆεtˆε0t ) = Σ

and finally ´ ³ ˆ F M = 1 (x0 x)−1 x0 Σx ˆ (x0 x)−1 . cov β T

Thus, the FM estimator of the standard error is also numerically equivalent to the OLS corrected standard error. Varying x If the xit vary through time, none of the three procedures are equal anymore, since the cross-sectional regressions ignore time-series variation in the xit . As an extreme example, suppose a scalar xit varies over time but not cross-sectionally, yit = α + xt β + εit ; i = 1, 2, ...N ; t = 1, 2, ...T.

The grand OLS regression is P P P ˜t N1 i yit ˜t yit tx it x ˆ P P β = = OLS ˜2t ˜2t it x tx

where x˜ = x − ET (x) denotes the demeaned variables. The estimate is driven by the covariance over time of xt with the cross-sectional average of the yit , which is sensible because all of the information in the sample lies in time variation. It is identical to a regression over time 233

C HAPTER 12

R EGRESSION - BASED TESTS OF LINEAR FACTOR MODELS

of cross-sectional averages. However, you can’t even run a cross-sectional estimate, since the right hand variable is constant across i. As a practical example, you might be interested in a CAPM specification in which the betas vary over time (β t ) but not across test assets. This sample still contains information about the CAPM: the time-variation in betas should be matched by time variation in expected returns. But any method based on cross-sectional ¥ regressions will completely miss it. In historical context, the Fama MacBeth procedure was also important because it allowed changing betas, which a single cross-sectional regression or a time-series regression test cannot easily handle.

12.4 1.

2. 3. 4.

Problems

When we express the CAPM in excess return form, can the test assets be differences between risky assets, Ri −Rj ? Can the market excess return also use a risky ¡ asset, or must ¢ it be relative to a risk free rate? (Hint: start with E(Ri ) − Rf = β i,m E(Rm ) − Rf and see if you can get to the other forms. Betas must be regression coefficients.) Can you run the GRS test on a model that uses industrial production growth as a factor, E(Ri ) − Rf = β i,∆ip λip ? Fama and French (1997b) report that pricing errors are correlated with betas in a test of a factor pricing model on industry portfolios. How is this possible? We saw that a GLS cross-sectional regression of the CAPM passes through the market and riskfree rate by construction. Show that if the market return is an equally weighted portfolio of the test assets, then an OLS cross-sectional regression with an estimated intercept passes through the market return by construction. Does it also pass through the riskfree rate or origin?

234

Chapter 13. GMM for linear factor models in discount factor form 13.1

GMM on the pricing errors gives a cross-sectional regression

The first stage estimate is an OLS cross-sectional regression, and the second stage is a GLS regression, First stage : Second stage :

ˆb1 = (d0 d)−1 d0 ET (p) ˆb2 = (d0 S −1 d)d0 S −1 E(p).

Standard errors are the corresponding regression formulas, and the variance of the pricing errors are the standard regression formula for variance of a residual. Treating the constant a × 1 as a constant factor, the model is m = b0 f E(p) = E(mx).

or simply E(p) = E(xf 0 )b.

(189)

Keep in mind that p and x are N × 1 vectors of asset prices and payoffs respectively; f is a K × 1 vector of factors, and b is a K × 1 vector of parameters. I suppress the time indices mt+1 , ft+1 , xt+1, pt . The payoffs are typically returns or excess returns, including returns scaled by instruments. The prices are typically one (returns) zero (excess returns) or instruments. To implement GMM, we need to choose a set of moments. The obvious set of moments to use are the pricing errors, gT (b) = ET (xf 0 b − p).

This choice is natural but not necessary. You don’t have to use p = E(mx) with GMM, and you don’t have to use GMM with p = E(mx). You can (we will) use GMM on expected return-beta models, and you can use maximum likelihood on p = E(mx). It is a choice, and the results will depend on this choice of moments as well as the specification of the model. 235

C HAPTER 13

GMM FOR LINEAR FACTOR MODELS IN DISCOUNT FACTOR FORM

The GMM estimate is formed from min gT (b)0 W gT (b) b

with first order condition d0 W gT (b) = d0 W ET (xf 0 b − p) = 0

where d0 =

∂gT0 (b) = ET (f x0 ). ∂b

This is the second moment matrix of payoffs and factors. The first stage has W = I , the second stage has W = S −1 . Since this is a linear model, we can solve analytically for the GMM estimate, and it is First stage : Second stage :

ˆb1 = (d0 d)−1 d0 ET (p) ˆb2 = (d0 S −1 d)d0 S −1 ET (p).

The first stage estimate is an OLS cross-sectional regression of average prices on the second moment of payoff with factors, and the second stage estimate is a GLS cross-sectional regression. What could be more sensible? The model (13.189) says that average prices should be a linear function of the second moment of payoff with factors, so the estimate runs a linear regression. These are cross-sectional regressions since they operate across assets on sample averages. The “data points” in the regression are sample average prices (y) and second moments of payoffs with factors (x) across test assets. We are picking the parameter b to make the model fit explain the cross-section of asset prices as well as possible.

We find the distribution theory from the standard GMM standard error formulas (11.144) and (11.150). In the first stage, a = d0 . First stage : Second stage :

1 cov(ˆb1 ) = (d0 d)−1 d0 Sd(d0 d)−1 T 1 cov(ˆb2 ) = (d0 S −1 d)−1 . T

(13.190)

Unsurprisingly, these are exactly the formulas for OLS and GLS regression errors with error covariance S . The pricing errors are correlated across assets, since the payoffs are correlated. Therefore the OLS cross-sectional regression standard errors need to be corrected for correlation, as they are in (13.190) and one can pursue an efficient estimate as in GLS. The analogy is GLS is close, since S is the covariance matrix of E(p) − E(xf 0 )b; S is the covariance matrix of the “errors” in the cross-sectional regression. 236

S ECTION 13.2

T HE CASE OF EXCESS RETURNS

The covariance matrix of the pricing errors is, from (11.147), (11.151) and (11.152) h i ¡ ¢ ¡ ¢ First stage : T cov gT (ˆb) = I − d(d0 d)−1 d0 S I − d(d0 d)−1 d0 (13.191) h i Second stage : T cov gT (ˆb) = S − d(d0 S −1 d)−1 d0 .

These are obvious analogues to the standard regression formulas for the covariance matrix of regression residuals. The model test gT (b)0 cov(gT )−1 gT (b) ∼ χ2 (#moments − #parameters)

which specializes for the second-stage estimate as T gT (ˆb)0 S −1 gT (ˆb) ∼ χ2 (#moments − #parameters).

There is not much point in writing these out, other than to point out that the test is a quadratic form in the vector of pricing errors. It turns out that the χ2 test has the same value for first and second stage for this model, even though the parameter estimates, pricing errors and covariance matrix are not the same.

13.2

The case of excess returns

When mt+1 = a − b0 ft+1 and the test assets are excess returns, the GMM estimate is a GLS cross-sectional regression of average returns on the second moments of returns with factors, First stage : Second stage :

ˆb1 = (d0 d)−1 d0 ET (Re ) ˆb2 = (d0 S −1 d)d0 S −1 ET (Re ).

where d is the covariance matrix between returns and factors. The other formulas are the same. The analysis of the last section requires that at least one asset has a nonzero price. If all assets are excess returns then ˆb1 = (d0 d)−1 d0 ET (p) = 0. Linear factor models are most often applied to excess returns, so this case is important. The trouble is that in this case the mean discount factor is not identified. If E(mRe ) = 0 then E((2 × m)Re ) = 0. Analogously in expected return-beta models, if all test assets are excess returns, then we have no information on the level of the zero-beta rate. Writing out the model as m = a−b0 f , we cannot separately identify a and b so we have to choose some normalization. The choice is entirely one of convenience; lack of identification 237

C HAPTER 13

GMM FOR LINEAR FACTOR MODELS IN DISCOUNT FACTOR FORM

means precisely that the pricing errors do not depend on the choice of normalization. The easiest choice is a = 1. Then gT (b) = ET (mRe ) = ET (Re ) − E(Re f 0 )b.

We have d0 =

∂gT (b) = E(fRe0 ), ∂b0

the second moment matrix of returns and factors. The first order condition to min gT0 W gT is d0 W [d b + ET (Re )] = 0.

Then, the GMM estimates of b are ˆb1 = (d0 d)−1 d0 ET (Re ) ˆb2 = (d0 S −1 d)d0 S −1 ET (Re ).

First stage : Second stage :

The GMM estimate is a cross-sectional regression of mean excess returns on the second moments of returns with factors. From here on in, the distribution theory is unchanged from the last section. Mean returns on covariances

We can obtain a cross-sectional regression of mean excess returns on covariances, which are just a heartbeat away from betas, by choosing the normalization a = 1 + b0 E(f) rather than a = 1. Then, the model is m = 1 − b0 (f − E(f )) with mean E(m) = 1. The pricing errors are gT (b) = ET (mRe ) = ET (Re ) − ET (Re f˜0 )b

where I denote f˜ ≡ f − E(f). We have d0 =

∂gT (b) ˜ e0 ), = ET (fR ∂b0

which now denotes the covariance matrix of returns and factors. The first order condition to min gT0 W gT is now d0 W [d b + ET (Re )] = 0.

Then, the GMM estimates of b are First stage : Second stage :

ˆb1 = (d0 d)−1 d0 ET (Re ) ˆb2 = (d0 S −1 d)d0 S −1 ET (Re ).

238

S ECTION 13.3

H ORSE R ACES

The GMM estimate is a cross-sectional regression of expected excess returns on the covariance between returns and factors. Naturally, the model says that expected excess returns should be proportional to the covariance between returns and factors, and the estimate estimates that relation by a linear regression. The standard errors and variance of the pricing errors are the same as in (13.190) and (13.191), with d now representing the covariance matrix. The formulas are almost exactly identical to those of the cross-sectional regressions in section 12.2. The p = E(mx) formulation of the model for excess returns is equivalent to E(Re ) = −Cov(Re , f 0 )b; thus covariances enter in place of betas β .

There is one fly in the ointment; the mean of the factor E(f ) is estimated, and the distribution theory should recognize sampling variation induced by this fact, as we did for the fact that betas are generated regressors in the cross-sectional regressions of section 2.3. The distribution theory is straightforward, and a problem at the end of the chapter guides you through it. However, I think it is better to avoid the complication and just use the second moment approach, or some other non-sample dependent normalization for a. The pricing errors are identical – the whole point is that the normalization of a does not matter to the pricing errors. Therefore, the χ2 statistics are also identical. As you change the normalization for a, you change the estimate of b. Therefore, the only effect is to add a term in the sampling variance of the estimated parameter b.

13.3

Horse Races

How to test whether one set of factors drives out another. Test b2 = 0 in m = b01 f1 + b02 f2 using the standard error of ˆb2 , or the χ2 difference test. It’s often interesting to test whether one set of factors drives out another. For example, Chen Roll and Ross (1986) test whether their five macroeconomic factors price assets so well that one can ignore even the market return. Given the large number of factors that have been proposed, a statistical procedure for testing which factors survive in the presence of the others is desirable. In this framework, such a test is very easy. Start by estimating a general model m = b01 f1 + b02 f2 .

(192)

We want to know, given factors f1 , do we need the f2 to price assets – i.e. is b2 = 0? There are two ways to do this. First and most obviously, we have an asymptotic covariance matrix for [b1 b2 ], so we can form a t test (if b2 is scalar) or χ2 test for b2 = 0 by forming the statistic ˆb0 var(ˆb2 )−1ˆb2 ∼ χ2 2 #b2

239

C HAPTER 13

GMM FOR LINEAR FACTOR MODELS IN DISCOUNT FACTOR FORM

where #b2 is the number of elements in the b2 vector. This is a Wald test.. Second, we can estimate a restricted system m = b01 f1 . Since there are fewer free parameters and the same number of moments than in (13.192), we expect the criterion JT to rise. If we use the same weighting matrix, (usually the one estimated from the unrestricted model (13.192)) then the JT cannot in fact decline. But if b2 really is zero, it shouldn’t rise “much.” How much? The χ2 difference test answers that question; T JT (restricted) − T JT (unrestricted) ∼ χ2 (#of restrictions)

This is very much like a likelihood ratio test.

13.4

Testing for characteristics

How to check whether an asset pricing model drives out a characteristic such as size, book/market or volatility. Run cross sectional regressions of pricing errors on characteristics; use the formulas for covariance matrix of the pricing errors to create standard errors. It’s often interesting to characterize a model by checking whether the model drives out a characteristic. For example, portfolios organized by size or market capitalization show a wide dispersion in average returns (at least up to 1979). Small stocks gave higher average returns than large stocks. The size of the portfolio is a characteristic. A good asset pricing model should account for average returns by betas. It’s ok if a characteristic is associated with average returns, but in the end betas should drive out the characteristic; the alphas or pricing errors should not be associated with the characteristic. The original tests of the CAPM similarly checked whether the variance of the individual portfolio had anything to do with average returns once betas were included. Denote the characteristic of portfolio i by yi . An obvious idea is to include both betas and the characteristic in a multiple, cross-sectional regression, E(Rei ) = (α0 ) + β 0i λ + γyi + εi ; i = 1, 2, ...N

Alternatively, subtract βλ from both sides and consider a cross-sectional regression of alphas on the characteristic, αi = (α0 ) + γyi + εi ; i = 1, 2, ...N.

(The difference is whether you allow the presence of the size characteristic to affect the λ estimate or not.) We can always run such a regression, but we don’t want to use the OLS formulas for the sampling error of the estimates, since the errors εi are correlated across assets. Under the null that γ = 0, ε = α, so we can simply use the covariance matrix of the alphas to generate 240

S ECTION 13.5

T ESTING FOR PRICED FACTORS :

LAMBDAS OR B ’ S ?

standard errors of the γ . Let X denote the vector of characteristics, then the estimate is γˆ = (X 0 X)−1 X 0 α ˆ

with standard error σ(ˆ γ ) = (X 0 X)−1 X 0 cov(ˆ α)X(X 0 X)−1

At this point, simply use the formula for cov(ˆ α) or cov(gT ) as appropriate for the model that you tested. Sometimes, the characteristic is also estimated rather than being a fixed number such as the size rank of a size portfolio, and you’d like to include the sampling uncertainty of its estimation in the standard errors of γˆ. Let yti denote the time series whose mean E(yti ) determines the characteristic. Now, write the moment condition for the ith asset as gT = ET (mt+1 (b)xt+1 − pt − γyti ).

The estimate of γ tells you how the characteristic E(yi ) is associated with model pricing errors E(mt+1 (b)xt+1 − pt ). The GMM estimate of γ is E(y)0 W (E(mx) − p − γy) −1

γˆ = (ET0 (y)W ET (y))

ET0 (y)W gT

a OLS or GLS regression of the pricing errors on the estimated characteristics. The standard GMM formulas for the standard deviation of γ or the χ2 difference test for γ = 0 tell you whether the γ estimate is statistically significant, including the fact that E(y) must be estimated.

13.5

Testing for priced factors: lambdas or b’s?

bj asks whether factor j helps to price assets given the other factors. bj gives the multiple regression coefficient of m on fj given the other factors. λj asks whether factor j is priced, or whether its factor-mimicking portfolio carries a positive risk premium. λj gives the single regression coefficient of m on fj .

Therefore, when factors are correlated, one should test bj = 0 to see whether to include factor j given the other factors rather than test λj = 0. Expected return-beta models defined with single regression betas give rise to λ with multiple regression interpretation that one can use to test factor pricing.

241

C HAPTER 13

GMM FOR LINEAR FACTOR MODELS IN DISCOUNT FACTOR FORM

In the context of expected return-beta models, it has been more traditional to evaluate the relative strengths of models by testing the factor risk premia λ of additional factors, rather than test whether their b is zero. (The b’s are not the same as the β ’s. b are the regression coefficient of m on f, β are the regression coefficients of Ri on f.) To keep the equations simple, I’ll use mean-zero factors, excess returns, and normalize to E(m) = 1, since the mean of m is not identified with excess returns. The parameters b and λ are related by λ = E(f f 0 )b.

See section 6.3. Briefly, 0 = E(mRe ) = E [Re (1 − f 0 b)] E(Re ) = cov(Re , f 0 )b = cov(Re , f 0 )E(f f 0 )−1 E(f f 0 )b = β 0 λ.

Thus, when the factors are orthogonal, E(f f 0 ) is diagonal, and each λj = 0 if and only if the corresponding bj = 0. The distinction between b and λ only matters when the factors are correlated. Factors are often correlated however. λj captures whether factor fj is priced. We can write λ = E [f (f 0 b)] = −E(mf) to see that λ is (the negative of) the price that the discount factor m assigns to f . b captures whether factor fj is marginally useful in pricing assets, given the presence of other factors. If bj = 0, we can price assets just as well without factor fj as with it. λj is proportional to the single regression coefficient of m on f . λj = cov(m, fj ). λj = 0 asks the corresponding single regression coefficient question—“is factor j correlated with the true discount factor?” bj is the multiple regression coefficient of m on fj given all the other factors. This just follows from m = b0 f . (Regressions don’t have to have error terms!) A multiple regression coefficient β j in y = xβ + ε is the way to answer “does xj help to explain variation in y given the presence of the other x’s?” When you want to ask the question, “should I include factor j given the other factors?” you want to ask the multiple regression question.

For example, suppose the CAPM is true, which is the single factor model m = a − bRem

where Rem is the market excess return. Consider any other excess return Rex , positively correlated with Rem (x for extra). If we try a factor model with the spurious factor Rex , the answer is m = a − bRem + 0 × Rex . bx is obviously zero, indicating that adding this factor does not help to price assets.

However, since the correlation of Rex with Rem is positive, the beta of Rex on Rem is positive, Rex earns a positive expected excess return, and λx = E(Rex ) > 0. In the expected 242

S ECTION 13.5

T ESTING FOR PRICED FACTORS :

LAMBDAS OR B ’ S ?

return - beta model E(Rei ) = β im λm + β ix λx λm = E(Rem ) is unchanged by the addition of the spurious factor. However, since the factors Rem , Rex are correlated, the multiple regression betas of Rei on the factors change when we add the extra factor Rex . If β ix is positive, β im will decline from its single-regression value, so the new model explains the same expected return E(Rei ). The expected return beta model will indicate a risk premium for β x exposure, and many assets will have β x exposure (Rx for example!) even though factor Rx is spurious. In particular, Rex will of course have multiple regression coefficients β x,m = 0 and β x,x = 1, and its expected return will be entirely explained by the new factor x.

So, as usual, the answer depends on the question. If you want to know whether factor i is priced, look at λ (or E(mf i )). If you want to know whether factor i helps to price other assets, look at bi . This is not an issue about sampling error or testing. All moments above are population values. Of course, testing b = 0 is particularly easy in the GMM, p = E(mx) setup. But you can always test the same ideas in any expression of the model. In an expected return-beta model, estimate b by E(ff 0 )−1 λ and test the elements of that vector rather than λ itself. You can write an asset pricing model as ERe = β 0 λ and use the λ to test whether each factor can be dropped in the presence of the others, if you use single regression betas rather than multiple regression betas. In this case each λ is proportional to the corresponding b. Problem 2 at the end of this chapter helps you to work out this case. 13.5.1

Mean-variance frontier and performance evaluation

A GMM, p = E(mx) approach to testing whether a return expands the mean-variance frontier. Just test whether m = a + bR prices all returns. If there is no risk free rate, use two values of a. We often summarize asset return data by mean-variance frontiers. For example, a large literature has examined the desirability of international diversification in a mean-variance context. Stock returns from many countries are not perfectly correlated, so it looks like one can reduce portfolio variance a great deal for the same mean return by holding an internationally diversified portfolio. But is this real or just sampling error? Even if the value-weighted portfolio were ex-ante mean-variance efficient, an ex-post mean-variance frontier constructed from historical returns on the roughly NYSE stocks would leave the value-weighted portfolio well inside the ex-post frontier. So is “I should have bought Japanese stocks in 1960” (and sold them in 1990!) a signal that broad-based international diversification a good idea now, or is it simply 20/20 hindsight regret like “I should have bought Microsoft in 1982?” Sim243

C HAPTER 13

GMM FOR LINEAR FACTOR MODELS IN DISCOUNT FACTOR FORM

Frontiers intersect E(R)

1/E(m)

σ(R) Figure 27. Mean variance frontiers might intersect rather than coincide. ilarly, when evaluating fund managers, we want to know whether the manager is truly able to form a portfolio that beats mean-variance efficient passive portfolios, or whether better performance in sample is just due to luck. Since a factor model is true if and only if a linear combination of the factors (or factormimicking portfolios if the factors are not returns) is mean-variance efficient, one can interpret a test of any factor pricing model as a test whether a given return is on the mean-variance frontier. Section 12.1 showed how the Gibbons Ross and Shanken pricing error statistic can be interpreted as a test whether a given portfolio is on the mean-variance frontier, when returns and factors are i.i.d., and the GMM distribution theory of that test statistic allows us to extend the test to non-i.i.d. errors. A GMM, p = E(mx), m = a − bRp test analogously tests whether Rp is on the mean-variance frontier of the test assets. We may want to go one step further, and not just test whether a combination of a set of assets Rd (say, domestic assets) is on the mean-variance frontier, but whether the Rd assets span the mean-variance frontier of Rd and Ri (say, foreign or international) assets. The trouble is, that if there is no riskfree rate, the frontier generated by Rd might just intersect the frontier generated by Rd and Ri together, rather than span or coincide with the latter frontier, as shown in Figure 27. Testing that m = a − b0 Rd prices both Rd and Ri only checks for intersection.

244

S ECTION 13.6

P ROBLEMS

DeSantis (1992) and Chen and Knez (1992,1993) show how to test for spanning as opposed to intersection. For intersection, m = a − b0d Rd will price both Rd and Rf only for one value of a, or equivalently E(m) or choice of the intercept, as shown. If the frontiers coincide or span, then m = a + b0d Rd prices both Rd and Rf for any value of a. Thus,we can test for coincident frontiers by testing whether m = a + b0d Rd prices both Rd and Rf for two prespecified values of a simultaneously. To see how this work, start by noting that there must be at least two assets in Rd . If not, there is no mean-variance frontier of Rd assets; it is simply a point. If there are two assets in Rd ,Rd1 and Rd2 , then the mean-variance frontier of domestic assets connects them; they are each on the frontier. If they are both on the frontier, then there must be discount factors m1 = a1 − ˜b1 Rd1

and m2 = a2 − ˜b2 Rd2

and, of course, any linear combination, i ¤ h £ m = λa1 + (1 − λ)a2 − λ˜b1 Rd1 + (1 − λ)˜b2 Rd2 .

Equivalently, for any value of a, there is a discount factor of the form ¢ ¡ m = a − b1 Rd1 + b2 Rd2 . Thus, you can test for spanning with a JT test on the moments ¤ £ E (a1 − b10 Rd )Rd = 0 £ ¤ E (a1 − b10 Rd )Ri = 0 ¤ £ E (a2 − b20 Rd )Rd = 0 ¤ £ E (a2 − b20 Rd )Ri = 0

for any two fixed values of a1 , a2 .

13.6 1.

Problems

Work out the GMM distribution theory for the model m = 1 − b0 (f − E(f )) and test assets are excess returns. The distribution should recognize the fact that E(f ) is estimated in sample. To do this, set up · ¸ ET (Re − Re (f 0 − Ef 0 ) b) gT = ET (f − Ef ) 245

C HAPTER 13

GMM FOR LINEAR FACTOR MODELS IN DISCOUNT FACTOR FORM

aT =

2.

"

³ ´ # ˜ e0 ET fR 0 . 0 IK

The estimated parameters are b, E(f). You should end up with a formula for the standard error of b that resembles the Shanken correction (12.184), and an unchanged JT test. Show that if you use single regression betas, then the corresponding λ can be used to test for the marginal importance of factors. However, the λ are no longer the expected return of factor mimicking portfolios.

246

Chapter 14.

Maximum likelihood

Maximum likelihood is, like GMM, a general organizing principle that is a useful place to start when thinking about how to choose parameters and evaluate a model. It comes with an asymptotic distribution theory, which, like GMM, is a good place to start when you are unsure about how to treat various problems such as the fact that betas must be estimated in a cross-sectional regression. As we will see, maximum likelihood is a special case of GMM. Given a statistical description of the data, it prescribes which moments are statistically most informative. Given those moments, ML and GMM are the same. Thus, ML can be used to defend why one picks a certain set of moments, or for advice on which moments to pick if one is unsure. In this sense, maximum likelihood (paired with carefully chosen statistical models) justifies the regression tests above, as it justifies standard regressions. On the other hand, ML does not easily allow you to use other non-“efficient” moments, if you suspect that ML’s choices are not robust to misspecifications of the economic or statistical model. For example, ML will tell you how to do GLS, but it will not tell you how to adjust OLS standard errors for non-standard error terms. Hamilton (1994) p.142-148 and the appendix in Campbell Lo MacKinlay (1997) give nice summaries of maximum likelihood theory. Campbell Lo and MacKinlay’s Chapter 5 and 6 treat many more variations of regression based tests and maximum likelihood.

14.1

Maximum likelihood

The maximum likelihood principle says to pick the parameters that make the observed data most likely. Maximum likelihood estimates are asymptotically efficient. The information matrix gives the asymptotic standard errors of ML estimates. The maximum likelihood principle says to pick that set of parameters that makes the observed data most likely. This is not “the set of parameters that are most likely given the data” – in classical (as opposed to Bayesian) statistics, parameters are numbers, not random variables. To implement this idea, you first have to figure out what the probability of seeing a data set {xt } is, given the free parameters θ of a model. This probability distribution is called the likelihood function f({xt } ; θ). Then, the maximum likelihood principle says to pick ˆθ = arg max f ({xt } ; θ). {θ}

For reasons that will soon be obvious, it’s much easier to work with the log of this probability 247

C HAPTER 14

MAXIMUM LIKELIHOOD

distribution L({xt } ; θ) = ln f ({xt } ; θ),

Maximizing the log likelihood is the same thing as maximizing the likelihood. Finding the likelihood function isn’t always easy. In a time-series context, the best way to do it is often to first find the log conditional likelihood function f (xt |xt−1 , xt−2 , ...x0 ; θ), the chance of seeing xt+1 given xt , xt−1 , ... and given values for the parameters, . Since joint probability is the product of conditional probabilities, the log likelihood function is just the sum of the conditional log likelihood functions, L({xt } ; θ) =

T X t=1

ln f(xt |xt−1 , xt−2 ...x0 ; θ).

(193)

More concretely, we usually assume normal errors, so the likelihood function is T

L=−

1 X 0 −1 T ln (2π |Σ|) − ε Σ εt 2 2 t=1 t

(194)

where εt denotes a vector of shocks; εt = xt − E(xt |xt−1 , xt−2 ...x0 ; θ).

This expression gives a simple recipe for constructing a likelihood function. You usually start with a model that generates xt from errors, e.g. xt = ρxt−1 + εt . Invert that model to express the errors εt in terms of the data {xt } and plug in to (14.194). There is a small issue about how to start off a model such as (14.193). Ideally, the first observation should be the unconditional density, i.e. L({xt } ; θ) = ln f (x1 ; θ) + ln f (x2 |x1 ; θ) + ln f (x3 |x2 , x1 ; θ)...

However, it is usually hard to evaluate the unconditional density or the first terms with only a few lagged xs. Therefore, if as usual the conditional density can be expressed in terms of a finite number k of lags of xt , one often maximizes the conditional likelihood function (conditional on the first k observations), treating the first k observations as fixed rather than random variables. L({xt } ; θ) = ln f (xk+1 |xk , xk−1 ...x1 ; θ) + ln f(xk+2 |xk , xk−1... x2 ; θ) + ...

Alternatively, one can treat k pre-sample values {x0 , x−1 , ...x−k+1 } as additional parameters over which to maximize the likelihood function. Maximum likelihood estimators come with a useful asymptotic (i.e. approximate) distri248

S ECTION 14.2

ML IS GMM ON THE SCORES

bution theory. First, the distribution of the estimates is ˆθ∼N

Ã · ¸−1 ! ∂ 2L θ, − ∂θ∂θ0

(195)

If the likelihood L has a sharp peak at ˆθ, then we know a lot about the parameters, while if the peak is flat, other parameters are just as plausible. The maximum likelihood estimator is asymptotically efficient meaning that no other estimator can produce a smaller covariance matrix. The second derivative in (14.195) is known as the information matrix, I=−

T 1 ∂ 2L 1 X ∂ 2 ln f (xt+1 |xt , xt−1 , ...x0 ; θ) = − . T ∂θ∂θ0 T t=1 ∂θ∂θ0

(196)

(More precisely, the information matrix is defined as the expected value of the second partial, which is estimated with the sample value.) The information matrix can also be estimated as a product of first derivatives. The expression ¶µ ¶0 T µ 1 X ∂ ln f (xt+1 |xt , xt−1 , ...x0 ; θ) ∂ ln f(xt+1 |xt , xt−1 , ...x0 ; θ) I =− . T t=1 ∂θ ∂θ

converges to the same value as (14.196). (Hamilton 1994 p.429 gives a proof.) If we estimate a model restricting the parameters, the maximum value of the likelihood function will necessarily be lower. However, if the restriction is true, it shouldn’t be that much lower. This intuition is captured in the likelihood ratio test 2(Lunrestricted − Lrestricted )∼χ2number of restrictions

(197)

The form and idea of this test is much like the χ2 difference test for GMM objectives that we met in section 11.1.

14.2

ML is GMM on the scores

ML is a special case of GMM. ML uses the information in the auxiliary statistical model to derive statistically most informative moment conditions. To see this fact, start with the first order conditions for maximizing a likelihood function T

∂L({xt } ; θ) X ∂ ln f (xt |xt−1 xt−2 ...; θ) = = 0. ∂θ ∂θ t=1

249

(198)

C HAPTER 14

MAXIMUM LIKELIHOOD

This is a GMM estimate. It is the sample counterpart to a population moment condition g(θ) = E

µ

∂ ln f (xt |xt−1 xt−2 ...; θ) ∂θ

¶

= 0.

(199)

The term ∂ ln f (xt |xt−1 xt−2 ...; θ)/∂θ is known as the “score.” It is a random variable, formed as a combination of current and past data (xt , xt−1 ...). Thus, maximum likelihood is a special case of GMM, a special choice of which moments to examine. For example, suppose that x follows an AR(1) with known variance, xt = ρxt−1 + εt ,

and suppose the error terms are i.i.d. normal random variables. Then, ln f (xt |xt−1 , xt−2 ...; ρ) = const. −

ε2t (xt − ρxt−1 )2 = const − 2σ 2 2σ 2

and the score is (xt − ρxt−1 ) xt−1 ∂ ln f (xt |xt−1 xt−2 ...; ρ) = . ∂ρ σ2

The first order condition for maximizing likelihood is T 1X (xt − ρxt−1 ) xt−1 = 0. T t=1

This expression is a moment condition, and you’ll recognize it as the OLS estimator of ρ, which we have already regarded as a case of GMM. The example shows another property of scores: The scores should be unforecastable. In the example, Et−1

·

¸ hε x i (xt − ρxt−1 ) xt−1 t t−1 = 0. = E t−1 σ2 σ2

(200)

Intuitively, if we used a combination of the x variables E(h(xt , xt−1 , ...)) = 0 that was predictable, we could form another moment – an instrument – that described the predictability of the h variable and use that moment to get more information about the parameters. To prove this property more generally, start with the fact that f (xt |xt−1 , xt−2 , ...; θ) is a conditional 250

S ECTION 14.3

WHEN FACTORS ARE RETURNS , ML PRESCRIBES A TIME - SERIES REGRESSION

density and therefore must integrate to one, Z 1 = f (xt |xt−1 , xt−2 , ...; θ)dxt Z ∂f (xt |xt−1 , xt−2 , ...; θ) dxt 0 = ∂θ Z ∂ ln f (xt |xt−1 , xt−2 , ...; θ) f (xt |xt−1 , xt−2 , ...; θ)dxt 0 = ∂θ · ¸ ∂ ln f (xt |xt−1 , xt−2 , ...; θ) 0 = Et−1 . ∂θ Furthermore, as you might expect, the GMM distribution theory formulas give the same result as the ML distribution, i.e., the information matrix is the asymptotic variance-covariance matrix. To show this fact, apply the GMM distribution theory (11.144) to (14.198). The derivative matrix is T ∂gT (θ) 1 X ∂ 2 ln f (xt |xt−1 xt−2 ...; θ) d= = =I 0 T t=1 ∂θ0 ∂θ∂θ

This is the second derivative expression of the information matrix. The S matrix is · ¸ ∂ ln f (xt |xt−1 xt−2 ...; θ) ∂ ln f (xt |xt−1 xt−2 ...; θ) 0 E =I ∂θ ∂θ The lead and lag terms in S are all zero since we showed above that scores should be unforecastable. This is the outer product definition of the information matrix. There is no a matrix, since the moments themselves are set to zero. The GMM asymptotic distribution of ˆ θ is therefore √ ¤ £ ¤ £ T (ˆθ − θ) → N 0, d−1 Sd−10 = N 0, I −1 .

We recover the inverse information matrix, as specified by the ML asymptotic distribution theory.

14.3

When factors are returns, ML prescribes a time-series regression

I add to the economic model E (Re ) = βE(f) a statistical assumption that the regression errors are independent over time and independent of the factors. ML then prescribes a timeseries regression with no constant. To prescribe a time series regression with a constant, we drop the model prediction α = 0. I show how the information matrix gives the same result as the OLS standard errors. 251

C HAPTER 14

MAXIMUM LIKELIHOOD

Given a linear factor model whose factors are also returns, as with the CAPM, ML prescribes a time-series regression test. To keep notation simple, I again treat a single factor f . The economic model is (201)

E (Re ) = βE(f )

Re is an N × 1 vector of test assets, and β is an N × 1 vector of regression coefficients of these assets on the factor (the market return Rem in the case of the CAPM).

To apply maximum likelihood, we need to add an explicit statistical model that fully describes the joint distribution of the data. I assume that the market return and regression errors are i.i.d. normal, i.e. Ret ft

(14.202)

= α + βft + εt = E(f ) + ut · ¸ µ· ¸ · ¸¶ εt 0 Σ 0 ∼N , ut 0 0 σ2u

(We can get by with non-normal factors, but it is easier not to present the general case.) Equation (14.202) has no content other than normality. The zero correlation between ut and εt identifies β as a regression coefficient. You can just write Re , Rem as a general bivariate normal, and you will get the same results. The economic model (14.201) implies restrictions on this statistical model. Taking expectations of (14.202), the CAPM implies that the intercepts α should all be zero. Again, this is also the only restriction that the CAPM places on the statistical model (14.202). The most principled way to apply maximum likelihood is to impose the null hypothesis throughout. Thus, we write the likelihood function imposing α = 0. To construct the likelihood function, we reduce the statistical model to independent error terms, and then add their log probability densities to get the likelihood function. T

L = (const.) −

T

1X e 1 X (ft − E(f ))2 (Rt − βft )0 Σ−1 (Ret − βft ) − 2 t=1 2 t=1 σ 2u

The estimates follow from the first order conditions, ∂L ∂β ∂L ∂E(f )

−1

= Σ

T X t=1

=

(Ret

ˆ= − βft ) ft = 0 ⇒ β

Ã T X t=1

ft2

!−1

T X

Ret ft

t=1

T T 1 X 1X [ ˆ (f − E(f)) = 0 ⇒ E(f) = λ = ft t σ 2u t=1 T t=1

(∂L/∂Σ and ∂L/∂σ2 also produce ML estimates of the covariance matrices, which turn out 252

S ECTION 14.3

WHEN FACTORS ARE RETURNS , ML PRESCRIBES A TIME - SERIES REGRESSION

to be the standard averages of squared residuals.) The ML estimate of β is the OLS regression without a constant. The null hypothesis says to leave out the constant, and the ML estimator uses that fact to avoid estimating a constant. Since the factor risk premium is equal to the market return, it’s not too surprising that the λ estimate is the same as that of the average market return. We know that the ML distribution theory must give the same result as the GMM distribution theory which we already derived in section 12.1, but it’s worth seeing it explicitly. The asymptotic standard errors follow from either estimate of the information matrix, for example T X ∂ 2L −1 = −Σ ft2 = 0 ∂β∂β 0 t=1

Thus, ˆ = cov(β)

1 1 1 1 Σ= Σ. T E(f 2 ) T E(f )2 + σ2 (f )

(203)

This is the standard OLS formula. We also want pricing error measurements, standard errors and tests. We can apply maximum likelihood to estimate an unconstrained model, containing intercepts, and then use Wald tests (estimate/standard error) to test the restriction that the intercepts are zero. We can also use the unconstrained model to run the likelihood ratio test. The unconstrained likelihood function is T

1X e L = (const.) − (R − α − βft )0 Σ−1 (Rte − α − βft ) + ... 2 t=1 t

(I ignore the term in the factor, since it will again just tell us to use the sample mean to estimate the factor risk premium.) The estimates are now ∂L ∂α

= Σ−1

∂L ∂β

= Σ−1

T X t=1

ˆ T (ft ) (Ret − α − βft ) = 0 ⇒ α ˆ = ET (Rte ) − βE

T X

e ˆ = covT (Rt , ft ) (Ret − α − βft ) ft = 0 ⇒ β σ2T (ft ) t=1

Unsurprisingly, the maximum likelihood estimates of α and β are the OLS estimates, with a constant. The inverse of the information matrix gives the asymptotic distribution of these estimates. Since they are just OLS estimates, we’re going to get the OLS standard errors, but it’s worth 253

C HAPTER 14 seeing it come out of ML.

MAXIMUM LIKELIHOOD

−1

∂ 2L · ¸ − £ ¤ α ∂ ∂ α β β

=

·

=

1 2 σ (f)

Σ−1 Σ−1 E(f ) −1 Σ E(f ) Σ−1 E(f 2 )

¸−1

¸

⊗Σ

·

E(f 2 ) E(f ) E(f ) 1

ˆ are thus The covariance matrices of α ˆ and β cov(ˆ α) = ˆ = cov(β)

" µ ¶2 # E(f ) 1 1+ Σ T σ(f ) 1 1 Σ. T σ 2 (f )

(14.204)

These are just the usual OLS standard errors, which we derived in section 12.1 as a special case of GMM standard errors for the OLS time-series regressions when errors are uncorrelated over time and independent of the factors, or by specializing σ2 (X 0 X)−1 . You cannot just invert ∂ 2 L/∂α∂α0 to find the covariance of α ˆ . That attempt would give ˆ , which would be wrong. You have to invert the entire just Σ as the covariance matrix of α information matrix to get the standard error of any parameter. Otherwise, you are ignoring ˆ . In fact, what I presented is really the effect that estimating β has on the distribution of α ˆ is independent of α ˆ and wrong, since we also must estimate Σ. However, it turns out that Σ ˆ β – the information matrix is block-diagonal – so the top left two elements of the true inverse information matrix are the same as I have written here. The variance of βˆ in (14.204) is larger than it is in (14.203) was when we impose the null of no constant. ML uses all the information it can to produce efficient estimates – estimates with the smallest possible covariance matrix. The ratio of the two formulas is equal to the familiar term 1 + E(f )2 /σ2 (f ). In annual data for the CAPM, σ(Rem ) = 16%, E(Rem ) = 8%, means that unrestricted estimate (14.204) has a variance 25% larger than the restricted estimate (14.203), so the gain in efficiency can be important. In monthly data, however the gain is smaller since variance and mean both scale with the horizon. We can also view this fact as a warning: ML will ruthlessly exploit the null hypothesis and do things like running regressions without a constant in order to get any small improvement in efficiency. We can use these covariance matrices to construct a Wald (estimate/standard error) test the restriction of the model that the alphas are all zero, Ã µ ¶2 !−1 E(f) T 1+ α ˆ 0 Σ−1 α ˆ ∼χ2N . (205) σ(f) 254

S ECTION 14.4

WHEN FACTORS ARE NOT EXCESS RETURNS , ML PRESCRIBES A CROSS - SECTIONAL REGRESSION

Again, we already derived this χ2 test in (12.168), and its finite sample F counterpart, the GRS F test (12.169). The other test of the restrictions is the likelihood ratio test (14.197). Quite generally, likelihood ratio tests are asymptotically equivalent to Wald tests, and so gives the same result. Showing it in this case is not worth the algebra.

14.4

When factors are not excess returns, ML prescribes a cross-sectional regression

If the factors are not returns, we don’t have a choice between time-series and cross-sectional regression, since the intercepts are not zero. As you might suspect, ML prescribes a crosssectional regression in this case. The factor model, expressed in expected return beta form, is E(Rei ) = αi + β 0i λ; i = 1, 2, ..N

(206)

The betas are defined from time-series regressions 0 i Rei t = ai + β i ft + εt

(207)

The intercepts ai in the time-series regressions need not be zero, since the model does not apply to the factors. They are not unrestricted, however. Taking expectations of the timeseries regression (14.207) and comparing it to (14.206) (as we did to derive the restriction α = 0 for the time-series regression), the restriction α = 0 implies ai = β 0i (λ − E(ft ))

(208)

Plugging into (14.207), the time series regressions must be of the restricted form Rtei = β 0i λ + β 0i [ft − E(ft )] + εit .

(209)

In this form, you can see that β 0i λ determines the mean return. Since there are fewer factors than returns, this is a restriction on the regression (14.209). Stack assets i = 1, 2, ...N to a vector; and introduce the auxiliary statistical model that the errors and factors are i.i.d. normal and uncorrelated with each other. Then, the restricted model is Rte = Bλ + B [ft − E(ft )] + εt ft = E(f) + ut · ¸ µ ¶ εt Σ 0 ∼ N 0, ut 0 V

where B denotes a N × K matrix of regression coefficients of the N assets on the K factors. 255

C HAPTER 14

MAXIMUM LIKELIHOOD

The likelihood function is T

L = (const.) − εt

T

1 X 0 −1 1 X 0 −1 εt Σ εt − u V ut 2 t=1 2 t=1 t

= Rte − B [λ + ft − E(f)] ; ut = ft − E(f).

Maximizing the likelihood function, ∂L ∂E(f ) ∂L ∂λ

: 0=

T X

0

BΣ

t=1

: 0 = B0

−1

T X t=1

(Ret −B [λ

+ ft − E(f )]) +

T X t=1

V −1 (ft − E(f))

Σ−1 (Ret − B [λ + ft − E(f )])

The solution to this pair of equations is [) = ET (ft ) E(f ¡ ¢ ˆ = B 0 Σ−1 B −1 B 0 Σ−1 ET (Re ) . λ t

(14.210) (14.211)

The maximum likelihood estimate of the factor risk premium is a GLS cross-sectional regression of average returns on betas.

The maximum likelihood estimates of the regression coefficients B are again not the same as the standard OLS formulas. Again, ML imposes the null to improve efficiency. ∂L ∂B ˆ B

:

T X t=1

=

T X

Σ−1 (Rte −B [λ + ft − E(f)]) [λ + ft − E(f )]0 = 0 Ret

t=1

[ft + λ − E(f)]

0

Ã T X t=1

(14.212) 0

[ft + λ − E(f)] [ft + λ − E(f )]

!−1

This is true, even though the B are defined in the theory as population regression coefficients. (The matrix notation hides a lot here! If you want to rederive thesePformulas, it’s helpful to start with scalar parameters, e.g. Bij , and to think of it as ∂L/∂θ = Tt=1 (∂L/∂εt )0 ∂εt /∂θ. ) Therefore, to really implement ML, you have to solve (14.211) and (14.212) simultaneously ˆ, B ˆ , along with Σ ˆ whose ML estimate is the usual second moment matrix of the residfor λ ˆ , run an OLS cross-sectional uals. This can usually be done iteratively: Start with OLS B ˆ , form Σ ˆ , and iterate. regression for λ

14.5

Problems

256

S ECTION 14.5 1.

2.

P ROBLEMS

Why do we use restricted ML when the factor is a return, but unrestricted ML when the factor is not a return? To see why, try to formulate a ML estimator based on an unrestricted regression when factors are not returns, equation (12.166). Add pricing errors αi to the regression as we did for the unrestricted regression in the case that factors are returns, and then find ML estimators for B, λ, α, E(f ). (Treat V and Σ as known to make the problem easier.) Instead of writing a regression, build up the ML for the CAPM a little more formally. Write the statistical model as just the assumption that individual returns and the market return are jointly normal, ¸ · ¸ µ· ¶ Re E(Re ) Σ cov(Rem , Re0 ) , ∼N Rem σ2m E(Rem ) cov(Rem , Re ) The model’s restriction is E(Re ) = γcov(Rem , Re ).

Estimate γ and show that this is the same time-series estimator as we derived by presupposing a regression.

257

Chapter 15. Time series, cross-section, and GMM/DF tests of linear factor models The GMM/DF, time-series and cross-sectional regression procedures and distribution theory are similar, but not identical. Cross-sectional regressions on betas are not the same thing as cross sectional regressions on second moments. Cross-sectional regressions weighted by the residual covariance matrix are not the same thing as cross-sectional regressions weighted by the spectral density matrix. GLS cross-sectional regressions and second stage GMM have a theoretical efficiency advantage over OLS cross sectional regressions and first stage GMM, but how important is this advantage, and is it outweighed by worse finite-sample performance? The time-series regression, as ML estimate, has a potential gain in efficiency when returns are factors and the residuals are i.i.d. normal. Why does ML prescribe a time-series regression when the return is a factor and a cross-sectional regression when the return is not a factor? The time-series regression seems to ignore pricing errors and estimate the model by entirely different moments. How does adding one test asset make such a seemingly dramatic difference to the procedure? Finally, and perhaps most importantly, the GMM/discount factor approach is still a “new” procedure. Many authors still do not trust it. It is important to verify that it produces similar results and well-behaved test statistics in the setups of the classic regression tests. To address these questions, I first apply the various methods to a classic empirical question. How do time-series regression, cross-sectional regression and GMM/stochastic discount factor compare when applied to a test of the CAPM on CRSP size portfolios? I find that three methods produce almost exactly the same results for this classic exercise. They produce almost exactly the same estimates, standard errors, t-statistics and χ2 statistics that the pricing errors are jointly zero. Then I conduct a Monte Carlo and Bootstrap evaluation. Again, I find little difference between the methods. The estimates, standard errors, and size and power of tests are almost identical across methods. The Bootstrap does reveal that the traditional i.i.d. assumption generates χ2 statistics with about 1/2 the correct size – they reject half as often as they should under the null. Simple GMM corrections to the distribution theory repair this size defect. Also, you can ruin any estimate and test with a bad spectral density matrix estimate. I try an estimate with 24 lags and no Newey-West weights. It is singular in the data sample and many Monte Carlo replications. Interestingly, this singularity has minor effects on standard errors, but causes disasters when you use the spectral density matrix to weight a second-stage GMM. I also find that second stage “efficient” GMM is only very slightly more efficient than first stage GMM, but is somewhat less robust; it is more sensitive to the poor spectral density matrix and its asymptotic standard errors can be slightly misleading. As OLS is often better 258

S ECTION 15.1

T HREE APPROACHES TO THE CAPM IN SIZE PORTFOLIOS

than GLS, despite the theoretical efficiency advantage of GLS, first-stage GMM may be better than second stage GMM in many applications. This section should give comfort that the apparently “new” GMM/discount factor formulation is almost exactly the same as traditional methods in the traditional setup. There is a widespread impression that GMM has difficulty in small samples. The literature on the small sample properties of GMM (for example, Ferson and Foerster, 1994, Fuhrer, Moore, and Schuh, 1995) naturally tries hard setups, with highly nonlinear models, highly persistent and heteroskedastic errors, conditioning information, potentially weak instruments and so forth. Nobody would write a paper trying GMM in a simple situation such as this one, correctly foreseeing that the answer would not be very interesting. Unfortunately, many readers take from this literature a mistaken impression that GMM always has difficulty in finite samples, even in very standard setups. This is not the case. The point of the GMM/discount factor method, of course, is not a new way to handle the simple i.i.d. normal CAPM problems, which are already handled efficiently by regression techniques. The point of the GMM/discount factor method is its ability to transparently handle situations that are very hard with expected return - beta models and ML techniques, including the incorporation of conditioning information and nonlinear models. With the reassurance of this section, we can proceed to those more exiting applications. Cochrane (2000) presents a more in-depth analysis, including estimation and Monte Carlo evaluation of individual pricing error estimates and tests. Jagannathan and Wang (2000) compare the GMM/discount factor approach to classic regression tests analytically. They show that the parameter estimates, standard errors and χ2 statistics are asymptotically identical to those of an expected return- beta cross-sectional regression when the factor is not a return.

15.1

Three approaches to the CAPM in size portfolios

The time-series approach sends the expected return - beta line through the market return, ignoring other assets. The OLS cross -sectional regression minimizes the sum of squared pricing errors, so allows some market pricing error to fit other assets better. The GLS crosssectional regression weights pricing errors by the residual covariance matrix, so reduces to the time-series regression when the factor is a return and is included in the test assets. The GMM/discount factor estimates, standard errors and χ2 statistics are very close to time-series and cross-sectional regression estimates in this classic setup. Time series and cross section

Figures 28 and 29 illustrate the difference between time-series and cross-sectional regressions, in an evaluation of the CAPM on monthly size portfolios.

259

C HAPTER 15

T IME SERIES , CROSS - SECTION , AND GMM/DF TESTS OF LINEAR FACTOR MODELS

Figure 28 presents the time-series regression. The time-series regression estimates the factor risk premium from the average of the factor, ignoring any information in the other ˆ = ET (Rem ). Thus, a time-series regression draws the expected return-beta line assets, λ across assets by making it fit precisely on two points, the market return and the riskfree rate– The market and riskfree rate have zero estimated pricing error in every sample. (The far right portfolios are the smallest firm portfolios, and their positive pricing errors are the small firm anomaly – this data set is the first serious failure of the CAPM. I come back to the substantive issue in Chapter 20.) The time-series regression is the ML estimator in this case, since the factor is a return. Why does ML ignore all the information in the test asset average returns, and estimate the factor premium from the average factor return only? The answer lies in the structure that we told ML to assume when looking at the data. When we write Re = a + βft + εt and ε independent of f , we tell ML that a sample of returns already includes the same sample of the factor, plus extra noise. Thus, the sample of test asset returns cannot possibly tell ML anything more than the sample of the factor alone about the mean of the factor. Second, we tell ML that the factor risk premium equals the mean of the factor, so it may not consider the possibility that the two are different in trying to match the data.

Figure 28. Average excess returns vs. betas on CRSP size portfolios, 1926-1998. The line gives the predicted average return from the time-series regression, E(Re ) = βE(Rem ). The OLS cross-sectional regression in Figure 29 draws the expected return-beta line by 260

S ECTION 15.1

T HREE APPROACHES TO THE CAPM IN SIZE PORTFOLIOS

Figure 29. Average excess returns vs betas of CRSP size portfolios 1926-1998, and the fit of cross-sectional regressions. minimizing the squared pricing error across all assets. Therefore, it allows some pricing error for the market return, if by doing so the pricing errors on other assets can be reduced. Thus, the OLS cross-sectional regression gives some pricing error to the market return in order to lower the pricing errors of the other portfolios. When the factor is not also a return, ML prescribes a cross-sectional regression. ML still [ = ET (ft ). ignores anything but the factor data in estimating the mean of the factor–E(f) However, ML is now allowed to us a different parameter for the factor risk premium that fits average returns to betas, which it does by cross-sectional regression. However, ML is a GLS cross sectional regression, not an OLS cross-sectional regression. The GLS cross-sectional regression in Figure 29 is almost exactly identical to the time-series regression result – it passes right through the origin and the market return ignoring all the other pricing errors. The GLS cross-sectional regression ¡ ¢ ˆ = β 0 Σ−1 β −1 β 0 Σ−1 ET (Re ). λ

weights the various portfolios by the inverse of the residual covariance matrix Σ. If we include the market return as a test asset, it obviously has no residual variance–Rem = 0+1×Rem t t +0– so the GLS estimate pays exclusive attention to it in fitting the market line. The same thing 261

C HAPTER 15

T IME SERIES , CROSS - SECTION , AND GMM/DF TESTS OF LINEAR FACTOR MODELS

happens if the test assets span the factors – if a linear combination of the test assets is equal to the factor and hence has no residual variance. The size portfolios nearly span the market return, so the GLS cross-sectional regression is visually indistinguishable from the timeseries regression in this case. This observation wraps up one mystery, why ML seemed so different when the factor is and is not a return. As the residual variance of a portfolio goes to zero, the GLS regression pays more and more attention to that portfolio, until you have achieved the same result as a time-series regression. If we allow a free constant in the OLS cross-sectional regression, thus allowing a pricing error for the risk free rate, you can see from Figure 29 that the OLS cross-sectional regression line will fit the size portfolios even better, though allowing a pricing error in the risk free rate as well as the market return. However, a free intercept in an OLS regression on excess returns puts no weight at all on the intercept pricing error. It is a better idea to include the riskfree rate as a test asset, either directly by doing the whole thing in levels of returns rather than excess returns or by adding E(Re ) = 0, β = 0 to the cross-sectional regression. The GLS cross-sectional regression will notice that the T-bill rate has no residual variance and so will send the line right through the origin, as it does for the market return. GMM/discount factor first and second stage

Figure 30 illustrates the GMM/discount factor estimate with the same data. The horizontal axis is the second moment of returns and factors rather than beta, but you would not know it from the placement of the dots. The first stage estimate is an OLS cross-sectional regression of average returns on second moments. It minimizes the sum of squared pricing errors, and so produces pricing errors almost exactly equal to those of the OLS cross-sectional regression of returns on betas. The second stage estimate minimizes pricing errors weighted by the spectral density matrix. The spectral density matrix is not the same as the residual covariance matrix, so the second stage GMM does not go through the market portfolio as does the GLS cross-sectional regression. In fact, the slope of the line is slightly higher for the second stage estimate. (The spectral density matrix of the discount factor formulation does not reduce to the residual covariance matrix even if we assume the regression model, the asset pricing model is true, and factors and residuals are i.i.d. normal. In particular, when the market is a test asset, the GLS cross-sectional regression focuses all attention on the market portfolio but the second stage GMM/DF does not do so. The parameter b is related to λ by b = λ/E(Rem2 ). The other assets still are useful in determining the parameter b, even though Given the market return and the regression model Rtei = β i Rtem + εit , seeing the other assets does not help to determine the mean of the market return, ) Overall, the figures do not suggest any strong reason to prefer first and second stage GMM/discount factor, time-series, OLS or GLS cross sectional regression in this standard model and data set. The results are affected by the choice of method. In particular, the size of the small firm anomaly is substantially affected by how one draws the market line. But 262

S ECTION 15.1

T HREE APPROACHES TO THE CAPM IN SIZE PORTFOLIOS

Figure 30. Average excess return vs. predicted value of 10 CRSP size portfolios, 1926-1998, based on GMM/SDF estimate. The model predicts E(Re ) = bE(Re Rem ). The second stage estimate of b uses a spectral density estimate with zero lags. the graphs and analysis do not strongly suggest that any method is better than any other for purposes other than fishing for the answer one wants. Parameter estimates, standard errors, and tests

Table c1 presents the parameter estimates and standard errors from time-series, crosssection, and GMM/discount factor approach in the CAPM size portfolio test illustrated by Figures 28 and 29. The main parameter to be estimated is the slope of the lines in the above figures, the market price of risk λ in the expected return-beta model and the relation between mean returns and second moments b in the stochastic discount factor model. The big point of Table c1 is that the GMM/discount factor estimate and standard errors behave very similarly to the traditional estimates and standard errors. The rows compare results with various methods of calculating the spectral density matrix. i.i.d. imposes no serial correlation and regression errors independent of right hand variables, and is identical to the Maximum Likelihood based formulas. The 0 lag estimate allows conditional heteroskedasticity, but no correlation of residuals. The 3 lag, Newey West estimate is a sensible correction for short order autocorrelation. I include the 24 lag spectral density matrix to show how things can go wrong if you use a ridiculous spectral density matrix. 263

C HAPTER 15

Estimate i.i.d. 0 lags 3 lags, NW 24 lags

T IME SERIES , CROSS - SECTION , AND GMM/DF TESTS OF LINEAR FACTOR MODELS

TimeSeries 0.66 0.18 (3.67) 0.18 (3.67) 0.20 (3.30) 0.16 (4.13)

Beta model λ Cross section OLS 0.71 0.20 (3.55) 0.19 (3.74) 0.21 (3. 38) 0.16 (4.44)

GLS 0.66 0.18 (3.67) 0.18 (3.67) 0.20 (3.30) 0.16 (4.13)

1st stage 2.35

GMM/DF b 2nd stage Est. Std. Err.

0.63 (3.73) 0.69 (3.41) 1.00 (2.35)

2.46 2.39 2.15

0.61 (4.03) 0.64 (3.73) 0.69 (3.12)

Table c1. Parameter estimates and standard errors. Estimates are shown in italic, standard errors in regular type, and t-statistics in parentheses. The time-series estimate is the mean market return in percent per month. The cross-sectional estimate is the slope coefficient λ in E(Re ) = βλ. The GMM estimate is the parameter b in E(Re ) = E(Re f )b. CRSP monthly data 1926-1998. “Lags” gives the number of lags in the spectral density matrix. “NW” uses Newey-West weighting in the spectral density matrix. The OLS cross-sectional estimate 0.71 is a little higher than the mean market return 0.66, in order to better fit all of the assets, as seen in Figure 29. The GLS cross-sectional estimate is almost exactly the same as the mean market return, and the GLS standard errors are almost exactly the same as the time-series standard errors. The Shanken correction for generated regressors is very important to standard errors of the cross-sectional regressions. Without the Σf term in the standard deviation of λ (12.184)– i.e. treating the β as fixed right hand variables – the standard errors come out to 0.07 for OLS and √ 0.00 for GLS – far less than the correct 0.20 and 0.18 shown in the table, and far less than σ/ T . The b estimates are not directly comparable to the risk premium estimates, but it is easy to translate their units. Applying the discount factor model with normalization a = 1 to the market return itself, b=

E(Rem ) . E(Rem2 )

With E(Rem ) = 0.66% and σ(Rem ) = 5.47%, we have 100 × b = 100 (0.66) /(0.662 + 5.472 ) = 2.17. The entries in Table c1 are close to this magnitude. Most are slightly larger, as is the OLS cross-sectional regression, in order to better fit the other portfolios. The t-statistics are quite close across methods, which is another way to correct the units. The second-stage GMM/DF estimates (as well as standard errors) depend on which spectral density weighting matrix is used as a weighting matrix. The results are quite similar for all the sensible spectral density estimates. The 24 lag spectral density matrix starts to produce unusual estimates. This spectral density estimate will cause lots of problems below. Table c2 presents the χ2 and F statistics that test whether the pricing errors are jointly significant. The OLS and GLS cross-sectional regression, and the first and second stage GMM/discount factor tests give exactly the same χ2 statistic, though the individual pricing 264

S ECTION 15.2

MONTE CARLO AND B OOTSTRAP

errors and covariance matrix are not the same so I do not present them separately. The big point of Table c2 is that the GMM/discount factor method gives almost exactly the same result as the cross-sectional regression.

i.i.d. GRS F 0 lags 3 lags NW 24 lags

Time series χ2(10) % p 8.5 58 0.8 59 10.5 40 11.0 36 -432 -100

Cross section χ2(9) %p 8.5 49

GMM/DF χ2(9) % p

10.6 11.1 7.6

10.5 11.1 7.7

31 27 57

31 27 57

Table c2. χ2 tests that all pricing errors are jointly equal to zero. For the time-series regression, the GRS F test gives almost exactly the same rejection probability as does the asymptotic χ2 test. Apparently, the advantages of a statistic that is valid in finite samples is not that important in this data set. The χ2 tests for the time-series case without the i.i.d. assumption are a bit more conservative, with 30-40% p value rather than almost 60%. However, this difference is not large. The one exception is the χ2 test using 24 lags and no weights in the spectral density matrix. That matrix turns out not to be positive definite in this sample, with disastrous results for the χ2 statistic. (Somewhat surprisingly, the CAPM is not rejected. This is because the small firm effect vanishes in the latter part of the sample. I discuss this fact further in Chapter 20. See in particular Figure 28.) Looking across the rows, the χ2 statistic is almost exactly the same for each method. The cross-sectional regression and GMM/DF estimate have one lower degree of freedom (the market premium is estimated from the cross-section rather than from the market return), and so show slightly greater rejection probabilities. For a given spectral density estimation technique, the cross-sectional regression and the GMM/DF approach give almost exactly the same χ2 values and rejection probabilities. The 24 lag spectral density matrix is a disaster as usual. In this case, it is a greater disaster for the time-series test than for the cross-section or GMM/discount factor test. It turns out not to be positive definite, so the sample pricing errors ˆ 0 cov(ˆ α)−1 α ˆ produce a nonsensical negative value of α

15.2

Monte Carlo and Bootstrap

The parameter distribution for the time-series regression estimate is quite similar to that from the GMM/discount factor estimate. The size and power of χ2 test statistics is nearly identical for time-series regression test 265

C HAPTER 15

T IME SERIES , CROSS - SECTION , AND GMM/DF TESTS OF LINEAR FACTOR MODELS

and the GMM/discount factor test. A bad spectral density matrix can ruin either time-series or GMM/discount factor estimates and tests. There is enough serial correlation and heteroskedasticity in the data that conventional i.i.d. formulas produce test statistics with about 1/2 the correct size. If you want to do classic regression tests, you should correct the distribution theory rather than use the ML i.i.d. distributions. Econometrics is not just about sensible point estimates, it is about sampling variability of those estimates, and whether standard error formulas correctly capture that sampling variability. How well do the various standard error and test statistic formulas capture the true sampling distribution of the estimates? To answer this question I conduct two Monte Carlos and two bootstraps. I conduct one each under the null that the CAPM is correct, to study size, and one each under the alternative that the CAPM is false, to study power. The Monte Carlo experiments follow the standard ML assumption that returns and factors are i.i.d. normally distributed, and the factors and residuals are independent as well as uncorrelated. I generate artificial samples of the market return from an i.i.d. normal, using the sample mean and variance of the value-weighted return. I then generate artificial size em + εit , using the sample residual covaridecile returns under the null by Rei t = 0 + β i Rt ance matrix Σ to draw i.i.d. normal residuals εit and the sample regression coefficients β i . To generate data under the alternative, I add the sample αi . draw 5000 artificial samples. I try a long sample of 876 months, matching the CRSP sample analyzed above. I also draw a short sample of 240 months or 20 years, which is about as short as one should dare try to test a factor model. The bootstraps check whether non-normalities, autocorrelation, heteroskedasticity, and non-independence of factors and residuals matters to the sampling distribution in this data set. I do a block-bootstrap, resampling the data in groups of three months with replacement, to preserve the short-order autocorrelation and persistent heteroskedasticity in the data. To impose the CAPM, I draw the market return and residuals in the time-series regression, and then compute artificial data on decile portfolio returns by Rtei = 0 + β i Rtem + εit . To study the alternative, I simply redraw all the data in groups of three. Of course, the actual data may display conditioning information not displayed by this bootstrap, such as predictability and conditional heteroskedasticity based on additional variables such as the dividend/price ratio, lagged squared returns, or implied volatilities. The first-stage GMM/discount factor and OLS cross-sectional regression are nearly identical in every artificial sample, as the GLS cross-sectional regression is nearly identical to the time-series regression in every sample. Therefore, the important question is to compare the time series regression – which is ML with i.i.d. normal returns and factors – to the first and second stage GMM/DF procedure. For this reason and to save space, I do not include the cross-sectional regressions in the Monte Carlo and bootstrap. 266

S ECTION 15.2

MONTE CARLO AND B OOTSTRAP

χ2 tests

Table 6c presents the χ2 tests of the hypothesis that all pricing errors are zero under the null that the CAPM is true, and Table 7c presents the χ2 tests under the null that the CAPM is false. Each table presents the percentage of the 5000 artificial data sets in which the χ2 tests rejected the null at the indicated level. The central point of these tables is that the GMM/discount factor test performs almost exactly the same way as the time-series test. Compare the GMM/DF entry to its corresponding Time series entry; they are all nearly identical. Neither the small efficiency advantage of time-series vs. cross section, nor the difference between betas and second moments seems to make any difference to the sampling distribution.

Sample size: level (%): i.i.d. 0 lags 3 lags, NW 24 lags

Monte Carlo Time series GMM/DF 240 876 240 876 5 5 1 5 5 1 7.5 6.0 1.1 7.7 6.1 1.1 7.5 6.3 1.0 10.7 6.5 1.4 9.7 6.6 1.3 25 39 32 25 41 31

Block-Bootstrap Time series GMM/DF 240 876 240 876 5 5 1 5 5 1 6.0 2.8 0.6 7.7 4.3 1.0 6.6 3.7 0.9 10.5 5.4 1.3 9.5 5.3 1.3 23 38 31 24 41 32

Table 6c. Size. Probability of rejection for χ2 statistics under the null that all pricing errors are zero

Sample size: level (%): i.i.d. 0 lags 3 lags, NW 24 lags

Monte Carlo Time-Series GMM/DF 240 876 240 876 5 1 5 1 17 48 26 17 48 26 17 50 27 22 49 27 21 51 29 29 60 53 29 66 57

Block-Bootstrap Time-Series GMM/DF 240 876 240 876 5 1 5 1 11 40 18 15 54 28 14 55 29 18 57 31 17 59 33 27 63 56 29 68 60

Table 7c. Power. Probability of rejection for χ2 statistics under the null that the CAPM is false, and the true means of the decile portfolio returns are equal to their sample means. Start with the Monte Carlo evaluation of the time-series test in Table 6c. The i.i.d. and 0 lag distributions produce nearly exact rejection probabilities in the long sample and slightly too many (7.5%) rejections in the short sample. Moving down, GMM distributions here correct for things that aren’t there. This has a small but noticeable effect on the sensible 3 lag test, which rejects slightly too often under this null. Naturally, this is worse for the short sample, but looking across the rows, the time-series and discount factor tests are nearly identical in every case. The variation across technique is almost zero, given the spectral 267

C HAPTER 15

T IME SERIES , CROSS - SECTION , AND GMM/DF TESTS OF LINEAR FACTOR MODELS

density estimate. The 24 lag unweighted spectral density is the usual disaster, rejecting far too often. It is singular in many samples. In the long sample, the 1% tail of this distribution occurs at a χ2 value of 440 rather than the 23.2 of the χ2(10) distribution! The long sample block-bootstrap in the right half of the tables shows even in this simple setup how i.i.d. normal assumptions can be misleading. The traditional i.i.d. χ2 test has almost half the correct size – it rejects a 10% test 6% of the time, a 5% test 2.8% of the time and a 1% test 0.6% of the time. Removing the assumption that returns and factors are independent, going from i.i.d. to 0 lags, brings about half of the size distortion back, while adding one of the sensible autocorrelation corrections does the rest. In each row, the time-series and GMM/DF methods produce almost exactly the same results again. The 24 lag spectral density matrices are a disaster as usual. Table 7c shows the rejection probabilities under the alternative. The most striking feature of the table is that the GMM/discount factor test gives almost exactly the same rejection probability as the time-series test, for each choice of spectral density estimation technique. When there is a difference, the GMM/discount factor test rejects slightly more often. The 24 lag tests reject most often, but this is not surprising given that they reject almost as often under the null. Parameter estimates and standard errors

Table c5 presents the sampling variation of the λ and b estimates. The rows and columns ˆ , σ(ˆb), and in italic font, give the variation of the estimated λ or b across the market σ(λ) 5000 artificial samples. The remaining rows and columns give the average across samples of the standard errors. The presence of pricing errors has little effect on the estimated b or λ and their standard errors, so I only present results under the null that the CAPM is true. The parameters are not directly comparable – the b parameter includes the variance as well as the mean of the factor, and ET (Rem ) is the natural GMM estimate of the mean market return as it is the Time-series estimate of the factor risk premium. Still, it is interesting to know and to compare how well the two methods do at estimating their central parameter.

268

S ECTION 15.2

Time series T=876: ˆ σ(λ), σ(ˆb) i.i.d. 0 lags 3 lags NW 24 lags T=240: ˆ , σ(ˆb) σ(λ) i.i.d. 0 lags 3 lags NW 24 lags

MONTE CARLO AND B OOTSTRAP

Monte Carlo GMM/DF 1st 2nd stage stage σ(ˆb) E (s.e.)

0.19 0.18 0.18 0.18 0.18

0.64

0.35 0.35 0.35 0.35 0.29

1.25

0.65 0.65 0.62

1.23 1.22 1.04

0.61 0.62 130

1.24 1.26 191

Time series

Block-Bootstrap GMM/DF 1st 2nd stage stage σ(ˆb) E (s.e.) 0.69

0.60 0.59 0.27

0.20 0.18 0.18 0.19 0.19

1.40

1.14 1.11 0.69

0.37 0.35 0.35 0.36 0.31

0.63 0.67 0.66

1.24 1.31 1.15

0.67 0.67 1724

0.60 0.62 0.24

1.45 1.48 893

1.15 1.14 0.75

Table 5. Monte Carlo and block-bootstrap evaluation of the sampling variability of parameter estimates b and λ. The Monte Carlo redraws 5000 artificial data sets of length T=876 from a random normal assuming that the CAPM is true. The blockbootstrap redraws the data in groups of 3 with replacement. The row and columns ˆ and σ(ˆb) and using italic font give the variation across samples of the marked σ(λ) estimated λ and b. The remaining entries of “Time series” “1st stage” and “E (s.e.)” columns in roman font give the average value of the computed standard error of the parameter estimate, where the average is taken over the 5000 samples. The central message of this table is that the GMM/DF estimates behave almost exactly as the time-series estimate, and the asymptotic standard error formulas almost exactly capture the sampling variation of the estimates. The second stage GMM/DF estimate is a little bit more efficient at the cost of slightly misleading standard errors. Start with the long sample and the first column. All of the standard error formulas give essentially identical and correct results for the time-series estimate. Estimating the sample mean is not rocket science. The first stage GMM/DF estimator in the second column behaves the same way, except the usually troublesome 24 lag unweighted estimate. The second stage GMM/DF estimate in the third and fourth columns uses the inverse spectral density matrix to weight, and so the estimator depends on the choice of spectral density estimate. The sensible spectral density estimates (not 24 lags) produce second-stage estimates that vary less than the first-stage estimates, 0.61 − 0.62 rather than 0.64. Second stage GMM is more efficient, meaning that it produces estimates with smaller sampling variation. However, the table shows that the efficiency gain is quite small, so not much is lost if one prefers first stage OLS estimates. The sensible spectral density estimates produce second-stage standard errors that again almost exactly capture the sampling variation of the 269

C HAPTER 15

T IME SERIES , CROSS - SECTION , AND GMM/DF TESTS OF LINEAR FACTOR MODELS

estimated parameters. The 24 lag unweighted estimate produces hugely variable estimates and artificially small standard errors. Using bad or even singular spectral density estimates seems to have a secondary effect on standard error calculations, but using its inverse as a weighting matrix can have a dramatic effect on estimation. With the block-bootstrap in the right hand side of Table 5c, the time-series estimate is slightly more volatile as a result of the slight autocorrelation in the market return. The i.i.d. and zero lag formulas do not capture this effect, but the GMM standard errors that allow autocorrelation do pick it up. However, this is a very minor effect as there is very little autocorrelation in the market return. The effect is more pronounced in the first stage GMM/DF estimate, since the smaller firm portfolios depart more from the normal i.i.d. assumption. The true variation is 0.69, but standard errors that ignore autocorrelation only produce 0.63. The standard errors that correct for autocorrelation are nearly exact. In the second-stage GMM/DF, the sensible spectral density estimates again produce slightly more efficient estimates than the first stage, with variation of 0.67 rather than 0.69. This comes at a cost, though, that the asymptotic standard errors are a bit less reliable. In the shorter sample, we see that standard errors for the mean market return in the Time series column are all quite accurate, except the usual 24 lag case. In the GMM/DF case, we see that the actual sampling variability of the b estimate is no longer smaller for the second stage. The second stage estimate is not more efficient in this “small” sample. Furthermore, while the first stage standard errors are still decently accurate, the second stage standard errors substantially understate the true sampling variability of the parameter estimate. They represent a hoped-for efficiency that is not present in the small sample. Even in this simple setup, first-stage GMM is clearly a better choice for estimating the central parameter, and hence for examining individual pricing errors and their pattern across assets.

270

Chapter 16.

Which method?

Of course, the point of GMM/discount factor methods is not a gain in efficiency or simplicity in a traditional setup – linear factor model, i.i.d. normally distributed returns and factors, etc. It’s hard to beat the efficiency or simplicity of regression methods in those setups. The point of the GMM/discount factor approach is that it allows a simple technique for evaluating nonlinear or otherwise complex models, for including conditioning information while not requiring the econometrician to see everything that the agent sees, and for allowing the researcher to circumvent inevitable model misspecifications or simplifications and data problems by keeping the econometrics focused on interesting issues. The alternative is usually some form of maximum likelihood. This is much harder in most circumstances, since you have to write down a complete statistical model for the conditional distribution of your data. Just evaluating, let alone maximizing, the likelihood function is often challenging. Whole series of papers are written just on the econometric issues of particular cases, for example how to maximize the likelihood functions of specific classes of univariate continuous time models for the short interest rate. Of course, there is no necessary pairing of GMM with the discount factor expression of a model, and ML with the expected return-beta formulation. Many studies pair discount factor expressions of the model with ML, and many others evaluate expected return-beta model by GMM, as we have done in adjusting regression standard errors for non-i.i.d. residuals. Advanced empirical asset pricing faces an enduring tension between these two philosophies. The choice essentially involves tradeoffs between statistical efficiency, the effects of misspecification of both the economic and statistical models, and the clarity and economic interpretability of the results. There are situations in which it’s better to trade some small efficiency gains for the robustness of simpler procedures or more easily interpretable moments; OLS can be better than GLS. The central reason is specification errors; the fact that our statistical and economic models are at best quantitative parables. There are other situations in which one may really need to squeeze every last drop out of the data, intuitive moments are statistically very inefficient, and more intensive maximum-likelihood approaches are more appropriate. Unfortunately, the environments are complex, and differ from case to case. We don’t have universal theorems from statistical theory or generally applicable Monte Carlo evidence. Specification errors by their nature resist quantitative modeling – if you knew how to model them, they wouldn’t be there. We can only think about the lessons of past experiences. In my experience, in the limited range of applications I have worked with, a GMM approach based on simple easily interpretable moments has proved far more fruitful than formal maximum likelihood. In addition, I have found first stage GMM – OLS cross sectional regressions – to be more trustworthy than second-stage GMM, in any case where there was a substantial difference between the two approaches. The rest of this chapter collects some thoughts on the choice between formal ML and less formal GMM, focusing on economically interesting rather than statistically informative moments. 271

CHAPTER 16

WHICH METHOD ?

“ML” vs. “GMM”

The debate is often stated as a choice between “maximum likelihood” and “GMM.” This is a bad way to put the issue. ML is a special case of GMM: it suggests a particular choice of moments that are statistically optimal in a well-defined sense. Given the set of moments, the distribution theories are identical. Also, there is no such thing as “the” GMM estimate. GMM is a flexible tool; you can use any aT matrix and gT moments that you want to use. For example, we saw how to use GMM to derive the asymptotic distribution of the standard time-series regression estimator with autocorrelated returns. The moments in this case were not the pricing errors. It’s all GMM; the issue is the choice of moments. Both ML and GMM are tools that a thoughtful researcher can use in learning what the data says about a given asset pricing model, rather than as stone tablets giving precise directions that lead to truth if followed literally. If followed literally and thoughtlessly, both ML and GMM can lead to horrendous results. The choice is between moments selected by an auxiliary statistical model, even if completely economically uninterpretable, and moments selected for their economic or data summary interpretation, even if not statistically efficient. ML is often ignored

As we have seen, ML plus the assumption of normal i.i.d. disturbances leads to easily interpretable time-series or cross-sectional regressions, empirical procedures that are close to the economic content of the model. However, asset returns are not normally distributed or i.i.d.. They have fatter tails than a normal, they are heteroskedastic (times of high and times of low volatility), they are autocorrelated, and predictable from a variety of variables. If one were to take seriously the ML philosophy and its quest for efficiency, one should model these features of returns. The result would be a different likelihood function, and its scores would prescribe different moment conditions than the familiar and intuitive time-series or cross-sectional regressions. Interestingly, few empirical workers do this. (The exceptions tend to be papers whose primary point is illustration of econometric technique rather than empirical findings.) ML seems to be fine when it suggests easily interpretable regressions; when it suggests something else, people use the regressions anyway. For example, ML prescribes that one estimate β s without a constant. β s are almost universally estimated with a constant. Researchers often run cross-sectional regressions rather than time-series regressions, even when the factors are returns. ML specifies a GLS cross-sectional regression, but many empirical workers use OLS cross-sectional regressions instead, distrusting the GLS weighting matrix. Time-series regressions are almost universally run with a constant, though ML prescribes a regression with no constant. The true ML formulas for GLS regressions require one to iterate between non-OLS formulas for betas, covariance matrix estimate and the cross-sectional regression estimate. Empirical applications usually use the unconstrained estimates of all these quantities. And of course, any of the regression tests continue to be run at all, with ML justifications, despite the fact that returns are not i.i.d. The 272

regressions came first, and the maximum likelihood formalization came later. If we had to assume that returns had a gamma distribution to justify the regressions, it’s a sure bet that we would make that “assumption” behind ML instead of the normal i.i.d. assumption! The reason must be that researchers feel that omitting some of the information in the null hypothesis, the estimation and test is more robust, though some efficiency is lost if the null economic and statistical models are exactly correct. Researchers must not really believe that their null hypotheses, statistical and economic, are exactly correct. They want to produce estimates and tests that are robust to reasonable model mis-specifications. They also want to produce estimates and tests that are easily interpretable, that capture intuitively clear stylized facts in the data, and that relate directly to the economic concepts of the model. Such estimates are persuasive in large part because the reader can see that they are robust. (And following this train of thought, one might want to pursue estimation strategies that are even more robust than OLS, since OLS places a lot of weight on outliers. For example, Chen and Ready 1997 claim that Fama and French’s 1993 size and value effects depend crucially on a few outliers.) ML does not necessarily produce robust or easily interpretable estimates. It wasn’t designed to do so. The point and advertisement of ML is that it provides efficient estimates; it uses every scrap of information in the statistical and economic model in the quest for efficiency. It does the “right” efficient thing if model is true. It does not necessarily do the “reasonable” thing for “approximate” models. OLS vs. GLS cross-sectional regressions

One place in which this argument crystallizes is in the choice between OLS and GLS cross-sectional regressions, or equivalently between first and second stage GMM. The last chapter can lead to a mistaken impression that the doesn’t matter that much. This is true to some extent in that simple environment, but not in more complex environments. For example, Fama and French (1997) report important correlations between betas and pricing errors in a time-series test of a three-factor model on industry portfolios. This correlation cannot happen with an OLS cross-sectional estimate, as the cross-sectional estimate sets the cross-sectional correlation between right hand variables (betas) and error terms (pricing errors) to zero by construction. First stage estimates seem to work better in factor pricing models based on macroeconomic data. For example, Figure 5 presents the first stage estimate of the consumption-based model. The second-stage estimate produced much larger individual pricing errors, because by so doing it could lower pricing errors of portfolios with strong long-short positions required by the spectral density matrix. The same thing happened in the investment based factor pricing model Cochrane (1996), and the scaled consumption-based model of Lettau and Ludvigson (2000). Authors as far back as Fama and MacBeth (1973) have preferred OLS cross-sectional regressions, distrusting the GLS weights. GLS and second-stage GMM gain their asymptotic efficiency when the covariance and spectral density matrices have converged to their population values. GLS and second stage GMM use these matrices to find well-measured portfolios; portfolios with small residual 273

CHAPTER 16

WHICH METHOD ?

variance for GLS, and small variance of discounted return for GMM. The danger is that these quantities are poorly estimated in a finite sample, that sample minimum-variance portfolios bear little relation to population minimum-variance portfolios. This by itself should not create too much of a problem for a perfect model, one that prices all portfolios. But an imperfect model that does a very good job of pricing a basic set of portfolios may do a poor job of pricing strange linear combinations of those portfolios, especially combinations that involve strong long and short positions, positions that really are outside the payoff space given transactions, margin, and short sales constraints. Thus, the danger is the interaction between spurious sample minimum-variance portfolios and the specification errors of the model. Interestingly, Kandel and Stambaugh (1995) and Roll and Ross (1995) argue for GLS cross-sectional regressions also as a result of model misspecification. They start by observing that show that so long as there is any misspecification at all – so long as the pricing errors are not exactly zero; so long as the market proxy is not exactly on the mean-variance frontier – then there are portfolios that produce arbitrarily good and arbitrarily bad fits in plots of expected returns vs. betas. Since even a perfect model leaves pricing errors in sample, this is always true in samples. It’s easy to see the basic argument. Take a portfolio long the positive alpha securities and short the negative alpha securities; it will have a really big alpha! More precisely, if the original securities follow E(Re ) = α + λβ,

then consider portfolios of the original securities formed from a non-singular matrix A. They follow E(ARe ) = Aα + λAβ.

You can make all these portfolios have the same β with Aβ =constant, and then they will have a spread in alphas. You will see a plot in which all the portfolios have the same beta but the average returns are spread up and down. Conversely, you can pick A to make the expected return-beta plot look as good as you want. GLS has an important feature in this situation: the GLS cross-sectional regression is independent of such repackaging of portfolios. If you transform a set of returns Re to ARe , then the OLS cross-sectional regression is transformed from

to

¡ ¢ ˆ = β 0 β −1 β 0 E (Re ) λ ¡ ¢ ˆ = β 0 A0 Aβ −1 β 0 A0 AE (Re ) . λ

This does depend on the repackaging A. However, the residual covariance matrix of ARe is 274

AΣA0 , so the GLS regression ¡ ¢ ˆ = β 0 Σ−1 β −1 β 0 Σ−1 E (Re ) λ

is not affected so long as A is full rank and therefore does not throw away information ¡ ¢ ¡ ¢ ˆ = β 0 A0 (AΣA0 )−1 Aβ −1 β 0 A0 (A0 ΣA)−1 AE (Re ) = β 0 Σ−1 β −1 β 0 Σ−1 E (Re ) . λ

(The spectral density matrix and second stage estimate shares this property in GMM estimates. These are not the only weighting matrix choices that are invariant to portfolios. For example, Hansen and Jagannathan’s 1997 suggestion of the return second moment matrix has the same property.) This is a fact, but it does not show that OLS chooses a particularly good or bad set of portfolios. Perhaps you don’t think that GLS’ choice of portfolios is particularly informative. In this case, you use OLS precisely to focus attention on a particular set of economically interesting portfolios. The choice depends subtly on what you want your test to accomplish. If you want to prove the model wrong, then GLS helps you to focus on the most informative portfolios for proving the model wrong. That is exactly what an efficient test is supposed to do. But many models are wrong, but still pretty darn good. It is a shame to throw out the information that the model does a good job of pricing an interesting set of portfolios. The sensible compromise would seem to be to report the OLS estimate on “interesting” portfolios, and also to report the GLS test statistic that shows the model to be rejected. That is, in fact, the typical collection of facts. Additional examples of trading off efficiency for robustness

Here are some additional examples of situations in which it has turned out to be wise to trade off some apparent efficiency for robustness to model misspecifications. Low frequency time-series models. In estimating time-series modelsPsuch as the AR(1), maximum likelihood minimizes one-step ahead forecast error variance, ε2t . But any timeseries model is only an approximation, and the researcher’s objective may not be one-step ahead forecasting. For example, in making sense of the yield on long term bonds, one is interested in the long-run behavior of the short rate of interest. In estimating the magnitude of long-horizon univariate mean reversion in stock returns, we want to know only the sum of autocorrelations or moving average coefficients. (We will study this application in section 20.335.) The approximate model that generates the smallest one-step ahead forecast error variance may be quite different from the model that best matches long-run autocorrelations. ML can pick the wrong model and make very bad predictions for long-run responses. (Cochrane 1986 contains a more detailed analysis of this point in the context of long-horizon GDP forecasting.) Lucas’ money demand estimate. Lucas (1988) is a gem of an example. Lucas was in-

275

CHAPTER 16

WHICH METHOD ?

terested in estimating the income elasticity of money demand. Money and income trend upwards over time and over business cycles, but also have some high-frequency movement that looks like noise. If you run a regression in log-levels, mt = a + byt + εt

you get a sensible coefficient of about b = 1, but you find that the error term is strongly serially correlated. Following standard advice, most researchers run GLS, which amounts pretty much to first-differencing the data, mt − mt−1 = b(yt − yt−1 ) + η t .

This error term passes its Durbin-Watson statistic, but the b estimate is much lower, which doesn’t make much economic sense, and, worse, is unstable, depending a lot on time period and data definitions. Lucas realized that the regression in differences threw out all of the information in the data, which was in the trend, and focused on the high-frequency noise. Therefore, the regression in levels, with standard errors corrected for correlation of the error term, is the right one to look at. Of course, GLS and ML didn’t know there was any “noise” in the data, which is why they threw out the baby and kept the bathwater. Again, ML ruthlessly exploits the null for efficiency, and has no way of knowing what is “reasonable” or “intuitive.” Stochastic singularities and calibration. Models of the term structure of interest rates (we will study these models in section 19) and real business cycle models in macroeconomics give even more stark examples. These models are stochastically singular. They generate predictions for many time series from a few shocks, so the models predict that there are combinations of the time series that leave no error term. Even though the models have rich and interesting implications, ML will seize on this economically uninteresting singularity, refuse to estimate parameters, and reject any model of this form.

The simplest example of the situation is the linear-quadratic permanent income model paired with an AR(1) specification for income. The model is yt ct − ct−1

= ρyt−1 + εt = (Et − Et−1 )

∞ 1 X j 1 εt β yt+j = 1 − β j=0 (1 − βρ) (1 − β)

This model generates all sorts of important and economically interesting predictions for the joint process of consumption and income (and asset prices). Consumption should be roughly a random walk, and should respond only to permanent income changes; investment should be more volatile than income and income more volatile than consumption. Since there is only one shock and two series, however, the model taken literally predicts a deterministic relation between consumption and income; it predicts ct − ct−1 =

rβ (yt − ρyt−1 ) . 1 − βρ

276

ML will notice that this is the statistically most informative prediction of the model. There is no error term! In any real data set there is no configuration of the parameters r, β, ρ that make this restriction hold, data point for data point. The probability of observing a data set {ct , yt } is exactly zero, and the log likelihood function is −∞ for any set of parameters. ML says to throw the model out. The popular affine models of the term structure of interest rates act the same way. They specify that all yields at any moment in time are deterministic functions of a few state variables. Such models can capture much of the important qualitative behavior of the term structure, including rising, falling and humped shapes, and the information in the term structure for future movements in yields and the volatility of yields. They are very useful for derivative pricing. But it is never the case in actual yield data that yields of all maturities are exact functions of K yields. Actual data on N yields always require N shocks. Again, a ML approach reports a −∞ log likelihood function for any set of parameters.

Addressing model mis-specification

The ML philosophy offers an answer to model mis-specification: specify the right model, and then do ML. If regression errors are correlated, model and estimate the covariance matrix and do GLS. If you are worried about proxy errors in the pricing factor, short sales costs or other transactions costs so that model predictions for extreme long-short positions should not be relied on, time-aggregation or mismeasurement of consumption data, non-normal or non-i.i.d. returns, time-varying betas and factor risk premia, additional pricing factors and so on–don’t chat about them, write them down, and then do ML. Following this lead, researchers have added “measurement errors” to real business cycle models (Sargent 1989 is a classic example) and affine yield models in order to break the stochastic singularity (I discuss this case a bit more in section 19.6). The trouble is, of course, that the assumed structure of the measurement errors now drives what moments ML pays attention to. And seriously modeling and estimating the measurement errors takes us further away from the economically interesting parts of the model. (Measurement error augmented models will often wind up specifying sensible moments, but by assuming ad-hoc processes for measurement error, such as i.i.d. errors. Why not just specify the sensible moments in the first place?) More generally, authors tend not to follow this advice, in part because it is ultimately infeasible. Economics necessarily studies quantitative parables rather than completely specified models. It would be nice if we could write down completely specified models, if we could quantitatively describe all the possible economic and statistical model and specification errors, but we can’t. The GMM framework, used judiciously, allows us to evaluate misspecified models. It allows us to direct that the statistical effort focus on the “interesting” predictions while ignoring the fact that the world does not match the “uninteresting” simplifications. For example, ML only gives us a choice of OLS, whose standard errors are wrong, or GLS, which we may not trust in small samples or which may focus on uninteresting parts of the data. GMM allows 277

CHAPTER 16

WHICH METHOD ?

us to keep an OLS estimate, but correct the standard errors for non-i.i.d. distributions. More generally, GMM allows one to specify an economically interesting set of moments, or a set of moments that one feels will be robust to misspecifications of the economic or statistical model, without having to spell out exactly what is the source of model mis-specification that makes those moments “optimal” or even “interesting” and “robust.” It allows one to accept the lower “efficiency” of the estimates under some sets of statistical assumptions, in return for such robustness. At the same time, the GMM framework allows us to flexibly incorporate statistical model misspecifications in the distribution theory. For example, knowing that returns are not i.i.d. normal, one may want to use the time series regression technique to estimate betas anyway. This estimate is not inconsistent, but the standard errors that ML formulas pump out under this assumption are inconsistent. GMM gives a flexible way to derive at least and asymptotic set of corrections for statistical model misspecifications of the time-series regression coefficient. Similarly, a pooled time-series cross-sectional OLS regression is not inconsistent, but standard errors that ignore cross-correlation of error terms are far too small. The “calibration” of real business cycle models is often really nothing more than a GMM parameter estimate, using economically sensible moments such as average output growth, consumption/output ratios etc. to avoid the stochastic singularity that would doom a ML approach. (Kydland and Prescott’s 1982 idea that empirical microeconomics would provide accurate parameter estimates for macroeconomic and financial models has pretty much vanished.) Calibration exercises usually do not compute standard errors, nor do they report any distribution theory associated with the “evaluation” stage when one compares the model’s predicted second moments with those in the data. Following Burnside, Eichenbaum and Rebelo (1993) however, it’s easy enough to calculate such a distribution theory – to evaluate whether the difference between predicted “second moments” and actual moments is large compared to sampling variation, including the variation induced by parameter estimation in the same sample – by listing the first and second moments together in the gT vector. “Used judiciously” is an important qualification. Many GMM estimations and tests suffer from lack of thought in the choice of moments, test assets and instruments. For example, early GMM papers tended to pick assets and especially instruments pretty much at random. Industry portfolios have almost no variation in average returns to explain. Authors often included many lags of returns and consumption growth as instruments to test a consumptionbased model. However, the 7th lag of returns really doesn’t predict much about future returns given lags 1-12, and the first-order serial correlation in seasonally adjusted, ex-post revised consumption growth may be economically uninteresting. More recent work tends to emphasize a few well-chosen assets and instruments that capture important and economically interesting features of the data.. Auxiliary model

ML requires an auxiliary statistical model. For example, in the classic ML formalization of regression tests, we had to stopped to assume that returns and factors are jointly i.i.d. normal. As the auxiliary statistical model becomes more and more complex and hence realistic, 278

more and more effort is devoted to estimating the auxiliary statistical model. ML has no way of knowing that some parameters – a, b; β, λ, risk aversion γ – are more “important” than others – Σ, and parameters describing time-varying conditional moments of returns. A very convenient feature of GMM is that it does not require such an auxiliary statistical model. For example, in studying GMM we went straight from p = E(mx) to moment conditions, estimates, and distribution theory. This is an important saving of the researcher’s and the reader’s time, effort and attention. Finite sample distributions

Many authors say they prefer regression tests and the GRS statistic in particular because it has a finite sample distribution theory, and they distrust the finite-sample performance of the GMM asymptotic distribution theory. This argument does not have much force. The finite sample distribution only holds if returns really are normal and i.i.d., and if the factor is perfectly measured. Since these assumptions do not hold, it is not obvious that a finite-sample distribution that ignores non-i.i.d. returns will be a better approximation than an asymptotic distribution that corrects for them. All approaches give essentially the same answers in the classic setup of i.i.d. returns. The issue is how the various techniques perform in more complex setups, especially with conditioning information, and here there are no analytic finite-sample distributions. In addition, once you have picked the estimation method – how you will generate a number from the data; or which moments you will use – finding its finite sample distribution, given an auxiliary statistical model, is simple. Just run a Monte Carlo or bootstrap. Thus, picking an estimation method because it delivers analytic formulas for a finite sample distribution (under false assumptions) should be a thing of the past. Analytic formulas for finite sample distributions are useful for comparing estimation methods and arguing about statistical properties of estimators, but they are not necessary for the empiricists’ main task. Finite sample quality of asymptotic distributions, and “nonparametric” estimates

Several investigations (Ferson and Foerster 1994, Hansen, Heaton, and Yaron 1996) have found cases in which the GMM asymptotic distribution theory is a poor approximation to a finite-sample distribution theory. This is especially true when one asks “non-parametric” corrections for autocorrelation or heteroskedasticity to provide large corrections and when the number of moments is large compared to the sample size, or if the moments one uses for GMM turn out to be very inefficient (Fuhrer, Moore, and Schuh 1995) which can happen if you put in a lot of instruments with low forecast power The ML distribution is the same as GMM, conditional on the choice of moments, but typical implementations of ML also use the parametric time-series model to simplify estimates of the terms in the distribution theory as well as to derive the likelihood function. If this is the case – if the “nonparametric” estimates of the GMM distribution theory perform poorly in a finite sample, while the “parametric” ML distribution works well – there is no reason not to use a parametric time series model to estimate the terms in the GMM 279

CHAPTER 16

WHICH METHOD ?

P distribution as well. For example, rather than calculate ∞ j=−∞ E(ut ut−j ) from a large sum ofPautocorrelations, you can model ut = ρut−1 + εt , estimate ρ, and then calculate 1+ρ j 2 σ2 (u) ∞ j=−∞ ρ = σ (u) 1−ρ . Section 11.7 discussed this idea in more detail.

The case for ML

In the classic setup, the efficiency gain of ML over GMM on the pricing errors is tiny. However, several studies have found cases in which the statistically motivated choice of moments suggested by ML has important efficiency advantages. For example, Jacquier, Polson and Rossi (1994) study the estimation of a time-series model with stochastic volatility. This is a model of the form dSt /St dVt

= µdt + Vt dZ1t = µV (Vt )dt + σ(Vt )dZ2t ,

(16.213)

and S is observed but V is not. The obvious and easily interpretable moments include the autocorrelation of squared returns, or the autocorrelation of the absolute value of returns. However, Jacquier, Polson and Rossi find that the resulting estimates are far less efficient than those resulting from the ML scores. Of course, this study presumes that the model (16.213) really is exactly true. Whether the uninterpretable scores or the interpretable moments really perform better to give an approximate model of the form (16.213), given some other data-generating mechanism is open to discussion. Even in the canonical OLS vs. GLS case, a wildly heteroskedastic error covariance matrix can mean that OLS spends all its effort fitting unimportant data points. A “judicious” application of GMM (OLS) in this case would require at least some transformation of units so that OLS is not wildly inefficient. Statistical philosophy

The history of empirical work that has been persuasive – that has changed people’s understanding of the facts in the data and which economic models understand those facts – looks a lot different than the statistical theory preached in econometrics textbooks. The CAPM was taught and believed in and used for years despite formal statistical rejections. It only fell by the wayside when other, coherent views of the world were offered in the multifactor models. And the multifactor models are also rejected! It seems that “it takes a model to beat a model,” not a rejection. Even when evaluating a specific model, most of the interesting calculations come from examining specific alternatives rather than overall pricing error tests. The original CAPM tests focused on whether the intercept in a cross-sectional regression was higher or lower than the risk free rate, and whether individual variance entered into cross-sectional regressions. The CAPM fell when it was found that characteristics such as size and book/market do enter cross-sectional regressions, not when generic pricing error tests rejected. 280

Influential empirical work tells a story. The most efficient procedure does not seem to convince people if they cannot transparently see what stylized facts in the data drive the result. A test of a model that focuses on its ability to account for the cross section of average returns of interesting portfolios will in the end be much more persuasive than one that (say) focuses on the model’s ability to explain the fifth moment of the second portfolio, even if ML finds the latter moment much more statistically informative. Most recently, Fama and French (1988b) and (1993) are good examples of empirical work that changed many people’s minds, in this case that long-horizon returns really are predictable, and that we need a multifactor model rather than the CAPM to understand the cross-section of average returns. These papers are not stunning statistically: long horizon predictability is on the edge of statistical significance, and the multifactor model is rejected by the GRS test. But these papers made clear what stylized and robust facts in the data drive the results, and why those facts are economically sensible. For example, the 1993 paper focused on tables of average returns and betas. Those tables showed strong variation in average returns that was not matched by variation in market betas, yet was matched by variation in betas on new factors. There is no place in statistical theory for such a table, but it is much more persuasive than a table of χ2 values for pricing error tests. On the other hand, I can think of no case in which the application of a clever statistical models to wring the last ounce of efficiency out of a dataset, changing t statistics from 1.5 to 2.5, substantially changed the way people think about an issue. Statistical testing is one of many questions we ask in evaluating theories, and usually not the most important one. This is not a philosophical or normative statement; it is a positive or empirical description of the process by which the profession has moved from theory to theory. Think of the kind of questions people ask when presented with a theory and accompanying empirical work. They usually start by thinking hard about the theory itself. What is the central part of the economic model or explanation? Is it internally consistent? Do the assumptions make sense? Then, when we get to the empirical work, how were the numbers produced? Are the data definitions sensible? Are the concepts in the data decent proxies for the concepts in the model? (There’s not much room in statistical theory for that question!) Are the model predictions robust to the inevitable simplifications? Does the result hinge on power utility vs. another functional form? What happens if you add a little measurement error, or if agents have an information advantage, etc.? What are the identification assumptions, and do they make any sense – why is y on the left and x on the right rather than the other way around? Finally, someone in the back of the room might raise his hand and ask, “if the data were generated by a draw of i.i.d. normal random variables over and over again, how often would you come up with a number this big or bigger?” That’s an interesting and important check on the overall believability of the results. But it is not necessarily the first check, and certainly not the last and decisive check. Many models are kept that have economically interesting but statistically rejectable results, and many more models are quickly forgotten that have strong statistics but just do not tell as clean a story. The classical theory of hypothesis testing, its Bayesian alternative, or the underlying hypothesis-testing view of the philosophy of science are miserable descriptions of the way 281

CHAPTER 16

WHICH METHOD ?

science in general and economics in particular proceed from theory to theory. And this is probably a good thing too. Given the non-experimental nature of our data, the inevitable fishing biases of many researchers examining the same data, and the unavoidable fact that our theories are really quantitative parables more than literal descriptions of the way the data are generated, the way the profession settles on new theories makes a good deal of sense. Classical statistics requires that nobody ever looked at the data before specifying the model. Yet more regressions have been run than there are data points in the CRSP database. Bayesian econometrics can in principle incorporate the information of previous researchers, yet it never applied in this way – each study starts anew with a “uninformative” prior. Statistical theory draws a sharp distinction between the model – which we know is right; utility is exactly power; and the parameters which we estimate. But this distinction isn’t true; we are just as uncertain about functional forms as we are about parameters. A distribution theory at bottom tries to ask an unknowable question: If we turned the clock back to 1947 and reran the postwar period 1000 times, in how many of those alternative histories would (say) the average S&P500 return be greater than 9%? It’s pretty amazing in fact that a statistician can purport to give any answer at all to such a question, having observed only one history. These paragraphs do not contain original ideas, and they mirror changes in the philosophy of science more broadly. 50 years ago, the reigning philosophy of science focused on the idea that scientists provide rejectable hypotheses. This idea runs through philosophical writings exemplified by Popper (1959), classical statistical decision theory, and mirrored in economics by Friedman (1953) However, this methodology contains an important inconsistency. Though researchers are supposed to let the data decide, writers on methodology do not look at how actual theories evolved. It was, as in Friedman’s title, a “Methodology of positive economics,” not a “positive methodology of economics.” Why should methodology be normative, a result of philosophical speculation, and not an empirical discipline like everything else. In a very famous book, Kuhn (1970) looked at the actual history of scientific revolutions, and found that the actual process had very little to do with the formal methodology. McCloskey (1983, 1998) has gone even further, examining the “rhetoric” of economics; the kinds of arguments that persuaded people to change their minds about economic theories. Needless to say, the largest t-statistic did not win! Kuhn’s and especially McCloskey’s ideas are not popular in the finance and economics professions. Precisely, they are not popular in how people talk about their work, though they describe well how people actually do their work. Most people in the fields cling to the normative, rejectable-hypothesis view of methodology. But we need not suppose that they would be popular. The ideas of economics and finance are not popular among the agents in the models. How many stock market investors even know what a random walk or the CAPM is, let alone believing those models have even a grain of truth? Why should the agents in the models of how scientific ideas evolve have an intuitive understanding of the models? “As if” rationality can apply to us as well! Philosophical debates aside, a researcher who wants his ideas to be convincing, as well as right, would do well to study how ideas have in the past convinced people, rather than just study a statistical decision theorist’s ideas about how ideas should convince people. Kuhn, 282

and, in economics, McCloskey have done that, and their histories are worth reading. In the end, statistical properties may be a poor way to choose statistical methods. Summary

The bottom line is simple: It’s ok to do a first stage or simple GMM estimate rather than an explicit maximum likelihood estimate and test. Many people (and, unfortunately, many journal referees) seem to think that nothing less than a full maximum likelihood estimate and test is acceptable. This section is long in order to counter that impression; to argue that at least in many cases of practical importance, a simple first stage GMM approach, focusing on economically interpretable moments, can be adequately efficient, robust to model misspecifications, and ultimately more persuasive.

283

PART III Bonds and options

284

The term structure of interest rates and derivative pricing use closely related techniques. As you might have expected, I present both issues in a discount factor context. All models come down to a specification of the discount factor. The discount factor specifications in term structure and option pricing models are quite simple. So far, we have focused on returns, which reduces the pricing problem to a one-period or instantaneous problem. Pricing bonds and options forces us to start thinking about chaining together the one-period or instantaneous representations to get a prediction for prices of longlived securities. Taking this step is very important, and I forecast that we will see much more multiperiod analysis in stocks as well, studying price and stream of payoffs rather than returns. This step rather than the discount factor accounts for the mathematical complexity of some term structure and option pricing models. There are two standard ways to go from instantaneous or return representations to prices. First, we can chain the discount factors together, finding from a one period discount factor mt,t+1 a long-term discount factor mt,t+j = mt,t+1 mt+1,t+2 ...mt+j−1 mt+j that can price a j period payoff. In continuous time, we will find the discount factor increments dΛ satisfy the instantaneous pricing equation 0 = Et [d (ΛP )], and then solve its stochastic differential equation to find its level Λt+j in order to price a j− period payoff as Pt = Et [Λt+j /Λt xt+j ]. Second, we can chain the prices together. Conceptually, this is the same as chaining returns Rt,t+j = Rt,t+1 Rt+1,t+2 ..Rt+j−1,t+j instead of chaining together the discount factors. From 0 = Et [d (ΛP )], we find a differential equation for the prices, and solve that back. We’ll use both methods to solve interest rate and option pricing models.

285

Chapter 17.

Option pricing

Options are a very interesting and useful set of instruments, as you will see in the background section. In thinking about their value, we will adopt an extremely relative pricing approach. Our objective will be to find out a value for the option, taking as given the values of other securities, and in particular the price of the stock on which the option is written and an interest rate.

17.1 17.1.1

Background Definitions and payoffs

A call option gives you the right to buy a stock for a specified strike price on a specified expiration date. The call option payoff is CT = max(ST − X, 0).

Portfolios of options are called strategies. A straddle – a put and a call at the same strike price – is a bet on volatility Options allow you to buy and sell pieces of the return distribution. Before studying option prices, we need to start by understanding option payoffs. A call option gives you the right, but not the obligation, to buy a stock (or other “underlying” asset) for a specified strike price (X) on (or before) the expiration date (T). European options can only be exercised on the expiration date. American options can be exercised anytime before as well as on the expiration date. I will only treat European options. A put option gives the right to sell a stock at a specified strike price on (or before) the expiration date. I’ll use the standard notation, C = Ct = CT = S = St = ST = X=

call price today call payoff = value at expiration (T). stock price today stock price at expiration strike price

Our objective is to find the price C . The general framework is (of course) C = E(mx) where x denotes the option’s payoff. The option’s payoff is the same things as its value at expiration. If the stock has risen above the strike price, then the option is worth the difference between stock and strike. If the stock has fallen below the strike price, it expires worthless. 286

S ECTION 17.1

BACKGROUND

Thus, the option payoff is ½

ST − X if ST ≥ X if ST ≤ X 0 = max(ST − X, 0).

Call payoff = CT

A put works the opposite way: It gains value as the stock falls below the strike price, since the right to sell it at a high price is more and more valuable. Put payoff = PT = max(X − ST , 0). It’s easiest to keep track of options by a graph of their value as a function of stock price. Figure 31 graphs the payoffs from buying calls and puts, and the corresponding short positions, which are called writing call and put options. One of the easiest mistakes to make is to confuse the payoff, with the profit, which is the value at expiration less the cost of buying the option. I drew in profit lines, payoff - cost, to emphasize this difference. Call

Put Payoff

Profit

ST

ST

Write Call

Write Put ST

ST

Figure 31. Payoff diagrams for simple option strategies. Right away, you can see some of the interesting features of options. A call option allows you a huge positive beta. Typical at-the-money options (strike price = current stock price) give a beta of about 10; meaning that the option is equivalent to borrowing $10 to invest $11 in the stock. However, your losses are limited to the cost of the option, which is paid upfront. Options are obviously very useful for trading. Imagine how difficult it would be to buy stock on such huge margin, and how difficult it would be to make sure people paid if the bet went 287

CHAPTER 17

O PTION PRICING

bad. Options solve this problem. No wonder that active options trading started only a year or two after the first stocks started trading. The huge beta also means that options are very useful for hedging. If you have a large illiquid portfolio, you can offset the risks very cheaply with options. Finally, options allow you to shape the distribution of returns in interesting and sometimes dangerous ways. For example, if you buy a 20% out of the money put option as well as a stock, you have bought “catastrophe insurance” for your stock portfolio, at what turns out to be a remarkably small price. You cut off the left tail of the return distribution, at a small cost to the mean of the overall distribution. On the other side, by writing out of the money put options, you can earn a small fee year in and year out, only once in a while experiencing a huge loss. You have a large probability of a small gain and a small probability of a large loss. You are providing catastrophe insurance to the market, and it works much like, say, writing earthquake insurance. The distribution of returns from this strategy is extremely non-normal, and thus statistical evaluation of its properties will be difficult. This strategy is tempting to a portfolio manager who is being evaluated only by the statistics of his achieved return. If he writes far out of the money options in addition to investing in an index, the chance of beating the index for one or even five years is extremely high. If the catastrophe does happen and he loses a billion dollars or so, the worst you can do is fire him. (Your contract with the manager is always a call option.) This is why portfolio management contracts are not purely statistical, but also write down what kind of investments can and cannot be made. Straddle

Call payoff Put payoff P+C Straddle profit

Stock Price

Make money if stock ends up here

Figure 32. Payoff diagram for a straddle.

288

S ECTION 17.1

BACKGROUND

Portfolios of put and call options are called strategies, and have additional interesting properties. Figure 32 graphs the payoff of a straddle, which combines a put and call at the same strike price. This strategy pays off if the stock goes up or goes down. It loses money if the stock does not move. Thus the straddle is a bet on volatility. Of course, everyone else understands this, and will bid the put and call prices up until the straddle earns only an equilibrium rate of return. Thus, you invest in a straddle if you think that stock volatility is higher than everyone else thinks it will be. Options allow efficient markets and random walks to operate on the second and higher moments of stocks as well as their overall direction! You can also see quickly that volatility will be a central parameter in option prices. The higher the volatility, the higher both put and call prices. More generally, by combining options of various strikes, you can buy and sell any piece of the return distribution. A complete set of options – call options on every strike price – is equivalent to complete markets, i.e., it allows you to form payoffs that depend on the terminal stock price in any way; you can form any payoff of the form f (ST ).

289

CHAPTER 17 17.1.2

O PTION PRICING

Prices: one period analysis

I use the law of one price – existence of a discount factor – and no-arbitrage – existence of a positive discount factor – to characterize option prices. The results are: 1) Put-call parity: P = C − S + X/Rf . 2) Arbitrage bounds, best summarized by Figure 34 3) The proposition that you should never exercise an American call option early on a stock that pays no dividends. The arbitrage bounds are a linear program, and this procedure can be used to find them in more complex situations where clever identification of arbitrage portfolios may fail. We have a set of interesting payoffs. Now what can we say about their prices – their values at dates before expiration? Obviously, p = E(mx) as always. We have learned about x, now we have to think about m. We can start by imposing little structure – the law of one price and the absence of arbitrage, or, equivalently, the existence of some discount factor or a positive discount factor. . In the case of options, these two principles actually do tell you a good deal about the option price. Put-Call parity

The law of one price, or the existence of some discount factor that prices stock, bond, and a call option, allows us to deduce the value of a put in terms of the price of the stock, bond, and call. Consider the following two strategies: 1) Hold a call, write a put, same strike price. 2) Hold stock, promise to pay X. The payoffs of these two strategies are the same, as shown in Figure 33. Equivalently, the payoffs are related by PT = CT − ST + X.

Thus, so long as the law of one price holds, the prices of left and right hand sides must be equal. Applying E(m·) to both sides for any m, P = C − S + X/Rf .

(The price of ST is S . The price of the payoff X is X/Rf .) Arbitrage bounds

If we add the absence of arbitrage, or equivalently the restriction that the discount factor must be positive, we can deduce bounds on the call option price without needing to know the put price. In this case, it is easiest to cleverly notice arbitrage portfolios – situations in which 290

S ECTION 17.1

BACKGROUND

Buy call X

ST

=

+ Write put

Buy stock =

+ Borrow strike

-X

Figure 33. Put call parity. portfolio A dominates portfolio B. Then, either directly from the definition of no-arbitrage or from A > B, m > 0 ⇒ E(mA) > E(mB), you can deduce that the price of A must be greater than the price of B. The arbitrage portfolios are 1. 2.

3.

CT > 0 ⇒ C > 0. The call payoff is positive so the call price must be positive. CT ≥ ST − X ⇒ C ≥ S − X/Rf . The call payoff is better than hold stock - pay strike for sure, so the call price is greater than holding the stock and borrowing the strike, i.e. promising to pay it for sure. CT ≤ ST ⇒ C ≤ S . The call payoff is worse than stock payoff (because you have to pay the strike price. Thus, the call price is less than stock price.

Figure 34 summarizes these arbitrage bounds on the call option value. We have gotten somewhere – we have restricted the range of the option prices. However, the arbitrage bounds are too large to be of much use. Obviously, we need to learn more about the discount factor than pure arbitrage or m > 0 will allow. We could retreat to economic models, e.g. use the CAPM or other explicit discount factor model. Option pricing is famous because we don’t have to do that. Instead, if we open up dynamic trading–the requirement that the discount factor price the stock and bond at every date to expiration–it turns out that we can sometimes determine the discount factor and hence the option value precisely. This presentation is unsettling for two reasons. First, you may worry that you will not be clever enough to dream up dominating portfolios in more complex circumstances. Second, you may worry that we have not dreamed up all of the arbitrage portfolios in this circumstance. Perhaps there is another one lurking out there, which would reduce the unsettling large size of the bounds. It leaves us hungry for a constructive technique for finding arbi291

CHAPTER 17

O PTION PRICING

C Call value Today Call value C In here

X/Rf

S Stock value today

Figure 34. Arbitrage bounds for a call option trage bounds that would be guaranteed to work in general situations, and to find the tightest arbitrage bound. We want to know Ct = E(mt,T xcT ) wherexcT = max(ST −X, 0) denotes the call payoff, we want to use information in the observed stock and bond prices to learn about the option price, and we want to impose the absence of arbitrage. We can capture this search with the following problem: (214)

f max Ct = Et (mxC T ) s.t. m > 0, St = Et (mST ), 1 = Et (mR ) m

and the corresponding minimization. The first constraint implements absence of arbitrage. The second and third use the information in the stock and bond price to learn what we can about the option price. Write 17.214 out in state notation, max Ct =

{m(s)}

X

π(s)m(s)xC T (s) s.t. m(s) > 0, St =

s

X s

π(s)m(s)ST (s), 1 =

X

π(s)m(s)Rf

s

This is a linear program – a linear objective and linear constraints. In situations where you do not know the answer, you can calculate arbitrage bounds – and know you have them all – by solving this linear program. I don’t know how you would begin to check that for every portfolio A whose payoff dominates B , the price of A is greater than the price of B . The discount factor method lets you construct the arbitrage bounds. 292

S ECTION 17.2

BLACK -S CHOLES FORMULA

Early exercise

By applying the absence of arbitrage, we can show quickly that you should never exercise an American call option on a stock that pays no dividends before the expiration date. This is a lovely illustration because such a simple principle leads to a result that isn’t initially obvious. Follow the table: Payoffs Price Rf > 1

CT = max(ST − X, 0) C C

≥ ≥ ≥

ST − X S − X/Rf S−X

S − X is what you get if you exercise now. The value of the call is greater than this value, because you can delay paying the strike, and because exercising early loses the option value. Put-call parity let us concentrate on call options; this fact lets us concentrate on European options.

17.2

Black-Scholes formula

Write a process for stock and bond, then use Λ∗ to price the option. The Black-Scholes formula 17.220 results. You can either solve for the finite-horizon discount factor ΛT /Λ0 and find the call option price by taking the expectation C0 = E0 (ΛT /Λ0 xC T ), or you can find a differential equation for the call option price and solve it backward. Our objective, again, is to learn as much as we can about the value of an option, given the value of the underlying stock and bond. The one-period analysis led only to arbitrage bounds, at which point we had to start thinking about discount factor models. Now, we allow intermediate trading, which means we really are thinking about dynamic multiperiod asset pricing. The standard approach to the Black-Scholes formula rests on explicitly constructing portfolios: at each date we cleverly construct a portfolio of stock and bond that replicates the instantaneous payoff of the option; we reason that the price of the option must equal the price of the replicating portfolio. Instead, I follow the discount factor approach. The law of one price is the same thing as the existence of a discount factor. Thus, rather than construct law-of-one-price replicating portfolios, construct at each date a discount factor that prices the stock and bond, and use that discount factor to price the option. the discount factor approach shows how thinking of the world in terms of a discount factor is equivalent in the result and as easy in the calculation as other approaches. This case shows some of the interest and engineering complexity of continuous time models. Though at each instant the analysis is trivial law of one price, chaining it together 293

CHAPTER 17

O PTION PRICING

over time is not trivial. The call option payoff is CT = max(ST − X, 0)

where X denotes the strike price X and ST denotes the stock price on the expiration date T . The underlying stock follows dS = µdt + σdz. S

There is also a money market security that pays the real interest rate rdt. We want a discount factor that prices the stock and bond. All such discount factors are of the form m = x∗ + w, E(xw) = 0. In continuous time, all such discount factors are of the form (µ − r) dΛ = −rdt − dz − σ w dw; E(dwdz) = 0. Λ σ

(You can check that this set of discount factors does in fact price the stock and interest rate, or take a quick look back at section 4.3.) Now we price the call option with this discount factor, and show that the Black-Scholes equation results. Importantly, the choice of discount factor via choice of σ w dw turns out to have no effect on the resulting option price. Every discount factor that prices the stock and interest rate gives the same value for the option price. The option is therefore priced using the law of one price alone. There are two paths to follow. Either we solve the discount factor forward, and then find the call value by C = E(mxC ), or we characterize the price path and solve it backwards from expiration. 17.2.1

Method 1: Price using discount factor

Let us use the discount factor to price the option directly: ¾ Z ½ ΛT ΛT C0 = Et max (ST − X, 0) = max (ST − X, 0) df (ΛT , ST ) Λt Λt where ΛT and ST are solutions to dS S dΛ Λ

= µdt + σdz = −rdt −

(17.215)

µ−r dz − σ w dw. σ

I simplify the algebra by setting σ w dw to zero, anticipating that it does not matter. You can 294

S ECTION 17.2

BLACK -S CHOLES FORMULA

reason that since S does not depend on dw, CT depends only on ST , so C will only depend on S , and dw will have no effect on the answer. If this isn’t good enough, a problem asks you to include the dw, trace through the remaining steps and verify that the answer does not in fact depend on dw. “Solving” a stochastic differential equation such as (17.215) means finding the distribution of the random variables ST and ΛT , using information as of date 0. This is just what we do with difference equations. For example, if we solve xt+1 = ρxt + εt+1 with ε normal P forward to xT = ρT x0 + Tj=1 ρT −j εj , we have expressed xT as a normally distributed ranP dom variable with mean ρT x0 and standard deviation Tj=1 ρ2(T −j) . In the continuous time case, it turns out that we can solve some nonlinear specifications as well. Integrals of dz give us shocks, as integrals of dt give us deterministic functions of time. We can find analytical expressions for the solutions equations of the form (17.215). Start with the stochastic differential equation dY = µY dt + σ Y dz. Y

(216)

Write d ln Y =

µ ¶ 1 1 dY 1 2 2 − σ dY = µ − dt + σ Y dZ Y Y 2Y2 2 Y

Integrating from 0 to T , (17.216) has solution µ ¶ σ2Y ln YT = ln Y0 + µY − T + σY (zT − z0 ) 2

(217)

zT − z0 is a normally distributed random variable with´mean zero and variance T . Thus, ln Y ³ 2 is conditionally normal with mean ln Y0 + µY − σ2Y T and variance σ2Y T. You can check this solution by differentiating it – don’t forget the second derivative terms.

Applying the solution (17.217) to (17.215), we have µ ¶ √ σ2 ln ST = ln S0 + µ − T + σ Tε 2 Ã µ ¶2 ! 1 µ−r µ − r√ ln ΛT = ln Λ0 − r + Tε T− 2 σ σ

(17.218)

where the random variable ε is ε=

zT − z0 √ ∼ N (0, 1) . T

Having found the joint distribution of stock and discount factor, we evaluate the call 295

CHAPTER 17

O PTION PRICING

option value by doing the integral corresponding to the expectation, C0

Z

=

∞

ST =X Z ∞

=

ST =X

ΛT (ST − X) df (ΛT , ST ) Λt ΛT (ε) (ST (ε) − X) df (ε) Λt

(17.219)

We know the joint distribution of the terminal stock price ST and discount factor ΛT on the right hand side, so we have all the information we need to calculate this integral. This example has enough structure that we can find an analytical formula. In more general circumstances, you may have to resort to numerical methods. At the most basic level, you can simulate the Λ, S process forward and then take the integral by summing over many such simulations. Doing the integral

Start by breaking up the integral (17.219) into two terms, Z

C0 =

∞

ΛT (ε) ST (ε) df (ε) − Λt

ST =X

Z

∞

ST =X

ΛT (ε) X df (ε) . Λt

ST and ΛT are both exponential functions of ε. The normal distribution is also an exponential function of ε. Thus, we can approach this integral exactly as we approach the expectation of a lognormal; we can merge the two exponentials in ε into one term, and express the result as integrals against a normal distribution. Here we go. Plug in (17.218) for ST , ΛT , and simplify the exponentials in terms of ε, C0

=

Z

∞

´ ³ √ 2 T − µ−r − r+ 12 ( µ−r Tε σ ) σ

e

ST =X Z ∞

−X Z = S0 −X

´ ³ √ 2 T − µ−r − r+ 12 ( µ−r Tε σ ) σ

e

ST =X h ∞

e

ST =X Z ∞

1 2 S0 e(µ− 2 σ )T +σ

ST =X

´ ³ √ 2 T − µ−r − r+ 12 ( µ−r Tε σ ) σ

Now add the normal distribution formula for f(ε), f (ε) =

1 − 1 ε2 e 2 . 2π

296

f (ε)dε

f (ε) dε

³ ´i √ 2 µ−r− 12 σ2 +( µ−r T +(σ− µ−r σ ) σ ) Tε

e

√ Tε

f (ε)

f(ε)dε

S ECTION 17.2

BLACK -S CHOLES FORMULA

The result is C0

Z ∞ h ³ ´i √ 2 1 2 1 µ−r− 12 σ 2 +( µ−r T +(σ− µ−r σ ) σ ) T ε− 2 ε √ = e dε S0 2π ST =X Z ∞ i h √ 2 1 T − µ−r − r+ 12 ( µ−r T ε− 12 ε2 σ ) σ −√ X e dε 2π S =X Z ∞T √ 2 µ−r 1 1 = √ S0 e− 2 [ε−(σ− σ ) T ] dε 2π ST =X Z ∞ √ 2 µ−r 1 1 −rT − √ Xe e− 2 (ε+ σ T ) dε. 2π ST =X

Notice that the integrals have the form of a normal distribution with nonzero mean. The lower bound ST = X is, in terms of ε, µ ¶ √ σ2 ln X = ln ST = ln S0 + µ − T + σ Tε 2 ³ ln X − ln S0 − µ − √ ε= σ T

σ2 2

´

T .

Finally, we can express definite integrals against a normal distribution by the cumulative normal, Z ∞ 2 1 1 √ e− 2 (ε−µ) dε = Φ (µ − a) 2π a i.e., Φ() is the area under the left tail of the normal distribution.

C0

´ ³ 2 ¶ ln X − ln S0 − µ − σ2 T µ √ µ − r √ + σ− = S0 Φ − T σ σ T −t ´ ³ 2 ln X − ln S0 − µ − σ2 T √ µ−r √ −Xe−r(T −t) Φ − T − σ σ T

Simplifying, we get the Black-Scholes formula Ã Ã ¤ ! ¤ ! £ £ ln S0 /X + r + 12 σ 2 T ln S0 /X + r − 12 σ 2 T −rT √ √ C0 = S0 Φ Φ − Xe . (220) σ T σ T 297

CHAPTER 17 17.2.2

O PTION PRICING

Method 2: Derive Black-Scholes differential equation

Rather than solve the discount factor forward and then integrate, we can solve the price backwards from expiration. The instantaneous or expected return formulation of a pricing model amounts to a differential equation for prices. Guess that the solution for the call price is a function of stock price and time to expiration, Ct = C(S, t). Use Ito’s lemma to find derivatives of C(S, t), 1 dC = Ct dt + CS dS + CSS dS 2 2 · ¸ 1 dC = Ct + CS Sµ + CSS S 2 σ 2 dt + CS Sσdz 2

Plugging into the basic asset pricing equation 0 = Et (dΛC) = CEt dΛ + ΛEt dC + Et dΛdC,

using Et (dΛ/Λ) = −rdt and canceling Λdt, we get 1 0 = −rC + Ct + CS Sµ + CSS S 2 σ2 − S (µ − r) CS 2

or, 1 0 = −rC + Ct + SrCS + CSS S 2 σ 2 . 2

(221)

This is the Black-Scholes differential equation for the option price. We now know a differential equation for the price function C(S, t). We know the value of this function at expiration, C(ST , T ) = max(ST − X, 0). The remaining task is to solve this differential equation backwards through time. Conceptually, and numerically, this is easy. Express the differential equation as −

∂C(S, t) 1 ∂ 2 C(S, t) 2 2 ∂C(S, t) = −rC(S, t) + Sr + S σ . ∂t ∂S 2 ∂S 2

At any point in time, you know the values of C(S, t) for all S – for example, you can store them on a grid for S . Then, you can take the first and second derivatives with respect to S and form the quantity on the right hand side at each value of S . Now, you can find the option price at any value of S , one instant earlier in time. This differential equation, solved with boundary condition C = max {ST − X, 0}

has an analytic solution – the familiar formula (17.220). One standard way to solve differential equations is to guess and check; and by taking derivatives you can check that (17.220) 298

S ECTION 17.3

P ROBLEMS

does satisfy (17.221). Black and Scholes solved the differential equation with a fairly complicated Fourier transform method. The more elegant Feynman-Kac solution amounts to showing that solutions of the partial differential equation (17.221) can be represented as integrals of the form that we already derived independently as in (17.219). (See Duffie 1992 p.87)

17.3 1.

2.

Problems

We showed that you should never exercise an American call early if there are no dividends. Is the same true for American puts, or are there circumstances in which it is optimal to exercise American puts early? Retrace the steps in the integral derivation of the Black-Scholes formula and show that the dw does not affect the final result.

299

Chapter 18. Option pricing without perfect replication 18.1

On the edges of arbitrage

The Black-Scholes formula is justly famous and launched a thousand techniques for option pricing. The principle of no-arbitrage pricing is obvious, but its application leads to many subtle and unanticipated pricing relationships. However, in many practical situations, the law of one price arguments that we used in the Black-Scholes formula break down. If options really were redundant, it is unlikely that they would be traded as separate assets. It really is easy to synthesize forward rates from zero-coupon bonds, and forward rates are not separately traded or quoted. We really cannot trade continuously, and trying to do so would drown a strategy in transactions costs. As a practical example, at the time of the 1987 stock market crash, several prominent funds were trying to follow “portfolio insurance” strategies, essentially synthesizing put options by systematically selling stocks as prices declined. During the time of the crash, however, they found that the markets just dried up – they were unable to sell as prices plummeted. We model this situation mathematically as a Poisson jump, a discontinuous movement in prices. In the face of such jumps the call option payoff is not perfectly hedged by a portfolio of stock and bond, and cannot be priced as such. Generalizations of the stochastic setup lead to the same result. If the interest rate or stock volatility are stochastic, we do not have securities that allow us to perfectly hedge the corresponding shocks, so the law of one price again breaks down. In addition, many options are written on underlying securities that are not traded, or not traded continually and with sufficient liquidity. Real options in particular – the option to build a factory in a particular location – are not based on a tradeable underlying security, so the logic behind Black-Scholes pricing does not apply. Executives are specifically forbidden to short stock in order to hedge executive options. Furthermore, applications of option pricing formulas to trading activities seem to suffer a strange inconsistency. We imagine that the stock and bond are perfectly priced and perfectly liquid – available for perfect hedging. Then, we search for options that are priced incorrectly as trading opportunities. If the options can be priced wrong, why can’t the stock and bond be priced wrong? We should treat all assets symmetrically in evaluating trading opportunities. Trading opportunities also involve risk, and a theory that pretends they are arbitrage opportunities does not help much to quantify that risk. In all of these situations, an unavoidable “basis risk” creeps in between the option payoff and the best possible hedge portfolio. Holding the option entails some risk, and the value of the option depends on the “market price” of that risk – the covariance of the risk with an 300

S ECTION 18.2

O NE - PERIOD GOOD DEAL BOUNDS

appropriate discount factor. Nonetheless, we would like not to give up and go back to the consumption-based model, factor models, or other “absolute” methods that try to price all assets. We are still willing to take as given the prices of lots of assets in determining the price of an option, and in particular assets that will be used to hedge the option. We can form an “approximate hedge” or portfolio of basis assets “closest to” the focus payoff, and we can hedge most of the option’s risk with that approximate hedge. Then, the uncertainty about the option payoff is reduced only to figuring out the price of the residual. In addition, since the residuals are small, we might be able to say a lot about option prices with much weaker restrictions on the discount factor than those suggested by absolute models. In this chapter, I survey “good deal” option price bounds, a technique that Jesus SaáRequejo and I (1999) advocated for this situation. The good deal bounds amount to systematically searching over all possible assignments of the “market price of risk” of the residual, constraining the total market price of risk to a reasonable value, and imposing no arbitrage opportunities, to find upper and lower bounds on the option price. It is not equivalent to pricing options with pure Sharpe ratio arguments. The concluding section of this chapter surveys some alternative and additional techniques.

18.2

One-period good deal bounds

We want to price the payoff xC , for example, xC = max(ST − K, 0) for a call option. We have in hand a N −dimensional vector of basis payoffs x, whose prices p we can observe, for example the stock and bond. The good deal bound finds the minimum and maximum value of xC by searching over all positive discount factors that price the basis assets and have limited volatility: C = max E(mxC ) s.t. p = E(mx), m ≥ 0, σ2 (m) ≤ h/Rf {m}

(222)

The corresponding minimization yields the lower bound C . This is a one-period discretetime problem. The Black-Scholes formula does not apply because you can’t trade between the price and payoff periods. The first constraint on the discount factor imposes the price of the basis assets. We want to do as much relative pricing as possible; we want to extend what we know about the prices of x to price xC , without worrying about where the prices of x come from. The second constraint imposes the absence of arbitrage. This problem without the last constraint yields the arbitrage bounds that we studied in section 2. In most situations, the arbitrage bounds are too wide to be of much use. The last is an additional constraint on discount factors, and the extra content of gooddeal vs. arbitrage bounds. It is a relatively weak restriction. We could obtain closer bounds 301

C HAPTER 18

O PTION PRICING WITHOUT PERFECT REPLICATION

on prices with more information about the discount factor. In particular, if we know the correlation of the discount factor with the payoff xC we could price the option a lot better! As m > 0 means that no portfolios priced by m may display an arbitrage opportunity, σ2 (m) ≤ h/Rf means that no portfolio priced by m may have a Sharpe ratio greater than h. Recall E(mRe ) = 0 implies E(m)E(Re ) = −ρσ(m)σ(Re ) and |ρ| ≤ 1.

It is a central advantage of a discount factor approach that we can easily impose both the discount factor volatility constraint and positivity, merging the lessons of factor models and option pricing models. The prices and payoffs generated by discount factors that satisfy both m ≥ 0 and σ(m) ≤ h/Rf do more than rule out arbitrage opportunities and high Sharpe ratios. I’ll treat the case that there is a riskfree rate, so we can write E(m) = 1/Rf . In this case, it is more convenient to express the volatility constraint as a second moment, so the bound (18.222) becomes ¡ ¢ C = min E (m xc ) s.t. p = E (mx) , E m2 ≤ A2 , m ≥ 0, {m}

(223)

where A2 ≡ (1 +h2 )/Rf 2 . The problem is a standard minimization with two inequality constraints. Hence we find a solution by trying all the combinations of binding and nonbinding constraints, in order of their ease of calculation. 1) Assume the volatility constraint binds and the positivity constraint is slack. This one is very easy to calculate, since we will find analytic formulas for the solution. If the resulting discount factor m is nonnegative, this is the solution. If not, 2) assume that the volatility constraint is slack and the positivity constraint binds. This is the classic arbitrage bound. Find the minimum variance discount factor that generates the arbitrage bound. If this discount factor satisfies the volatility constraint, this is the solution. If not, 3) solve the problem with both constraints binding. 18.2.1

Volatility constraint binds, positivity constraint is slack

If the positivity constraint is slack, the problem reduces to ¡ ¢ C = min E(m xc ) s.t. p = E (mx) , E m2 ≤ A2 . {m}

(224)

We could solve this problem directly, choosing m in each state with Lagrange multipliers on the constraints. But as with the mean-variance frontier it is much more elegant to set up orthogonal decompositions and then let the solution pop out. Figure 35 describes the idea. X denotes the space of payoffs of portfolios of the basis assets x, a stock and a bond in the classic Black-Scholes setup. Though graphed as a line, X is typically a larger space. We know all prices in X , but the payoff xc that we wish to value does not lie in X . 302

S ECTION 18.2

O NE - PERIOD GOOD DEAL BOUNDS

Start by decomposing the focus payoff xc into an approximate hedge x ˆc and a residual w, xc x ˆc w

= x ˆc + w, ≡ proj(xc |X) = E(xc x0 )E(xx0 )−1 x, ˆc . ≡ xc − x

(18.225)

We know the price of x ˆc . We want to bound the price of the residual w to learn as much as we can about the price of xc . All discount factors that price x – that satisfy p = E(mx)– lie in the plane through x∗ . As we sweep through these discount factors, we generate any price from −∞ to ∞ for the residual w and hence payoff xc . All positive discount factors m > 0 lie in the intersection of the m plane and the positive orthant – the triangular region. Discount factors m in this range generate a limited range of prices for the focus payoff – the arbitrage bounds. Since second moment defines distance in Figure 35, the set of discount factors that satisfies the volatility constraint E(m2 ) ≤ A2 lies inside a sphere around the origin. The circle in Figure 35 shows the intersection of this sphere with the set of discount factors. This restricted range of discount factors will produce a restricted range of values for the residual w and hence a restricted range of values for the focus payoff xc . In the situation I have drawn, the positivity constraint is slack, since the E(m2 ) ≤ A2 circle lies entirely in the positive orthant.

We want to find the discount factors in the circle that minimize or maximize the price of the residual w. The more a discount factor points in the w direction, the larger a price E(mw) it assigns to the residual. Obviously, the discount factors that maximize or minimize the price of w point as much as possible towards and away from w. If you add any movement ε orthogonal to w, this increases discount factor volatility without changing the price of w. Hence, the discount factor that generates the lower bound is m = x∗ − vw

(226)

where v=

s

A2 − E(x∗2 ) E(w2 )

(227)

is picked to just satisfy the volatility constraint. The bound is C = E(mxc ) = E(x∗ xc ) − vE(w2 )

(228)

The upper bound is given by v = −v

The first term in equation (18.228) is the value of the approximate hedge portfolio, and can be written several ways, including E(x∗ xc ) = E(x∗ x ˆc ) = E(mˆ xc )

303

(229)

C HAPTER 18

O PTION PRICING WITHOUT PERFECT REPLICATION

xc

X

w

x xˆ c

x * + vw

E (m 2 ) < A2

x*

x * −vw

m>0

Figure 35. Construction of a discount factor to solve the one-period good deal bound when the positivity constraint is slack. for any discount factor m that prices basis assets. (Don’t forget, E(xy) = E[x proj(y|X)].) The second term in equation (18.228) is the lowest possible price of the residual w consistent with the discount factor volatility bound: vE(w2 ) = E(vw w) = E[(x∗ + vw)w] = E(mw).

For calculations you can substitute the definitions of x∗ and w in equation (18.228) to obtain an explicit, if not very pretty, formula: C = p0 E(xx0 )−1 E(xxc ) −

p p A2 − p0 E(xx0 )−1 p E(xc2 ) − E(xc x0 )E(xx0 )−1 E(xxc ). (230)

304

S ECTION 18.2

O NE - PERIOD GOOD DEAL BOUNDS

The upper bound C is the same formula with a + sign in front of the square root. Using (18.226), check whether the discount factor is positive in every state of nature. If so, this is the good-deal bound, and the positivity constraint is slack. If not, proceed to the next step. If you prefer an algebraic and slightly more formal argument, start by noticing that any discount factor that satisfies p = E(mx) can be decomposed as m = x∗ + vw + ε

where E(x∗ w) = E(x∗ ε) = E(wε). Check these properties from the definition of w and ε; this is just like R = R∗ + wRe∗ + n. Our minimization problem is then min E(mxc ) s.t. E(m2 ) ≤ A2 ¡ ¢ min E [((x∗ + vw + ε) (ˆ xc + w)] s.t. E x∗2 + v2 E(w2 ) + E(ε2 ) ≤ A2 {v,ε} ¡ ¢ min E(x∗ x ˆc ) + vE(w2 ) s.t. E x∗2 + v2 E(w2 ) + E(ε2 ) ≤ A2 {v,ε}

{v,ε}

q 2 −E(x∗2 ) The solution is ε = 0 and v = ± A E(w . 2) 18.2.2

Both constraints bind

Next, I find the bounds when both constraints bind. Though this is the third step in the procedure, it is easiest to describe this case first. Introducing Lagrange multipliers, the problem is C = min

max E (m xc ) + λ0 [E (mx) − p] +

{m>0} {λ,δ>0}

¤ δ £ ¡ 2¢ E m − A2 2

The first order conditions yield a discount factor that is a truncated linear combination of the payoffs, ¶ · c µ c ¸+ x + λ0 x x + λ0 x ,0 = − m = max − . δ δ

(231)

The last equality defines the []+ notation for truncation. In finance terms, this is a call option with zero strike price. You can derive (18.231) by introducing a Kuhn-Tucker multiplier π(s)ν(s) on m > 0 305

C HAPTER 18

O PTION PRICING WITHOUT PERFECT REPLICATION

and taking partial derivatives with respect to m in each state, # " X X 0 c C = min π(s)m(s)x (s) + λ π(s)m(s)x(s) − p {m}

s

s

# X δ X 2 2 + π(s)m(s) − A + π(s)ν(s)m(s) 2 s s "

1 ∂ : xc (s) + λ0 x(s) + δm(s) + ν(s) = 0 π(s) ∂s

(232)

If the positivity constraint is slack, the Kuhn-Tucker multiplier ν(s) is zero, m(s) = −

xc (s) + λ0 x(s) . δ

If the positivity constraint binds, then m(s) = 0, and ν(s) is just enough to make (18.232) hold. In sum, we have (18.231). We could plug expression (18.231) into the constraints, and solve numerically for Lagrange multipliers λ and δ that enforce the constraints. Alas, this procedure requires the solution of a system of nonlinear equations in (λ, δ ), which is often a numerically difficult or unstable problem. Hansen, Heaton and Luttmer (1995) show how to recast the problem as a maximization, which is numerically much easier. Interchanging min and max, C = max

min E(m xc ) + λ0 [E (mx) − p] +

{λ,δ>0} {m>0}

¤ δ £ ¡ 2¢ E m − A2 . 2

(233)

The inner minimization yields the same first order conditions (18.231). Plugging those firstorder conditions into the outer maximization of (18.233) and simplifying, we obtain ( · c ¸+2 ) δ δ x + λ0 x C = max E − − (234) − λ0 p− A2 . 2 δ 2 {λ,δ>0} You can search numerically over (λ, δ) to find the solution to this problem. The upper bound is found by replacing max with min and replacing δ > 0 with δ < 0. 18.2.3

Positivity binds, volatility is slack

If the volatility constraint is slack and the positivity constraint binds, the problem reduces to C = min E (m xc ) s.t. p = E (mx) , m > 0. {m}

306

(235)

S ECTION 18.2

O NE - PERIOD GOOD DEAL BOUNDS

These are the arbitrage bounds. We found these bounds in section 2 for a call option by just being clever. If you can’t be clever, (18.235) is a linear program. We still have to check that the discount factor volatility constraint can be satisfied at the arbitrage bound. Denote the lower arbitrage bound by Cl . The minimum variance (second moment) discount factor that generates the arbitrage bound Cl solves · ¸ µ · ¸¶ p x E(m2 )min = min E(m2 ) s.t =E m , m > 0. Cl xc {m} Using the same conjugate method, this problem is equivalent to o n +2 − 2v0 p−2µCl . E(m2 )min = max −E [− (µxc + v0 x)] {v,µ}

Again, search numerically for (v,µ) to solve this problem. If E(m2 )min ≤ A, Cl is the solution to the good-deal bound; if not we proceed with the case that both constraints are binding described above. 18.2.4

Application to Black-Scholes

The natural first exercise with this technique is to see how it applies in the Black-Scholes world. Keep in mind, this is the Black-Scholes world with no intermediate trading; compare the results to the arbitrage bounds, not to the Black-Scholes formula. Figure 36, taken from Cochrane and Saá-Requejo (1999) presents the upper and lower good-deal bounds for a call option on the S&P500 index with strike price K = $100, and three months to expiration. We used parameter values E(R) = 13%, σ(R) = 16% for the stock index return and an riskfree rate Rf = 5%. The discount factor volatility constraint is twice the historical market Sharpe ratio, h = 2 × E(R − Rf )/σ(R) = 1.0. To take the expectations required in the formula, we evaluated integrals against the lognormal stock distribution. The figure includes the lower arbitrage bounds C ≥ 0, C ≥ K/Rf . The upper arbitrage bound states that C ≤ S , but this 45◦ line is too far up to fit on the vertical scale and still see anything else. As in many practical situations, the arbitrage bounds are so wide that they are of little use. The upper good-deal bound is much tighter than the upper arbitrage bound. For example, if the stock price is $95, the entire range of option prices between the upper bound of $2 and the upper arbitrage bound of $95 is ruled out. The lower good-deal bound is the same as the lower arbitrage bound for stock prices less than about $90 and greater than about $110. In this range, the positivity constraint binds and the volatility constraint is slack. This range shows that it is important to impose both volatility and positivity constraints. Good deal bounds are not just the imposition of low Sharpe ratios on options. (I emphasize it because this point causes a lot of confusion.) The volatility bound alone admits negative prices. A free out of the money call option is like a lottery ticket: it is an arbitrage opportunity, but its expected return/standard deviation ratio is terrible, because the standard deviation is so high. A Sharpe ratio criterion alone will not rule it out. 307

C HAPTER 18

O PTION PRICING WITHOUT PERFECT REPLICATION

Figure 36. Good deal option price bounds as a function of stock price. Options have three months to expiration and strike price K = $100. The bounds assume no trading until expiration, and a discount factor volatility bound h = 1.0 corresponding to twice the market Sharpe ratio. The stock is lognormally distributed with parameters calibrated to an index option. In between $90 and $110, the good-deal bound improves on the lower arbitrage bound. It also improves on a bound that only imposes only the volatility constraint. In this region, both positivity and volatility constraints bind. This fact has an interesting implication: Not all values outside the good-deal bounds imply high Sharpe ratios or arbitrage opportunities. Such values might be generated by a positive but highly volatile discount factor, and generated by another less volatile but sometimes negative discount factor, but no discount factor generates these values that is simultaneously nonnegative and respects the volatility constraint. It makes sense rule out these values. If we know that an investor will invest in any arbitrage opportunity or take any Sharpe ratio greater than h, then we know that his unique marginal utility satisfies both restrictions. He would find a utility-improving trade for values outside the good-deal bounds, even though those values may not imply a high Sharpe ratio, an arbitrage opportunity, or any other simple portfolio interpretation. The right thing to do is to intersect restrictions on the discount factor. Simple portfolio 308

S ECTION 18.3

MULTIPLE PERIODS AND CONTINUOUS TIME

interpretations, while historically important, are likely to fall by the wayside as we add more discount factor restrictions or intersect simple ones.

18.3

Multiple periods and continuous time

Now, on to the interesting case. Option pricing is all about dynamic hedging, even if imperfect dynamic hedging. Good deal bounds would be of little use if we could only apply them to one-period environments. 18.3.1

The bounds are recursive

The central fact that makes good deal bounds tractable in dynamic environments is that the bounds are recursive. Today’s bound can be calculated as the minimum price of tomorrow’s bound, just as today’s option price can be calculated as the value of tomorrow’s option price. To see that the bounds are recursive, consider a two-period version of the problem, C0 =

min

{m1 , m2 }

E0 (m1 m2 xc2 ) s.t.

pt = Et (mt+1 pt+1 ); Et (m2t+1 ) ≤ A2t , mt+1 > 0, t = 0, 1.

(236)

This two period problem is equivalent to a series of one period problems, in which the C0 problem finds the lowest price of the C1 lower bound, C 1 = min E1 (m2 xc2 ) ; C 0 = min E0 (m1 C 1 ) {m2 }

{m1 }

subject to (18.236). Why? The solution to the two-period problem min E0 (m1 E1 (m2 xc )) must minimize E1 (m2 xc ) in each state of nature at time 1. If not, you could lower E1 (m2 xc ) without affecting the constraints, and lower the objective. Note that this recursive property only holds if we impose m > 0. If m1 < 0 were possible we would want to maximize E1 (m2 xc ) in some states of nature. 18.3.2

Basis risk and real options

The general case leads to some dense formulas, so a simple example will let us understand the idea most simply. Let’s value a European call option on an event V that is not a traded asset, but is correlated with a traded asset that can be used as an approximate hedge. This situation is common with real options and nonfinancial options and describes some financial options on illiquid assets. 309

C HAPTER 18

O PTION PRICING WITHOUT PERFECT REPLICATION

The terminal payoff is xcT = max(VT − K, 0).

Model the joint evolution of the traded asset S and the event V on which the option is written as dS S dV V

= µS dt + σS dz, = µV dt + σ V z dz + σ V w dw.

The dw risk cannot be hedged by the S asset, so the market price of dw risk – its correlation with the discount factor – will matter to the option price. We are looking for a discount factor that prices S and rf , has instantaneous volatility A, and generates the largest or smallest price for the option. Hence, it will have the largest loading on dw possible. By analogy with the one period case (18.226), you can quickly see that the discount factor will have the form q dΛ∗ dΛ = ± A2 − h2S dw Λ Λ∗ dΛ∗ = −rdt − hS dz Λ∗ µS − r hS = . σS dΛ∗ /Λ∗ is the familiar analogue to x∗ that prices stock and bond. We add a loading on the orthogonal shock dw just sufficient to satisfy the constraint Et (dΛ2 /Λ2 ) = A2 . One of ± will generate the upper bound, and one will generate the lower bound.

Now that we have the discount factor, the good deal bound is given by ¸ · ΛT C t = Et max(VT − K) . Λt St , Vt , and Λt are all diffusions with constant coefficients. Therefore, ST , VT and ΛT are jointly lognormally distributed, so the double integral defining the expectation is straightforward to perform, and works very similarly to the integral we evaluated to solve the BlackScholes formula in section 2.1. (If you get stuck, see Cochrane and Saá-Requejo 1999 for the algebra.)

The result is C or C = V0 e

ηT

µ ¶ µ ¶ 1 √ 1 √ −rT φ d + σ V T − Ke φ d − σV T 2 2

310

(237)

S ECTION 18.3

MULTIPLE PERIODS AND CONTINUOUS TIME

where φ(·) denotes the left tail of the normal distribution and dV 2 = σ 2V z + σ2V w V2 ln(V0 /K) + (η + r) T √ d ≡ σV T s Ã !# " p A2 − 1 1 − ρ2 σV η ≡ hV − hS ρ − a h2S

σ 2V

≡ Et

µS − r µ −r ; hV ≡ V σS σ µ ¶ V dV dS σV z , ρ ≡ corr = V S σV ½ +1 upper bound . a = −1 lower bound

hS

≡

This expression is exactly the Black-Scholes formula with the addition of the η term. µV enters the formula because the event V may not grow at the same rate as the asset S . Obviously, the correlation ρ between V shocks and asset shocks enters the formula, and as this correlation declines, the bounds widen. The bounds also widen as the volatility constraint A becomes larger relative to the asset Sharpe ratios hS . Market prices of risk

Continuous-time pricing problems are often specified in terms of “market prices of risk” rather than discount factors. This is the instantaneous Sharpe ratio that an asset must earn if it loads on a specific shock. If an asset has a price process P that loads on a shock σdw, then its expected return must be ¶ µ dP dΛ f − r dt = −σEt dw Et P Λ with Sharpe ratio f Et dP P − r dt = −Et λ= σ

µ

¶ dΛ dw . Λ

I have introduced the common notation λ for the market price of risk. Thus, problems are often attacked by making assumptions about λ directly and then proceeding from Et

dP − rf dt = λσ. P

In this language, the market price of stock risk is hS and can be measured by observing the stock, and does not matter when you can price by arbitrage (notice it is missing from the Black-Scholes formula). Our problem comes down to choosing the market price of dw 311

C HAPTER 18

O PTION PRICING WITHOUT PERFECT REPLICATION

risk, which cannot be measured by observing a traded asset, in such a way as qto minimize or maximize the option price, subject to a constraint that the total price of risk 18.3.3

h2S + λ2 ≤ A.

Continuous time

Now, a more systematic expression of the same ideas in continuous time. As in the option pricing case in the last chapter and the term structure case in the next chapter, we will obtain a differential characterization. To actually get prices, we have either to solve the discount factor forward, or to find a differential equation for prices which we solve backward. Basis assets

In place of E(x), E(xx0) etc., model the price processes of an nS -dimensional vector of basis assets by a diffusion, dS = µS (S, V, t)dt + σS (S, V, t)dz; E(dz dz 0 ) = I S

(238)

Rather than complicate£the notation, understand division to operate element by element on ¤ vectors, e.g., dS/S = dS1 /S1 dS2 /S2 · · · . The basis assets may pay dividends at rate D(S, V, t)dt. V represents an nV -dimensional vector of additional state variables that follow

dV = µV (S, V, t)dt + σ V z (S, V, t)dz + σV w (S, V, t)dw; E(dw dw0 ) = I; E(dw dz 0 ) = 0. (239)

This could include a stochastic stock volatility or stochastic interest rate – classic cases in which the Black-Scholes replication breaks down. Again, I keep it simple by assuming there is a risk free rate r(S, V, t)dt. The problem

We want to value an asset that pays continuous dividends at rate xc (S, V, t)dt and with a terminal payment xcT (S, V, T ). Now we must choose a discount factor process to minimize the asset’s value Ct =

min

{Λs , t<s≤T }

Et

Z

T

s=t

Λs c x ds + Et Λt s

µ

ΛT c x Λt T

¶

(240)

subject to the constraints that 1) the discount factor prices the basis assets S, r at each moment in time, 2) the instantaneous volatility of the discount factor process is less than a prespecified value A2 and 3) the discount factor is positive Λs > 0, t ≤ s ≤ T .

One period at a time; differential statement

312

S ECTION 18.3

MULTIPLE PERIODS AND CONTINUOUS TIME

Since the problem is recursive, we can study how to move one step back in time, C t Λt = min Et {Λs }

Z

t+∆t

s=t

or, for small time intervals,

¡ ¢ Λs xcs ds + Et Λt+∆t C t+∆t

C t Λt = min Et {xc ∆t + (C t + ∆C) (Λt + ∆Λ)} . {∆Λ}

Letting ∆t → 0, we can write the objective in differential form, 0=

xct Et [d (ΛC)] dt + min , C {dΛ} ΛC

(241)

subject to the constraints. We can also write (18.241) as dC xct Et + dt − rf dt = − min Et C C {dΛ}

µ

dΛ dC Λ C

¶

.

(242)

Since the second and third terms on the left hand side are fixed, the condition sensibly tells us to find the lowest value C by maximizing the drift of the bound at each date. You should recognize the form of (18.241) and (18.242) as the basic pricing equations in continuous time, relating expected returns to covariance with discount factors. Constraints

Now we express the constraints. As in the discrete time case, we orthogonalize the discount factor in m = x∗ + ε form, and then the solution pops out. Any discount factor that prices the basis assets is of the form dΛ∗ dΛ = ∗ − vdw Λ Λ

(243)

where dΛ∗ Λ∗ µ ˜S

≡ −rdt − µ ˜ 0S Σ−1 S σ S dz ≡ µS +

D − r; ΣS = σS σ0S . S

and v is a 1 × nV matrix. We can add shocks orthogonal to dw if we like, but they will have no effect on the answer; the minimization will say to set such loadings to zero. The volatility constraint is 1 dΛ2 Et 2 ≤ A2 dt Λ

313

C HAPTER 18

O PTION PRICING WITHOUT PERFECT REPLICATION

and hence, using (18.243), vv0 ≤ A2 −

1 dΛ∗2 Et ∗2 = A2 − µ ˜ 0S Σ−1 ˜S . S µ dt Λ

(244)

By expressing the constraints via (18.243) and (18.244), we have again reduced the problem of choosing the stochastic process for Λ to the choice of loadings v on the noises dw with unknown values, subject to a quadratic constraint on vv0 . Since we are picking differentials and have ruled out jumps, the positivity constraint is slack so long as Λ > 0. Market prices of risk

Using equation (18.243), v is the vector of market prices of risks of the dw shocks – the expected return that any asset must offer if its shocks are dw: −

1 E dt

µ

dΛ dw Λ

¶

= v.

Thus, the problem is equivalent to: find at each date the assignment of market prices of risk to the dw shocks that minimizes (maximizes) the focus payoff value, subject to the constraint that the total (sum of squared) market price of risk is bounded by A2 . Now, we’re ready to follow the usual steps. We can characterize a differential equation for the option price that must be solved back from expiration, or we can try to solve the discount factor forward and take an expectation. Solutions: the discount factor and bound drift at each instant

We can start by characterizing the bound’s process, just as the basis assets follow (18.238). This step is exactly the instantaneous analogue of the one-period bound without a positivity constraint, so remember that logic if the equations start to get a bit forbidding. Guess that lower bound C follows a diffusion process, and figure out what the coefficients must look like. Write dC = µC (S, V, t)dt + σCz (S, V, t)dz + σCw (S, V, t)dw. C

(245)

σCz and σ Cw capture the stochastic evolution of the bound over the next instant – the analogues to E(xxc ), etc. that were inputs to the one period problem. Therefore, a differential or moment-to-moment characterization of the bound will tell us µC and dΛin terms of σCz and σ Cw .

Theorem:The lower bound discount factor Λt follows dΛ∗ dΛ = ∗ − v dw Λ Λ

314

(246)

S ECTION 18.3

MULTIPLE PERIODS AND CONTINUOUS TIME

and µC , σ Cz and σCw satisfy the restriction ¶ µ ∗ xc 1 dΛ µC + − r = − Et σCz dz + vσ0Cw C dt Λ∗

(247)

where v=

r

A2 −

σCw 1 dΛ∗2 Et ∗2 q dt Λ σ Cw σ 0

(248)

Cw

The upper bound process C t and discount factor Λt have the same representation with v = −v. This theorem has the same geometric interpretation as shown in Figure 35. dΛ∗ /Λ∗ is the combination of basis asset shocks that prices the basis assets by construction, in analogy to x∗ . The term σCw dw corresponds to the error w, and σCw σ0Cw corresponds to E(w2 ). The proposition looks a little different because now we choose a vector v rather than a number. We could define a residual σCw dw and then the problem would reduce to choosing a number, the loading of dΛ on this residual. It is not convenient to do so in this case since σCw potentially changes over time. In the geometry of Figure 35, the w direction may change over time. The algebraic proof just follows this logic. Proof: Substituting equation (18.243) into the problem (18.241) in order to impose the pricing constraint, the problem is · ¸ µ ¶ µ ∗2 ¶ xc 1 d(Λ∗ C) dC dΛ 0 2 0 = dt + Et s.t. vv ≤ A − Et − min vEt dw . C Λ∗ C C dt Λ∗2 {v}

Using equation (18.245) for dC/C in the last term, the problem is · ¸ µ ∗2 ¶ xc 1 1 d(Λ∗ C) dΛ 0 0 2 E 0= + Et vσ s.t. vv ≤ A − − min . t Cw ∗ C dt Λ C dt Λ∗2 {v}

(249)

This is a linear objective in v with a quadratic constraint. Therefore, as long as σCw 6= 0, the constraint binds and the optimal v is given by (18.248). v = −v gives the maximum since σ Cw σ0Cw > 0. Plugging the optimal value for v in (18.249) gives · ¸ xc 1 d (Λ∗ C) 0= + Et − vσ 0Cw . C dt Λ∗ C For clarity, and exploiting the fact that dΛ∗ does not load on dw, write the middle term as ¶ · ¸ µ ∗ 1 d (Λ∗ C) dΛ 1 Et σCz dz = µC − r + Et dt Λ∗ C dt Λ∗ 315

C HAPTER 18

O PTION PRICING WITHOUT PERFECT REPLICATION

If σ Cw = 0, any v leads to the same price bound. In this case we can most simply ¥ take v = 0. As in the discrete-time case, we can plug in the definition of Λ∗ to obtain explicit, if less intuitive, expressions for the optimal discount factor and the resulting lower bound, q σCw dΛ = −rdt + µ ˜ 0S Σ−1 σ dz − A2 − µ ˜ 0S Σ−1 ˜S q dw (250) S S S µ Λ σ Cw σ 0Cw µC +

xc −r =µ ˜ 0S Σ−1 S σ S σ Cz + C

q q A2 − µ ˜ 0S Σ−1 µ ˜ σCw σ0Cw . S S

(251)

A partial differential equation

Now we are ready to apply the standard method; find a partial differential equation and solve it backwards to find the price at any date. The method proceeds exactly as for the BlackScholes formula: Guess a solution C(S, V, t). Use Ito’s lemma to derive expressions for µC and σ Cz , σCw in terms of the partial derivatives of C(S, V, t). Substitute these expressions into restriction (18.251). The result is ugly, but straightforward to evaluate numerically. Just like the Black-Scholes partial differential equation, it expresses the time derivative ∂C/∂t in terms of derivatives with respect to state variables, and thus can be used to work back from a terminal period. Theorem. The lower bound C(S, V, t) is the solution to the partial differential equation xc − rC +

∂C + ∂t

+

X ∂ 2C 1 X ∂2C 1 X ∂ 2C Si Sj σSi σ 0Sj + (σ V zi σ 0V zj + σ V wj σ0V wj ) + Si σSi σ0V zj = 2 i,j ∂Si ∂Sj 2 i,j ∂Vi ∂Vj ∂S ∂V i j i,j

=

µ

D −r S

¶0

q q ¢ ¡ 0 −1 (SC S ) + µ ˜ S ΣS σ S σ0V z − µ0V C V + A2 − µ ˜ 0S Σ−1 ˜ S C 0V σV w σ 0V w C V S µ

subject to the boundary conditions provided by the focus asset payoff xcT . C V denotes the vector with typical element ∂C/∂Vj and (SC S ) denotes the vector with typical element Si ∂C/∂Si . Replacing + with − before the square root gives the partial differential equation satisfied by the upper bound. The discount factor

316

S ECTION 18.4

EXTENSIONS , OTHER APPROACHES , AND BIBLIOGRAPHY

In general, the Λ process (18.246) or (18.250) depends on the parameters σCw . Hence, without solving the above partial differential equation we do not know how to spread the loading of dΛ across the multiple sources of risk dw whose risk prices we do not observe. Equivalently, we do not know how to optimally spread the total market price of risk across the elements of dw. Thus, in general we cannot use the integration approach– solve the discount factor forward – to find the bound by µ ¶ Z T Λs c ΛT c C t = Et xs ds + Et x . Λt T s=t Λt However, if there is only one shock dw, then we don’t have to worry about how the loading of dΛ spreads across multiple sources of risk. v can be determined simply by the volatility constraint. In this special case, dw and σCw are scalars. Hence equation (18.246) simplifies as follows: Theorem: In the special case that there is only one extra noise dw driving the V process, we can find the lower bound discount factor Λ from directly from q dΛ = −rdt − µ ˜ 0S Σ−1 σ dz − A2 − µ ˜ 0S Σ−1 ˜ S dw. (252) S S S µ Λ

I used this characterization to solve for the case of a non-traded underlying in the last section. In some applications, the loading of dΛ on multiple shocks dw may be constant over time. In such cases, one can again construct the discount factor and solve for bounds by (possibly numerical) integration, avoiding the solution of a partial differential equation.

18.4

Extensions, other approaches, and bibliography

The roots of the good deal good deal idea go a long way back. Ross (1976) bounded APT residuals by assuming that no portfolio can have more than twice the market Sharpe ratio, and I used the corresponding idea that discount factor volatility should be bounded to generate a robust approximate APT in Chapter 9.4. Good deal bounds apply the same idea to option payoffs. However, the good deal bounds also impose positive discount factors, and this constraint is important in an option pricing context. We also study dynamic models that chain discount factors together as in the option pricing literature. The one-period good-deal bound is the dual to the Hansen-Jagannathan (1991) bound with positivity – Hansen and Jagannathan study the minimum variance of positive discount factors that correctly price a given set of assets. The good deal bound interchanges the position of the option pricing equation and the variance of the discount factor. The techniques for solving the bound, therefore, are exactly those of the Hansen-Jagannathan bound in this one-period setup. There is nothing magic about discount factor volatility. This kind of problem needs weak 317

C HAPTER 18

O PTION PRICING WITHOUT PERFECT REPLICATION

but credible discount factor restrictions that lead to tractable and usefully tight bounds. Several other similar restrictions have been proposed in the literature. 1) Levy (1985) and Constantinides (1998) assume that the discount factor declines monotonically with a state variable; marginal utility should decline with wealth. 2) The good deal bounds allow the worst case that marginal utility growth is perfectly correlated with a portfolio of basis and focus assets. In many cases one could credibly impose a sharper limit than −1 ≤ ρ ≤ 1 on this correlation to obtain tighter bounds. 3) Bernardo and Ledoit (1999) use the restriction a ≥ m ≥ b to sharpen the no-arbitrage restriction ∞ ≥ m > 0. They show that this restriction has a beautiful portfolio interpretation – a < m < b corresponds to limited “gain - loss ratios” just as σ(m)/E(m) corresponds to limited Sharpe ratios. Define [Re ]+ = max(Re , 0) and [Re ]− = − min(−Re , 0) as the “gains and losses” of an excess return Re . Then, max e e

{R ∈R

[Re ]+ sup(m) min e − = e } [R ] {m:0=E(mR )} inf(m)

(253)

(The sup and inf ignore measure zero states.) This is exactly analogous to max e e

{R ∈R

|E(Re )| σ(m) = min } σ(Re ) {m:0=E(mRe )} E(m)

and hints at an interesting restatement of asset pricing theory in L1 with sup norm rather than L2 with second moment norm. Since m ≥ a, the call option price generated by this restriction in a one-period model is strictly greater than the lower arbitrage bound generated by m = 0; as in this case, the gain-loss bound can improve on the good-deal bound. 4) Bernardo and Ledoit also suggest a ≥ m/y ≥ b where y is an explicit discount factor model such as the consumption-based model or CAPM, as a way of imposing a “weak implication” of that particular model. These alternatives are really not competitors. Add all the discount factor restrictions that are appropriate and useful for a given problem. This exercise seems to me a strong case for discount factor methods as opposed to portfolio methods. The combination of positivity and volatility constraints on the discount factor leads to a sharper bound than the intersection of no-arbitrage and limited Sharpe ratios. I don’t know of a simple portfolio characterization of the set of prices that are ruled out by the good deal bound when both constraints bind. The same will be true as we add, say, gain-loss restrictions, monotonicity restrictions, etc. In continuous time, option pricing and terms structure problems increasingly feature assumptions about the “market price of risk” of the non-traded shocks. The good-deal bounds treat these rather formally; they choose the market prices of risks at each instant to minimize or maximize the option price subject to a constraint that the total market price of risk is less 318

S ECTION 18.5

P ROBLEMS

than a reasonable value, compared to the Sharpe ratios of other trading opportunities. One needn’t be this formal. Many empirical implementations of option pricing and term structure models feature unbelievable sizes and time-variation in market prices of risk. Just imposing sensible values for the market prices of risk, and trying on a range of sensible values may be good enough for many practical situations. The continuous-time treatment has not yet been extended to the important case of jumps rather than diffusion processes. With jumps, both the positivity and volatility constraints will bind.

18.5 1.

Problems

Prove (18.253), max e e

{R ∈R

2.

[Re ]+ sup(m) min e − = e } [R ] {m:0=E(mR )} inf(m)

Start with a finite state space. Binomial models are very popular in option pricing. This simple problem illustrates the technique. A stock currently selling at S will either rise to ST = uS with probability πu or decline to ST = dS with probability π d , paying no dividends in the interim. There is a constant gross interest rate Rf . (a) (b) (c) (d)

Find a discount factor that prices stock and bond. This means, find its value in each state of nature. Use the discount factor to price a call option one step before expiration. Express your results as an expected value using risk-neutral probabilities Do the same thing two steps before expiration. Cox, Ross and Rubinstein (1979) derive these formulas by setting up a hedge portfolio of stocks and bonds, and finding portfolio weights to exactly synthesize the option. Rederive your result with this method.

319

Chapter 19. rates

Term structure of interest

Term structure models are particularly simple, since bond prices are just the expected value of the discount factor. In equations, the price at time t of a zero coupon bond that comes (j) due at time t + j is Pt = Et (mt,t+j ). Thus, once you specify a time-series process for a one-period discount factor mt,t+1 , you can in principle find the price of any bond by chaining (j) together the discount factors and finding Pt = Et (mt,t+1 mt+1,t+2 ...mt+j−1,t+j ). As with option pricing models, this chaining together can be hard to do, and much of the analytical machinery in term structure models centers on this technical question. As with option pricing models, there are two equivalent ways to do the chaining together: Solve the discount factor forward and take an integral, or find a partial differential equation for prices and solve it backwards from the maturity date.

19.1

Definitions and notation

A quick introduction to bonds, yields, holding period returns, forward rates, and swaps. (N)

= log price of N period zero-coupon bond at time t.

(N)

= − N1 p(N) = log yield.

pt y

(N)

(N−1)

hprt+1 = pt+1 hpr =

dP (N,t) P

(N→N+1) ft

f(N, t) =

19.1.1

=

−

(N)

− pt

= log holding period return.

1 ∂P (N,t) P ∂N dt

(N) pt

−

(N+1) pt

(N,t) − P1 ∂P∂N

= instantaneous return.

= forward rate.

= instantaneous forward rate.

Bonds

The simplest fixed-income instrument is a zero-coupon bond. A zero-coupon bond is a promise to pay one dollar (a nominal bond) or one unit of the consumption good (a real (3) bond) on a specified date. I use a superscript in parentheses to denote maturity: Pt is the price of a three year zero-coupon bond. I will suppress the t subscript when it isn’t necessary. (N)

(N)

I denote logs by lowercase symbols, pt = ln Pt . The log price has a nice interpretation. If the price of a one-year zero coupon bond is 0.95, i.e. 95/c per dollar face value, the log price is ln(0.95) = −0.051. This means that the bond sells at a 5% discount. Logs also 320

S ECTION 19.1

D EFINITIONS AND NOTATION

give the continuously compounded rate. If we write erN = 1/P (N) then the continuously compounded rate is rN = − ln P (N) .

Coupon bonds are common in practice. For example, a $100 face value 10 year coupon bond may pay $5 every year for 10 years and $100 at 10 years. (Coupon bonds are often issued with semiannual or more frequent payments, $2.50 every six months for example.) We price coupon bonds by considering them as a portfolio of zeros.

Yield. The yield of a bond is the fictional, constant, known, annual, interest rate that justifies the quoted price of a bond, assuming that the bond does not default. It is not the rate of return of the bond. From this definition, the yield of a zero coupon bond is the number Y (N) that satisfies

Hence

1 P (N) = £ ¤N . (N) Y 1 1 (N) Y (N) = £ = − p(N) . ¤ N1 ; y N P (N)

The latter expression nicely connects yields and prices. If the price of a 4 year bond is -0.20 or a 20% discount, that is 5% discount per year, or a yield of 5%. The yield of any stream of cash flows is the number Y that satisfies P =

N X CFj j=1

Yj

.

In general, you have to search for the value Y that solves this equation, given the cash flows and the price. So long as all cash flows are positive, this is fairly easy to do. As you can see, the yield is just a convenient way to quote the price. In using yields we make no assumptions. We do not assume that actual interest rates are known or constant; we do not assume the actual bond is default-free. Bonds that may default trade at lower prices or higher yields than bonds that are less likely to default. This only means a higher return if the bond happens not to default. 19.1.2

Holding Period Returns

If you buy an N period bond and then sell it—it has now become an N −1 period bond—you achieve a return of (N−1)

(N)

HP Rt+1 =

P $back = t+1 (N) $paid Pt

321

(254)

CHAPTER 19

T ERM STRUCTURE OF INTEREST RATES

or, of course, (N)

(N−1)

hprt+1 = pt+1

(N)

− pt

.

We date this return (from t to t + 1) as t + 1 because that is when you find out its value. If this is confusing, take the time to write returns as HP Rt→t+1 and then you’ll never get lost. In continuous time, we can easily find the instantaneous holding period return of bonds with fixed maturity date P (T, t) hpr =

P (T, t + ∆) − P (T, t) P (T, t)

and, taking the limit, hpr =

dP (T, t) . P

However, it’s nicer to look for a bond pricing function P (N, t) that fixes the maturity rather than the date. As in (19.254), we have to account for the fact that you sell bonds that have shorter maturity than you buy. hpr = =

P (N − ∆, t + ∆) − P (N, t) P (N, t) P (N − ∆, t + ∆) − P (N, t + ∆) + P (N, t + ∆) − P (N, t) P (N, t)

and, taking the limit hpr =

19.1.3

1 ∂P (N, t) dP (N, t) − dt P P ∂N

(255)

Forward rate

The forward rate is defined as the rate at which you can contract today to borrow or lend money starting at period N, to be paid back at period N + 1. You can synthesize a forward contract from a spectrum of zero coupon bonds, so the forward rate can be derived from the prices of zero-coupon bonds. Here’s how. Suppose you buy one N period zero and simultaneously sell x N + 1 period zero coupon bonds. Let’s track your cash flow at every date. 322

S ECTION 19.1

D EFINITIONS AND NOTATION

Buy N-Period zero −P (N) 1

Today 0: Time N: Time N+1:

Sell x N+1 Period zeros +xP (N+1) -x

Net cash flow xP (N+1) − P (N) 1 -x

Now, choose x so that today’s cash flow is zero: x=

P (N) P (N+1)

You pay or get nothing today, you get $1.00 at N , and you pay P (N) /P (N+1) at N + 1. You have synthesized a contract signed today for a loan from N to N + 1—a forward rate! Thus, (N)

(N→N+1)

Ft

= Forward rate at t for N → N + 1 =

Pt

(N+1)

Pt

and of course (N→N+1)

ft

(N)

= pt

(N+1)

− pt

(256)

. (N)

People sometimes identify forward rates by the initial date, ft , and sometimes by the (N+1) ending date, ft . I use the arrow notation when I want to be really clear about dating a return. Forward rates have the lovely property that you can always express a bond price as its discounted present value using forward rates, (N)

(N)

pt

= pt

(N−1)

− pt

(N−1→N)

= −ft (1)

(yt

(0→1)

= ft

(N−1)

+ pt

(N−2)

− pt

(N−2→N−1)

− ft

(2)

(1)

(1)

− ...pt − pt + pt (1→2)

− ... − ft

(1)

− yt

of course), so (N)

Pt

(N) pt

=e

= e−

P N−1

(j→j+1) j=0 ft

N−1 Y

=

j=0

−1

(j→j+1)

Ft

.

Intuitively, the price today must be equal to the present value of the payoff at rates you can lock in today. In continuous time, we can define the instantaneous forward rate f(N, t) = −

∂p(Nt ) 1 ∂P (N, t) =− P ∂N ∂N

(257)

Then, forward rates have the same property that you can express today’s price as a discounted 323

CHAPTER 19

T ERM STRUCTURE OF INTEREST RATES

value using the forward rate, p(N, t) = −

Z

P (N, t) = e−

N

f(x, t)dx

x=0 RN x=0

f (x,t)dx

.

Equations 19.256 and 19.257 express forward rates as derivatives of the price vs. maturity curve. Since yield is related to price, we can relate forward rates to the yield curve directly. Differentiating the definition of yield y(N, t) = −p(N, t)/N, 1 1 1 1 ∂p(N, t) ∂y(N, t) = 2 p(N, t) − = − y(N, t) + f(N, t) ∂N N N ∂N N N

Thus, f(N, t) = y(N, t) + N

∂y(N, t) . ∂N

In the discrete case, (19.256) implies (N→N+1)

ft

19.1.4

(N)

= −N yt

(N+1)

+ (N + 1)yt

(N+1)

= yt

´ ³ (N+1) (N) . + N yt − yt

Swaps and options

Swaps are an increasingly popular fixed income instrument. The simplest example is a fixedfor-floating swap. Party A may have issued a 10 year fixed coupon bond. Party B may have issued a 10 year variable rate bond – a bond that promises to pay the current one year rate. (For example, if the current rate is 5%, the variable rate issuer would pay $5 for every $100 of face value. A long-term variable rate bond is the same thing as rolling over one-period debt.) They may be unhappy with these choices. For example, the fixed-rate payer may not want to be exposed to interest rate risk that the present value of his promised payments rises if interest rates decline. The variable-rate issuer may want to take on this interest rate risk, betting that rates will rise or to hedge other commitments. If they are unhappy with these choices, they can swap the payments. The fixed rate issuer pays off the variable rate coupons, and the variable rate issuer pays off the fixed rate coupons. Obviously, only the difference between fixed and variable rate actually changes hands. Swapping the payments is much safer than swapping the bonds. If one party defaults, the other can drop out of the contract, losing the difference in price resulting from intermediate interest rate changes, but not losing the principal. For this reason, and because they match the patterns of cashflows that companies usually want to hedge, swaps have become very popular tools for managing interest rate risk. Foreign exchange swaps are also popular: Party A may swap dollar payments for party B’s yen payments. Obviously, you don’t need to have issued 324

S ECTION 19.2

Y IELD CURVE AND EXPECTATIONS HYPOTHESIS

the underlying bonds to enter into a swap contract – you simply pay or receive the difference between the variable rate and the fixed rate each period. The value of a pure floating rate bond is always exactly one. The value of a fixed rate bond varies. Swaps are set up so no money changes hands initially, and the fixed rate is calibrated so that the present value of the fixed payments is exactly one. Thus, the “swap rate” is the same thing as a the yield on a comparable coupon bond. Many fixed income securities contain options, and explicit options on fixed income securities are also popular. The simplest example is a call option. The issuer may have the right to buy the bonds back at a specified price. Typically, he will do this if interest rates fall a great deal, making a bond without this option more valuable. Home mortgages contain an interesting prepayment option: if interest rates decline, the homeowner can pay off the loan at face value, and refinance. Options on swaps also exist; you can buy the right to enter into a swap contract at a future date. Pricing all of these securities is one of the tasks of term structure modeling.

19.2

Yield curve and expectations hypothesis

The expectations hypothesis is three equivalent statements about the pattern of yields across maturity, 1. 2. 3.

The N period yield is the average of expected future one-period yields The forward rate equals the expected future spot rate. The expected holding period returns are equal on bonds of all maturities. The expectations hypothesis is not quite the same thing as risk neutrality, since it ignores 1/2σ 2 terms that arise when you move from logs to levels. The yield curve is a plot of yields of zero coupon bonds as a function of their maturity. Usually, long-term bond yields are higher than short therm bond yields – a rising yield curve – but sometimes short yields are higher than long yields – an inverted yield curve. The yield curve sometimes has humps or other shapes as well. The expectations hypothesis is the classic theory for understanding the shape of the yield curve. More generally, we want to think about the evolution of yields – the expected value and conditional variance of next period’s yields. This is obviously the central ingredients for portfolio theory, hedging, derivative pricing, and economic explanation. The expectations hypothesis is the traditional benchmark for thinking about the expected value of future yields. We can state the expectations hypothesis in three mathematically equivalent forms:

325

CHAPTER 19 1.

2.

3.

T ERM STRUCTURE OF INTEREST RATES

The N period yield is the average of expected future one-period yields ³ ´ 1 (N) (1) (1) (1) (1) yt = Et yt + yt+1 + yt+2 + ... + yt+N−1 (+ risk premium) N The forward rate equals the expected future spot rate ³ ´ (1) ftN→N+1 = Et yt+N (+ risk premium)

(258)

(259)

The expected holding period returns are equal on bonds of all maturities (N)

(1)

Et (hprt+1 ) = yt

(+ risk premium).

(260)

You can see how the expectations hypothesis explains the shape of the yield curve. If the yield curve is upward sloping – long term bond yields are higher than short term bond yields – the expectations hypothesis says this is because short term rates are expected to rise in the future. You can view the expectations hypothesis as a response to a classic misconception. If long term yields are 10% but short term yields are 5%, an unsophisticated investor might think that long-term bonds are a better investment. The expectations hypothesis shows how this may not be true. Future short rates are expected to rise: this means that you will roll over the short term bonds at a really high rate, say 20%, giving the same long-term return. When the short term interest rates rise in the future, long-term bond prices decline. Thus, the long-term bonds will only give a 5% rate of return for the first year. You can see from the third statement that the expectations hypothesis is roughly the same as risk-neutrality. If we had said that the expected level of returns was equal across maturities, that would be the same as risk-neutrality. The expectations hypothesis specifies that the expected log return is equal across maturities. This is typically a close approximation to risk 1 2 neutrality, but not the same thing. If returns are log-normal, then E(R) = eE(r)+ 2 σ (r) . If mean returns are about 10% or 0.1 and the standard deviation of returns is about 0.1, then 1/2σ 2 is about 0.005, which is very small but not zero. We could easily specify risk-neutrality in the third expression of the expectations hypothesis, but then it would not imply the other two – 12 σ 2 terms would crop up. The intuition of the third form is clear: risk-neutral investors will adjust positions until the expected one-period returns are equal on all securities. Any two ways of getting money from t to t + 1 must offer the same expected return. The second form adapts the same idea to the choice of locking in a forward contract vs. waiting and borrowing and lending at the spot rate. Risk-neutral investors will load up on one or the other contract until the expected returns are the same. Any two ways of getting money from t + N to t + N + 1 must give the same expected return. The first form reflects a choice between two ways of getting money from t to N. You can buy a N period bond, or roll-over N one-period bonds. Risk neutral investors will choose one over the other strategy until the expected N − period return is the same. 326

S ECTION 19.3

T ERM STRUCTURE MODELS – A DISCRETE - TIME INTRODUCTION

The three forms are mathematically equivalent. If every way of getting money from t to t + 1 gives the same expected return, then so must every way of getting money from t + 1 to t + 2, and, chaining these together, every way of getting money from t to t + 2. For example, let’s show that forward rate = expected future spot rate implies the yield curve. Start by writing (1)

ftN−1→N = Et (yt+N−1 ).

Add these up over N, ³ ´ (1) (1) (1) (1) ft0→1 + ft1→2 + ... + ftN−2→N−1 + ftN−1→N = Et yt + yt+1 + yt+2 + ... + yt+N−1 .

The right hand side is already what we’re looking for. Write the left hand side in terms of the definition of forward rates, remembering P (0) = 1 so p(0) = 0, ´ ³ ´ ³ ´ ³ (0) (1) (1) (2) (N−1) (N) + pt − pt + ... + pt ft0→1 + ft1→2 + ... + ftN−2→N−1 + ftN−1→N = pt − pt − pt (N)

= −pt

(N)

= Nyt

.

You can show all three forms are equivalent by following similar arguments. (This is a great problem.) It is common to add a constant risk premium and still refer to the resulting model as the expectations hypothesis, and I include a risk premium in parentheses to remind you of this idea. One end of each of the three statements does imply more risk than the other. A forward rate is known while the future spot rate is not. Long-term bond returns are more volatile than short term bond returns. Rolling over short term real bonds is a riskier long-term investment than buying a long term real bond. If real rates are constant, and the bonds are nominal, then the converse can hold: short term real rates can adapt to inflation, so rolling over short nominal bonds can be a safer long-term real investment than long-term nominal bonds. These risks will generate expected return premia if they covary with the discount factor, and our theory should reflect this fact. If you allow an arbitrary, time-varying risk premium, the model is a tautology, of course. Thus, the entire content of the “expectations hypothesis” augmented with risk premia is in the restrictions that the risk premium is constant over time. We will see that the constant risk premium model does not do that well empirically. One of the main points of term structure models is – or at least ought to be – to quantify the size and movement over time in the risk premium.

19.3

Term structure models – a discrete-time introduction

327

CHAPTER 19

T ERM STRUCTURE OF INTEREST RATES

Term structure models specify the evolution of the short rate and potentially other state variables, and the prices of bonds of various maturities at any given time as a function of the short rate and other state variables. I examine a very simple example based on an AR(1) for the short rate and the expectations hypothesis, which gives a geometric pattern for the yield curve. A good way to generate term structure models is to write down a process for the discount factor, and then price bonds as the conditional mean of the discount factor. This procedure guarantees the absence of arbitrage. I give a very simple example of an AR(1) model for the log discount factor, which also results in geometric yield curves. . A natural place to start in modeling the term structure is to model yields statistically. You might run regressions of changes in yields on the levels of lagged yields, and derive a model of the mean and volatility of yield changes. You would likely start with a factor analysis of yield changes and express the covariance matrix of yields in terms of a few large factors that describe their common movement. The trouble with this approach is that you can quite easily reach a statistical representation of yields that admits an arbitrage opportunity, and you would not want to use such a statistical characterization for economic understanding of yields, for portfolio formation, or for derivative pricing. For example, a statistical analysis would strongly suggest that a first factor should be a “level” factor, in which all yields move up and down together. It turns out that this assumption violates arbitrage: the long-maturity yield must converge to a constant. (More precisely, the long-term forward rate, if it exists, must never fall.) How do you model yields without arbitrage? An obvious solution is to use the discount factor existence theorem: Write a statistical model for a positive discount factor, and find bond prices as the expectation of this discount factor. Such a model will be, by construction, arbitrage free. Conversely, any arbitrage-free distribution of yields can be captured by some positive discount factor, so you don’t lose any generality with this approach. 19.3.1

A term structure model based on the expectations hypothesis

We can use the expectations hypothesis to give the easiest example of a term structure model. (This one does not start from a discount factor and so may not be arbitrage-free.) Suppose the one-period yield follows an AR(1),

(1)

(1)

yt+1 − δ = ρ(yt

− δ) + εt+1 .

Now, we can use the expectations hypothesis (19.258) to calculate yields on bonds of all 328

S ECTION 19.3

T ERM STRUCTURE MODELS – A DISCRETE - TIME INTRODUCTION

maturities as a function of today’s one period yield, (2)

yt

i 1 h (1) (1) Et yt + yt+1 2 i 1 h (1) (1) yt + δ + ρ(yt − δ) = 2 1 + ρ (1) (yt − δ) = δ+ 2

=

Continuing in this way, ³ ´ 1 1 − ρN+1 (1) (N) yt − δ = (yt − δ). N 1−ρ

(261)

You can see some features that will recur throughout the term structure models. First, the model (19.261) can describe some movements in the yield curve over time. If the short rate is below its mean, then there is a smoothly upward sloping yield curve. Long term bond yields are higher as short rates are expected to increase in the future. If the short rate is above its mean, we get a smoothly inverted yield curve. This particular model cannot produce humps or other interesting shapes that we sometimes see in the term structure. Second, this model (N) (1) predicts no average slope of the term structure – E(yt ) = E(yt ) = δ . In fact, the average term structure seems to slope up slightly. Third, all bond yields move together in the model. If we were to stack the yields up in a VAR representation, it would be ³ ´ (1) (1) yt+1 − δ = ρ yt − δ + εt+1 ³ ´ (2) (2) 1+ρ yt+1 − δ = ρ yt − δ + 2 εt+1 . ... ³ ´ N +1 (N) (N) + N1 1−ρ yt+1 − δ = ρ yt − δ 1−ρ εt+1 (1)

(You can write the right hand variable in terms of yt if you want – any one yield carries the same information as any other.) The error terms are all the same. We can add more factors to the short rate process, to improve on this prediction, but most tractable term structure models maintain less factors than there are bonds, so some perfect factor structure is a common prediction of term structure models. Finally, this model has a problem in that the short rate, following an AR(1), can be negative. Since people can always hold cash, nominal short rates are never negative, so we want to start with a short rate process that does not have this feature. With this simple model in hand, you can see some obvious directions for generalization. First, we will want more complex driving processes than an AR(1). For example, a humpshape in the conditionally expected short rate will result in a hump-shaped yield curve. If there are multiple state variables driving the short rate, then we will have multiple factors driving the yield curve which will also result in more interesting shapes. We also want processes that keep the short rate positive in all states of nature. Second, we will want to add 329

CHAPTER 19

T ERM STRUCTURE OF INTEREST RATES

some “market prices of risk” – some risk premia. This will allow us to get average yield curves to not be flat, and time-varying risk premia seem to be part of the yield data. The yield curve literature proceeds in exactly this way: specify a short rate process and the risk premia, and find the prices of long term bonds. The trick is to specify sufficiently complex assumptions to be interesting, but preserve our ability to solve the models. 19.3.2

The simplest discrete time model

The simplest nontrivial model I can think of is to let the log of the discount factor follow an AR(1) with normally distributed shocks. I write the AR(1) for the log rather than the level in order to make sure the discount factor is positive, precluding arbitrage. Log discount factors are typically slightly negative, so I denote the unconditional mean E (ln m) = −δ (ln mt+1 + δ) = ρ (ln mt + δ) + εt+1 .

In turn, you can think of this discount factor model as arising from a consumption based power utility model with normal errors. mt+1 ct+1 − ct

¶γ Ct+1 = e Ct = ρ(ct − ct−1 ) + δ t+1 . −δ

µ

The term structure literature has only started to think whether the empirically successful discount factor processes can be connected empirically back to macroeconomic events in this way. From this discount factor, we can find bond prices and yields. This is easy because the conditional mean and variance of an AR(1) are easy to find. (I am following the strategy of solving the discount factor forward rather than solving the price backward.) We need (1)

yt

(2)

yt

(1)

= −pt = − ln Et (eln mt+1 ) 1 (2) 1 = − pt = − ln Et (eln mt+1 +ln mt+2 ) 2 2

and so on. Iterating the AR(1) forward, (ln mt+2 + δ) = ρ2 (ln mt + δ) + ρεt+1 + εt+2 (ln mt+3 + δ) = ρ3 (ln mt + δ) + ρ2 εt+1 + ρεt+2 + εt+3

so (ln mt+1 + δ) + (ln mt+2 + δ) = (ρ + ρ2 ) (ln mt + δ) + (1 + ρ)εt+1 + εt+2

330

S ECTION 19.3

T ERM STRUCTURE MODELS – A DISCRETE - TIME INTRODUCTION

(ln mt+1 + δ) + (ln mt+2 + δ) + (ln mt+3 + δ) = (ρ + ρ2 + ρ3 ) (ln mt + δ) + (1 + ρ + ρ2 )εt+1 + (1 + ρ)εt+2 + εt+3 . 1

2

Using the rule for a lognormal E(ex ) = eE(x)+ 2 σx , we have finally (1)

yt

(2)

yt

(3)

yt

1 = δ − ρ(ln mt + δ) − σ2ε 2 (ρ + ρ2 ) 1 + (1 + ρ)2 2 (ln mt + δ) − σε = δ− 2 4 (ρ + ρ2 + ρ3 ) 1 + (1 + ρ)2 + (1 + ρ + ρ2 )2 2 (ln mt + δ) − σε . = δ− 3 6

Notice all yields move as linear functions of a single state variable, ln mt + δ . Therefore, we can substitute out the discount factor and express the yields on bonds of any maturity as functions of the yields on bonds of one maturity. Which one we choose is arbitrary, but it’s conventional to use the shortest interest rate as the state variable. With E(y (1) ) = δ − 12 σ2ε , we can write our term structure model as h i (1) (1) yt − E(y (1) ) = ρ yt−1 − E(y (1) ) + εt (19.262) ´ 1 + (1 + ρ)2 (ρ + ρ2 ) ³ (1) (2) yt − E(y(1) ) − σ2ε yt = δ− 2 4 ´ 1 + (1 + ρ)2 + (1 + ρ + ρ2 )2 (ρ + ρ2 + ρ3 ) ³ (1) (3) yt − E(y (1) ) − σ2ε . yt = δ− 3 6 j ∞ X ´ X ρ 1 − ρN ³ (1) (1 + ρk )2 2 (N) yt − E(y (1) ) + σε = δ− yt N 1−ρ 2j j=1 k=1

This is the form in which term structure models are usually written – an evolution equation for the short rate process (together, in general, with other factors or other yields used to identify those factors), and then longer rates written as functions of the short rate, or the other factors. This is not a very realistic term structure model. In the data, the average yield curve – (N) the plot of {E[yt ]} versus N – is slightly upward sloping. The average yield curve from this model is slightly downward sloping as the σ2ε terms pile up. The effect is not large; with (2) (1) (3) (1) ρ = 0.9 and σ ε = 0.02, I find E(yt ) = E(yt ) − 0.02% and E(yt ) = E(yt ) − 0.06%. Still, it does not slope up. More importantly, this model only produces smoothly upward sloping or downward sloping term structures. For example, with ρ = 0.9, the first three terms multiplying the one period rate in (19.262) are 0. 86, 0. 81, 0.78. Two, three and four period bonds move exactly with one-period bonds using these coefficients. This model shows no conditional heteroskedasticity – the conditional variance of yield changes is always the same. The term structure data show times of high and low volatility, and times of high yields and high yield spreads seem to track these changes in volatility Finally, this model shows a 331

CHAPTER 19

T ERM STRUCTURE OF INTEREST RATES

weakness of almost all term structure models – all the yields move together; they follow an exact one-factor structure, they are perfectly conditionally correlated. The solution, of course, is to specify more complex discount rate processes that give rise to more interesting term structures.

19.4

Continuous time term structure models

The basic steps. 1.

Write a time-series model for the discount factor, typically in the form dΛ = −rdt − σ Λ (·)dz Λ dr = µr (·)dt + σ r (·)dz

2.

Solve the discount factor model forward and take expectations, to find bond prices (N)

Pt

3.

= Et

µ

Λt+N Λt

¶

.

Alternatively, from the basic pricing equation 0 = E [d(ΛP )] we can find a differential equation that the price must follow, ∂P 1 ∂P 2 ∂P ∂P µr + − rP = σr σΛ. σr − 2 ∂r 2 ∂r ∂N ∂r (0)

You can solve this back from PN = 1. I contrast the discount factor approach to the market price of risk and arbitrage pricing approaches. Term structure models are usually more convenient in continuous time. As with the last model, I specify a discount factor process and then find bond prices. A wide and popular class of term structure models are based on discount factor process of the form dΛ Λ dr

= −rdt − σΛ (·)dz = µr (·)dt + σ r (·)dz

332

(19.263)

S ECTION 19.4

CONTINUOUS TIME TERM STRUCTURE MODELS

This specification is analogous to a discrete time model of the form mt+1 xt+1

= xt + σεt+1 = ρxt + εt+1 .

This is a convenient representation since the state variable x carries the mean discount factor information. The r variable starts out as a state variable for the drift of the discount factor. However, you can see quickly that it will become the short rate process since Et (dΛ/Λ) = −rtf dt. The dots (·) remind you that these terms can be functions of state variables. Of course, we can add orthogonal components to the discount factor with no effect on the bond prices. Thus, the perfect correlation between interest rate and discount factor shocks is not essential. Term structure models differ in the specification of the functional forms for µr , σ r , σ Λ . We will study three famous examples, the Vasicek model, the Cox-Ingersoll-Ross model and the general affine specification. The first two are Vasicek

CIR

dΛ Λ dr

dΛ Λ dr

= −rdt − σΛ dz

(19.264)

= φ(¯ r − r)dt + σ r dz

√ = −rdt − σΛ Λdz √ = φ(¯ r − r)dt + σr rdz

(19.265)

The Vasicek model is quite similar to the AR(1) we studied in the last section. The CIR model adds the square root terms in the volatility. This specification captures the fact in US data that higher interest rates seem to be more volatile. In the other direction, it keeps the level of r guarantee that the square root the interest rate from falling below zero. (We need σr ≤ 2φ¯ process does not get stuck at zero.) Having specified a discount factor process, it’s a simple matter to find bond prices once again, µ ¶ Λt+N (N) Pt = Et . Λt We can simply solve the discount factor forward and take the expectation. We can also use the instantaneous pricing condition 0 = E(d(ΛP )) to find a partial differential equation for prices, and solve that backward. Both methods naturally adapt to pricing term structure derivatives – call options on bonds, interest rate floors or caps, “swaptions” that give you the right to enter a swap, and so forth. We simply put any payoff xC that depends on interest rates or interest rate state variables 333

CHAPTER 19

T ERM STRUCTURE OF INTEREST RATES

inside the expectation (N) Pt

= Et

Z

∞

s=t

Λs C x (s)ds. Λt

Alternatively, the price of such options will also be a function of the state variables that drive the term structure, so we can solve the bond pricing differential equation backwards using the option payoff rather than one as the boundary condition. 19.4.1

Expectation approach

As with the Black-Scholes option pricing model, we can think of solving the discount factor forward, and then taking the expectation. We can write the solution8 to (19.263) RT RT 1 2 ΛT = e− s=0 (rs + 2 σΛs )ds − s=0 σΛs dz Λ0

and thus, (N)

P0

³ RT ´ RT 1 2 = E0 e− s=0 (rs + 2 σΛs )ds − s=0 σΛs dz .

(266)

For example, in a riskless economy σΛ = 0, we obtain the continuous-time present value formula, (N)

P0

= e−

RT

s=0

rs ds.

With a constant interest rate r, P0 = e−rT .

In more interesting situations, solving the Λ equation forward and taking the expectation analytically is not so easy. Conceptually and numerically, it is easy, of course. Just simulate the system (19.263) forward a few thousand times, and take the average. 8

If this is mysterious, write first

d ln Λ =

µ ¶ 1 dΛ2 dΛ 1 2 − σ = − r + dt − σ Λ dz Λ 2 Λ2 2 Λ

and then integrate both sides from zero to T .

334

S ECTION 19.4 19.4.2

CONTINUOUS TIME TERM STRUCTURE MODELS

Differential equation approach

Recall the basic pricing equation for a security with price S and no dividends is µ

Et

dS S

¶

− rdt = −Et

µ

dS dΛ S Λ

¶

.

(267)

The left hand side is the expected excess return. As we guessed an option price C(S, t) and used (19.267) to derive a differential equation for the call option price, so we will guess a bond price P (N, t) and use this equation to derive a differential equation for the bond price. If we specified bonds by their maturity date T , P (t, T ), we could apply (19.267) directly. However, it’s nicer to look for a bond pricing function P (N, t) that fixes the maturity rather than the date. Equation (19.255) gives the holding period return for this case, which adds an extra term to correct for the fact that you sell younger bonds than you buy, return =

1 ∂P (N, t) dP (N, t) − dt P P ∂N

Thus, the fundamental pricing equation, applied to the price of bonds of given maturity P (N, t) is Et

µ

dP P

¶

−

µ

¶ µ ¶ dP dΛ 1 ∂P (N, t) + r dt = −Et . P ∂N P Λ

(268)

Now, we’re ready to find a differential equation for the bond price, just as we did for the option price to derive the Black-Scholes formula. Guess that all the time dependence comes through the state variable r, so P (N, r). Using Ito’s lemma dP =

µ

¶ ∂P 1 ∂ 2P 2 ∂P µr + σ r dz. σr dt + 2 ∂r 2 ∂r ∂r

(There is no ∂P/∂N because we defined the P function as the price of bonds with constant maturity N . If we had defined P (T, r) as the price of bonds that come due on date T , then there would be a −∂P/∂T term. This would take the place of the ∂P/∂N term in (19.268).) Plugging in to (19.268) and canceling dt, we obtain the fundamental differential equation for bonds, ∂P 1 ∂P 2 ∂P ∂P µr + − rP = σr σΛ. σr − 2 ∂r 2 ∂r ∂N ∂r

(269)

All you have to do is specify the functions µr (·), σr (·), σ Λ (·) and solve the differential equation. 335

CHAPTER 19 19.4.3

T ERM STRUCTURE OF INTEREST RATES

Market price of risk and risk-neutral dynamic approaches

The bond pricing differential equation (19.269) is conventionally derived without discount factors. One conventional approach is to write the short rate process dr = µr (·)dt + σ r (·)dz , and then specify that any asset whose payoffs have shocks σr dz must offer a Sharpe ratio of λ(·). We would then write ∂P 1 ∂P 2 ∂P ∂P µ + − rP = σr λ. σ − ∂r r 2 ∂r2 r ∂N ∂r

With λ = σΛ , this is just (19.269) of course. (If the discount factor and shock are imperfectly correlated, then λ = σ Λ ρ.) Different authors use “market price of risk” in different ways. Cox Ingersoll and Ross (1985, p.398) warn against modeling the right hand side as ∂P/∂r ψ (·) directly; this specification could lead to a positive expected return when σr = 0 and hence an infinite Sharpe ratio or arbitrage opportunity. By generating expected returns as the covariance of payoff shocks and discount factor shocks, we naturally avoid this mistake and other subtle ways to introduce arbitrage opportunities without realizing that you have done so. A second conventional approach is to use an alternative interest rate and discount factor process, dΛ Λ dr

= −rdt

(19.270)

= (µr − σ r λ)dt + σr dzr .

If we use this alternative process, we obtain 1 ∂P 2 ∂P ∂P (µr − σr λ) + − rP = 0 σ − ∂r 2 ∂r2 r ∂N

which is of course the same thing. This is the “risk neutral probability” approach, since the drift term in (19.270) is not the true drift. Since (19.270) gives the same prices, we can find and represent the bond price via the integral h RT i (N) Pt = Et∗ e− s=0 rs ds where E ∗ represents expectation with respect to the risk-neutral process defined in (19.270) rather than the true probabilities defined by the process (19.263).

When we derive the model from a discount factor, the single discount factor carries two pieces of information. The drift or conditional mean of the discount factor gives the short rate of interest, while the covariance of the discount factor shocks with asset payoff shocks generates expected returns or “market prices of risk.” I find it useful to write the discount factor model to keep the term structure connected with the rest of asset pricing, and to remind 336

S ECTION 19.5

T HREE LINEAR TERM STRUCTURE MODELS

myself where “market prices of risk” come from, and reasonable values for their magnitude. Of course, the result is the same no matter which method you follow. The fact that there are fewer factors than bonds means that once you have one bond price you can derive all the others by “no arbitrage” arguments and make this look like option pricing. Some derivations of term structure models follow this approach, setting up arbitrage portfolios. 19.4.4

Solving the bond price differential equation

Now we have to solve the partial differential equation (19.269) subject to the boundary condition P (N = 0, r) = 1. Solving this equation is straightforward conceptually and numerically. Express (19.269) as ∂P 1 ∂ 2P 2 ∂P = (µr − σ r σ Λ ) + σ − rP. ∂N ∂r 2 ∂r2 r

We can start at N = 0 on a grid of r, and P (0, r) = 1. For fixed N , we can work to one step larger N by evaluating the derivatives on the right hand side. The first step is P (∆N, r) =

∂P ∆N = −r∆N. ∂N

At the second step, ∂P/∂r = ∆N , ∂ 2 P/∂r2 = 0, so P (2∆N, r) = ∆N 2 (µr − σr σ Λ ) − r2 ∆N 2 .

Now the derivatives of µr and σr with respect to r will start to enter, and we let the computer take it from here. Analytic solutions only exist in special cases, which we study next.

19.5

Three linear term structure models

I solve the Vasicek, Cox Ingersoll Ross, and Affine model. Each model gives a linear function for log bond prices and yields, for example,

ln P (N, r) = A(N ) − B(N )r

As we have seen, term structure models are easy in principle and numerically: specify a discount factor process and find its conditional expectation or solve the bond pricing partial differential equation back from maturity. In practice, the computations are hard. I present 337

CHAPTER 19

T ERM STRUCTURE OF INTEREST RATES

next three famous special cases of term structure models – specifications for the discount factor process – that allow analytical or quickly calculable solutions. Analytical or close to analytical solutions are still important, because we have not yet found good techniques for reverse-engineering the term structure. We know how to start with a discount factor process and find bond prices. We don’t know how to start with the characteristics of bond prices that we want to model and construct an appropriate discount factor process. (At least one whose parameters do not change every day, as in the “arbitrage free” literature.) Thus, in evaluating term structure models, we will have to do lots of the “forward” calculations – from assumed discount factor model to bond prices – and it is important that we should be able to do them quickly. Obviously, this field would be dramatically changed if we could find a way to reverseengineer the term structure to directly estimate the discount factor process. Also, the ad-hoc time-series models of the discount factor obviously need to be connected to macroeconomics; if not to consumption, then at least to inflation, marginal products of capital, and macroeconomic variables used by the Federal Reserve in its interventions in the interest rate markets. 19.5.1

Vasicek model via by pde

The Vasicek (1977) model is a special case that allows a fairly easy analytic solution. The method is the same as the more complex analytic solution in the CIR and affine classes, but the algebra is easier, so this is a good place to start. The Vasicek discount factor process is dΛ Λ dr

= −rdt − σΛ dz = φ(¯ r − r)dt + σr dz

Using this process in the basic bond differential equation (19.269), we obtain 1 ∂ 2 P 2 ∂P ∂P ∂P φ(¯ r − r) + − rP = σr σΛ . σ − ∂r 2 dr2 r ∂N ∂r

(271)

I’ll solve this equation with the usual unsatisfying non-constructive technique – guess the functional form of the answer and show it’s right. I guess that log yields and hence log prices are a linear function of the short rate, P (N, r) = eA(N)−B(N)r .

(272)

I take the partial derivatives required in (19.271) and see if I can find A(N ) and B(N ) to make (19.271) work. The result is a set of ordinary differential equations for A(N) and B(N ), and these are of a particularly simple form that can be solved by integration. I solve 338

S ECTION 19.5

T HREE LINEAR TERM STRUCTURE MODELS

them, subject to the boundary condition imposed by P (0, r) = 1. The result is ¢ 1¡ 1 − e−φN φ ¶ µ 2 σ2 1 σr σ r σΛ − r ¯ (N − B(N)) − r B(N )2 . A(N ) = + 2 2φ φ 4φ

B(N ) =

(19.273) (19.274)

The exponential form of (19.272) means that log prices and log yields are linear functions of the interest rate, p(N, r) = A(N ) − B(N )r A(N ) B(N ) + r. y(N, r) = − N N

Solving the pde: details.

The boundary condition P (0, r) = 1 will be satisfied if A(0) − B(0)r = 0.

Since this must hold for every r, we will need A(0) = 0; B(0) = 0.

Given the guess (19.272), the derivatives that appear in (19.271) are 1 ∂Pr P ∂r 1 ∂2P P ∂r2 1 ∂P P ∂N

= −B(N ) = B(N )2 = A0 (N) − B 0 (N )r.

Substituting these derivatives in (19.271), 1 −B(N )φ(¯ r − r) + B(N)2 σ2r − A0 (N ) + B 0 (N )r − r = −B(N )σr σΛ . 2

This equation has to hold for every r, so the terms multiplying r and the constant terms must separately be zero. 1 B(N )2 σ 2r − (φ¯ r − σ r σΛ ) B(N ) 2 B 0 (N) = 1 − B(N)φ. A0 (N) =

(19.275)

We can solve this pair of ordinary differential equations by simple integration. The second 339

CHAPTER 19

T ERM STRUCTURE OF INTEREST RATES

one is dB dN Z dB 1 − φB

= 1 − φB = dN

1 − ln (1 − φB) = N φ

and hence B(N ) =

¢ 1¡ 1 − e−φN . φ

(276)

Note B(0) = 0 so we did not need a constant in the integration. We solve the first equation in (19.275) by simply integrating it, and choosing the constant to set A(0) = 0. Here we go. A0 (N ) = A(N ) = A(N ) = A(N ) =

1 B(N )2 σ 2r − (φ¯ r − σ r σ Λ ) B(N ) 2 Z 2 Z σr B(N)2 dN − (φ¯ r − σ r σΛ ) B(N )dN + C 2 µ ¶Z Z ¡ ¡ ¢ ¢ σ r σΛ σ 2r −φN −2φN 1 − 2e 1 − e−φN dN + C dN − r¯ − +e 2 φ 2φ µ ¶ µ ¶µ ¶ σ 2r e−2φN 2e−φN σr σ Λ e−φN − N + − r ¯ − N + +C φ 2φ φ φ 2φ2

We pick the constant of integration to give A(0) = 0. You can do this explicitly, or figure out directly that the result is achieved by subtracting one from the e−φN terms, Ã ¢ ¡ −2φN ¢! µ ¢! ¡ ¡ −φN ¶Ã e −1 −1 2 e−φN − 1 e σ 2r σr σΛρ − A(N ) = 2 N + − r¯ − N+ φ 2φ φ φ 2φ Now, we just have to make it pretty. I’m aiming for the form given in (19.274). Note ¢ 1 ¡ −φN + e−2φN 2 1 − 2e φ e−2φN − 1 1 − e−φN + φB(N)2 = 2 φ φ −2φN e − 1 φB(N )2 − 2B(N ) = φ B(N)2

=

340

S ECTION 19.5

T HREE LINEAR TERM STRUCTURE MODELS

Then ¶ µ µ ¶ σ2r σ r σΛ φ 2 B(N ) + B(N ) − r ¯ − N − 2B(N ) − (N − B(N )) 2 φ 2φ2 µ ¶ σ2 σr σΛ σ2 − r2 (N − B(N )) . A(N ) = − r B(N)2 − r¯ − 4φ φ 2φ

A(N ) =

We’re done. 19.5.2

Vasicek model by expectation

What if we solve the discount rate forward and take an expectation instead? The Vasicek model is simple enough that we can follow this approach as well, and get the same analytic solution. The same methods work for the other models, but the algebra gets steadily worse. The model is dΛ Λ dr

= −rdt − σΛ dz = φ(¯ r − r)dt + σr dz.

(19.277) (19.278)

The bond price is (N) P0

= E0

µ

ΛN Λ0

¶

(279)

I use 0 and N rather than t and t + N to save a little bit on notation. To find the expectation in (19.279), we have to solve the system (19.277)-(19.278) forward. The steps are simple, though the algebra is a bit daunting. First, we solve r forward. Then, we solve Λ forward. ln Λt turns out to be conditionally normal, so the expectation in (19.279) is the expectation of a lognormal. Collecting terms in the resulting expectation that depend on r0 as the B(N ) term, and the constant term as the A(N) term, we find the same solution as (19.273)-(19.274). The interest rate is just an AR(1). By analogy with a discrete time AR(1) you can guess that its solution is rt =

Z

t

s=0

e−φ(t−s) σr dzs + e−φt r0 + (1 − e−φt )¯ r.

To derive this solution, define r˜ by r˜t = eφt (rt − r¯).

341

(280)

CHAPTER 19

T ERM STRUCTURE OF INTEREST RATES

Then, d˜ rt d˜ rt d˜ rt d˜ rt

= = = =

φ˜ rt dt + eφt drt φ˜ rt dt + eφt φ(¯ r − r)dt + eφt σ r dzt φ˜ rt dt − eφt φe−φt r˜t dt + eφt σr dzt eφt σr dzt .

This equation is easy to solve, r˜t − r˜0

= σr

Z

t

eφs dzs

s=0 t

φt

e (rt − r¯) − (r0 − r¯) = σr

Z

s=0

−φt

rt − r¯ = e

eφs dzs (r0 − r¯) + σ r

Z

t

e−φ(t−s) dzs . s=0

And we have (19.280). Now, we solve the discount factor process forward. It isn’t pretty, but it is straightforward. d ln Λt ln Λt − ln Λ0

dΛ 1 dΛ2 1 − = −(rt + σ 2Λ )dt − σ Λ dzt Λ 2 Λ2 2 Z t Z t 1 2 = − (rs + σΛ )ds − σΛ dzs . 2 s=0 s=0 =

Plugging in the interest rate solution (19.280), ¶ ¸ Z t Z t ·µZ s 1 ln Λt − ln Λ0 = − e−φ(s−u) σr dzu + e−φs (r0 − r¯) + r¯ + σ 2Λ ds − σΛ dzs 2 s=0 u=0 s=0 Interchanging the order of the first integral, evaluating the easy ds integrals and rearranging, ¸ ·µ ¸ ¶ Z t Z t ·Z t Z t 1 = −σΛ dzs − σr e−φ(s−u) ds dzu − r¯ + σ2Λ t + (r0 − r¯) e−φs ds 2 s=0 u=0 s=u s=0 µ ¶ ¸ Z t · ´ ³ −φt σr 1 1 − e = − 1 − e−φ(t−u) dzu − r¯ + σ 2Λ t − (r0 − r¯) . (19.281) σΛ + φ 2 φ u=0 The first integral has a deterministic function of time u. This gives rise to a normally distributed random variable – it’s just a weighted sum of independent normals dzu : µ Z t ¶ Z t 2 f(u)dzu ∼ N 0, f (u)du . u=0

u=0

342

S ECTION 19.5

T HREE LINEAR TERM STRUCTURE MODELS

Thus, ln Λt − ln Λ0 is normally distributed with mean given by the second set of terms in (19.281) and variance var0 (ln Λt − ln Λ0 ) = · ´¸2 σr ³ 1 − e−φ(t−u) du σΛ + φ u=0 # ¶2 µ ¶ Z t "µ σr σr σr σ 2r −2φ(t−u) −φ(t−u) = −2 + 2e σΛ + du σΛ + e φ φ φ φ u=0 µ µ ¶2 ¶ ¢ ¢ σ2 ¡ σr σr σr ¡ 1 − e−φt + r3 1 − e−2φt . (19.282) t − 2 2 σΛ + = σΛ + φ φ φ 2φ =

Z

t

Since we have the distribution of ΛN we are ready to take the expectation. ¡ ¢ 1 ln P (N, 0) = ln E0 eln ΛN −ln Λ0 = E0 (ln ΛN − ln Λ0 ) + σ 20 (ln ΛN − ln Λ0 ) . 2

Plugging in the mean from (19.281) and the variance from (19.282) (N)

ln P0

·µ ¶ ¸ 1 1 − e−φN r¯ + σ2Λ N + (r0 − r¯) (19.283) 2 φ ¶2 µ ¶ µ ¡ ¢ ¢ σ2 ¡ 1 σr σr σr −2φN + σΛ N − 2 + σ Λ 1 − e−φN + r3 1 − e(19.284) + 2 φ φ φ 4φ

= −

All that remains is to make it pretty. To compare it with our previous result, we want to express it in the form ln P (N, r0 ) = A(N ) − B(N )r0 . The coefficient on r0 (19.283) is B(N) =

1 − e−φN , φ

the same expression we derived from the partial differential equation. To simplify the constant term, recall that (19.285) implies 1 − e−2φN = −φB(N )2 + 2B(N ). φ

343

(285)

CHAPTER 19

T ERM STRUCTURE OF INTEREST RATES

Thus, the constant term (the terms that do not multiply r0 ) in (19.283) is ·µ ¶ ¸ 1 2 1 − e−φN A(N ) = − r¯ + σΛ N − r¯ 2 φ ¶2 µ ¶ µ ¡ ¢ ¢ σ2 ¡ 1 σr σr σr + σΛ N − 2 + σ Λ 1 − e−φN + r3 1 − e−2φN + 2 φ φ φ 4φ ·µ ¶ ¸ 1 A(N ) = − r¯ + σ2Λ N − r¯B(N ) 2 ¶2 ¶ µ µ ¢ 1 σr σr σr σ2 ¡ + + σΛ N − + σ Λ B(N) − r2 φB(N )2 − 2B(N) 2 φ φ φ 4φ ¶ µ 2 σ 2r σr 1 σr − r¯ (N − B(N )) − 2 φB(N )2 . + σΛ A(N ) = 2 φ2 φ 4φ Again, this is the same expression we derived from the partial differential equation. This integration is usually expressed under the risk-neutral measure. If we write the riskneutral process dΛ Λ dr

= −rdt = [φ(¯ r − r) − σ r σ Λ ] dt + σr dz.

Then the bond price is (N)

P0

= Ee−

RN

s=0

rs ds

.

The result is the same, of course. 19.5.3

Cox Ingersoll Ross Model

For the Cox-Ingersoll-Ross (1985) model dΛ Λ dr

√ = −rdt − σΛ rdz √ = φ(¯ r − r)dt + σr rdz

our differential equation (19.269) becomes 1 ∂ 2P 2 ∂P ∂P ∂P φ(¯ r − r) + − rP = σr σ Λ r. σr r − 2 ∂r 2 ∂r ∂N ∂r

(286)

Guess again that log prices are a linear function of the short rate, P (N, r) = eA(N)−B(N)r .

344

(287)

S ECTION 19.5

T HREE LINEAR TERM STRUCTURE MODELS

Substituting the derivatives of (19.287) into (19.286), 1 −B(N)φ(¯ r − r) + B(N )2 σ 2r r − A0 (N ) + B 0 (N )r − r = −B(N )σ r σ Λ r. 2

Again, the coefficients on the constant and on the terms in r must separately be zero, 1 B 0 (N ) = 1 − σ2r B(N )2 − (σ r σ Λ + φ) B(N ) 2 A0 (N ) = −B(N)φ¯ r.

(19.288)

The ordinary differential equations (19.288) are quite similar to the Vasicek case, (19.275). However, now the variance terms multiply an r, so the B(N ) differential equation has the extra B(N)2 term. We can still solve both differential equations, though the algebra is a little bit more complicated. The result is ¢ ¡ 2 1 − eγN B(N) = (γ + φ + σr σΛ )(eγN − 1) + 2γ µ µ ¶ ¶ φ¯ r 2γ A(N) = 2 ln + ψN σ 2r ψ(eγN − 1) + 2γ where γ ψ

q = (φ + σ r σ Λ )2 + 2σ2r = φ + σΛ σr + γ.

The CIR model can also be solved by expectation. In fact, this is how Cox Ingersoll and Ross (1985) actually solve it – their marginal value of wealth JW is the same thing as the discount factor. However, where the interest rate in the Vasicek model was a simple conditional normal, the interest rate now has a non-central χ2 distribution, so taking the integral is a little messier. 19.5.4

Multifactor affine models

The Vasicek and CIR models are special cases of the affine class of term structure models (Duffie and Kan 1996, Dai and Singleton 1999). These models allow multiple factors, meaning all bond yields are not just a function of the short rate. Affine models maintain the convenient form that log bond prices are linear functions of the state variables. This means that we can take K bond yields themselves as the state variables, and the yields will reveal anything of interest in the hidden state variables. The short rate and its volatility will be forecast by lagged short rates but also by lagged long rates or interest rate spreads. My presentation and notation is similar to Dai and Singleton’s, but as usual I add the discount factor explicitly. 345

CHAPTER 19

T ERM STRUCTURE OF INTEREST RATES

Here is the affine model setup: dy r dΛ Λ dwi

= φ (¯ y − y) dt + Σdw = δ0 + δ0y

(19.289) (19.290)

= −rdt − b0Λ dw q = αi + β 0i ydzi ; E(dzi dzj ) = 0.

(19.291) (19.292)

Equation (19.289) describes the evolution of the state variables. In the end, yields will be linear functions of the state variables, so we can take the state variables to be yields; thus we use the letter y . y denotes a K− dimensional vector of state variables. φ is now a K × K matrix, y¯ is a K− dimensional vector, Σ is a K × K matrix. Equation (19.290) describes the mean of the discount factor or short rate as a linear function of the state variables. Equation (19.291) is the discount factor. bΛ is a K−dimensional vector that describes how the discount factor responds to the K shocks. The more Λ responds to a shock, the higher the market price of risk of that shock. Equation (19.292) describes the shocks dw. The functional form nests the CIR square root type models if αi = 0 and the Vasicek type Gaussian process if β i = 0. You can’t pick αi and β i arbitrarily, as you have to make sure that αi + β 0i y > 0 for all values of y that the process can attain. Dai and Singleton characterize this “admissibility” criterion. We find bond prices in the affine setup following exactly the same steps as for the Vasicek and CIR models. Again, we guess that prices are linear functions of the state variables y . 0

P (N, y) = eA(N)−B(N) y .

We apply Ito’s lemma to this guess, and substitute in the basic bond pricing equation (19.268). We obtain ordinary differential equations that A(N ) and B(N) must satisfy, ¶ Xµ 1 0 ∂B(N ) 2 0 = −φ B(N ) − B(N)i bΛi + [Σ B(N )]i β i + δ (19.293) ∂N 2 i µ ¶ X 1 ∂A(N ) 2 = y − δ 0 . (19.294) B(N )i bΛi + [Σ0 B(N )]i αi − B(N )0 φ¯ ∂N 2 i I use the notation [x]i to denote the ith element of a vector x. As with the CIR and Vasicek models, these are ordinary differential equations that can be solved by integration starting with A(0) = 0, B(0) = 0. While they do not always have analytical solutions, they are quick to solve numerically – much quicker than solving a partial differential equation. Derivation

To derive (19.294) and (19.293), we start with the basic bond pricing equation (19.268), which I repeat here, ¶ µ µ ¶ µ ¶ dP dP dΛ 1 ∂P + r dt = −Et Et (295) − . P P ∂N P Λ 346

S ECTION 19.5

T HREE LINEAR TERM STRUCTURE MODELS

We need dP/P, 1 ∂P 0 1 1 0 ∂ 2P dP = dy dy + dy. P P ∂y 2P ∂y∂y0

The derivatives are 1 ∂P P ∂y 1 ∂ 2P P ∂y∂y0 1 ∂P P ∂N

= −B(N ) = B(N)B 0 (N ) =

∂A(N ) ∂B(N ) 0 − y. ∂N ∂N

Thus, the first term (19.295) is µ ¶ 1 dP Et y − y) dt + Et (dw0 Σ0 B(N )B 0 (N )Σdw) = −B(N )0 φ (¯ P 2 Et (dwi dwj ) = 0, which allows us to simplify the last term. If w1 w2 = 0, then, ¸· ¸ · X ¤ b1 b1 b1 b2 £ w1 (w0 bb0 w) = w1 w2 b2i wi2 . = b21 w12 + b22 w22 = b2 b1 b2 b2 w2

Applying the same algebra to our case, X X ¢ 2 2¡ Et (dw0 Σ0 B(N)B 0 (N )Σdw) = [Σ0 B(N )]i dwi2 = [Σ0 B(N )]i αi + β 0i y dt. i

i

I use the notation [x]i to denote the ith element of the K−dimensional vector x. In sum, we have µ ¶ ¢ 1X 0 dP 2¡ Et y − y) dt + [Σ B(N)]i αi + β 0i y dt. (296) = −B(N )0 φ (¯ P 2 i The right hand side term in (19.295) is µ ¶ dP dΛ −Et = −B(N)0 dwdw0 bΛ P Λ ¡ ¢ dwdw0 is a diagonal matrix with elements αi + β 0i y . Thus, µ ¶ X ¡ ¢ dP dΛ −Et B(N )i bΛi αi + β 0i y =− P Λ i 347

(297)

CHAPTER 19

T ERM STRUCTURE OF INTEREST RATES

Now, substituting (19.296) and (19.297) in (19.295), along with the easier ∂P/∂N central term, we get µ ¶ ¢ 1X 0 ∂A(N ) ∂B(N ) 0 2¡ 0 0 0 − −B(N ) φ (¯ y − y) + [Σ B(N)]i αi + β i y − y + δ0 + δ y 2 i ∂N ∂N X ¡ ¢ = − B(N )i bΛi αi + β 0i y . i

Once again, the terms on the constant and each yi must separately be zero. The constant term: X 1X 0 ∂A(N) 2 − δ0 = − −B(N )0 φ¯ y+ [Σ B(N)]i αi − B(N)i bΛi αi . 2 i ∂N i µ ¶ ∂A(N ) X 1 2 = y − δ0 B(N )i bΛi + [Σ0 B(N )]i αi − B(N)0 φ¯ ∂N 2 i

The terms multiplying y : B(N )0 φy +

X 1X 0 ∂B(N ) 0 2 [Σ B(N )]i β 0i y + y − δ0y = − B(N )i bΛi β 0i y. 2 i ∂N i

Taking the transpose and solving,

¶ Xµ 1 ∂B(N) 2 = −φ0 B(N ) − B(N )i bΛi + [Σ0 B(N)]i β i + δ. ∂N 2 i

19.6

Bibliography and comments

The choice of discrete vs. continuous time is really one of convenience. Campbell, Lo and MacKinlay (1997) give a discrete-time treatment, showing that bond prices are linear functions of the state variables even in a discrete-time two-parameter square root model. Models also don’t have to be affine. Constantinides (1992) is a nice discrete time model; its discount factor is driven by the squared value of AR(1) state variables. It gives closed form solutions for bond prices. The bond prices are not linear functions of the state variables, but it is the existence of closed forms rather than linearity that makes affine models so attractive. It allows for both signs of the term premium, as we seem to see in the data. So far most of the term structure literature has emphasized the risk-neutral probabilities, rarely making any reference to the separation between drifts and market prices of risk. This was not a serious shortcoming for option pricing uses, for which modeling the volatilities is much more important than for modeling the drifts, and to draw smooth yield curves across 348

S ECTION 19.6

BIBLIOGRAPHY AND COMMENTS

maturities. However, it makes the models unsuitable for bond portfolio analysis and other uses. Many models imply high and time-varying market prices of risk or conditional Sharpe ratios. Recently, Duffee (1999) and Duarte (2000) have started the important task of specifying term structure models that fit the empirical facts about expected returns in term structure models. In particular, they try to fit the Fama-Bliss (1986) and Campbell and Shiller (1991) regressions that relate expected returns to the slope of the term structure (see Chapter 20), while maintaining the tractability of affine models. Term structure models used in finance amount to regressions of interest rates on lagged interest rates. Macroeconomists also run regressions of interest rates on a wide variety of variables, including lagged interest rates, but also lagged inflation, output, unemployment, exchange rates, and so forth. They often interpret these equations as the Federal Reserve’s policy-making rule for setting short rates as a function of macroeconomic conditions. This interpretation is particularly clear in the Taylor rule literature (Taylor 1999) and monetary VAR literature, see Christiano Eichenbaum and Evans (1999), Cochrane (1994) for surveys. Someone, it would seem, is missing important right hand variables. The criticism of finance models is stinging when we only use the short rate as a state variable. Multifactor models are more subtle. If any variable forecasts future interest rates, then it becomes a state variable, and it should be revealed by bond yields. Thus, bond yields should completely drive out any other macroeconomic state variables as interest rate forecasters. They don’t, which is an interesting observation. In addition, there is an extensive literature that studies yields from a purely statistical point of view, Gallant and Tauchen (1997) for example, and a literature that studies high frequency behavior in the federal funds market, for example Hamilton (1996). Obviously, these three literatures need to become integrated. Balduzzi, Bertola and Foresi (1996) consider a model based on the Federal Funds target, and Piazzesi (2000) integrated a careful specification of high-frequency moves in the Federal funds rate into a term structure model. The models studied here are all based on diffusions with rather slow-moving state variables. These models generate one-day ahead densities that are almost exactly normal. In fact, as Johannes (2000) points out, one-day ahead densities have much fatter tails than normal distributions predict. This behavior could be modeled by fast-moving state variables. However, it is more natural to think of this behavior as generated by a jump process, and Johannes nicely fits a combined jump-diffusion for yields. This specification can change pricing and hedging characteristics of term structure models significantly. All of the term structure models in this chapter describe many bond yields as a function of a few state variables. This is a reasonable approximation to the data. Almost all of the variance of yields can be described in terms of a few factors, typically a “level” “slope” and “hump” factor. Knez, Litterman and Scheinkman (1994) make the point with a formal maximum likelihood factor analysis, but you can see the point with a simple eigenvalue decomposition of log yields. 349

CHAPTER 19

σ 6.36 0.61 0.10 0.08 0.07

1 0.45 -0.75 0.47 0.10 0.07

T ERM STRUCTURE OF INTEREST RATES

2 0.45 -0.21 -0.62 -0.49 -0.36

Maturity 3 0.45 0.12 -0.41 0.39 0.68

4 0.44 0.36 0.11 0.55 -0.60

5 0.44 0.50 0.46 -0.55 0.21

“Level” “Slope” “Hump”

Eigenvalue decomposition of the covariance matrix of zero coupon bond yields, 1952-1997. The first column gives the square root of the eigenvalues. The columns marked 1-5 give the eigenvectors corresponding to 1-5 year zero coupon bond yields. I decomposed the covariance matrix as Σ = QΛQ0 ; σ2 gives the diagonal entries in Λ and the rest of the table gives the entries of Q. With this decomposition, we can say that bond yields are generated by y = QΛ1/2 ε; E(εε0 ) = I , thus Q give “loadings” on the shocks ε. Not only is the variance of yields well described by a factor model, but the information in current yields about future yields – the expected changes in yields and the conditional volatility of yields – is well captured by one level and a few spreads as well. It is a good approximation, but it is an approximation. Actual bond prices do not exactly follow any smooth yield curve, and the covariance matrix of actual bond yields does not have an exact K factor structure – the remaining eigenvalues are not zero. Hence you cannot estimate a term structure model directly by maximum likelihood; you either have to estimate the models by GMM, forcing the estimate to ignore the stochastic singularity, or you have to add distasteful measurement errors. As always, the importance of an approximation depends on how you use the model. If you take the model literally, a bond whose price deviates by one basis point is an arbitrage opportunity. In fact, it is at best a good Sharpe ratio, but a K factor model will not tell you how good – it won’t quantify the risk involved in using the model for trading purposes. Hedging strategies calculated from K-factor models may be sensitive to small deviations as well. One solution has been to pick different parameters at each point in time (Ho and Lee 1986). This approach is useful for derivative pricing, but is obviously not a satisfactory solution. Models in which the whole yield curve is a state variable, Kennedy (1994), Santa Clara and Sornette (1999), are another interesting response to the problem, and potentially provide a realistic description of the data. The market price of interest rate risk reflects the market price of real interest rate changes and the market price of inflation – or whatever real factors are correlated with inflation. The relative contributions of inflation and real rates in interest rate changes is very important for the nature of the risks that bondholders face. For example, if real rates are constant and nominal rates change on inflation news, then short term bonds are the safest real long 350

S ECTION 19.7

P ROBLEMS

term investment. If inflation is constant and nominal rates change on real rate news, then long term bonds are the safest long term investment. The data seem to suggest a change in regime between the 1970s and 1990s: in the 70’s, most interest rate changes were due to inflation, while the opposite seems true now. Despite all these provocative thoughts, though, little empirical work has been done that usefully separates interest rate risk premia into real and inflation premium components. Buraschi and Jiltsov (1999) is one recent effort in this direction, but a lot more remains to be done.

19.7 1.

2.

3.

4. 5.

Problems

Complete the proof that each of the three statements of the expectations hypothesis implies the other. Is this also true if we add a constant risk premium? Are the risk premia in each of the three statements of the yield curve of the same sign? Under the expectations hypothesis, if long-term yields are higher than short term yields, does this mean that future long term rates should go up, down, or stay the same? (Hint: a plot of the expected log bond prices over time will really help here.) (N) (1) Start by assuming risk neutrality, E(HP Rt+1 ) = Yt for all maturities N . Try to derive the other representations of the expectations hypothesis. Now you see why we specify that the expected log returns are equal. Look at (19.266) and show that adding orthogonal dw to the discount factor has no effect on bond pricing formulas. Look at (19.266) and show that P = e−rT if interest rates are constant, i.e. if dΛ/Λ = −rdt + σΛ dz .

351

PART IV Empirical survey

352

S ECTION 19.7

P ROBLEMS

This part is a brief attempt to survey some of the central empirical issues that have driven recent thinking about financial economics for, and which are driving the development of our theoretical understanding of the nature of risk and risk premia. This part draws heavily on two previous review articles, Cochrane (1998), (1999a) and on Cochrane and Hansen (1992). Fama’s (1970) and (1991) efficient market reviews are classic and detailed reviews of much of the underlying empirical literature, focusing on crosssectional questions. Campbell (1999, 2000) and Kocheralkota (1996) are good surveys of the equity premium literature.

353

Chapter 20. Expected returns in the time-series and cross-section The first revolution in finance started the modern field. Peaking in the early 1970s, this revolution established the CAPM, random walk, efficient markets, portfolio based view of the world. The pillars of this view are: 1.

2.

The CAPM is a good measure of risk and thus a good explanation why some stocks, portfolios, strategies or funds (assets, generically) earn higher average returns than others. Returns are unpredictable. In particular, (a)

3.

Stock returns are close to unpredictable. Prices are close to random walks; expected returns do not vary greatly through time. “Technical analysis” that tries to divine future returns from past price and volume data is nearly useless. Any apparent predictability is either a statistical artifact which will quickly vanish out of sample, or cannot be exploited after transactions costs. The near unpredictably of stock returns is simply stated, but its implications are many and subtle. (Malkiel 1990 is a classic and easily readable introduction.) It also remains widely ignored, and therefore is the source of lots of wasted trading activity. (b) Bond returns are nearly unpredictable. This is the expectations model of the term structure. If long term bond yields are higher than short term yields – if the yield curve is upward sloping – this does not mean that expected long-term bond returns are any higher than those on short term bonds. Rather, it means that short term interest rates are expected to rise in the future, so you expect to earn about the same amount on short term or long term bonds at any horizon. (c) Foreign exchange bets are not predictable. If a country has higher interest rates than are available in the U.S. for bonds of a similar risk class, its exchange rate is expected to depreciate. After you convert your investment back to dollars, you expect to make the same amount of money holding foreign or domestic bonds. (d) Stock market volatility does not change much through time. Not only are returns close to unpredictable, they are nearly identically distributed as well. Professional managers do not reliably outperform simple indices and passive portfolios once one corrects for risk (beta). While some do better than the market in any given year, some do worse, and the outcomes look very much like good and bad luck. Managers who do well in one year are not more likely to do better than average the next year. The average actively-managed fund does about 1% worse than the market index. The more actively a fund trades, the lower returns to investors. Together, these views reflected a guiding principle that asset markets are, to a good approximation, informationally efficient. (Fama 1970, 1991.) This statement means that market prices already contain most information about fundamental value. Informational efficiency 354

in turn derives from competition. The business of discovering information about the value of traded assets is extremely competitive, so there are no easy quick profits to be made, as there are not in every other well-established and competitive industry. The only way to earn large returns is by taking on additional risk. These statements are not doctrinaire beliefs. Rather, they summarize the findings of a quarter-century of extensive and careful empirical work. However, every single one of them has now been extensively revised by a new generation of empirical research. Now, it seems that : 1.

2.

There are assets, portfolios, funds, and strategies whose average returns cannot be explained by their market betas. Multifactor models dominate the empirical description, performance attribution, and explanation of average returns. Returns are predictable. In particular, (a)

3.

Variables including the dividend/price ratio and term premium can in fact predict substantial amounts of stock return variation. This phenomenon occurs over business-cycle and longer horizons. Daily, weekly and monthly stock returns are still close to unpredictable, and “technical” systems for predicting such movements are still close to useless after transactions costs. (b) Bond returns are predictable. Though the expectations model works well in the long run, a steeply upward sloping yield curve means that expected returns on long term bonds are higher than on short term bonds for the next year. (c) Foreign exchange returns are predictable. If you buy bonds in country whose interest rates are unusually higher than those in the U.S., you expect a greater return, even after converting back to dollars. (d) Stock market volatility does in fact change through time. Conditional second moments vary through time as well as first moments. Means and variances do not seem to move in lockstep, so conditional Sharpe ratios vary through time. Some funds seem to outperform simple indices, even after controlling for risk through market betas. Fund returns are also slightly predictable: past winning funds seem to do better in the future, and past losing funds seem to do worse than average in the future. For a while, this seemed to indicate that there is some persistent skill in active management. However, we now see that multifactor performance attribution models explain most fund persistence: funds earn persistent returns by following fairly mechanical “styles,” not by persistent skill at stock selection (Carhart 1997). Again, these views summarize a large body of empirical work. The strength and interpretation of many results are hotly debated. This new view of the facts need not overturn the view that markets are reasonable competitive and therefore reasonably efficient. It does substantially enlarge our view of what activities provide rewards for holding risks, and it challenges our economic understanding of those risk premia. As of the early 1970s, asset pricing theory anticipated the possibility and 355

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

even probability that expected returns should vary over time and that covariances past market betas would be important for understanding cross-sectional variation in expected returns. What took another 15 to 20 years was to see how important these long-anticipated theoretical possibilities are in the data.

20.1

Time-series predictability

I start by looking at patterns in expected returns over time in large market indices, and then look at patterns in expected returns across stocks. 20.1.1

Stocks

Dividend-price ratios forecast excess returns on stocks. Regression coefficients and R2 rise with the forecast horizon. This is a result of the fact that the forecasting variable is persistent. Table 1 gives a simple example of market return predictability. “Low” prices relative to dividends forecast higher subsequent returns. The one-year horizon 0.17 R2 is not particularly remarkable. However, at longer and longer horizons larger and larger fractions of return variation are forecastable. At a 5 year horizon 60% of the variation in stock returns is forecastable ahead of time from the price/divided ratio! (Fama and French 1988.) Horizon k (years) 1 2 3 5

Rt→t+k = a + b(Dt /Pt ) b σ(b) R2 5.3 (2.0) 0.15 10 (3.1) 0.23 15 (4.0) 0.37 33 (5.8) 0.60

Dt+k /Dt = a + b(Dt /Pt ) b σ(b) R2 2.0 (1.1) 0.06 2.5 (2.1) 0.06 2.4 (2.1) 0.06 4.7 (2.4) 0.12

Table 1. OLS regressions of percent excess returns (value weighted NYSE - treasury bill rate) and real dividend growth on the percent VW dividend/price ratio. Rt→t+k indicates the k year return. Standard errors in parenthesis use GMM to correct for heteroskedasticity and serial correlation. Sample 1947-1996. One can object to dividends as the divisor for prices. However, ratios formed with just about any sensible divisor works about as well, including earnings, book value, and moving averages of past prices. Many other variables forecast excess returns, including the term spread between long and short term bonds, the default spread, the level of the T-bill rate, (Fama and French 1989,) 356

S ECTION 20.1

T IME - SERIES PREDICTABILITY

the detrended T-bill rate, and the earnings/dividend ratio (Lamont 1998). Macro variables forecast stock returns as well, including the investment/capital ratio (Cochrane 1991) and the consumption/wealth ratio (Lettau and Ludvigson 2000). Most of these variables are correlated with each other and correlated with or forecast business cycles. This fact suggests a natural explanation, emphasized by Fama and French (1999): Expected returns vary over business cycles; it takes a higher risk premium to get people to hold stocks at the bottom of a recession. When expected returns go up, prices go down. We see the low prices, followed by the higher returns expected and required by the market. (Regressions do not have to have causes on the right and effects on the left. You run regressions with the variable orthogonal to the error on the right, and that is the case here since the error is a forecasting error. This is like a regression of actual weather on a weather forecast.) Table LL, adapted from Lettau and Ludvigson (2000) compares several of these variables. At a one year horizon, both the consumption/wealth ratio and the detrended T bill rate forecast returns, with R2 of 0.18 and 0.10 respectively. At the one year horizon, these variables are more important than the dividend/price and dividend/earnings ratios, and their presence cuts the dividend ratio coefficients in half. However, the d/p and d/e ratios are slower moving than the t bill rate and consumption/wealth ratio. They track decade-to-to decade movements more than business cycle movements. This means that their importance builds with horizon. By six years, the bulk of the return forecastability again comes from the dividend ratios, and it is their turn to cut down the cay and t-bill regression coefficients. The cay and d/e variables have not been that affected by the late 90s, while it has substantially cut down dividend yield forecastability. Horizon(years) 1 1 1 1 6 6 6 6

cay 6.7

5.4 12.4 5.9

d−p

d−e

0.14

0.08

0.07

-0.05

0.95

0.68

0.89

0.65

rrel

-4.5 -3.8 -5.10 1.36

R2 0.18 0.04 0.10 0.23 0.16 0.39 0.03 0.42

Table LL. Long-horizon return forecasts. The return variable is log excess returns on the S&P composite index. cay is Lettau and Ludvigson’s consumption to wealth ratio. d − p is the log dividend yield and e − p is the log earnings yield. rrel is a detrended short term interest rate. Sample 1952:4-1998:3. Source: Lettau and Ludvigson (2000) Table 5. I emphasize that excess returns are forecastable. We have to understand this as timevariation in the reward for risk, not time-varying interest rates. One naturally slips in to non-risk explanations for price variation; for example that the current stock market boom is due to life-cycle savings of the baby boomers. A factor like this does not reference risks; it 357

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

predicts that interest rates should move just as much as stock returns. Persistent d/p; Long horizons are not a separate phenomenon

The results at different horizons are not separate facts, but reflections of a single underlying phenomenon. If daily returns are very slightly predictable by a slow-moving variable, that predictability adds up over long horizons. For example, you can predict that the temperature in Chicago will rise about 1/3 degree per day in the springtime. This forecast explains very little of the day to day variation in temperature, but tracks almost all of the rise in temperature from January to July. Thus, the R2 rises with horizon. Thus, a central fact driving the predictability of returns is that the dividend price ratio is very persistent. Figure 37 plots the d/p ratio and you can see directly that it is extremely slow-moving. Below, I will estimate an AR(1) coefficient around 0.9 in annual data.

Figure 37. To see more precisely how the results at various horizons are linked, and how they result from the persistence of the d/p ratio, suppose that we forecast returns with a forecasting variable x, according to rt+1 xt+1

= axt + εt+1 = ρxt + δ t+1 .

(20.298) (20.299)

(0bviously, you demean the variables or put constants in the regressions.) Small values of b and R2 in (20.298) and a large coefficient ρ in (20.299) imply mathematically that the 358

S ECTION 20.1

T IME - SERIES PREDICTABILITY

long-horizon regression has a large regression coefficient b and large R2 . To see this, write rt+1 + rt+2 rt+1 + rt+2 + rt+3

= a(1 + ρ)xt + aδ t+1 + εt+1 + εt+2 = a(1 + ρ + ρ2 )xt + aρδ t+1 + aδ t+2 + εt+1 + εt+2 + εt+3 .

You can see that with ρ near one, the coefficients increase with horizon, almost linearly at first and then at a declining rate. The R2 are a little messier to work out, but also rise with horizon. The numerator in the long-horizon regression coefficient is E [(rt+1 + rt+2 + ... + rt+k ) xt ]

(300)

where the symbols represent deviations from their means. With stationary r and x, E(rt+j xt ) = E(rt+1 xt−j ), so this is the same moment as E [rt+1 (xt + xt−1 + xt−2 + ...)] ,

(301)

the numerator of a regression coefficient of one year returns on many lags of price dividend ratios. Of course, if you run a multiple regression of returns on lags of p/d, you quickly find that most lags past the first do not help the forecast power. (That statement would be exact in the AR(1) example.) This observation shows once again that one-year and multi-year forecastability are two sides of the same coin. It also suggests that on a purely statistical basis, there will not be a huge difference one-year return forecasts and multi-year return forecasts (correcting the latter for the serial correlation of the error term due to overlap). Hodrick (1991) comes to this conclusion in a careful Monte Carlo experiment, comparing moments of the form (20.300), (20.301) and E(rt+1 xt ). The multi-year regressions, or the implied multi year regressions from one-year forecasts with a slow moving right hand variable are thus mostly useful for illustrating the dramatic economic implications of forecastability, rather than as clever statistical tools that enhance power and allow us to distinguish previously foggy hypotheses. The slow movement of the price-dividend ratio means that on a purely statistical basis, return forecastability is a very open question. What we really know (see Figure 37) is that low prices relative to dividends and earnings in the 50’s preceded the boom market of the early 60’s; that the high price-dividend ratios of the mid-60’s preceded the poor returns of the 70’s; that the low price ratios of the mid-70’s preceded the current boom. We really have three data postwar data points; a once per generation change in expected returns. In addition, the last half of the 1990s has seen a historically unprecedented rise in stock prices and price/dividend ratios (or any other ratio). This rise has cut the postwar return forecasting regression coefficient in half. On the other hand, another crash or even just a decade of poor returns will restore the regression. Data back to the 1600s show the same pattern, but we are often uncomfortable making inferences from centuries-old data. 359

C HAPTER 20 20.1.2

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

Volatility

Price dividend ratios can only move at all if they forecast future returns, if they forecast future dividend growth, or if there is a bubble – if the price-dividend ratio is nonstationary and is expected to grow explosively. In the data, most variation in price-dividend ratios results from varying expected returns. “Excess volatility” – relative to constant discount rate present value models – is thus exactly the same phenomenon as forecastable long-horizon returns. I also derive the very useful price-dividend and return linearizations. Ignoring constants (means), pt − dt rt − Et−1 rt rt+1

= Et

∞ X j=1

ρj−1 (∆dt+j − rt+j )

∞ ∞ X X = (Et − Et−1 ) ρj ∆dt+j − ρj rt+j j=0

j=1

= ∆dt+1 − ρ(dt+1 − pt+1 ) + (dt − pt ).

The volatility test literature starting with Shiller (1981) and LeRoy and Porter (1981) (See Cochrane 1991 for a review) started out trying to make a completely different point. Predictability seems like a sideshow. The stunning fact about the stock market is its extraordinary volatility. On a typical day, the value of the U.S. capital stock changes by a full percentage point, and days of 2 or 3 percentage point changes are not uncommon. In a typical year it changes by 16 percentage points, and 30 percentage point changes are not uncommon. Worse, most of that volatility seems not to be accompanied by any important news about future returns and discount rates. 30% of the capital stock of the United States vanished in a year and nobody noticed? Surely, this observation shows directly that markets are “not efficient” – that prices do not correspond to the value of capital – without worrying about predictability? It turns out however, that “excess volatility” is exactly the same thing as return predictability. Any story you tell about prices that are “too high” or “too low” necessarily imply that subsequent returns will be too low or too high as prices rebound to their correct levels. When prices are high relative to dividends (or earnings, cashflow, book value or some other divisor), one of three things must be true: 1) Investors expect dividends to rise in the future. 2) Investors expect returns to be low in the future. Future cashflows are discounted at a lower than usual rate, leading to higher prices. 3) Investors expect prices to rise forever, giving an adequate return even if there are no dividends. This statement is not a theory, it is an identity: If the price-dividend ratio is high, either dividends must rise, prices must decline, or the price-dividend ratio must grow explosively The open question is, which option holds for 360

S ECTION 20.1

T IME - SERIES PREDICTABILITY

our stock market? Are prices high now because investors expect future earnings, dividends etc. to rise, because they expect low returns in the future, or because they expect prices to go on rising forever? Historically, we find that virtually all variation in price-dividend ratios has reflected varying expected excess returns. Exact present value identity

To document this statement, we need to relate current prices to future dividends and returns. Start with the identity −1 1 = Rt+1 Rt+1 = R−1 t+1

Pt+1 + Dt+1 Pt

(302)

and hence µ ¶ Pt Pt+1 Dt+1 −1 = Rt+1 1 + . Dt Dt+1 Dt

We can iterate this identity forward and take conditional expectations to obtain the identity Ã j ! ∞ X Y Pt −1 = Et Rt+k ∆Dt+k (303) Dt j=1 k=1

where ∆Dt ≡ Dt /Dt−1 . (We could iterate (20.302) forward to Ã j ! ∞ X Y −1 Pt = Rt+k Dt+j , j=1

k=1

but prices are not stationary, so we can’t find the variance of prices from a time-series average. Much of the early volatility test literature concerned stationarity problems. Equation (20.303) also requires a limiting ³Q condition´ that the price dividend ratio cannot explode faster than j −1 returns, limj→∞ Et k=1 Rt+k Pt+j /Dt+j . I come back to this condition below) Equation (??) shows that high prices must, mechanically, come from high future dividend growth or low future returns.

Approximate identity

The nonlinearity of (20.303) makes it hard to handle, and means that we cannot use simple time-series tools. You can linearize (20.303) directly with a Taylor expansion ( Cochrane 1991 takes this approach.) Campbell and Shiller (1988) approximate the one period return identity before iterating, which is algebraically simpler and is the most popular linearization. Start again from the obvious, −1 1 = R−1 t+1 Rt+1 = Rt+1

361

Pt+1 + Dt+1 . Pt

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

Multiplying both sides by Pt /Dt and massaging the result, µ ¶ Pt Pt+1 Dt+1 −1 = Rt+1 . 1+ Dt Dt+1 Dt Taking logs, and with lowercase letters denoting logs of uppercase letters, ¢ ¡ pt − dt = −rt+1 + ∆dt+1 + ln 1 + ept+1 −dt+1

Taking a Taylor expansion of the last term about a point P/D = ep−d µ ¶ P P pt − dt = −rt+1 + ∆dt+1 + ln 1 + + D P [pt+1 − dt+1 − (p − d)] D 1+ D (20.304) pt − dt = −rt+1 + ∆dt+1 + k + ρ (pt+1 − dt+1 ) . Since the average dividend yield is about 4% and average price/dividend ratio is about 25, ρ is a number very near one. I will use ρ = 0.96 for calculations, ρ=

1 P/D = ≈ 1 − D/P = 0.96. 1 + P/D 1 + D/P

Without the constant k, the equation can also apply to deviations from means or any other point. Now, iterating forward is easy, and results in the approximate identity pt − dt = const. +

∞ X j=1

ρj−1 (∆dt+j − rt+j ).

(305)

(Again, we need a condition that pt − dt does not explode faster than ρ−t , limj→∞ ρj (pt+j − dt+j ) = 0. I return to this condition below.) Since (20.305) holds ex-post, we can take conditional expectations and relate price-dividend ratios to ex-ante dividend growth and return forecasts pt − dt = const. + Et

∞ X j=1

ρj−1 (∆dt+j − rt+j ).

(306)

Now it is really easy to see that a high price-dividend ratio must be followed by high dividend growth ∆d, or low returns r. Which is it? Decomposing the variance of price-dividend ratios

To address this issue, equation (20.305) implies ∞ ∞ X X var(pt − dt ) = cov pt − dt , ρj−1 ∆dt+j − cov pt − dt , ρj−1 rt+j (307) j=1

j=1

362

S ECTION 20.1

T IME - SERIES PREDICTABILITY

In words, price-dividend ratios can only vary if they forecast changing dividend growth or of they forecast changing returns. (To derive 20.307 from (20.305), multiply both sides by (pt − dt ) − E(pt − dt ) and take expectations.) Notice that both terms on the right hand side of (20.307) are the numerators of exponentially weighted long-run regression coefficients. This is a powerful equation. At first glance, it would seem a reasonable approximation that returns are unforecastable (the “random walk” hypothesis) and that dividend growth is not forecastable either. But if this were the case, the price/dividend ratio would have to be a constant. Thus the fact that the price/dividend ratio varies at all means that either dividend growth or returns must be forecastable – that the world is not i.i.d. At a simple level, Table 1 includes regressions of long-horizon dividend growth on dividend/price ratios to match the return regressions. The coefficients in the dividend growth case are much smaller, typically one standard error from zero, and the R2 are tiny. Worse, the signs are wrong. To the extent that a high price-dividend ratio forecasts any change in dividends, it seems to forecast a small decline in dividends! Having seen equation (20.307), one is hungry for estimates. Table 2 presents some, taken from Cochrane (1991b). As one might suspect from Table 1, Table 2 shows that in the past almost all variation in price-dividend ratios is due to changing return forecasts. The elements do not have to be between 0 and 100%. For example, -34, 138 occurs because high prices seem to forecast lower real dividend growth (though this number is not statistically significant). Therefore they must and do forecast really low returns, and returns must account for more than 100% of price-dividend variation.

Real std. error Nominal std. error

Dividends -34 10 30 41

Returns 138 32 85 19

Table 2. Variance decomposition of value-weighted NYSE price-dividend ratio. Table entries are the percent of the variance of theP price-dividend ratio attributable j−1 ∆dt+j )/var(pt −dt ) to dividend and return forecasts, 100×cov(pt −dt , 15 j=1 ρ and similarly for returns. This observation solidifies one’s belief in price-dividend ratio forecasts of returns. Yes, the statistical evidence that price-dividend ratios forecast returns is weak, and many return forecasting variables have been tried and discarded, so selection bias is a big worry in forecasting regressions. But the price-dividend ratio (or price-earning, market to book, etc.) has a special status since it must forecast something. To believe that the price-dividend ratio is stationary and varies, but does not forecast returns, you have to believe that the price-dividend ratio does forecast dividends. Given this choice and Table 1, it seems a much firmer conclusion that it forecasts returns. 363

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

It is nonetheless an uncomfortable fact that almost all variation in price-dividend ratios is due to variation in expected excess returns. How nice it would be if high prices reflected expectations of higher future cashflows. Alas, that seems not to be the case. If not, it would be nice if high prices reflected lower interest rates. Again, that seems not to be the case. High prices reflect low risk premia, lower expected excess returns. Campbell’s return decomposition.

Campbell (1991) provides a similar decomposition for unexpected returns, ∞ ∞ X X rt − Et−1 rt = (Et − Et−1 ) ρj ∆dt+j − ρj rt+j . j=0

(308)

j=1

A positive shock to returns must come from a positive shock to forecast dividend growth, or to a negative shock to forecast returns. Since a positive shock to time t dividends is directly paid as a return, (the first sum starts at j = 0), Campbell finds some fraction of return variation is due to current dividends. However, once again, the bulk of index return variation comes from shocks to future returns, i.e. discount rates. To derive (20.308), start with the approximate identity (20.305), and move it back one period pt−1 − dt−1 = const. +

∞ X j=0

ρj (∆dt+j − rt+j ).

Now take innovations of both sides, 0 = (Et − Et−1 )

∞ X j=0

ρj (∆dt+j − rt+j ).

Pulling rt over to the left hand side, you obtain (20.308). (Problem 3 at the end of the chapter guides you through an alternative and more constructive derivation.) Cross-section

So far, we have concentrated on the index. One can apply the same analysis to firms. What causes the variation in price-dividend ratios, or, better book/market ratios (since dividends can be zero) across firms, or over time for a given firm? Vuolteenaho (2000) applies the same sort of analysis to individual stock data. He finds that as much as half of the variation in individual firm book/market ratios reflect expectations of future cashflows. Much of the expected cashflow variation is idiosyncratic, while the expected return variation is common, which is why variation in the index book/market ratio, like variation in the index dividend/price ratio, is almost all due to varying expected excess returns. Bubbles

364

S ECTION 20.1

T IME - SERIES PREDICTABILITY

In deriving the exact and linearized present value identities, I assumed an extra condition that the price-dividend ratio does not explode. Without that condition, and taking expectations of both sides, the exact identity reads ∞ X Pt = Et Dt j=1

Ã

j Y

k=1

−1 Rt+k ∆Dt+k

!

+ lim Et j→∞

Ã

j Y

k=1

R−1 t+k

!

Pt+j Dt+j

(309)

and the linearized identity reads pt − dt = const. + Et

∞ X j=1

ρj−1 (∆dt+j − rt+j ) + Et lim ρj (pt+j − dt+j ). j→∞

(310)

As you can see, the limits in the right hand sides of (20.309) and (20.310) are zero if the price-dividend ratio is stationary, or even bounded. For these terms not to be zero, the price dividend ratio must be expected to grow explosively, and faster than R or ρ−1 . Especially in the linearized form 20.310 you can see that stationary r, ∆d and implies stationary p − d if the last term is zero, and p − d is not stationary if the last term is not zero. Thus, you might want to rule out these terms just based on the view that price dividend ratios do not and are not expected to explode in this way. You can also invoke economic theory to rule them out. The last terms must be zero in an equilibrium of infinitely lived agents or altruistically linked generations. If wealth explodes, optimizing long-lived agents will consume more. Technically, this limiting condition is a first order condition for optimality just like the period to period first order condition. The presence of the last term also presents an arbitrage opportunity in complete markets, as you can short a security whose price contains the last term, buy the dividends separately and eat the difference right away. On the other hand, there are economic theories that permit the limiting terms – overlapping generations models, and they capture the interesting possibility of “rational bubbles” that many observers think they see in markets, and that have sparked a huge literature and a lot of controversy. An investor holds a security with a rational bubble not for any dividends, but on the expectation that someone else will pay even more for that security in the future. This does seem to capture the psychology of investors from the tulip bubble of 17th century Holland to the dot-com bubble of the millenial United States – why else would anyone buy Cisco systems at a price-earnings ratio of 217 and market capitalization 10 times that of General Motors in early 2000? A “rational bubble” imposes a little discipline on this centuries old psychological description, however, by insisting that the person who is expected to buy the security in the future also makes the same calculation. He must expect the price to rise even further. Continuing recursively, the price of a rational bubble must be expected to rise forever. A Ponzi scheme, in which everyone knows the game will end at some time, cannot rationally get off the ground. The expectation that prices will grow at more than a required rate of return forever does not 365

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

mean that sample paths do so. For example, consider the bubble process ( ) Pt R−1 γRPt prob = γP t R−1 Pt+1 = . R 1 prob = γPPttR−1 Figure 38 plots a realization of this process with γ = 1.2. This process yields an expected return R, and the dashed line graphs this expectation as of the first date. Its price is positive though it never pays dividends. It repeatedly grows with a high return γR for a while and then bursts back to one. The expected price always grows, though almost all sample paths do not do so.

Figure 38. Sample path from a simple bubble process. The solid line gives the bubble. The dashed line gives the expected value of the bubble as of time zero, i.e. pRt 0 . Infinity is a long time. It’s really hard to believe that prices will rise forever. The solar system will end at some point; any look at the geological and evolutionary history of the earth suggests that our species will be around a lot less than that. Thus, the infinity in the bubble must really be a parable for “a really long time.” But then the “rational” part of the bubble pops – it must hinge on the expectation that someone will be around to hold the bag; to buy a security without the expectation of dividends or further price increases. (The forever part of usual present value formulas is not similarly worrying because 99.99% of the value comes from the first few hundred years of dividends.) 366

S ECTION 20.1

T IME - SERIES PREDICTABILITY

Empirically, bubbles do not appear to be the reason for historical price-dividend ratio variation. First, price-dividend ratios do seem stationary. (Craine 1993 runs a unit root test with this conclusion.) Even if statistical tests are not decisive, as is expected for a slow moving series, or a series such as that plotted in Figure 38, it is hard to believe that pricedividend ratios can explode rather than revert back to their four-century average level of about 20 to 25. Second, Table 2 shows that return and dividend forecastability terms add up to 100% of the variance of price-dividend ratios. In a bubble, we would expect price variation not matched by any variation in expected returns or dividends, as is the case in Figure 38. I close with a warning: The word “bubble” is widely used to mean very different things. Some people seem to mean any large movement in prices. Others mean large movements in prices that do correspond to low or perhaps negative expected excess returns (I think this is what Shiller 2000 has in mind), i.e. any price movement not explained by a present value model with constant expected returns. 20.1.3

A simple model for digesting predictability

To unite the various predictability and return observations, I construct a simple VAR representation for returns, price growth, dividend growth, dividend price ratio. I start only with a slow moving expected return and unforecastable dividends. This specification implies that d/p ratios reveal expected returns. This specification implies return forecastability. To believe in a lower predictability of returns, you must either believe that dividend growth really is predictable, or that the d/p ratio is really much more persistent than it appears to be. This specification shows that small but persistent changes in expected returns add up to large price changes. We have isolated two important features of the long-horizon forecast phenomenon: dividend/price ratios are highly persistent, and dividend growth is essentially unforecastable. Starting with these two facts, a simple VAR representation can tie together many the predictability and volatility phenomena. Start by specifying a slow-moving state variable xt that drives expected returns, and unforecastable dividend growth, xt rt+1 ∆dt+1

= bxt−1 + δ t = xt + εrt+1 = εdt+1

(20.311) (20.312) (20.313)

All variables are demeaned logs. (The term structure models of Chapter 19 were of this 367

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

form.) From this specification, using the linearized present value identity and return, we can derive a VAR representation for prices, returns, dividends, and the dividend price ratio, (dt+1 − pt+1 ) = b(dt − pt ) + rt+1 ∆pt+1 ∆dt+1

δ t+1 1 − ρb

(20.314)

µ ¶ ρ δ t+1 = (1 − ρb) (dt − pt ) + εdt+1 − 1 − ρb µ ¶ 1 δ t+1 = (1 − b) (dt − pt ) + εdt+1 − 1 − ρb = εdt+1

(20.315) (20.316) (20.317)

Dividend-price ratio: Using the approximate present value identity, we can find the dividend price ratio dt − pt = Et

∞ X j=1

ρj−1 (Et rt+1 − Et dt+j ) =

xt . 1 − ρb

(318)

This equation makes precise my comments that the dividend price ratio reveals expected returns. Obviously, the feature that the dividend price ratio is exactly proportional to the expected return does not generalize. If dividend growth is also forecastable, then the dividendprice ratio is a combination of dividend growth and return forecasts. Actual return forecasting exercises can often benefit from cleaning up the dividend price ratio to focus on the implied return forecast. Returns: Since we know where the dividend/price ratio and dividends are going, we can figure out where returns are going. Use the return linearization (this is equivalent to (20.304)) Rt+1 rt+1

µ ¶ Pt+1 Dt+1 Pt = 1+ / Dt+1 Dt Dt = ρ(pt+1 − dt+1 ) + (dt+1 − dt ) − (pt − dt ).

(20.319)

Now, plug in the from (20.314) and (20.313) to get (20.315). Prices: Write pt+1 − pt = −(dt+1 − pt+1 ) + (dt − pt ) + (dt+1 − dt ).

(320)

Then, plugging in from (20.314) and (20.313), we get (20.316). We can back out parameters from the reduced form return - d/p VAR. (Any two equations carry all the information of this system.) Table RR presents some estimates. 368

S ECTION 20.1 Sample 27-98 48-98 27-92 48-92

a 0.16 0.14 0.28 0.27

T IME - SERIES PREDICTABILITY

a, D/P 4.7 4.0 6.7 6.2

b 0.92 0.97 0.82 0.87

σ(εr ) 19.2 15.0 19.0 14.5

σ(εdp ) 15.2 12.6 15.0 12.4

ρ(εr , εdp ) -0.72 -0.71 -0.69 -0.67

Table RR. Estimates of log excess return and log dividend-price ratio regressions, using annual CRSP data. r is the difference between the log value weighted return and the log treasury bill rate. The estimates are of the system rt+1 dt+1 − pt+1

= a(dt − pt ) + εrt+1 = b(dt − pt ) + εdp,t+1

and rt+1 = (a, D/P )

Dt + εt+1 Pt

I report both the more intuitive coefficients on the actual d/p ratio and the coefficients on the log d/p ratio, which is a more useful specification for our transformations. The two line up; a coefficient of 5 on Dt /Pt implies a coefficient of 5×D/P ≈ 0.25 on (Dt /Pt ) /(D/P ).

You can see that the parameters depend substantially on the sample. In particular, the dramatic returns of the late 1990s, despite low dividend yields, cut the postwar return forecast coefficients in half and the overall sample estimate by about one third. That dramatic decline in the d/p ratio also induces a very high apparent persistence in the d/p ratio, rising to a 0.97 estimate in the 48-98 sample. (Faced with an apparent trend in the data, an autoregression estimates a root near unity.) With these estimates in mind, given the considerations outlined below, I will make calculations using reduced form parameters b ρ σ(εr ) σ(εdp ) ρr,dp

= = = = =

0.9 0.96 15 12.5 −0.7

(20.321)

From these parameters, we can find the underlying parameters of (20.311)-(20.313). I com369

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

ment on each one below as it becomes useful. σ(δ) = σ(εdp )(1 − ρb) = 1.7 (20.322) q σ(εd ) = σ(εr + ρεdp ) = σ2 (εr ) + ρ2 σ 2 (εdp ) + 2ρσ(εr , εdp ) = 10.82

σ(εd , εdp ) = σ(εr εdp ) + ρσ 2 (εdp ) ρ(εr εdp )σ(εr ) + ρσ(εdp ) ρ(εd , εdp ) = = 0.139 σ(εd )

(20.323)

The size of the return forecasting coefficient.

Does the magnitude of the estimated predictability make sense? Given the statistical uncertainties, do other facts guide us to higher or lower predictability? The coefficient of the one year excess return on the dividend price ratio in Table 1 is about 5, and the estimates in Table RR vary from 4 to 6 depending on the sample. These values are surprisingly large. For example, a naive investor might think that dividend yields move onefor-one with returns; if they pay more dividends, you get more money. Before predictability, we would have explained that high dividend yield means that prices are low in anticipation of lower future dividends, leaving the expected return unchanged. Now we recognize the possibility of time-varying expected returns, but does it make sense that expected returns move even more than dividend yields? Return forecastability follows from the fact that dividends are not forecastable, and that the dividend/price ratio is highly but not completely persistent. We see this in the calculated coefficients of prices and returns on the dividend price ratio in (20.315) and (20.316). We derived rt+1 ∆pt+1

= (1 − ρb) (dt − pt ) + εrt+1 = (1 − b) (dt − pt ) + εpt+1

Since dividends are not forecastable, it is no surprise that the formulas for price growth and return are so similar. The return formula basically just adjusts for the fact that a higher dividend yield directly contributes to return by paying more dividends. To transform units to regressions on D/P, multiply by 25, e.g. rt+1 =

1 − ρb Dt + εrt+1 . D/P Pt

Suppose the d/p ratio were not persistent at all–b = 0. Then both return and price growth coefficients should be 1 in logs or about 25 in levels! If the d/p ratio is one percentage point above its average, we must forecast enough of a rise in prices to restore the d/p ratio to its average in one year. The average d/p ratio is about 4%, though, so prices and hence returns must rise by 25% to change the d/p ratio by one percentage point. d(D/P ) = −D/P d(P )/P .

Suppose instead that the d/p ratio were completely persistent i.e. a random walk with 370

S ECTION 20.1

T IME - SERIES PREDICTABILITY

b = 1. Then the return coefficient is 1 − ρ = 0.04, and about 1.0 in levels, while the price coefficient is 0. If the d/p ratio is one percent above average and expected to stay there, and dividends are not forecastable, then prices must not be forecast to change either. The return is one percentage point higher, because you get the higher dividends as well. Thus, the naive investor who expects dividend yield to move one for one with returns not only implicitly assumes that dividends are not forecastable – which turns out to be true – but also that the d/p ratio will stay put forever.

A persistence parameter b = 0.90 implies price and return regression coefficients of

1 − b = 0.10 1 − ρb = 1 − 0.96 × 0.90 = 0.14

or about 2.5 and 3.4 in levels. If the dividend yield is one percentage point high, and is expected to be 0.9 percentage points high in one year, then prices must increase by P/D × 0.1 percentage points in the next year. The return gets the additional dividend. This, fundamentally, is how zero forecastability of dividends implies that returns move more than one for one with the dividend yield. This is a little below the sample estimates in Table 1 and Table RR of 4-6. That is because in the sample, a high price seems to forecast even lower dividend growth – the wrong sign, which is hard to believe. To continue with a calibration that consistently captures the facts with no dividend forecastability, we either have to lower the persistence coefficient or lower the return forecasting coefficient from the values reported in Table RR. A persistence b = 0.8 implies a return coefficient (1− ρb) = (1− 0.96× 0.8) = 0.23 or in levels 0.23 ×25 = 5.75. However, given the uncertainties of dividend/price forecastability, it seems more sensible to continue calculations with b = 0.9 and corresponding return coefficient of 0.14,equal to the estimate in the 48-98 sample. Going in the other direction, statistical uncertainty, the recent runup in stocks despite low dividend yields, and the dramatic portfolio implications of time-varying returns for investors whose risks or risk aversion do not change over time all lead one to consider lower predictability. As we see from these calculations though, there are only two ways to make sense of lower predictability. You could follow the “new economy” advocates, and believe that this time, prices really are rising on advance news of dividend growth, even though prices have not forecast dividend growth in the past. If not, you have to believe that dividend price ratios are substantially more persistent than they have seemed in the postwar data. Much more persistent d/p is a tough road to follow, since D/P ratios already move incredibly slowly. Now, they basically change sign once a generation; high in the 50’s, low in the 60’s, high in the mid-70’s, and decreasing ever since (see Figure 37.) As a quantitative example, suppose the D/P ratio had an AR(1) coefficient of 0.96 in annual data. This means a half life of ln 0.5/ ln 0.96 = 17 years. In this case, the price coefficient would be coefficient 371

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

would be 1 − 0.96 1−b = =1 D/P 0.04

and the return coefficient would be 1 − 0.962 1 − ρb = ≈2 D/P 0.04

A one percentage point higher d/p ratio means that prices must rise 1 percentage point next year, so returns must be about 2 percentage points higher. A two for one movement of expected returns with the dividend yield thus seems about the lower bound for return predictability, so long as dividend growth remains unforecastable. Persistence, price volatility and expected returns

From the dividend-price ratio equation (20.314) we can find the volatility of the dividend price ratio and related it to the volatility and persistence of expected returns. σ(dt − pt ) =

1 σ (xt ) 1 − ρb

With b = 0.9, 1/(1 − ρb) = 1/(1 − 0.96 × 0.9) = 7. 4. Thus, the high persistence of expected returns means that a small expected return variation translates into a potentially very large price variation; or equivalently that very large price variations, unaccounted for by forecasts of dividend variation, can be explained by small variation in expected returns. Translating to levels, a one percentage point change in expected returns with persistence b = 0.9 corresponds to a 7.4% increase in price. The Gordon growth model is a classic and even simpler way to see this point. With constant dividend growth g and return r, the present value identity becomes P =

D . r−g

A price-dividend ratio of 25 means r − g = 0.04. Then, a one percentage point permanent change in expected return translates into a 25 percentage point change in price! This is an overstatement, since expected returns are not this persistent, but it allows you to clearly see the point. This point also shows that small market imperfections in expected returns can translate into substantial market imperfections in prices, if those expected return changes are persistent. We know markets cannot be perfectly efficient (Grossman and Stiglitz 1980). If they were perfectly efficient, there would be no traders around to make them efficient. Especially in situations where short sales or arbitrage are constrained by market frictions, prices of similar assets can be substantially different, while the expected returns of those assets are almost the same. For example the “closed end fund” puzzle (Thompson 1978) noted that baskets of securities sold for substantial price discounts relative to the sum of the individual securi372

S ECTION 20.1

T IME - SERIES PREDICTABILITY

ties. However, these price differentials persist for a long time. You can’t short the closed end funds to buy the securities and keep that short position on for years. 20.1.4

Mean-reversion

I introduce long-horizon return regressions and variance ratios. I show that they are related: each one picks up a string of small negative return autocorrelations. I show though that the direct evidence for mean reversion and Sharpe ratios that rise with horizon is weak. Long run regressions and variance ratios

The first evidence of long-run forecastability in the stock market did not come from d/p regressions, but rather from clever ways of looking at the long-run univariate properties of returns. Fama and French (1988a) ran regressions of long-horizon returns on past longhorizon returns, rt→t+k = a + bk rt−k→t + εt+k ,

(324)

basically updating classic autocorrelation tests from the 60s to long horizon data. They found negative and significant b coefficients: a string of good past returns forecasts bad future returns. Poterba and Summers (1988) considered a related “variance ratio” statistic. If stock returns are i.i.d., then the variance of long horizon returns should grow with the horizon var(rt→t+k ) = var(rt+1 + rt+2 + .. + rt+k ) = kvar(rt+1 ).

(325)

They computed the variance ratio statistic vk =

1 var(rt→t+k ) . k var(rt+1 )

They found variance ratios below one. Stocks, it would seem, really are safer for “long-run investors” who can “afford to wait out the ups and downs of the market,” common Wall Street advice, long maligned by academics. These two statistics are closely related, and reveal the same basic fact: stock returns have a string of small negative autocorrelations. To see this relation, write the variance ratio statistic

vk =

1 var k

³P k

j=1 rt+j

var(rt+1 )

´

=

k k X X |k − j| |k − j| ρj = 1 + 2 ρj , k k j=1

j=−k

373

(326)

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

and the regression coefficient in (20.324) k k X X 1 cov bk = rt+j , rt−j+1 var(rt→t+k ) j=1 j=1 =

k k 1 X |k − j| k var(rt+1 ) X |k − j| ρk+j = ρk+j . var(rt→t+k ) k vk k j=−k

j=−k

Both statistics are based on tent-shaped sums of autocorrelations, as illustrated by Figure 39. If there are many small negative autocorrelations which bring returns back slowly after a shock, these autocorrelations might be individually insignificant. Their sum might be economically and statistically significant, however, and these two statistics will reveal that fact by focusing on the sum of autocorrelations. The long-horizon regression weights emphasize the middle of the autocorrelation function, so a k year horizon long-horizon regression is comparable to a somewhat longer variance ratio.

Variance ratio weights Long horizon regression weights

Return autocorrelations Figure 39. Long horizon regression and variance ratio weights on autocorrelations. Moving average representation and mean reversion

The “mean-reversion” description of these statistics comes from their implications for where values go at long horizons following a shock. We can show that the square root of the variance ratio measures the long-horizon impact of a shock relative to its instantaneous impact – the extent to which values revert back towards their mean following a shock. You can always write returns as a moving average of their own shocks. From a regression of returns on past returns a(L)rt = εt

374

(327)

S ECTION 20.1

T IME - SERIES PREDICTABILITY

you can find the θj in rt =

∞ X

θj εt−j = θ(L)εt = a(L)−1 εt .

j=0

(Most simply, just simulate (20.327) forward.) The θj are the moving average representation or impulse-response function – they tell you the path of expected returns following a shock. Let vt represent the cumulative returns, or the log value of a dollar invested, ∆vt = rt . P Then, the partial sum kj=1 θj tells you the effect on invested wealth vt+k of a univariate return shock εt

Relating variance ratios, long-horizon regressions and moving averages for finite k is possible but not pretty. However, we can nicely relate the limiting response – where limk→∞ Et vt+k ends up after a shock – to the autocorrelations, and thus to the limit of the variance ratio statistic very simply as

1+2

∞ X j=1

2 ∞ X ρj = θ j /σ2ε .

(328)

j=0

If returns are i.i.d., the variance ratio is one at all horizons; all autocorrelations are zero, and all θ past the first are zero so the long-run price moves one for one with the shock. A longP string of small negative autocorrelations means a variance ratio less than one, and means ∞ j=0 θj < 1 so the long-run effect on price is lower than the impact effect - this is “mean-reversion.” The right hand equality of (20.328) follows by just taking the k → ∞ in (20.326). For the second equality, you can recognize in both expressions the spectral density of r at frequency zero. (Cochrane 1986 discusses these and other properties of variance ratios.)

Numbers

Table A1 presents an estimate of the variance of long-horizon returns and long-horizon return regressions. The long-horizon regressions do show some interesting mean reversion, especially in the 3-5 year range. However, that turns around at year 7 and disappears by year 10. The variance ratios do show some long-horizon stabilization. At year 10, the variance ratio is (16.3/19.8)2 = 0.68, and the long-run price impact of a shock is 16.8/19.8 = 0.85. The mean log return grows linearly with horizon whether returns are autocorrelated or not – E(r1 + r2 ) = 2E(r). If the variance also grows linearly with the horizon, as it does for non-autocorrelated returns, then the Sharpe ratio grows with the square root of horizon. If the variance grows more slowly than horizon, then the Sharpe ratio grows faster than the square root of the horizon. This is the fundamental question for whether stocks are (unconditionally) “safer for the long run.” Table A1 includes the long-horizon Sharpe ratios, and you can see that they do increase. 375

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

logs, 1926-1996. √ σ (rk ) / k βk √ Sharpe/ k

1 19.8 0.08 0.31

2 20.6 -0.15 0.30

3 19.7 -0.22 0.30

5 18.2 -0.04 0.31

7 16.5 0.24 0.36

10 16.3 0.08 0.39

Table A1. Mean reversion using logs, 1926-1996. r denotes the difference between the log value weighted NYSE return and the log treasury bill return. σ(rk ) = σ(rt→t+k ) is the variance of long-horizon returns. β k is the regression coefficient in rt→t+k = α + β k rt−k→t + εt+k . The Sharpe ratio is E(rt→t+k )/σ(rt→t+k ) You would not be to blame if you thought that the evidence of Table A1 was rather weak, especially compared with the dramatic dividend/price regressions. It is, and it is for this reason that most current evidence for predictability focuses on other variables such as the d/p ratio. In addition, Table A2 shows that the change from log returns to levels of returns, while having a small effect on long-horizon regressions, destroys any evidence for higher Sharpe ratios at long horizons. Table A3 shows the same results in the postwar period. Some of the negative long-horizon regression coefficients are negative and significant, but there are just as large positive coefficients, and no clear pattern. The variance ratios are flat or even rising with horizons, and the Sharpe ratios are flat or even declining with horizon. 1926-1996 √ levels σ (rk ) / k βk √ Sharpe/ k

1 20.6 0.02 0.41

2 22.3 -0.21 0.41

3 22.5 -0.22 0.41

5 24.9 -0.03 0.40

7 28.9 0.22 0.40

10 39.5 -0.63 0.38

Table A2. r denotes the difference between the gross (not log) long-horizon valueweighted NYSE return and the gross treasury bill return. 1947-1996 √ logs σ (rk ) / k βk √ Sharpe/ k 1947-1996 √ levels σ (rk ) / k βk √ Sharpe/ k

1 15.6 -0.10 0.44 1 17.1 -0.13 0.50

2 14.9 -0.29* 0.46 2 17.9 -0.33* 0.51

3 13.0 0.30* 0.51 3 16.8 0.30 0.55

5 13.9 0.30 0.46 5 21.9 0.25 0.48

7 15.0 0.17 0.41 7 29.3 0.13 0.41

10 15.6 -0.18 0.36 10 39.8 -0.25 0.37

Table A3. Mean-reversion in postwar data. In sum, the direct evidence for mean-reversion in index returns seems quite weak. I consider next whether indirect evidence, values of these statistics implied by other estimation 376

S ECTION 20.1

T IME - SERIES PREDICTABILITY

techniques, still indicate mean-reversion. (The mean-reversion of individual stock returns as examined by Fama and French (1988a) is somewhat stronger, and results in the stronger cross-sectional “reversal” effect described in section 2.5 below.) Keep in mind also that the unconditional Sharpe ratio does not in the end, drive investment decisions. Investment decisions are driven by the conditional moments of asset returns at any moment in time, using every information variable that there is. 20.1.5

Mean-reversion and forecastability20.335

I reconcile large forecastability from d/p ratios with a small mean reversion. I calculate the univariate return process implied by the simple VAR, and find that it displays little mean reversion. I show that if dividend shocks are uncorrelated with expected return shocks, there must be some mean reversion. If one rules out the small positive correlation in our samples, one gets a slightly higher estimate of univariate mean-reversion. I tie the strong negative correlation between return and d/p shocks to an essentially zero correlation between expected return and dividend growth shocks. How is it possible that variables such as the dividend price ratio forecast returns strongly, but there seems to be little evidence for mean reversion in stock returns? To answer this question, we have to connect the d/p regressions and the mean-reversion statistics. Forecastability from variables such as the dividend-price ratios is related to, but does not necessarily imply mean-reversion. (Campbell 1991 emphasizes this point.) Mean-reversion is about the univariate properties of the return series, forecasts of rt+j based on {rt , rt−1 , rt−2 ...}. Predictability is about the multivariate properties, forecasts of rt+j based on {xt , xt−1 , xt−2 , ...} as well as {rt , rt−1 , rt−2 ...}. Variables xt can forecast rt+1 , while {rt−j } fail to forecast rt+1 . As a simple example, suppose that returns are i.i.d., but you get to see tomorrow’s newspaper. You forecast returns with a variable xt = rt+1 , rt+1 xt+1

= xt = δ t+1 .

In this example, xt forecasts returns very well, but lagged returns do not forecast returns at all. To examine this issue, continue with the VAR representation built up from a slowly moving expected return and unforecastable dividends, (20.311)-(20.317). We want to find the univariate return process implied by this VAR: what would happen if you took infinite data from the system and ran a regression of returns on lagged returns? The answer, derived below, 377

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

is of the form rt =

1 − γL νt. 1 − bL

(329)

This is just the kind of process that can display slow mean-reversion or momentum. The moving average coefficients are rt = ν t − (γ − b)ν t−1 − b(γ − b)ν t−2 − b2 (γ − b)ν t−3 − b3 (γ − b)ν t−4 − ...

(330)

Thus, if γ > b, a positive return shock sets off a long string of small negative returns, which cumulatively bring the value back towards where it started. If γ < b, a positive shock sets off a string of small positive returns, which add “momentum” to the original increase in value. The long-run statistics are 2 µ ¶2 ∞ X X 1−γ 1+2 ρj = θj /σ2 (ν t ) = . 1−b j=0

Thus, if γ > b, returns will have a variance ratio below one, and if γ < b a variance ratio above one.

Now, what value of γ does our VAR predict? Is there a sensible structure of the VAR that generates substantial predictability but little mean-reversion? The general formula, derived below, is that γ solves ¡ ¢ 1 + b2 σ2 (εd ) + (1 + ρ2 )σ2 (εdp ) − 2(ρ + b)σ(εd , εdp ) 1 + γ2 = = 2q, γ bσ2 (εd ) + ρσ 2 (εdp ) − (ρ + b)σ(εd , εdp )

(331)

and hence, γ=q−

Case 1: No predictability.

p q 2 − 1.

If returns are not predictable in this system; if σ(δ) = 0 so σ(εdp ) = 0; then (20.331) specializes to 1 + b2 1 + γ2 = . γ b γ = b, so returns are not autocorrelated. Sensibly enough.

Case 2: Constant dividend growth.

Next, suppose that the case that dividend growth is constant; σ(εd ) = 0 and variation in expected returns is the only reason that returns vary at all. In this case, (20.331) specializes 378

S ECTION 20.1

T IME - SERIES PREDICTABILITY

quickly to 1 + ρ2 1 + γ2 = , γ ρ

and thus γ = ρ. This is a substantial amount of mean reversion. (γ − b) in (20.330) is then 0.96 − 0.90 = 0.06, so that each year j after a shock returns come back by 6 × bj percent of the original shock. The cumulative impact is that value ends up at (1 − γ)/(1 − b) = (1 − 0.96)/(1 − 0.9) = 0.4 or only 40% of the original shock. Case 3: Dividend growth uncorrelated with expected return shocks.

Pure variation in expected returns is of course not realistic. Dividends do vary. If we add dividend growth uncorrelated with expected return shocks – with σ(εdp , εd ) = 0– (20.331) specializes to 1 + b2 1 + ρ2 1 + γ2 bσ2 (εd ) ρσ2 (εdp ) = + = 2q 2 2 2 γ b bσ (εd ) + ρσ (εdp ) ρ bσ (εd ) + ρσ2 (εdp )

(332)

In this case, b < γ < ρ. There will be some mean reversion in returns – this model cannot generate γ ≤ b. However, the mean reversion in returns will be lower than with no dividend growth, because dividend growth obscures the information in ex-post returns about timevarying expected returns. (See (20.333).) How much lower depends on the parameters. Using the parameters (20.321), I find that (20.332) implies p γ = q − q 2 − 1 = 0.928.

Our baseline VAR with no correlation between dividend growth and expected return shocks thus generates a univariate return process that is slightly on the mean-reversion edge of uncorrelated. The long-run response to a shock is 1 − 0.928 1−γ = = 0.72 1−b 1 − 0.9

This is a lot less mean-reversion than 0.4, but still somewhat more mean reversion than we see in Tables A1-A3. This case is an important baseline worth stressing. If expected returns are positively correlated, realized returns are negatively autocorrelated. If (unchanged) expected dividends are discounted at a higher rate, today’s price falls. You can see this most easily by just looking at the return or its linearization, (20.319) rt+1 = ∆dt+1 − ρ(dt+1 − pt+1 ) + (dt − pt ).

(333)

The d − p ratio is proportional to expected returns. A positive shock to expected returns, uncorrelated with dividend growth, lowers actual returns. A little more deeply, look at the 379

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

return innovation identity (20.308),

rt − Et−1 rt = (Et − Et−1 )

∞ X j=0

ρj ∆dt+j −

∞ X j=1

ρj rt+j .

(334)

P j If expected returns (Et − Et−1 ) ∞ j=1 ρ rt+j increase, with no concurrent news about current or future dividends, then rt − Et−1 rt decreases.

This is the point to remark on a curious feature of the return - dividend/price VAR; the negative correlation between ex-post return shocks and dividend/price ratio shocks. All the estimates were around -0.7. At first glance such a strong correlation between VAR residuals seems strange. At second glance, it is expected. From (20.333) you can see that a positive innovation to the dividend price ratio will correspond to a negative return innovation, unless a striking dividend correlation gets in the way. More deeply, you can see the point in (20.334). Quantitatively, from (20.315), the return shock is related to the dividend growth shock and the expected return shock by εr = εd −

ρ δ = εd − ρεdp 1 − ρb

Thus, a zero correlation between the underlying dividend growth and expected return shocks, ρ(εd , δ) = 0 implies a negative covariance between return shocks and expected return shocks. σ(εr , δ) = −

ρ σ2 (δ) 1 − ρb

The correlation is a perfect −1 if there are no dividend growth shocks. At the parameters (??) σ(εdp ) = 12.5, σ(εr ) = 15, we obtain ρ(εr , δ) = ρ(εr , εdp ) = −

σ(εdp ) 12.5 ρ σ(δ) = −ρ = −0.96 × = −0.8. 1 − ρb σ(ε) σ(ε) 15

The slight 0.1 positive correlation between dividend growth and expected return shocks results (or, actually, results from) a slightly lower −0.7 specification for the correlation of return and d/p shocks. The strong negative correlation between return shocks and expected return shocks, expected from a low correlation between dividend growth shocks and expected return shocks, is crucial to the finding that returns are not particularly correlated despite predictability. Consider what would happen if the correlation ρ(εr , εdp ) = ρ(εr , δ) were zero. The expected return xt is slow moving. If it is high now, it has been high for a while, and there has likely been a series of good past returns. But it also will remain high for a while, leading to a period of high future returns. This is “momentum,” positive return autocorrelation, the opposite of mean-reversion. Case 4: Dividend growth shocks positively correlated with expected return shocks

As we have seen, the VAR with no correlation between expected return and dividend 380

S ECTION 20.1

T IME - SERIES PREDICTABILITY

growth shocks cannot deliver uncorrelated returns or positive “momentum” correlation patterns. At best, volatile dividend growth can obscure an underlying negative correlation pattern. However, looking at (20.333) or (20.334), you can see that adding dividend growth shocks positively correlated with expected return shocks could give us uncorrelated or positively correlated returns. The estimate in Table RR implied a slight positive correlation of dividend growth and expected return shocks, ρεd δ = 0.14 in (20.323). If we use that estimate in (20.331), we recover an estimate γ = 0.923;

1−γ = 0.77 1−ρ

This γ is quite close to b = 0.9, and the small mean reversion is more closely consistent with Tables A1-A3. Recall that point estimates as in Table 1 actually showed that a high d/p ratio forecast higher dividends – the wrong sign. This point estimate means that shocks to the d/p ratio and expected returns are positively correlated with shocks to expected dividend growth. If you generalize the VAR to allow such shocks, along with a richer specification allowing additional lags and variables, you find that VARs give point estimates with slight but very small mean reversion. (See Cochrane 1994 for a plot. The estimated univariate process has slight mean-reversion, with an impulse-response ending up at about 0.8 of its starting value, and no different from the direct estimate. Can we generate unforecastable returns in this system? To do so, we have to increase the between expected return shocks and dividend growth. Equating (20.331) to ¡ correlation ¢ 1 + b2 /b and solving for ρ(εd , εdp ), we obtain ρ(εd , εdp ) =

(1 − ρb) (ρ − b) σ(εdp ) = 0.51. (1 − b)2 (ρ + b) σ(εd )

This is possible, but not likely. Any positive correlation between dividend growth and expected return shocks strikes me as suspect. If anything, I would expect that since expected returns rise in “bad times” when risk or risk aversion increases, we should see a positive shock to expected returns associated with a negative shock to current or future dividend growth. Similarly, if we are going to allow dividend price ratios to forecast dividend growth, a high dividend price ratio should forecast lower dividends. Tying together all these thoughts, I think it’s reasonable to impose zero dividend forecastability and zero correlation between dividend growth and expected return shocks. This specification means that returns are really less forecastable than they seem in some samples. As we have seen, b = 0.9 and no dividend forecastability means that the coefficient of return on D/P is really about 3.4 rather than 5 or 6. This specification means that expected returns really account for 100% rather than 130% of the price-dividend variance. However, it also means that univariate mean reversion is slightly stronger than it seems in our sample. 381

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

This section started with the possibility that the implied mean reversion from a multivariate system could be a lot larger than that revealed by direct estimates. Instead, we end up by reconciling strong predictability and slight mean-reversion. How to find the univariate return representation

To find the implied univariate representation, we have to find a representation rt+1 = a(L)ν t

(335)

in which the a(L) is invertible. The Wold decomposition theorem tells us that there is a unique moving invertible moving average representation in which the ν t are the one-step ahead forecast error shocks, i.e. the errors in a regression model a(L)rt+1 = ν t+1 . Thus, if you find any invertible moving average representation, you know you have the right one. We can’t do it by simply manipulating the systems starting with (20.311), because they are expressed in terms of multivariate shocks, errors in regressions that include x. There are three fundamental representations of a time series: its Wold moving average representation, its autocorrelation function, and its spectral density. To find the univariate representation (20.335), you either calculate the autocorrelations E(rt rt−j ) from (20.311) and then try to recognize what process has that autocorrelation pattern, or you calculate the spectral density and try to recognize what process has that spectral density. In our simple setup, we can write the return-d/p VAR (20.314)-(20.315) as

rt+1 = (1 − ρb) (dt − pt ) + (εdt+1 − ρεdpt+1 ) (dt+1 − pt ) = b(dt − pt ) + εdpt+1

Then, write returns as rt+1 (1 − bL) rt+1 (1 − bL) rt+1

(1 − ρb) εdpt + (εdt+1 − ρεdpt+1 ) 1 − bL = (1 − ρb) εdpt + (εdt+1 − ρεdpt+1 ) − b (εdt − ρεdpt ) = (εdt+1 − ρεdpt+1 ) + (εdpt − bεdt ) (20.336) =

Here, you can see that rt must follow an ARMA(1,1) with one root equal to b and the other root to be determined. Write yt = (1 − bL)rt , and thus yt = (1 − γL)ν t . Then the autocovariances of y from (20.336) are ¢ ¡ 2 E(yt+1 ) = 1 + b2 σ2 (εd ) + (1 + ρ2 )σ2 (εdp ) − 2(ρ + b)σ(εd , εdp ) E(yt+1 yt ) = −bσ 2 (εd ) − ρσ 2 (εdp ) − (ρ + b)σ(εd , εdp )

while yt = (1 − γL)ν t implies 2 E(yt+1 ) =

¢ ¡ 1 + γ 2 σ2ν

E(yt+1 yt ) = −γσ2ν .

382

S ECTION 20.1

T IME - SERIES PREDICTABILITY

Hence, we can find γ from the condition ¡ ¢ 1 + b2 σ2 (εd ) + (1 + ρ2 )σ2 (εdp ) − 2(ρ + b)σ(εd , εdp ) 1 + γ2 = = 2q. γ bσ2 (εd ) + ρσ2 (εdp ) − (ρ + b)σ(εd , εdp ) The solution (the root less than one) is γ=q−

p q 2 − 1.

For more general processes, such as computations from an estimated VAR, it is better to approach the problem via the spectral density. This approach allows you to construct the uni¤0 £ variate representation directly without relying on cleverness. If you write yt = rt xt , the VAR is yt = A(L)η t . Then spectral density of returns Sr (z) is given by the top left element of Sy (z) = A(z)E(ηη0 )A(z −1 )0 with z = e−iω . Like the autocorrelation, the spectral density is the same object whether it comes from the univariate or multivariate representation. You can find the autocorrelations by (numerically) inverse-Fourier transforming the spectral density. The autocorrelations and spectral densities are directly revealing: a string of small negative autocorrelations or a dip in the spectral density near frequency zero correspond to mean-reversion; positive autocorrelations or a spectral density higher at frequency zero than elsewhere corresponds to momentum. To find the univariate, invertible moving average representation from the spectral density, you have to factor the spectral density Srr (z) = a(z)a(z) where a(z) is a polynomial with roots outside the unit circle, a(z) = (1 − γ 1 z)(1 − γ 2 z)...γ i < 1. Then, since a(L) is invertible, rt = a(L)εt σ2ε = 1 is the univariate representation of the return process. 20.1.6

Multivariate mean-reversion

I calculate the responses to multivariate rather than univariate shocks. In a multivariate system you can isolate expected return shocks and dividend growth shocks. The price response to expected return shocks is entirely stationary. We are left with a troubling set of facts: high price/dividend ratios strongly forecast low returns, yet high past returns do not seem to forecast low subsequent returns. Surely, there must be some sense in which “high prices” forecast lower subsequent returns? The resolution must involve dividends (or earnings, book value, or a similar divisor for prices). A price rise with no change in dividends results in lower subsequent returns. A price rise that comes with a dividend rise does not result in lower subsequent returns. A high return combines dividend news and price-dividend news, and so obscures the lower expected return message. In a more time-series language, instead of looking at the response to a univariate return shock – a return that was unanticipated based on lagged returns – let us look at the 383

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

responses to multivariate shocks – a return that was unanticipated based on lagged returns and dividends. This is easy to do in our simple VAR. We can simulate (20.314) -(20.317) forward and trace the responses to a dividend growth shock and an expected return (d/p ratio) shock. Figures 40 and 41 present the results of this calculation. (Cochrane 1994 presents a corresponding calculation using an unrestricted VAR, and the results are very similar.)

Figure 40. Responses to a one standard deviation (1.7%) negative expected return shock in the simple VAR. Start with Figure 40. The negative expected return shock raises prices and the p-d ratio immediately. We can identify such a shock in the data as a return shock with no contemporaneous movement in dividends. The p-d ratio then reverts to its mean. Dividends are not forecastable, so they show no immediate or eventual response to the expected return shock. It could be the case that prices move in advance of future dividends; if this were the case we would see dividends rising to meet higher prices after a return shock. Instead, prices show a long and complete reversion back to the level of dividends. This shock looks a lot like a negative yield shock to bonds: such a shock raises prices now so that bonds end up at the same maturity value despite a smaller expected return. The cumulative return “mean-reverts” even more than prices. For given prices, dividends are now smaller (smaller d-p) so returns deviate from their mean by more than price growth. 384

S ECTION 20.1

T IME - SERIES PREDICTABILITY

Figure 41. Responses to a one standard deviation (14%) dividend growth shock in the simple VAR. The cumulative return ends up below its previously expected value. Compare this value response to the univariate value response, which we calculated above ends up at about 0.8 of its time-1 response. The dividend shock raises prices and cumulative returns immediately and proportionally to dividends, so the price-dividend ratio does not change. Expected returns or the discount rate, reflected in any slope of the value line, do not change. If the world were i.i.d., this is the only kind of shock we would see, and dividend-price ratios would always be constant. Figure (40) and (41) plot the responses to “typical,” one standard deviation shocks. Thus you can see that actual returns are typically about half dividend shocks and half expected return shocks. That is why returns alone are a poor indicator of expected returns. In sum, at last we can see some rather dramatic “mean-reversion.” Good past returns by themselves are not a reliable signal of lower subsequent returns, because they contain substantial dividend growth noise. Good returns that do not include good dividends isolate an expected return shock. This does signal low subsequent returns. It sets off a completely transitory variation in prices.

385

C HAPTER 20 20.1.7

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

Cointegration and short vs. long-run volatility

If d − p, ∆p and ∆d then the long-run variance of ∆d and ∆p must be the same, longrun movements in d and p must be perfectly correlated, and d and p must end up in the same place after any shock. Thus, the patterns of predictability, volatility, univariate and multivariate mean-reversion really all just stem from these facts, the persistence of d − p and the near-unforecastability of ∆d. You might think that the facts about predictability depend on the exact structure of the VAR, including parameter estimates. In fact, most of what we have learned about predictability and mean reversion comes down to a few facts: the dividend-price ratio, returns, and dividend growth are all stationary; dividend growth is not (or at best weakly) forecastable, and dividend growth varies less than returns. These facts imply that the dividend and price responses to each shock are eventually equal in Figures (40) and (41). If d − p, ∆p and ∆d are stationary, then d and p must end up in the same place following a shock. The responses of a stationary variable (d − p) must die out. If dividends are not forecastable, then it must be the case that prices do all the adjustment following a price shock that does not affect dividends. Stationary d − p, ∆p and ∆d also implies that the variance of long-horizon ∆p must equal the variance of long-horizon ∆d. 1 1 var (pt+k − pt ) = lim var (dt+k − dt ) , k→∞ k k→∞ k lim

(337)

and the correlation of long-run price and dividend growth must approach one. These facts follow from the fact that the variance ratio of a stationary variable must approach zero, and d − p is stationary. Intuitively, long run price growth cannot be more volatile than long run dividend growth, or the long-run p − d ratio would not be stationary. Now, if dividend growth is not forecastable, its long run volatility is the same as its short run volatility – its variance ratio is one. Short run price growth is more volatile than short run dividend growth, so we conclude that prices must be mean-reverting; their variance ratio must be below one.

Quantitatively, this observation supports the magnitude of univariate mean reversion that we have found so far. Dividend growth has a short run, and thus long-run, standard deviation of about 10% per year, while returns and prices Thus, prices must p have a long-run variance ratio of about 2/3, or a long-run response to univariate shocks of 2/3 = 0.82.

The change in prices is not the same thing as the return, especially at long horizons, since returns include the intervening dividends. One can address this question with a slightly different accounting: define d as the dividend paid to a dollar investment. The resulting dividend series is still not predictable and has roughly the same volatility, so in this case we 386

S ECTION 20.1

T IME - SERIES PREDICTABILITY

get approximately the same result. The work of Lettau and Ludvigson (2000) suggests that we may get much more dramatic implications by including consumption data. The ratio of stock market values to consumption should also be stationary; if wealth were to explode people would surely consume more and vice versa. The ratio of dividends to aggregate consumption should also be stationary. Consumption growth seems independent at all horizons, and consumption growth is very stable, with roughly 1% annual standard deviation. For example, Lettau and Ludvigson 2000 find that none of the variables that forecast returns in Table LL – including d − p and a consumption to wealth ratio – forecast consumption growth at any horizon. These facts suggest that aggregate dividends are forecastable by the consumption/dividend ratio, and strongly so – the long-run volatility of aggregate dividend growth must be the 1% volatility of consumption growth, not the 10% short run volatility of dividend growth. These facts also mean that almost all of the 15% or more variation in annual stock market wealth must be transitory – the long run volatility of stock market value must be no more than the 1% consumption growth volatility! Again, total market value is not the same thing as price, price is not the same thing as cumulated return, and aggregate dividends are not the same thing as the dividend concept we have used so far (dividends paid to a dollar investment with dividends consumed), or dividends paid to a dollar investment with dividends reinvested. Lettau and Ludvigson show that the consumption/wealth ratio does forecast returns, but noone has yet worked out the mean-reversion implications of this fact. My statements about the implications of stationary d − p, ∆d, ∆p, r are developed in detail in Cochrane 1994. They are special cases of the representation theorems for cointegrated variables developed by Engel and Granger (1987). A regression of a difference like ∆p on a ratio like p − d is called the error-correction representation of a cointegrated system. Error correction regressions have subtly and dramatically changed almost all empirical work in finance and macroeconomics. The vast majority of the successful return forecasting regressions in this section, both time-series and cross-section, are error-correction regressions of one sort or another. Corporate finance is being redone with regressions of growth rates on ratios, as is macroeconomic forecasting. For example, the consumption/GDP ratio is a powerful forecaster of GDP growth. 20.1.8

Bonds

The expectations model of the term structure works well on average and for horizons of 4 years or greater. At the one year horizon, however, a forward rate 1 percentage point higher than the spot rate seems entirely to indicate a one percentage point higher expected excess return rather than a one percentage point rise in future interest rates.

387

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

The venerable expectations model of the term structure specifies that long term bond yields are equal to the average of expected future short term bond yields. As with the CAPM and random walk, the expectations model was the workhorse of empirical finance for a generation. And as with those other views, a new round of research has significantly modified the traditional view.

Maturity N 1 2 3 4 5

Avg. Return (N) E(hprt+1 ) 5.83 6.15 6.40 6.40 6.36

Std. error 0.42 0.54 0.69 0.85 0.98

Std. dev. (N) σ(hprt+1 ) 2.83 3.65 4.66 5.71 6.58

Table 4. Average continuously compounded (log) one-year holding period returns on zero-coupon bonds of varying maturity. Annual data from CRSP 1953-1997. Table 4 calculates the average return on bonds of different maturities. The expectations hypothesis seems to do pretty well. Average holding period returns do not seem very different across bond maturities, despite the increasing standard deviation of bond returns as maturity rises. The small increase in returns for long term bonds, equivalent to a slight average upward slope in the yield curve, is usually excused as a small “liquidity premium.” In fact, the curious pattern in Table 4 is that bonds do not share the high Sharpe ratios of stocks. Whatever factors account for the volatility of bond returns, they seem to have very small risk prices. Table 4 is again a tip of an iceberg of an illustrious career for the expectations hypothesis. Especially in times of great inflation and exchange rate instability, the expectations hypothesis does a very good first-order job. However, one can ask a more subtle question. Perhaps there are times when long term bonds can be forecast to do better, and other times when short term bonds are expected to do better. If the times even out, the unconditional averages in Table 4 will show no pattern. Equivalently, we might want to check whether a forward rate that is unusually high forecasts an unusual increase in spot rates. 388

S ECTION 20.1

N 1 2 3 4

T IME - SERIES PREDICTABILITY

Change in yields (1) (1) yt+N − yt = (N→N+1) (1) = a + b(ft − yt ) + εt+N 2 a σ(a) b σ(b) R 0.1 0.3 -0.10 0.36 -0.02 -0.01 0.4 0.37 0.33 0.005 -0.04 0.5 0.41 0.33 0.013 -0.3 0.5 0.77 0.31 0.11

Holding period returns (N+1) (1) hprt+1 − yt = (N→N+1) (1) = a + b(ft − yt ) + εt+1 2 a σ(a) b σ(b) R -0.1 0.3 1.10 0.36 0.16 -0.5 0.5 1.46 0.44 0.19 -0.4 0.8 1.30 0.54 0.10 -0.5 1.0 1.31 0.63 0.07

Table 5. Forecasts based on forward-spot spread. OLS regressions 1953-1997 annual data. Yields and returns in annual percentages. The left hand panel runs the change in the one year yield on the forward-spot spread. The right hand panel runs the one period excess return on the forward-spot spread. Table 5 gets at these issues, updating Fama and Bliss’ (1986) classic regression tests. (Campbell and Shiller 1991 and Campbell 1995 make the same point with regressions of yield changes on yield spreads.) The left hand panel presents a regression of the change in yields on the forward-spot spread. The expectations hypothesis predicts a coefficient of 1.0, since the forward rate should equal the expected future spot rate. At a one-year horizon we see instead coefficients near zero and a negative adjusted R2 . Forward rates one year out seem to have no predictive power whatsoever for changes in the spot rate one year from now. On the other hand, by 4 years out, we see coefficients within one standard error of 1.0. Thus, the expectations hypothesis seems to do poorly at short (1 year) horizons, but much better at longer horizons and on average (Table 4). If the yield expression of the expectations hypothesis does not work at one year horizons, then the expected return expression of the expectations hypothesis must not hold either – one must be able to forecast one year bond returns. To check this fact, the right hand panel of Table 5 runs regressions of the one year excess return on long-term bonds on the forward-spot spread. Here, the expectations hypothesis predicts a coefficient of zero: no signal (including the forward-spot spread) should be able to tell you that this is a particularly good time for long bonds vs. short bonds. As you can see, the coefficients in the right hand panel of Table 5 are all about 1.0. A high forward rate does not indicate that interest rates will be higher one year from now; it seems entirely to indicate that you will earn that much more holding long term bonds (The right hand panel is really not independent evidence, since the coefficients in the right and left hand panels of Table 5 are mechanically linked. For example 1.14 + (-0.14) = 1.0, and this holds as an accounting identity. Fama and Bliss call them “complementary regressions.”) Figures 42 and 43 provide a pictorial version of the results in Table 5. Suppose that the yield curve is upward sloping as in the left panel. What does this mean? A naive investor might think this pattern indicates that long-term bonds give a higher return than short term bonds. The expectations hypothesis denies this conclusion. If the expectations hypothesis were true, the forward rates plotted against maturity in the left hand panel would translate 389

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

one-for-one to the forecast of future spot rates in the right hand panel, as plotted in the line marked “Expectations model.” Rises in future short rates should lower bond prices, cutting off the one-period advantage of long-term bonds. The rising short rates would directly raise the multi-year advantage of short term bonds. We can calculate the actual forecast of future spot rates from the estimates in the left hand panel of Table 5, and these are given by the line market “Estimates” in Figure 43. The essence of the phenomenon is sluggish adjustment of the short rates. The short rates do eventually rise to meet the forward rate forecasts, but not as quickly as the forward rates predict that they should.

Figure 42. If the current yield curve is as plotted here.... As dividend growth should be forecastable so that returns are not forecastable, short-term yields should be forecastable so that returns are not forecastable. In fact, yield changes are almost unforecastable at a one year horizon, so, mechanically, bond returns are. We see this directly in the first row of the left hand panel of Table 5 for the one-period yield. It is an implication of the right hand panel as well. If

(N+1)

hprt+1

(1)

− yt

(N→N+1)

= 0 + 1(ft

390

(1)

− yt ) + εt+1

(338)

S ECTION 20.1

T IME - SERIES PREDICTABILITY

Figure 43. ...this is the forecast of future one year interest rates. The dashed line gives the forecast from the expectations hypothesis. The solid line is constructed from the estimates in Table 4. then, writing out the definition of holding period return and forward rate, (N)

(N+1)

pt+1 − pt

(1)

+ pt

(N) pt+1 (N) yt+1

(N)

= 0 + 1(pt = 0 = 0

(N+1)

− pt

(1)

+ pt ) + εt+1

(20.339)

(N) + 1(pt ) + εt+1 (N) + 1(yt ) − εt+1 /N

A coefficient of 1.0 in (20.338) is equivalent to yields or bond prices that follow random walks; yield changes that are completely unpredictable. Of course yields are stationary and not totally unpredictable. However, they move slowly. Thus, yield changes are very unpredictable at short horizons but much more predictable at long horizons. That is why the coefficients in the right hand panel of Table 5 build with horizon. If we did holding period return regressions at longer horizons, they would gradually approach the expectations hypothesis result. The roughly 1.0 coefficients in the right hand panel of Table 5 mean that a one percentage point increase in forward rate translates into a one percentage point increase in expected return. It seems that old fallacy of confusing bond yields with their expected returns also contains a grain of truth, at least for the first year. However, the one-for-one variation of expected returns with forward rates does not imply a one-for-one variation of expected returns 391

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

with yield spreads. Forward rates are related to the slope of the yield curve, (N→N+1)

ft

(1)

− yt

(N)

= pt

(N+1)

− pt (N)

(1)

− yt

(N+1)

(1)

= −N yt + (N + 1)yt − yt ´ ³ ´ ³ (N+1) (N) (N+1) (1) + yt = N yt − yt − yt

Thus, the forward-spot spread varies a more than the yield spread, so regression coefficients of holding period yields on yield spreads give coefficients greater than one. Expected returns move more than one-for-one with yield spreads. Campbell (1995) reports coefficients of excess returns on yield spreads that rise from one at a 2 month horizon to 5 at a 5 year horizon. The facts are analogous to the dividend/price regression. There, dividends were essentially unforecastable. This implied that a one percentage point change in dividend yield implied a 5 percentage point change in expected excess returns. Of course, there is risk: the R2 are all about 0.1-0.2, about the same values as the R2 from the dividend/price regression at a one year horizon, so this strategy will often go wrong. Still, 0.1-0.2 is not zero, so the strategy does pay off more often than not, in violation of the expectations hypothesis. Furthermore, the forward-spot spread is a slow moving variable, typically reversing sign once per business cycle. Thus, the R2 build with horizon as with the D/P regression, peaking in the 30% range (Fama and French 1989). The fact that the regressions in Table 5 run the change in yield on the forward-spot spread and the excess return on the forward-spot spread is very important. The overall level of (N) interest rates moves up and down a great deal but slowly over time. Thus, if you run yt+j = (N+1)

a + bft + εt+N , you will get a coefficient b almost exactly equal to 1.0 and a stupendous R2 , seemingly a stunning validation of the expectations hypothesis. If you run a regression of tomorrow’s temperature in Chicago on today’s temperature, the regression coefficient will be near 1.0 with a huge R2 as well, since the temperature varies a lot over the year. But today’s temperature is not a useful temperature forecast. To measure a temperature forecast we want to know if the forecast can predict the change in temperature. Is (forecast - today’s temperature) a good measure of (tomorrow’s temperature - today’s temperature)? Table 5 runs this regression.

The decomposition in (20.339) warns us of one of several econometric traps in this kind of regression. Notice that two of the three right hand variables are the same. Thus any mea(N+1) (1) and pt will induce a spurious common movement in left and right surement error in pt hand variables. In addition, since the variables are a triple difference, the difference may eliminate a common signal and isolate measurement error or noise. There are pure measurement errors in the bond data, and we seldom observe pure discount bonds of the exactly desired maturity. In addition, various liquidity and microstructure effects can influence the yields of particular bonds in ways that are not exploitable for typical investors. As an example of what this sort of “measurement error” can do, suppose all bond yields 392

S ECTION 20.1

T IME - SERIES PREDICTABILITY

are 5%, but there is one “error” in the two period bond price at time 1 – rather than being -10 it is -15. The table below tracks the effects of this error. It implies a blip of the one year forward rate in year one, and then a blip in the return from holding this bond from year one to year two. The price and forward rate “error” automatically turns in to a subsequent return when the “error” is corrected. If the price is real, of course, this is just the kind of event we want the regression to tell us about – the forward rate did not correspond to a change in future spot rate, so there was a large return; it was a price that was “out of line” and if you could trade on it, you should. But the regression will also pounce on measurement error in prices and indicate spurious returns. t (1) pt (2) pt (3) pt (i) yt , i 6= 2 (2) yt (1→2) ft (1→2) (1) ft − yt (2→1) (1) hprt − yt

0 -5 -10 -15 5 5 5 0 0

1 -5 -15 -15 5 7.5 10 5 0

2 -5 -10 -15 5 5 5 0 5

3 -5 -10 -15 5 5 5 0 0

Numerical example of the effect of measurement error in yields on yield regressions. 20.1.9

Foreign exchange

The expectations model works well on average. However, a foreign interest rate one percentage point higher than its usual differential with the US rate (equivalently, a one percentage point higher forward-spot spread) seems to indicate even more than one percentage point expected excess return; a further appreciation of the foreign currency. Suppose interest rates are higher in Germany than in the U.S. Does this mean that one can earn more money by investing in German bonds? There are several reasons that the answer might be no. First, of course is default risk. While not a big problem for German government bonds, Russia and other governments have defaulted on bonds in the past and may do so again. Second, and more important, is the risk of devaluation. If German interest rates are 10%, US interest rates are 5%, but the Euro falls 5% relative to the dollar during the year, you make no more money holding the German bonds despite their attractive interest rate. Since lots of investors are making this calculation, it is natural to conclude that an interest rate differential across countries on bonds of similar credit risk should reveal an expectation of currency devaluation. The logic is exactly the same as the “expectations hypothesis” in 393

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

the term structure. Initially attractive yield or interest rate differentials should be met by an offsetting event so that you make no more money on average in one country or another, or in one currency versus another. As with bonds, the expectations hypothesis is slightly different from pure risk neutrality since the expectation of the log is not the log of the expectation. Again, the size of the phenomena we study usually swamps this distinction. As with the expectations hypothesis in the term structure, the expected depreciation view ruled for many years, and still constitutes an important first-order understanding of interest rate differentials and exchange rates. For example, interest rates in east Asian currencies were very high on the eve of the currency collapses of 1997, and many banks were making tidy sums borrowing at 5% in dollars to lend at 20% in local currencies. This situation should lead one to suspect that traders expect a 15% devaluation, and most likely a small chance of a larger devaluation. That is, in this case, exactly what happened. Many observers and policy analysts who ought to know better often attribute high nominal interest rates in troubled countries to “tight monetary policy” that is “strangling the economy” to “defend the currency.” In fact, one’s first order guess should be that such high nominal rates reflect a large probability of inflation and devaluation – loose monetary and fiscal policy – and that they correspond to much lower real rates. Still, does a 5% interest rate differential correspond to an exactly 5% expected depreciation, or does some of it still represent a high expected return from holding debt in that country’s currency? Furthermore, while expected depreciation is clearly a large part of the story for high interest rates in countries that have constant high inflation or that may suffer spectacular depreciation of a pegged exchange rate, how does the story work for, say, the U.S. vs. Germany, where inflation rates diverge little, yet exchange rates fluctuate a surprisingly large amount? Table 6 presents the facts, as summarized by Hodrick (2000) and Engel (1996). The first row of Table 6 presents the average appreciation of the dollar against the indicated currency over the sample period. The dollar fell against DM, yen and Swiss Franc, but appreciated against the pound. The second row gives the average interest rate differential – the amount by which the foreign interest rate exceeds the US interest rate. According to the expectations hypothesis, these two numbers should be equal – interest rates should be higher in countries whose currencies depreciate against the dollar. The second row shows roughly the right pattern. Countries with steady long-term inflation have steadily higher interest rates, and steady depreciation. The numbers in the first and second rows are not exactly the same, but exchange rates are notoriously volatile so these averages are not well measured. Hodrick shows that the difference between the first and second rows is not statistically different from zero. This fact is exactly analogous to the fact of Table 4 that the expectations hypothesis works well “on average” for US bonds and is the tip of an iceberg of empirical successes for the expectations hypothesis as applied to currencies. As in the case of bonds, however, we can also ask whether times of temporarily higher or lower interest rate differentials correspond to times of above and below average depre394

S ECTION 20.1

T IME - SERIES PREDICTABILITY

ciation as they should. The third and fifth rows of Table 6 address this question, updating Hansen and Hodrick’s (1980) and Fama’s (1984) regression tests. The number here should be +1.0 in each case – an extra percentage point interest differential should correspond to one extra percentage point expected depreciation. As you can see, we have exactly the opposite pattern: a higher than usual interest rate abroad seems to lead, if anything to further appreciation. It seems that the old fallacy of confusing interest rate differentials across countries with expected returns, forgetting about depreciation, also contains a grain of truth. This is the “forward discount puzzle,” and takes its place alongside the forecastability of stock and bond returns. Of course it has produced a similar avalanche of academic work dissecting whether it is really there and if so, why. Hodrick (1987), Engel (1996), and Lewis (1995) provide surveys. The R2 shown in Table 6 are quite low. However, like D/P, the interest differential is a slow-moving forecasting variable, so the return forecast R2 build with horizon. Bekaert and Hodrick (1992) report that the R2 rise to the 30-40% range at six month horizons and then decline again. Still, taking advantage of this predictability, like the bond strategies described above, is quite risky. Mean appreciation Mean interest differential b, 1975-1989 R2 b, 1976-1996

DM -1.8 -3.9 -3.1 .026 -0.7

£ 3.6 2.1 -2.0 .033 -1.8

U -5.0 -3.7 -2.1 .034 -2.4

SF -3.0 -5.9 -2.6 .033 -1.3

Table 6. The first row gives the average appreciation of the dollar against the indicated currency, in percent per year. The second row gives the average interest differential – foreign interest rate less domestic interest rate, measured as the forward premium – the 30 day forward rate less the spot exchange rate. The third through fifth rows give the coefficients and R2 in a regression of exchange rate changes on the interest differential = forward premium, st+1 − st = a + b(ft − st ) + εt+1 = a + b(rtf − rtd ) + εt+1

where s = log spot exchange rate, f = forward rate, rf = foreign interest rate, rd = domestic interest rate. Source: Hodrick (1999) and Engel (1996). The puzzle does not say that one earns more by holding bonds from countries with higher interest rates than others. Average inflation, depreciation, and interest rate differentials line up as they should. If you just buy bonds with high interest rates, you end up with debt from Turkey and Brazil, whose currencies inflate and depreciate steadily. The puzzle does say that one earns more by holding bonds from countries whose interest rates are higher than usual relative to U.S. interest rates. However, the fact that the “usual” rate of depreciation and “usual” interest differential 395

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

varies through time, if they are well-defined concepts at all, may diminish if not eliminate the out-of-sample performance of trading rules based on these regressions. The foreign exchange regressions offer a particularly clear-cut case in which “Peso problems” can skew forecasting regressions. Lewis (1995) credits Milton Friedman for coining the term to explain why Mexican interest rates were persistently higher than U.S. interest rates in the early 1970’s even though the currency had been pegged for more than a decade. A small probability of a huge devaluation each period can correspond to a substantial interest differential. You will see long stretches of data in which the expectations hypothesis seems not to be satisfied, because the collapse does not occur in sample. The Peso subsequently collapsed, giving substantial weight to this view. Since “Peso problems” have become a generic term for the effects of small probabilities of large events on empirical work. Rietz (1988) offered a Peso problem explanation for the equity premium that investors are afraid of another great depression which has not happened in sample. Selling out of the money put options and earthquake insurance in Los Angeles are similar strategies whose average returns in a sample will be severely affected by rare events.

20.2

The Cross-section: CAPM and Multifactor Models

Having studied how average returns change over time, now we study how average returns change across different stocks or portfolios. 20.2.1

The CAPM

For a generation, portfolios with high average returns also had high betas. I illustrate with the size-based portfolios. The first tests of the CAPM such as Lintner (1965) were not a great success. If you plot or regress the average returns versus betas of individual stocks, you find a lot of dispersion, and the slope of the line is much too flat – it does not go through any plausible riskfree rate. Miller and Scholes (1972) diagnosed the problem. Betas are measured with error, and measurement error in right hand variables biases down regression coefficients. Fama and MacBeth (1973) and Black, Jensen and Scholes (1972) addressed the problem by grouping stocks into portfolios. Portfolio betas are better measured because the portfolio has lower residual variance. Also, individual stock betas vary over time as the size, leverage, and risks of the business change. Portfolio betas may be more stable over time, and hence easier to measure accurately. There is a second reason for portfolios. Individual stock returns are √ so volatile that you cannot reject the hypothesis that all average returns are the same. σ/ T is big when σ = 396

S ECTION 20.2

T HE C ROSS - SECTION : CAPM AND MULTIFACTOR MODELS

40 − 80%. By grouping stocks into portfolios based on some characteristic (other than firm name) related to average returns, you reduce the portfolio variance and thus make it possible to see average return deferences. Finally, I think much of the attachment to portfolios comes from a desire to more closely mimic what actual investors would do rather than simply form a test.

Fama and MacBeth and Black Jensen and Scholes formed their portfolios on betas. They found individual stock betas, formed stocks into portfolios based on their betas, and then estimated the portfolio’s beta in the following period. More recently, size, book/market, industry, and many other characteristics have been used to form portfolios. Ever since, the business of testing asset pricing models has been conducted in a simple loop: 1.

2. 3.

Find a characteristic that you think is associated with average returns. Sort stocks into portfolios based on the characteristic, and check that there is a difference in average returns between portfolios. Worry here about measurement, survival bias, fishing bias, and all the other things that can ruin a pretty picture out of sample. Compute betas for the portfolios, and check whether the average return spread is accounted for by the spread in betas. If not, you have an anomaly. Consider multiple betas. This is the traditional procedure, but econometrics textbooks urge you not to group data in this way. They urge you to use the characteristic as an instrument for the poorly measured right hand variable instead. It is an interesting and unexplored idea whether this instrumental variables approach could fruitfully bring us back to the examination of individual securities rather than portfolios. The CAPM proved stunningly successful in empirical work. Time after time, every strategy or characteristic that seemed to give high average returns turned out to also have high betas. Strategies that one might have thought gave high average returns (such as holding very volatile stocks) turned out not to have high average returns when they did not have high betas. To give some sense of that empirical work, Figure 44 presents a typical evaluation of the Capital Asset Pricing Model. (Chapter 15 presented some of the methodological issues surrounding this evaluation; here I focus on the facts.) I examine 10 portfolios of NYSE stocks sorted by size (total market capitalization), along with a portfolio of corporate bonds and long-term government bonds. As the spread along the vertical axis shows, there is a sizeable spread in average returns between large stocks (lower average return) and small stocks (higher average return), and also a large spread between stocks and bonds. The figure plots these average returns against market betas. You can see how the CAPM prediction fits: portfolios with higher average returns have higher betas. In particular, notice that the long term and corporate bonds have mean returns in line with their low betas, despite their standard deviations nearly as high as those of stocks. Comparing this graph with the similar Figure 5 of the consumption-based model back in Chapter 2, the CAPM fits very well. 397

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

Figure 44. The CAPM. Average returns vs. betas on the NYSE value-weighted portfolio for 10 size-sorted stock portfolios, government bonds, and corporate bonds, 1947-1996. The solid line draws the CAPM prediction by fitting the market proxy and treasury bill rates exactly (a time-series test). The dashed line draws the CAPM prediction by fitting an OLS cross-sectional regression to the displayed data points. The small firm portfolios are at the top right. The points far down and to the left are the government bond and treasury bill returns. In fact, Figure 44 captures one of the first significant failures of the CAPM. The smallest firms (the far right portfolio) seem to earn an average return a few percent too high given their betas. This is the celebrated “small-firm effect” (Banz 1981). Would that all failed economic theories worked so well! It is also atypical in that the estimated market line through the stock portfolios is steeper than predicted, while measurement error in betas usually means that the estimated market line is too flat. 20.2.2

Fama-French 3 factors

Book to market sorted portfolios show a large variation in average returns that is unrelated to market betas. The Fama and French 3 factor model successfully explains the average returns of the 25 size and book to market sorted portfolios with a 3 factor model, consisting of the market, a small minus big (SMB) portfolio and a high minus low (HML) portfolio. 398

S ECTION 20.2

T HE C ROSS - SECTION : CAPM AND MULTIFACTOR MODELS

In retrospect, it is surprising that the CAPM worked so well for so long. The assumptions on which it is built are very stylized and simplified. Asset pricing theory recognized at least since Merton (1971a,b) the theoretical possibility, indeed probability, that we should need factors, state variables or sources of priced risk beyond movements in the market portfolio in order to explain why some average returns are higher than others. The Fama - French model is one of the most popular multi-factor models that now dominate empirical research. Fama and French (1993) presents the model; Fama and French (1996) gives an excellent summary, and also shows how the 3 factor model performs in evaluating expected return puzzles beyond the size and value effects that motivated it. “Value” stocks have market values that are small relative to the accountant’s book value. (Book values essentially track past investment expenditures.) This category of stocks has given large average returns. “Growth” stocks are the opposite of value and have had low average returns. Since low prices relative to dividends, earnings or book value forecast times when the market return will be high, it is natural to suppose that these same signals forecast categories of stocks that will do well; the “value effect” is the cross-sectional analogy to price-ratio predictability in the time-series. High average returns are consistent with the CAPM, if these categories of stocks have high sensitivities to the market, high betas. However, small and especially value stocks seem to have abnormally high returns even after accounting for market beta. Conversely “growth” stocks seem to do systematically worse than their CAPM betas suggest. Figure 45 shows this value-size puzzle. It is just like Figure 44, except that the stocks are sorted into portfolios based on size and book-market ratio9 rather than size alone. As you can see, the highest portfolios have three times the average excess return of the lowest portfolios, and this variation has nothing at all to do with market betas. Figures 46 and 47 dig a little deeper to diagnose the problem, by connecting portfolios that have different size within the same book/market category, and different book/market within size category. As you can see, variation in size produces a variation in average returns that is positively related to variation in market betas, as we had in Figure 45. Variation in book/market ratio produces a variation in average return is negatively related to market beta. Because of this value effect, the CAPM is a disaster when confronted with these portfolios. (Since the size effect disappeared in 1980, it is likely that almost the whole story can be told with book/market effects alone.) To explain these patterns in average returns, Fama and French advocate a multifactor model with the market return, the return of small less big stocks (SMB) and the return of high book/market minus low book/market stocks (HML) as three factors. They show that variation in average returns of the 25 size and book/market portfolios can be explained by varying loadings (betas) on the latter two factors. (All their portfolios have betas close to one on the market portfolio. Thus, market beta explains the average return difference between 9

I thank Gene Fama for providing me with these data.

399

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

Figure 45. Average returns vs. market beta for 25 stock portfolios sorted on the basis of size and book/market ratio. stocks and bonds, but not across categories of stocks.) Figures 48 and 49 illustrate Fama and French’s results. The vertical axis is still the average return of the 25 size and book/market portfolios. Now, the horizontal axis is the predicted values from the Fama-French three factor model. The points should all lie on a 45◦ line if the model is correct. The points lie much closer to this prediction than they do in Figures 46 and 47. The worst fit is for the growth stocks (lowest line, left hand panel), for which there is little variation in average return despite large variation in size beta as one moves from small to large firms. 20.2.3

What are the size and value factors?

What are the macroeconomic risks for which the Fama-French factors are proxies or mimicking portfolios? There are hints of some sort of “distress” or “recession” factor at work. A central part of the Fama French model is the fact that these three pricing factors also explain a large part of the ex-post variation in the 25 portfolios – the R2 in time-series regressions are very high. In this sense, one can regard it as an APT rather than a macroeconomic 400

S ECTION 20.2

T HE C ROSS - SECTION : CAPM AND MULTIFACTOR MODELS

Figure 46. Average excess returns vs. market beta. Lines connect portfolios with different size category within book to market categories. factor model. The Fama-French model is not a tautology, despite the fact that factors and test portfolios are based on the same set of characteristics. We would like to understand the real, macroeconomic, aggregate, nondiversifiable risk that is proxied by the returns of the HML and SMB portfolios. Why are investors so concerned about holding stocks that do badly at the times that the HML (value less growth) and SMB (small-cap less large-cap) portfolios do badly, even though the market does not fall? Fama and French (1995) note that the typical “value” firm has a price that has been driven down from a long string of bad news, and is now in or near financial distress. Stocks bought on the verge of bankruptcy have come back more often than not, which generates the high average returns of this strategy. This observation suggests a natural interpretation of the value premium: If a credit crunch, liquidity crunch, flight to quality or similar financial event comes along, stocks in financial distress will do very badly, and this is just the sort of time at which one particularly does not want to hear that one’s stocks have become worthless! (One cannot count the “distress” of the individual firm as a “risk factor.” Such distress is idiosyncratic and can be diversified away. Only aggregate events that average investors care about can result in a risk premium.) Unfortunately, empirical support for this theory is weak, since the HML 401

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

Figure 47. Average excess returns vs. market beta. Lines connect portfolios with different book to market categories within size categories. portfolio does not covary strongly with other measures of aggregate financial distress. Still, it is a possible and not totally tested interpretation, since we have so few events of actual systematic financial stress in recent history. Heaton and Lucas’ (1997) results add to this story for the value effect. They note that the typical stockholder is the proprietor of a small, privately held business. Such an investor’s income is of course particularly sensitive to the kinds of financial events that cause distress among small firms and distressed value firms. Such an investor would therefore demand a substantial premium to hold value stocks, and would hold growth stocks despite a low premium. Lettau and Ludvigson (2000) (also discussed in the next section) document that HML has a time-varying beta on both the market return and on consumption. Thus, though there is (unfortunately) very little unconditional correlation between HML and recession measures, Lettau and Ludvigson document that HML is sensitive to bad news in bad times. Liew and Vassalou (1999) are an example of current attempts to link value and small firm returns to macroeconomic events. They find that in many countries counterparts to HML and SMB contain information above and beyond that in the market return for forecasting GDP growth. For example, they report a regression GDPt→t+1 = a + 0.065 M KTt−1→t + 0.058 HM Lt−1→t + εt+1

402

S ECTION 20.2

T HE C ROSS - SECTION : CAPM AND MULTIFACTOR MODELS

Figure 48. Average excess return vs. prediction of the Fama-French 3 factor model. Lines connect portfolios of different size categories within book to market category. GDPt→t+1 denotes the next year’s GDP growth and MKT, HML denote the previous year’s return on the market index and HML portfolio. Thus, a 10% HML return reflects a 1/2 percentage point rise in the GDP forecast.

On the other hand, one can ignore Fama and French’s motivation and regard the model as an arbitrage pricing theory. If the returns of the 25 size and book/market portfolios could be perfectly replicated by the returns of the 3 factor portfolios – if the R2 in the time-series regressions were 100% – then the multifactor model would have to hold exactly, in order to preclude arbitrage opportunities. In fact the R2 of Fama and French’s time-series regressions are all in the 90%-95% range, so extremely high Sharpe ratios for the residuals (which are portfolios) would have to be invoked for the model not to fit well. Equivalently, given the average returns from HML and SMB and the failure of the CAPM to explain those returns, there would be near-arbitrage opportunities if value and small stocks did not move together in the way described by the Fama-French model. One way to assess whether the three factors proxy for real macroeconomic risks is by checking whether the multifactor model prices additional portfolios, and especially portfolios that do not have high R2 values. Fama and French (1996) extend their analysis in this direction: They find that the SMB and HML portfolios comfortably explain strategies based on alternative price multiples (P/E, B/M), strategies based on 5 year sales growth (this is especially interesting since it is the only strategy that does not form portfolios based on price variables) and the tendency of 5 year returns to reverse. All of these strategies are not ex403

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

Figure 49. Average excess returns vs. predictions of the Fama-French 3 factor model. Lines connect portfolios of different book to market category within the same size category. plained by CAPM betas. However they all also produce portfolios with high R2 values in a time-series regression on the HML and SMB portfolios! This is good and bad news. It might mean that the model is a good APT; that the size and book/market characteristics describe the major sources of priced variation in all stocks. On the other hand it might mean that these extra sorts just haven’t identified other sources of priced variation in stock returns. (Fama and French also find that HML and SMB do not explain “momentum,” despite large R2 values. More on momentum later.) One’s first reaction may be that explaining portfolios sorted on the basis of size and book to market by factors sorted on the same basis is a tautology. This is not the case. For example, suppose that average returns were higher for stocks whose ticker symbols start later in the alphabet. (Maybe investors search for stocks alphabetically, so the later stocks are “overlooked.”) This need not trouble us if Z stocks happened to have higher betas. If not – if letter of the alphabet were a CAPM anomaly like book to market – however, it would not necessarily follow that letter based stock portfolios move together. Adding A-L and M-Z portfolios to the right hand side of a regression of the 26 A,B,C, etc. portfolios on the market portfolio need not (and probably does not) increase the R2 at all. The size and book to market premia are hard to measure, and seem to have declined substantially in recent years. But even if they decline back to CAPM values, Fama and French will still have found a surprisingly large source of common movement in stock returns. More to the point, in testing a model It is exactly the right thing to do to sort stocks 404

S ECTION 20.2

T HE C ROSS - SECTION : CAPM AND MULTIFACTOR MODELS

into portfolios based on characteristics related to expected returns. When Black Jensen and Scholes and Fama and MacBeth first tested the CAPM, they sorted stocks into portfolios based on betas, because betas are a good characteristic for sorting stocks into portfolios that have a spread in average returns. If your portfolios have no spread in average returns – if you just choose 25 random portfolios – then there will be nothing for the asset pricing model to test. In fact, despite the popularity of the Fama French 25, there is really no fundamental reason to sort portfolios based on 2 way or larger sorts of individual characteristics. You should use all the characteristics at hand that (believably!) indicate high or low average returns and simply sort stocks according to a one-dimensional measure of expected returns. The argument over the status of size and book/market factors continues, but the important point is that it does so. Faced with the spectacular failure of the CAPM documented in Figures and 4, one might have thought that any hope for a rational asset pricing theory was over. Now we are back where we were, examining small anomalies and arguing over refinements and interpretations of the theory. That is quite an accomplishment! 20.2.4

Macroeconomic factors

Labor income, industrial production, news variables and conditional asset pricing models have also all had some successes as multifactor models. I have focused on the size and value factors since they provide the most empirically successful multifactor model to date, and have therefore attracted much attention. Several authors have used macroeconomic variables as factors in order to examine directly the story that stock performance during bad macroeconomic times determines average returns. Jagannathan and Wang (1996) and Reyfman (1997) use labor income; Chen Roll and Ross (1986) use industrial production and inflation among other variables. Cochrane (1996) uses investment growth. All these authors find that average returns line up against betas calculated using these macroeconomic indicators. The factors are theoretically easier to motivate, but none explains the value and size portfolios as well as the (theoretically less solid, so far) size and value factors. Lettau and Ludvigson (2000) specify a macroeconomic model that does just as well as the Fama-French factors in explaining the 25 Fama-French portfolios. Their plots of actual average returns vs. model predictions show a relation as strong as those of Figures 48 and 49. Their model is mt+1 = a + b(cawt )∆ct+1

where caw is a measure of the consumption-wealth ratio. This is a “scaled factor model” of 405

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

the sort advocated in Chapter 8. You can think of it as capturing a time-varying risk aversion. This is a stunning result. Though Merton’s (1971a,b) theory says that variables which predict market returns should show up as factors which explain cross-sectional variation in average returns, surprisingly few papers have actually tried to see whether this is true, now that we do have variables that we think forecast the market return. Campbell (1996) and Ferson and Harvey (1999) are among the few exceptions. 20.2.5

Momentum and reversal

Sorting stocks based on past performance, you find that a portfolio that buys long-term losers and sells long-term winners does better than the opposite – individual stock long-term returns mean-revert. This “reversal” effect makes sense given return predictability and meanreversion, and is explained by the Fama-French 3 factor model. However, a portfolio that buys short-term winners and sells short-term losers also does well – “momentum.” This effect is a puzzle. Since a string of good returns gives a high price, it is not surprising that stocks that do well for a long time (and hence build up a high price) subsequently do poorly, and stocks that do poorly for a long time (and hence dwindle down to a low price, market value, or market to book ratio) subsequently do well Table 3, taken from Fama and French (1996) reveals that this is in fact the case. (As usual, this table is the tip of an iceberg of research on these effects, starting with DeBont and Thaler 1985 and Jagadeesh and Titman 1993.)

Strategy Reversal Momentum Reversal Momentum

Period 6307-9312 6307-9312 3101-6302 3101-6302

Portfolio Formation Months 60-13 12-2 60-13 12-2

Average Return, 10-1 (Monthly %) -0.74 +1.31 -1.61 +0.38

Table 3. Average monthly returns from reversal and momentum strategies. Each month, allocate all NYSE firms on CRSP to 10 portfolios based on their performance during the “portfolio formation months” interval. For example, 60-13 forms portfolios based on returns from 5 years ago to 1 year, 1 month ago. Then buy the best-performing decile portfolio and short the worst-performing decile portfolio. Source: Fama and French (1996) Table VI. Reversal

406

S ECTION 20.2

T HE C ROSS - SECTION : CAPM AND MULTIFACTOR MODELS

Here is the “reversal” strategy. Each month, allocate all stocks to 10 portfolios based on performance in year -5 to year -1. Then, buy the best-performing portfolio and short the worst-performing portfolio. The first row of Table 3 shows that this strategy earns a hefty -0.74% monthly return10 . Past long-term losers come back and past winners do badly. This is a cross-sectional counterpart to the mean-reversion that we studied in section 1.4. Fama and French (1998a) already found substantial mean-reversion – negative long-horizon return autocorrelations – in disaggregated stock portfolios, so one would expect this phenomenon. Spreads in average returns should correspond to spreads in betas. Fama and French verify that these portfolio returns are explained by their 3 factor model. Past losers have a high HML beta; they move together with value stocks, and so inherit the value stock premium. Momentum

The second row of Table 3 tracks the average monthly return from a “momentum” strategy. Each month, allocate all stocks to 10 portfolios based on performance in the last year. Now, quite surprisingly, the winners continue to win, and the losers continue to lose, so that buying the winners and shorting the losers generates a positive 1.31% monthly return. At every moment there is a most-studied anomaly, and momentum is that anomaly as I write. It is not explained by the Fama French 3 factor model. The past losers have low prices and tend to move with value stocks. Hence the model predicts they should have high average returns, not low average returns. Momentum stocks move together, as do value and small stocks so a “momentum factor” works to “explain” momentum portfolio returns. This is so obviously ad-hoc (i.e. an APT factor that will only explain returns of portfolios organized on the same characteristic as the factor) that nobody wants to add it as a risk factor. A momentum factor is more palatable as a performance attribution factor – to say that a fund did well by following a momentum strategy rather than by stock picking ability, leaving aside why a momentum strategy should work. Carhart (1997) uses it in this way to show that similar momentum behavior in fund returns is due to momentum in the underlying stocks rather than persistent stock-picking skill. Momentum may be explained as just a new way of looking at an old phenomenon, the small apparent predictability of monthly individual stock returns. A tiny regression R2 for forecasting monthly returns of 0.0025 (1/4%) is more than adequate to generate the momentum results of Table 3. The key is the large standard deviation of individual stock returns, typically 40% or more at an annual basis. The average return of the best performing decile of a normal distribution is 1.76 standard deviations above the mean11 , so the winning moFama and French do not provide direct measures of standard deviations for these portfolios. One can infer however from the betas, R2 values and standard deviation of market and factor portfolios that the standard deviations are roughly 1-2 times that of the market return, so that Sharpe ratios of these strategies are comparable to that of the market return. 11 We’re looking for 10

R∞ rf (r)dr E(r|r ≥ x) = Rx∞ x f (r)dr

407

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

mentum portfolio typically went up about 80% in the previous year, and the typical losing portfolio went down about 60% per year. Only a small amount of continuation will give a 1% monthly return when multiplied by such large√past returns. To be precise, the monthly indi2 vidual stock standard deviation is about 40%/ √ 12 ≈ 12% . If the R is 0.0025, the standard deviation of the predictable part of returns is 0.0025 × 12% = 0.6%. Hence, the decile predicted to perform best will earn 1.76 × 0.6% ≈ 1% above the mean. Since the strategy buys the winners and shorts the losers, an R2 of 0.0025 implies that one should earn a 2% monthly return by the momentum strategy – more even than the 1.3% shown in Table 3. Lewellen (2000) offers a related explanation for momentum coming from small cross-correlations of returns. We have known at least since Fama (1965) that monthly and higher frequency stock returns have slight, statistically significant predictability with R2 in the 0.01 range. However, such small though statistically significant high frequency predictability, especially in small stock returns, has also since the 1960s always failed to yield exploitable profits after one accounts for transactions costs, thin trading, high short sale costs and other microstructure issues. Hence, one naturally worries whether momentum is really exploitable after transactions costs. Momentum does require frequent trading. The portfolios in Table 3 are reformed every month. Annual winners and losers will not change that often, but the winning and losing portfolio must still be turned over at least once per year. Carhart (1996) calculates transactions costs and concludes that momentum is not exploitable after those costs are taken into account. Moskowitz and Grinblatt (1999) note that most of the apparent gains come from short positions in small, illiquid stocks, positions that also have high transactions costs. They also find that a large part of momentum profits come from short positions taken November, anticipating tax-loss selling in December. This sounds a lot more like a small microstructure glitch rather than a central parable for risk and return in asset markets. Table 3 already shows that the momentum effect essentially disappears in the earlier data sample, while reversal is even stronger in that sample. Ahn, Boudoukh, Richardson, and Whitelaw (1999) show that apparent momentum in international index returns is missing from the futures markets, also suggesting a microstructure explanation. Of course, it is possible that a small positive autocorrelation is there and related to some risk. However, it is hard to generate real positive autocorrelation in realized returns. As we saw extensively in section 20.335, a slow and persistent variation in expected returns most naturally generates negative autocorrelation in realized returns. News that expected returns are higher means future dividends are discounted at a higher rate, so today’s price and return declines. The only way to overturn this prediction is to suppose that expected return where x is defined as the top 10th cutoff, Z

∞

f (r)dr =

x

1 . 10

With a normal distribution, x = 1.2816σ and E(r|r ≥ x) = 1.755σ.

408

S ECTION 20.3

S UMMARY AND INTERPRETATION

shocks are positively correlated with shocks to current or expected future dividend growth. A convincing story for such correlation has not yet been constructed. On the other hand, the required positive correlation is very small and not very persistent.

20.3

Summary and interpretation

While the list of new facts appears long, similar patters show up in every case. Prices reveal slow-moving market expectations of subsequent excess returns, because potential offsetting events seem sluggish or absent. The patterns suggest that there are substantial expected return premia for taking on risks of recession and financial stress unrelated to the market return. Magnifying glasses

The effects are not completely new. We knew since the 1960s that high frequency returns are slightly predictable, with R2 of 0.01 to 0.1 in daily to monthly returns. These effects were dismissed because there didn’t seem to be much that one could do about them. A 51/49 bet is not very attractive, especially if there is any transactions cost. Also, the increased Sharpe ratio one can obtain by exploiting predictability is directly related to the forecast R2 , so tiny R2 , even if exploitable, did not seem like an important phenomenon. Many of the new facts amount to clever magnifying glasses, ways of making small facts economically interesting. For forecasting market returns, we now realize that R2 rise with horizon when the forecasting variables are slow-moving. Hence small R2 at high frequency can mean really substantial R2 , in the 30-50% range, at longer horizons. Equivalently, we realize that small expected return variation can add up to striking price variation if the expected return variation is persistent. For momentum effects, the ability to sort stocks and funds into momentum-based portfolios means that incredibly small predictability times portfolios with huge past returns gives important subsequent returns. Dogs that did not bark

In each case, an apparent difference in yield should give rise to an offsetting movement, but seems not to do so. Something should be predictable so that returns are not predictable, and it isn’t. The d/p forecasts of the market return were driven by the fact that dividends should be predictable, so that returns are not. Instead, dividend growth seems nearly unpredictable. As we saw, this fact and the speed of the d/p mean reversion imply the observed magnitude of return predictability. The term structure forecasts of bond returns were driven by the fact that bond yields should be predictable, so that returns are not. Instead, yields seem nearly unpredictable at the one year horizon. This fact means that the forward rate moves one for one with expected returns, and that a one percentage point increase in yield spread signals as much as a 5 percentage point increase in expected return. Exchange rates should be forecastable so that foreign exchange returns are not. Instead, 409

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

a one percentage point increase in interest rate abroad seems to signal a greater than one percentage point increase in expected return. Prices reveal expected returns

If expected returns rise, prices are driven down, since future dividends or other cash flows are discounted at a higher rate. A “low” price, then, can reveal a market expectation of a high expected or required return. Most of our results come from this effect. Low price/dividend, price/earnings, price/book values signal times when the market as a whole will have high average returns. Low market value (price times shares) relative to book value signals securities or portfolios that earn high average returns. The “small firm” effect derives from low prices – other measures of size such as number of employees or book value alone have no predictive power for returns (Berk 1997). The “5 year reversal” effect derives from the fact that 5 years of poor returns lead to a low price. A high long-term bond yield means that the price of long term bonds is “low,” and this seems to signal a time of good long-term bonds returns. A high foreign interest rate means a low price on foreign bonds, and this seems to indicate good returns on the foreign bonds. The most natural intepretatation of all these effects is that the expected or required return – the risk premium – on individual securities as well as the market as a whole varies slowly over time. Thus we can track market expectations of returns by watching price/dividend, price/earnings or book/market ratios. Macroeconomic risks

The price-based patterns in time-series and cross-sectional expected returns suggest a premium for holding risks related to recession and economy-wide financial distress. All of the forecasting variables are connected to macroeconomic activity (Fama and French 1989). The dividend price ratio is highly correlated with the default spread and rises in bad times. The term spread forecasts bond and stock returns, and is also one of the best recession forecasters. It rises steeply at the bottoms of recessions, and is inverted at the top of a boom. Thus, return forecasts are high at the bottom of business cycles and low at the top of booms. “Value” and “small-cap” stocks are typically distressed. Formal quantitative and empirically successful economic models of the recession and distress premia are still in their infancy (I think Campbell and Cochrane 1999 is a good start), but the story is at least plausible, and the effects have been expected by theorists for a generation. To make this point come to life, think concretely about what you have to do to take advantage of the value or predictability strategies. You have to buy stocks or long-term bonds at the bottom, when stock prices are low after a long and depressing bear market; in the bottom of a recession or financial panic; a time when long-term bond prices and corporate bond prices are unusually low. This is a time when few people have the guts (the risktolerance) or the wallet to buy risky stocks or risky long-term bonds. Looking across stocks rather than over time, you have to invest in “value” or small market capitalization companies, dogs by any standards. These are companies with years of poor past returns, years of poor 410

S ECTION 20.3

S UMMARY AND INTERPRETATION

sales, companies on the edge of bankruptcy, far off of any list of popular stocks to buy. Then, you have to sell stocks and long term bonds in good times, when stock prices are high relative to dividends, earnings and other multiples, when the yield curve is flat or inverted so that long term bond prices are high. You have to sell the popular “growth” stocks with good past returns, good sales and earnings growth. I’m going on a bit here to counter the widespread impression, best crystallized by Shiller (2000) that high price earnings ratios must signal “irrational exuberance.” Perhaps, but is it just a coincidence that this exuberance comes at the top of an unprecedented economic expansion, a time when the average investor is surely feeling less risk averse than ever, and willing to hold stocks despite historically low risk premia? I don’t know the answer, but the rational explanation is surely not totally impossible! Is it just a coincidence that we are finding premia just where a generation of theorists said we ought to – in recessions, credit crunches, bad labor markets, investment opportunity set variables, and so forth? This line of explanation for the foreign exchange puzzle is still a bit farther off, though there are recent attempts to make economic sense of the puzzle (See Engel’s 1996 survey; Atkeson, Alvarez and Kehoe 1999 is a recent example.) At a verbal level, the strategy leads you to invest in countries with high interest rates. High interest rates are often a sign of monetary instability or other economic trouble, and thus may mean that the investments are be more exposed to the risks of global financial stress or a global recession than are investments in the bonds of countries with low interest rates, who are typically enjoying better times. Overall, the new view of finance amounts to a profound change. We have to get used to the fact that most returns and price variation come from variation in risk premia, not variation in expected cash flows, interest rates, etc. Most interesting variation in priced risk comes from non-market factors. These are easy to say, but profoundly change our view of the world. Doubts

Momentum is, so far, unlike all the other results. The underlying phenomenon is a small predictability of high frequency returns. However, the price-based phenomena make this predictability important by noting that, with a slow-moving forecasting variable, the R2 build over horizon. Momentum is based on a fast-moving forecast variable – the last year’s return. Therefore the R2 decline with horizon. Instead, momentum makes the tiny autocorrelation of high frequency returns significant by forming portfolios of extreme winners and losers, so a small continuation of huge past returns gives a large current return. All the other results are easily digestible as a slow, business-cycle related time-varying expected return. This specification gives negative autocorrelation (unless we add a distasteful positive correlation of expected return and dividend shocks) and so does not explain momentum. Momentum returns have also not yet been linked to business cycles or financial distress in even the informal way that I suggested for the price-based strategies. Thus, it still lacks much of a plausible economic interpretation. To me, this adds weight to the view that it isn’t there, it isn’t exploitable, or it represents a small illiquidity (tax-loss selling of small illiquid stocks) that will be quickly remedied once a few traders understand it. In the entire history of finance there has always been an anomaly-du-jour, and momentum is it right now. We will have to wait to 411

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

see how it is resolved. Many of the anomalous risk premia seem to be declining over time. The small firm effect completely disappeared in 1980; you can date this as the publication of the first small firm effect papers or the founding of small firm mutual funds that made diversified portfolios of small stocks available to average investors. To emphasize this point, Figure 50 plots size portfolio average returns vs. beta in the period since 1979. You can see that not only has the small firm premium disappeared, the size-related variation in beta and expected return has disappeared.

Figure 50. Average returns vs. market betas. CRSP size portfolios less treasury bill rate, monthly data 1979-1998. The value premium has been cut roughly in half in the 1990s, √ and 1990 is roughly the date of widespread popularization of the value effect, though σ/ T leaves a lot of room for error here. As you saw in Table RR, the last 5 years of high market returns have cut the estimated return predictability from the dividend-price ratio in half. These facts suggest an uncomfortable implication: that at least some of the premium the new strategies yielded in the past was due to the fact that they were simply overlooked or are artifacts of data-dredging. Since they are hard to measure, one is tempted to put less emphasis on these average 412

S ECTION 20.4

P ROBLEMS

returns. However, they are crucial to our interpretation of the facts. The CAPM is perfectly consistent with the fact that there are additional sources of common variation. For example, it was long understood that stocks in the same industry move together; the fact that value or small stocks also move together need not cause a ripple. The surprise is that investors seem to earn an average return premium for holding these additional sources of common movement, whereas the CAPM predicts that (given beta) they should have no effect on a portfolio’s average returns.

20.4 1.

2. 3.

Problems

Does equation (20.308) condition down to information sets coarser than that observed by agents? Or must we assume that whatever VAR is used by the econometrician contains all information seen by agents? Show that the two regressions in Table 5 are complementary – that the coefficients add up to one, mechanically, in sample. Derive the return innovation decomposition (20.319), directly. Write the return rt = ∆dt + ρ (pt − dt ) − (pt−1 − dt−1 )

Apply Et − Et−1 to both sides, rt − Et−1 rt = (Et − Et−1 ) ∆dt + ρ (Et − Et−1 ) (pt − dt ) .

4. 5.

(340)

Use the price-dividend identity and iterate forward to obtain (20.308). Find the univariate representation and mean-reversion statistics for prices implied by the simple VAR and the three dividend examples. Find the univariate return representation from a general return forecasting VAR. rt+1 = axt + εrt+1 xt+1 = bxt + εxt+1

6.

7.

Find the correlation between return and x shocks necessary to generate uncorrelated returns. Show that stationary xt − yt , ∆xt , ∆yt imply that xt and yt must have the same variance ratio and long-run differences must become perfectly correlated. Start by showing that the long run variance limk→∞ var(xt+k − xt )/k for any stationary variable must be zero. Apply that fact to xt − yt . Compute the long-horizon regression coefficients and R2 in the VAR (20.311)-(20.317). Show that the R2 do indeed rise with horizon. Do coefficients and R2 rise forever, or do they turn around at some point?

413

Chapter 21. Equity premium puzzle and consumption-based models The original specification of the consumption-based model was not a great success, as we saw in Chapter 1. Still, it is in some sense the only model we have. The central task of financial economics is to figure out what are the real risks that drive asset prices and expected returns. Something like the consumption-based model – investors’ first order conditions for savings and portfolio choice – has to be the starting point. Rather than dream up models, test them and reject them, financial economists since the work of Mehra and Prescott (1986) and Hansen and Jagannathan (1991) have been able to work backwards to some extent, characterizing the properties that discount factors must have in order to explain asset return data. Among other things, we learned that the discount factor had to be extremely volatile, while not too conditionally volatile; the riskfree rate or conditional mean had to be pretty steady. This knowledge is now leading to a much more successful set of variations on the consumption-based model.

21.1 21.1.1

Equity premium puzzles The basic equity premium/riskfree rate puzzle

The postwar US market Sharpe ratio is about 0.5 – an 8% return and 16% standard deviation. The basic Hansen-Jagannathan bound

σ(m) E(Re ) ≤ ≈ γσ(∆c) σ(Re ) E(m)

implies σ(m) ≥ 50% on an annual basis, requiring huge risk aversion or consumption growth volatility. The average risk free rate is about 1%, so E(m) ≈ 0.99. High risk aversion with power utility implies a very high riskfree rate, or requires a negative subjective discount factor. Interest rates are quite stable over time and across countries, so Et (m) varies little. High risk aversion with power utility implies that interest rates are very volatile. In Chapter 1, we derived the basic Hansen-Jagannathan (1991) bounds. These are characterizations of the discount factors that price a given set of asset returns. Manipulating 414

S ECTION 21.1

E QUITY PREMIUM PUZZLES

0 = E(mRe ) we found |E(Re )| σ(m) ≥ . E(m) σ(Re )

(341)

In continuous time, or as an approximation in discrete time, we found that time-separable utility implies γσ(∆c) ≥

|E(Re )| σ(Re )

(342)

where γ = −cu00 /u0 is the local curvature of the utility function, and risk aversion coefficient for the power case. Equity premium puzzle

The postwar mean value weighted NYSE is about 8% per year over the T-bill rate, with a standard deviation of about 16%. Thus, the market Sharpe ratio E(Re )/σ(Re ) is about 0.5 for an annual investment horizon. If there were a constant risk free rate, E(m) = 1/Rf

would nail down E(m). The T-bill rate is not very risky, so E(m) is not far from the inverse of the mean T-bill rate, or about E(m) ≈ 0.99. Thus, these basic facts about the mean and variance of stocks and bonds imply σ(m) > 0.5. The volatility of the discount factor must be about 50% of its level in annual data! Per capita consumption growth has standard deviation about 1% per year. With log utility, that implies σ(m) = 0.01 = 1% which off by a factor of 50. To match the equity premium we need γ > 50,which seems a huge level of risk aversion. Equivalently, a log utility investor with consumption growth of 1% and facing a 0.5 Sharpe ratio should be investing dramatically more in the stock market, borrowing to do so. He should invest so much that his wealth and hence consumption growth does vary by 50% each year. Correlation puzzle

The bound takes the extreme possibility that consumption and stock returns are perfectly correlated. They are not, in the data. Correlations are hard to measure, since they are sensitive to data definition, timing, time-aggregation, and so forth. Still, the correlation of annual stock returns and nondurable plus services consumption growth in postwar U.S. data is no more than about 0.2. If we use this information as well – if we characterize the mean and standard deviation of all discount factors that have correlation less than 0.2 with the market return – the calculation becomes |E(Re )| 1 1 σ(m) ¯ ≥ ¯¯ = 0.5 = 2.5 e ¯ E(m) 0.2 ρm,Re σ(R )

with σ(m) ≈ γσ(∆c), we now need a risk aversion coefficient of 250! 415

C HAPTER 21

E QUITY PREMIUM PUZZLE AND CONSUMPTION - BASED MODELS

Here is a classier way to state the correlation puzzle. Remember that proj(m|X) should price assets just as well as m itself. Now, m = proj(m|X)+ε and σ 2 (m) = σ2 (proj (m|X))+ σ2 (ε). Some of the early resolutions of the equity premium puzzle ended up adding noise uncorrelated with asset payoffs to the discount factor. This modification increased discount factor volatility and satisfied the bound. But as you can see, adding ε increases σ 2 (m) with no effect whatsoever on the model’s ability to price assets. As you add ε, the correlation between m and asset returns declines. A bound with correlation, or equivalently comparing σ2 (proj(m|X)) rather than σ 2 (m) to the bound avoids this trap. Average interest rates and subjective discount factors

It has been traditional to use risk aversion numbers of 1 to 5 or so, but perhaps this is tradition, not fact. What’s wrong with γ = 50 to 250? The most basic piece of evidence for low γ comes from the relation between consumption growth and interest rates. " µ ¶−γ # Ct+1 f Rt = Et (mt+1 ) = Et β Ct or, in continuous time, rtf = δ + γEt (∆c) .

(343)

We can take unconditional expectations to compare these equations with average interest rates and consumption growth. Average real interest rates are also about 1% Thus, γ = 50 to 250 with a typical δ such as δ = 0.01 implies a very high riskfree rate, of 50 − 250%. To get a reasonable interest rate, we have to use a subjective discount factor δ = −0.5 to −2.5, or −50% to −250%. That’s not impossible – present values can converge with negative discount rates (Kocherlakota 1990) – but it does not seem reasonable. People prefer earlier consumption, not later consumption. Interest rate variation and the conditional mean of the discount factor

Again, however, maybe we’re being too doctrinaire. What evidence is there against γ = 50 − 250 with corresponding δ = −0.5 to −2.5?

Real interest rates are not only low on average, they are also relatively stable over time and across countries. γ = 50 in equation (21.343) means that a country or a boom time with consumption growth 1 percentage point higher than normal must have real interest rates 50 percentage points higher than normal, and consumption 1 percentage point lower than normal should be accompanied by real interest rates of 50 percentage points lower than normal– you pay them 48% to keep your money. We don’t see anything like this. γ = 50 to 250 in a time-separable utility function implies that consumers are essentially unwilling to substitute (expected) consumption over time, so huge interest rate variation must force them to make the small variations in consumption growth that we do see. This level of aversion to intertemporal substitution is too large. For example, think about what interest

416

S ECTION 21.1

E QUITY PREMIUM PUZZLES

rate you need to convince someone to skip a vacation. Take a family with $50,000 per year consumption, and which spends $2,500 (5%) on an annual vacation. If interest rates are good enough, though, the family can be persuaded to skip this year’s vacation and go on a much more lavish vacation next year. The required interest rate is ($52, 500/$47, 500)γ − 1. For γ = 250 that is an interest rate of 3 × 1011 ! For γ = 50, we still need an interest rate of 14, 800%. I think most of us would give in and defer the vacation for somewhat lower interest rates! A reasonable willingness to substitute intertemporally is central to most macroeconomic models that try to capture output, investment, consumption, etc. dynamics. As always, we can express the observation as a desired characteristic of the discount factor. Though mt+1 must vary a lot, its conditional mean Et (mt+1 ) = 1/Rtf must not vary much. You can get variance in two ways – variance in the conditional mean and variance in the unexpected component; var(x) = var [Et (x)] + var [x − Et (x)]. The fact that interest rates are stable means that almost all of the 50% or more unconditional variance must come from the second term. The power functional form is really not an issue. To get past the equity premium and these related puzzles, we will have to introduce other arguments to the marginal utility function – some non-separability. One important key will be to introduce some non-separability that distinguishes intertemporal substitution from risk aversion. 21.1.2

Variations

Just raising the interest rate will not help, as all-stock portfolios have high Sharpe ratios too. Uninsured individual risk is not an obvious solution. Individual consumption is not volatile enough to satisfy the bounds, and is less correlated with stock returns than aggregate consumption. The average return in postwar data may overstate the true expected return; a target of 3-4% is not unreasonable. Is the interest rate “too low”?

A large literature has tried to explain the equity premium puzzle by introducing frictions that make treasury bills “money-like” and so argue that the short-term interest rate is artificially low. (Aiyagari and Gertler 1991 is an example). However, high Sharpe ratios are pervasive in financial markets. Portfolios long small stocks and short big stocks, or long value (high book/market) and short growth stocks, give Sharpe ratios of 0.5 or more as well. Individual shocks

Maybe we should abandon the representative agent assumption. Individual income shocks 417

C HAPTER 21

E QUITY PREMIUM PUZZLE AND CONSUMPTION - BASED MODELS

are not perfectly insured, so individual income and consumption is much more volatile than aggregate consumption. Furthermore, through most of the sample, only a small portion of the population held any stocks at all. This line of argument faces a steep uphill battle. The basic pricing equation applies to each consumer. Individual income growth may be more volatile than the aggregate, but it’s not credible that any individual’s consumption growth varies by 50% -250% per year! Keep in mind, this is nondurable and services consumption and the flow of services from durables, not durables purchases. Furthermore, individual consumption growth is likely to be less correlated with stock returns than is aggregate consumption growth, and the more volatile it is, the less correlated. As a simple example, write individual consumption equal to aggregate consumption plus an idiosyncratic shock, uncorrelated with economywide variables, ∆cit = ∆cat + εit .

Hence, ¢ ¡ cov(∆cit , rt ) = cov ∆cat + εit , rt = cov (∆cat , rt ) .

As we add more idiosyncratic variation, the correlation of consumption with the any aggregate such as stock returns declines in exact proportion so that the asset pricing implications are completely unaffected. Luck and a lower target

One nagging doubt is that a large part of the U.S. postwar average stock return may represent good luck rather than ex-ante expected return. First of all, the standard deviation of stock returns √ is so high that standard errors are surprisingly large. Using the standard√formula σ/ T , the standard error of average stock returns in 50 years of data is about 16/ 50 ≈ 2.3. This fact means that a two-standard error confidence interval for the expected return extends from about 3% to about 13%! This is a pervasive, simple, but√surprisingly under-appreciated problem in empirical asset pricing. In 20 years of data, 16/ 20 = 3.6 so we can barely say that an 8% average return is above zero. 5 year performance averages √ of something like a stock return are close to meaningless on a statistical basis, since 16/ 5 = 7. 2. (This is one reason that many funds are held to tracking error limits relative to a benchmark. You may be able to measure performance relative to a benchmark, even if your return and the benchmark are both very volatile. √ If σ(Ri − Rm ) is small, then σ(Ri − Rm )/ T can be small, even if σ(Ri ) and σ(Rm ) are large.) However, large standard errors can argue that the equity premium is really higher than the postwar return. Several other arguments suggest a bias – that a substantial part of the 8% average excess return of the last 50 years was good luck, and that the true equity premium is more like 3-4%. 418

S ECTION 21.1

E QUITY PREMIUM PUZZLES

Brown, Goetzmann and Ross (1995) suggest that the U.S. data suffer from selection bias. One of the reasons that I write this book in the U.S., and that the data has been collected from the U.S., is precisely because U.S. stock returns and growth have been so good for the last 50 - 100 years. One way to address this question is to look at other samples. Average returns were a lot lower in the U.S. before WWII. In Shiller’s (1989) annual data from 1871-1940, the S&P500 average excess return was only 4.1% However, Campbell (1999, table 1) looks across countries for which we have stock market data from 1970-1995, and finds the average equity premium practically the same as that for the U.S. in that period. The other countries averaged a 4.6% excess return while the U.S. had a 4.4% average excess return in that period. On the other hand, Campbell’s countries are Canada, Japan, Australia and Western Europe. These probably shared a lot of the U.S. “good luck” in the postwar period. There are lots of countries for which we don’t have data, and usually because returns were very low in those countries. As Brown, Goetzmann and Ross (1995) put it, “Looking back over the history of the London or the New York stock markets can be extraordinarily comforting to an investor – equities appear to have provided a substantial premium over bonds, and markets appear to have recovered nicely after huge crashes. ... Less comforting is the past history of other major markets: Russia, China, Germany and Japan. Each of these markets has had one or more major interruptions that prevent their inclusion in long term studies” [my emphasis]. Think of the things that didn’t happen in the last 50 years. We had no banking panics, and no depressions; no civil wars, no constitutional crises; we did not lose the cold war, no missiles were fired over Berlin, Cuba, Korea or Vietnam. If any of these things had happened, we might well have seen a calamitous decline in stock values, and I would not be writing about the equity premium puzzle. A view that stocks are subject to occasional and highly non-normal crashes – world wars, great depressions, etc. – makes sampling uncertainty even larger, and means that the average return from any sample that does not include a crash will be larger than the actual average return – the Peso Problem again (Reitz 1988). Fama and French (2000) notice that the price/dividend ratio is low at the beginning of the sample and high at the end. Much of that is luck–the dividend yield is stationary in the very long run, with slow-moving variation through good and bad times. We can understand their alternative calculation most easily using the return linearization, rt+1 = ∆dt+1 + (dt − pt ) − ρ(dt+1 − pt+1 ).

Then, imposing the view that the dividend price ratio is stationary, we can estimate the average return as E (rt+1 ) = E (∆dt+1 ) + (1 − ρ)E(dt − pt ).

The right hand expression gives an estimate of the unconditional average return on stocks equal to 3.4%. This differes from the sample average return of 9% because, the d/p ratio 419

C HAPTER 21

E QUITY PREMIUM PUZZLE AND CONSUMPTION - BASED MODELS

declined dramatically in the postwar sample. Here is the fundamental issue: Was it clear to people in 1947 (or 1871, or whenever one starts the sample) and throughout the period that the average return on stocks would be 8% greater than that of bonds, subject only to the 16% year to year variation? Given that knowledge, would investors have changed their portfolios, or would they have stayed pat, patiently explaining that these average returns are earned in exchange for risk that they are not prepared to take? If people expected these mean returns, then we face a tremendous challenge of explaining why people did not buy more stocks. This is the basic assumption and challenge of the equity premium puzzle. But phrased this way, the answer is not so clear. I don’t think it was obvious in 1947 that the United States would not slip back into depression, or another world war, but would instead experience a half century of economic growth and stock returns never before seen in human history. 8% seems like an extremely – maybe even irrationally – exuberant expectation for stock returns as of 1947, or 1871. (You can ask the same question, by the way, about value effects, market timing, or other puzzles we try to explain. Only if you can reasonably believe that people understood the average returns and shied away because of the risks does it make sense to explain the puzzles by risk rather than luck. Only in that case with the return premia continue anyway!) This consideration mitigates, but cannot totally solve the equity premium puzzle. Even a 3% equity premium is tough to understand with 1% consumption volatility. If the premium is 3%, the Sharpe ratio is 3/16 ≈ 0.2, so we still need risk aversion of 20, and 100 if we include correlation. 20-100 is a lot better than 50-250, but is still quite a challenge. 21.1.3

Predictability and the equity premium

The Sharpe ratio varies over time. This means that discount factor volatility must vary over time. Since consumption volatility does not seem to vary over time, this suggests that risk aversion must vary over time – a conditional equity premium puzzle. Conventional portfolio calculations suggest that people are not terribly risk averse. These calculations implicitly assume that consumption moves proportionally to wealth, and inherits the large wealth volatility. If stock returns mean-revert, E(Re )/σ(Re ) and hence σ(m)/E(m) rises faster than the square root of the horizon. Consumption growth is roughly i.i.d., so σ(∆c) rises about with the square root of horizon. Thus, mean-reversion means that the equity premium puzzle is even worse for long-horizon investors and long-horizon returns. We have traced the implications of the unconditional Sharpe ratio, and of low and relatively constant interest rates. The predictability of stock returns also has important implications for discount factors. 420

S ECTION 21.1

E QUITY PREMIUM PUZZLES

Heteroskedasticity in the discount factor–conditional equity premium puzzle

The Hansen-Jagannathan bound applies conditionally of course, ¡ e ¢ Et Rt+1 1 σ t (mt+1 ) . e ) = − ρ (Re , m σt (Rt+1 ) E t+1 t (mt+1 ) t t+1 Mean returns are predictable, and the standard deviation of returns varies over time. So far, however, the two moments are forecasted by different sets of variables and at different horizons – d/p, term premium, etc. forecast the mean at long horizons; past squared returns and implied volatility forecast the variance at shorter horizons – and these variables move at different times. Hence, it seems that the conditional Sharpe ratio on the left hand side moves over time. (Glosten, Jagannathan and Runkle 1993, French Schwert and Stambaugh 1987, Yan 2000 find some co-movements in conditional mean and variance, but do not find that all movement in one moment is matched by movement in the other.) On the right hand side, the conditional mean discount factor equals the risk free rate and so must be relatively stable over time. Time-varying conditional correlations are a possibility, but hard to interpret. Thus, the predictability of returns strongly suggests that the discount factor must be conditionally heteroskedastic – σt (mt+1 ) must vary through time. Certainly the discount factors on the volatility bound, or the mimicking portfolios for discount factors, both of which have ρ = 1, must have time-varying volatility. In the standard time-separable model, σt (mt+1 ) = γ t σ t (∆ct+1 ). Thus, we need either time-varying consumption risk or time-varying curvature; loosely speaking a time-varying risk aversion. The data don’t show much evidence of conditional heteroskedasticity in consumption growth, leading one to favor a time-varying risk aversion. However, this is a case in which high risk aversion helps: if γ is sufficiently high, a small and perhaps statistically hard to measure amount of consumption heteroskedasticity can generate a lot of discount factor heteroskedasticity. (Kandel and Stambaugh 1990 follow this approach to explain predictability.) Capm, portfolios and consumption

The equity premium puzzle is centrally about the smoothness of consumption. This is why it was not noticed as a major puzzle in the early development of financial theory. In turn, the smoothness of consumption is centrally related to the predictability of returns. In standard portfolio analyses, there is no puzzle that people with normal levels of risk aversion do not want to hold far more stocks. From the usual first order condition and with Λ = VW (W ) we can also write the Hansen-Jagannathan bound in terms of wealth, analogously to (21.342), ¯ ¯ ¯E(r) − rf ¯ −W VW W ≤ σ (∆w) (344) σ(r) VW The quantity −W WW W /VW is in fact the measure of risk aversion corresponding to most survey and introspection evidence, since it represents aversion to bets on wealth rather than 421

C HAPTER 21

E QUITY PREMIUM PUZZLE AND CONSUMPTION - BASED MODELS

to bets on consumption. (They can be the same for power utility, but not in general.) For an investor who holds the market, σ (∆w) is the standard deviation of the stock return, about 16%. With a market Sharpe ratio of 0.5, we find the lower bound on risk aversion, 0.5 −W VW W ≈ 3. = VW 0.16

Furthermore, the correlation between wealth and the stock market is one in this calculation, so no correlation puzzle crops up to raise the required risk aversion. This is the heart of the oft-cited Friend and Blume (1975) calculation of risk aversion, one source of the idea that 3-5 is about the right level of risk aversion rather than 50 or 250. The Achilles heel is the hidden simplifying assumption that returns are independent over time, and the investor has no other source of income, so no variables other than wealth show up in its marginal value VW . In such an i.i.d. world, consumption moves one-for-one with wealth, and σ (∆c) = σ (∆w). If your wealth doubles and nothing else has changed, you double consumption. This calculation thus hides a consumption-based “model,” and the model has the drastically counterfactual implication that consumption growth has a 16% standard deviation! All this calculation has done is say that “in a model in which consumption has a 16% volatility like stock returns, we don’t need high risk aversion to explain the equity premium.” Hence the central point – the equity premium is about consumption smoothness. Just looking at wealth and portfolios, you do not notice anything unusual. In the same way, retreating to the CAPM or factor models doesn’t solve the puzzle either. The CAPM is a specialization of the consumption-based model, not alternatives, and thus hide an equity premium puzzle. For example, I derived the CAPM above as a consequence of log utility. With log utility, you have to believe that properly measured consumption growth has a 50% per year standard deviation! That testable implication is right there in the model, though often ignored. Most implementations of the CAPM take the market premium as given (ignoring the link to consumption in the model’s derivation) and estimate the market premium as a free parameter. The equity premium puzzle asks whether the market premium itself makes any sense. The long-run equity premium puzzle

The fact that annual consumption is much smoother than wealth is an important piece of information. In the long-run, consumption must move one-for-one with wealth, so consumption and wealth volatility must be the same. Therefore, we know that the world is very far from i.i.d., so predictability will be an important issue in understanding risk premia. Predictability can imply mean reversion and Sharpe ratios that rise faster than the square root of horizon. Thus, E(Ret→t+k ) σ(mt→t+k ) ¢≤ ¡ e ≈ γσ(∆ct→t+k ). E(mt→t+k ) σ Rt→t+k

422

S ECTION 21.2

N EW MODELS

If stocks do mean-revert, then discount factor volatility must increase faster than the square root of the horizon. Consumption growth is close to i.i.d.,so the volatility of consumption growth only increases with the square root of horizon. Thus mean-reversion implies that the equity premium puzzle is even worse at long investment horizons.

21.2

New models

We want to end up with a model that explains a high market Sharpe ratio, and the high level and volatility of stock returns, with low and relatively constant interest rates, roughly i.i.d. consumption growth with small volatility, and that explains the predictability of excess returns – the fact that high prices today correspond to low excess returns in the future. Eventually, we would like the model to explain the predictability of bond and foreign exchange returns as well, the time-varying volatility of stock returns and the cross-sectional variation of expected returns, and it would be nice if in addition to fitting all of the facts, people in the models did not display unusually high aversion to wealth bets. I start with a general outline of the features that most models that address these puzzles share. Then, I focus on two models, the Campbell-Cochrane (1999) habit persistence model and the Constantinides and Duffie (1996) model with uninsured idiosyncratic risks. The mechanisms we uncover in these models apply to a large class. The Campbell-Cochrane model is a representative from the literature that attacks the equity premium by modifying the representative agent’s preferences. The Constantinides and Duffie model is a representative of the literature that attacks the equity premium by modeling uninsured idiosyncratic risks, market frictions, and limited participation. 21.2.1

Outlines of new models

Additional state variables are the natural route to solving the empirical puzzles. Investors must not be particularly scared of the wealth or consumption effects of holding stocks, but of the fact that stocks do badly at particular times, or in particular states of nature. Broadly speaking, most solutions introduce something like a “recession” state variable. This fact makes stocks different, and more feared, than pure wealth bets, whose risk is unrelated to the state of the economy. In the ICAPM way of looking at things, we get models of this sort by specifying things so there is an additional recession state variables z in the value function V (W, z). Then, expected returns are

−W VW W E(r) − r = cov VW f

µ

423

¶ zVW z dW ,r + cov(z, r). W VW

(345)

C HAPTER 21

E QUITY PREMIUM PUZZLE AND CONSUMPTION - BASED MODELS

In a utility framework, we add other arguments to the utility function u(C, z), so −CuCC E(r) − r = cov uC f

µ

¶ zuCz dC ,r + cov(z, r). C uC

(346)

The extra utility function arguments must enter non-separably. If u(C, z) = f (C) + g(z), then uCz = 0. All utility function modifications are of this sort – they add extra goods like leisure, nonseparability over time in the form of habit persistence, or nonseparability across states of nature so that consumption if it rains affects marginal utility if it shines. The lesson of the equity premium literature is that the second term must account for essentially all of the market premium. Since the cross-sectional work surveyed in Chapter 20 seemed to point to something like a recession factor as the primary determinant of cross-sectional variation in expected returns, a gratifying unity seems close at hand – and a fundamental revision of the CAPM-i.i.d. view of the source of risk prices. The predictability of returns – emphasized by the dramatic contrast between consumption and wealth volatility at short horizons – suggests a natural source of state variables. Unfortunately, the sign is wrong. The fact that stocks go up when their expected subsequent returns are low means that stocks, like bonds, are good hedges for shocks to their own opportunity sets. Therefore, adding the effects of predictability typically lowers expected returns. (The “typically” in this sentence is important. The sign of this effect – the sign of zVW z – does depend on the utility function and environment. For example, there is no risk premium for log utility.) Thus, we need an additional state variable, and one strong enough to not only explain the equity premium, given that the first terms in (21.345) and (21.346) are not up to the job, but one stronger still to overcome the effects of predictability. Recessions are times of low prices and high expected returns. We want a model in which recessions are bad times, so that investors fear bad stock returns in recessions. But high expected returns are good times for a pure Merton investor. Thus, the other state variable(s) that describe a recession – high risk aversion, low labor income, high labor income uncertainty, liquidity, etc. – must overcome the “good times” of high expected returns and indicate that times really are bad after all. 21.2.2

Habits

A natural explanation for the predictability of returns from price/dividend ratios is that people get less risk averse as consumption and wealth increase in a boom, and more risk averse as consumption and wealth decrease in a recession. We can’t tie risk aversion to the level of consumption and wealth, since that increases over time while equity premia have not declined. Thus, to pursue this idea, we must specify a model in which risk aversion depends on the level of consumption or wealth relative to some “trend” or the recent past. Following this idea, Campbell and Cochrane (1999) specify that people slowly develop habits for higher or lower consumption. Thus, the “habits” form the “trend” in consumption. 424

S ECTION 21.2

N EW MODELS

The idea is not implausible. Anyone who has had a large pizza dinner or smoked a cigarette knows that what you consumed yesterday can have an impact on how you feel about more consumption today. Might a similar mechanism apply for consumption in general and at a longer time horizon? Perhaps we get used to an accustomed standard of living, so a fall in consumption hurts after a few years of good times, even though the same level of consumption might have seemed very pleasant if it arrived after years of bad times. This thought can at least explain the perception that recessions are awful events, even though a recession year may be just the second or third best year in human history rather than the absolute best. Law, custom and social insurance also insure against falls in consumption as much as low levels of consumption. The Model

We model an endowment economy with i.i.d. consumption growth. ∆ct+1 = g + vt+1 ; vt+1 ∼ i.i.d. N (0, σ 2 ).

We replace the utility function u(C) with u(C − X) where X denotes the level of habits. E

∞ X t=0

δt

(Ct − Xt )1−γ − 1 . 1−γ

Habits should move slowly in response to consumption, something like xt ≈ λ

∞ X

φj ct−j

(347)

j=0

or, equivalently xt = φxt−1 + λct .

(348)

(Small letters denote the logs of large letters throughout this section, ct = ln Ct , etc.) Rather than letting habit itself follow an AR(1) we let the “surplus consumption ratio” of consumption to habit follow an AR(1): St =

Ct − Xt Ct

st+1 = (1 − φ)¯ s + φst + λ (st ) (ct+1 − ct − g)

(349)

Since s contains c and x, this equation also specifies how x responds to c, and it is locally the same as (21.347). We also allow consumption to affect habit differently in different states by 425

C HAPTER 21

E QUITY PREMIUM PUZZLE AND CONSUMPTION - BASED MODELS

specifying an square root type process rather than a simple AR(1), 1p 1 − 2 (st − s¯) − 1 S¯r γ S¯ = σ 1−φ

λ(st ) =

(21.350) (21.351)

The extra complication of (21.349) rather than (21.347) means consumption is always above habit, since S = es > 0. Other habit models can give consumption below habit which leads to infinite or imaginary marginal utility. St becomes the single state variable in this economy. Time-varying expected returns, price-dividend ratios, etc. are all functions of this state variable.

Marginal utility is uc (Ct , Xt ) = (Ct − Xt )−γ = St−γ Ct−γ .

The model assumes an external habit – each individual’s habit is determined by everyone else’s consumption, as in Abel’s (1990) “keeping up with the Joneses” specification. This is mostly a technical simplification, since it allows us to ignore terms by which current consumption affect future habits; the opposite specification gives very similar results (see problem 2). With marginal utility, we now have a discount factor. µ ¶−γ St+1 Ct+1 uc (Ct+1 , Xt+1 ) =δ Mt+1 ≡ δ . uc (Ct , Xt ) St Ct Since we have a stochastic process for S and C , and each is lognormal, we can evaluate the conditional mean of the discount factor to evaluate the riskfree rate 1 rtf = − ln Et (Mt+1 ) = − ln(δ) + γg − γ(1 − φ). 2

(352)

We gave up on analytic solutions and evaluated the price-dividend ratio as a function of the state variable by iteration on a grid: ¶¸ · µ Pt Ct+1 Pt+1 (st ) = Et Mt+1 (st+1 ) 1+ Ct Ct Ct+1 With price-dividend ratios, we can calculate returns, expected returns, etc. How does it work – equity premium and predictability

We choose parameters, simulate 100,000 artificial data points, and report standard statistics and tests in artificial data. The parameters g = 1.89, σ = 1.50, rf = 0.94 match their values in postwar data. The parameter φ = 0.87 matches the autocorrelation of the pricedividend ratio and the choice γ = 2.00 matches the postwar Sharpe ratio. δ = 0.89, S¯ = 0.57 follow from the model. 426

S ECTION 21.2

N EW MODELS

Table 2cc presents means and standard deviations predicted by the model. The model replicates the postwar Sharpe ratio, with a constant 0.94% risk free rate and a reasonable subjective discount factor δ < 1. Of course, we picked the parameters to do this, but given the above equity premium discussion it’s already an achievement that we are able to pick any parameters to hit these moments. Some models can replicate the Sharpe ratio, but do not replicate the level of expected returns and return volatility. E = 1% and σ = 2% will give the right Sharpe ratio, but this model predicts the right levels as well. The model also gets the level of the price-dividend ratio about right. Table 2cc. Means and standard deviations of simulated and historical data. Statistic E(R − Rf )/σ(R − Rf ) E(r − rf ) σ(r − rf ) exp[E(p − d)] σ(p − d)

Consumption claim 0.50 6.64 15.2 18.3 0.27

Dividend claim 6.52 20.0 18.7 0.29

Postwar data 0.50 6.69 15.7 24.7 0.26

The model is simulated at a monthly frequency; statistics are calculated from artificial time-averaged data at an annual frequency. Asterisks (*) denote statistics that model parameters were chosen to replicate. All returns are annual percentages. Table 5cc shows how the artificial data match the predictability of returns from pricedividend ratios. The paper goes on, and shows how the model matches the volatility test result that almost all return variation is due to variation in expected excess returns, the “leverage effect” of higher volatility after a big price decline, and several related phenomena. Table 5cc. Long-horizon return regressions Horizon (Years) 1 2 3 5 7

Cons. claim 10×coef. R2 -2.0 0.13 -3.7 0.23 -5.1 0.32 -7.5 0.46 -9.4 0.55

Postwar data 10×coef. R2 -2.6 0.18 -4.3 0.27 -5.4 0.37 -9.0 0.55 -12.1 0.65

How does it work?

How does this model get around all the equity premium - riskfree rate difficulties described above, and explain predictability as well? When a consumer has a habit, local curvature depends on how far consumption is above 427

C HAPTER 21

E QUITY PREMIUM PUZZLE AND CONSUMPTION - BASED MODELS

the habit, as well as the power γ, ηt ≡

γ −Ct ucc (Ct − Xt ) = . uc (Ct − Xt ) St

As consumption falls toward habit, people become much less willing to tolerate further falls in consumption; they become very risk averse. Thus a low power coefficient γ can still mean a high, and time-varying curvature. Recall our fundamental equation for the Sharpe ratio, Et (r) − rtf = η t σt (∆c)corrt (∆c, r). σ t (r)

High curvature ηt means that the model can explain the equity premium, and curvature ηt that varies over time as consumption rises in booms and falls toward habit in recessions means that the model can explain a time-varying and countercyclical (high in recessions, low in booms) Sharpe ratio, despite constant consumption volatility σt (∆c) and correlation corrt (∆c, r). So far so good, but didn’t we just learn that raising curvature implies high and timevarying interest rates? This model gets around interest rate problems with precautionary saving. Suppose we are in a bad time, in which consumption is low relative to habit. People want to borrow against future, higher, consumption, and this force should drive up interest rates. (In fact, many habit models have very volatile interest rates.) However, people are also much more risk averse when consumption is low. This consideration induces them to save more, in order to build up assets against the event that tomorrow might be even worse. This desire to save drives down interest rates. Our λ(s) specification makes these two forces exactly offset, leading to constant real rates. The precautionary saving motive also makes the model more plausibly consistent with variation in consumption growth across time and countries. Adding (21.351) to (21.352), we can write 1 ³ γ ´2 2 rf = ρ + γg − σ 2 S¯

The power coefficient γ = 2 controls the relation between consumption growth and interest rates, while the curvature coefficient γ/St controls the risk premium. Thus this habit model allows high “risk aversion” with low “aversion to intertemporal substitution,” and it is consistent with the consumption and interest rate data. As advertised, this model explains the equity premium and predictability by fundamentally changing the story for why consumers are afraid of holding stocks. The k− period stochastic discount factor is µ ¶−γ St+k Ct+k k Mt→t+k = δ . St Ct

covariances with S shocks now drive average returns as well as covariances with C shocks. S = (C − X)/C is a recession indicator – it is low after several quarters of consumption 428

S ECTION 21.2

N EW MODELS

declines and high in booms. While (Ct+k /Ct )−γ and (St+k /St )−γ enter symmetrically in the formula, the volatility of (Ct+k /Ct )−γ with γ = 2 is so low that it accounts for essentially no risk premia. Therefore, it must be true, and it is, that variation in (St+k /St )−γ is much larger, and accounts for nearly all risk premia. In the Merton language of (21.345) and (21.346), variation across assets in expected returns is driven by variation across assets in covariances with recessions far more than by variation across assets in covariances with consumption growth. At short horizons, shocks to St+1 and Ct+1 move together, so the distinction between a recession state variable and consumption risk is minor; one can regard S as an amplification mechanism for consumption risks in marginal utility. dS/∂C ≈ 50, so this amplification generates the required volatility of the discount factor. At long horizons, however, St+k becomes less and less conditionally correlated with Ct+k . St+k depends on Ct+k relative to its recent past, but the overall level of consumption may be high or low. Therefore, investors fear stocks because they do badly in occasional serious recessions, times of recent belt-tightening. These risks are at the long run unrelated to the risks of long-run average consumption growth. As another way to digest how this model works, we can substitute in the s process from (21.349) and write the marginal rate of substitution as Mt+1

=

ln Mt+1

= = =

ln Mt+1

µ

¶−γ St+1 Ct+1 δ St Ct ln δ − γ (st+1 − st ) − γ(ct+1 − ct ) {ln δ − γ(1 − φ)¯ s} + {γ (1 − φ) st + γgλ (st )} − γ [λ (st ) + 1] (ct+1 − ct ) a + b(st ) + d(st )(ct+1 − ct )

Up to the question of logs vs. levels, this is a “scaled factor model” of the form we studied in Chapter 8. It still is a consumption-based model, but the sensitivity of the discount factor to consumption changes over time. The long-run equity premium is even more of a puzzle. Most recession state variables, such as recessions, labor, and instruments for time-varying expected returns (“shifts in the investment opportunity set”) are stationary. Hence, the standard deviation of their growth rates eventually stops growing with horizon. At a long enough horizon, the standard deviation of the discount factor is dominated by the standard deviation of the consumption growth term, and we return to the equity premium puzzle at a long enough run. Since this model produces predictability of the right sign, it produces a long run equity premium puzzle. How it manages this feat with a stationary state variable St is subtle (and we didn’t notice it until the penultimate draft!) The answer is that while St is stationary, St−γ −γ is not. St has a fat tail approaching zero so the conditional variance of St+k grows without bound. While the distinction between stationary S and nonstationary S −γ seems initially minor, it 429

C HAPTER 21

E QUITY PREMIUM PUZZLE AND CONSUMPTION - BASED MODELS

is in fact central. Any model that wishes to explain the equity premium at long and short runs by means of an additional, stationary state variable must find some similar transformation so that the volatility of the stochastic discount factor remains high at long horizons. This model does have high risk aversion. The utility curvature and value function curvature are both high. Many authors require that a “solution” of the equity premium puzzle display low risk aversion. This is a laudable goal, and no current model has attained it. No current model generates the equity premium with a low and relatively constant interest rate, low risk aversion, and the right pattern of predictability – high prices forecast low returns, not high returns, and consumption is roughly a random walk. Constantinides (1990) and Boldrin, Christiano and Fisher (1997) are habit models with a large equity premium and low risk aversion, but they don’t get the pattern of predictability right. Boldrin, Christiano and Fisher have highly variable interest rates to keep consumption from being predictable. Constantinides (1990) has a constant interest rate, but consumption growth that is serially correlated, so consumption rises to meet i.i.d. wealth growth. The long-run equity premium is solved with counterfactually high long-run consumption volatility. To get a high equity premium with low risk aversion, we need to find some crucial characteristic that separates stock returns from wealth bets. This is a difficult task. After all, what are stocks if not a bet? The answer must be some additional state variable. Stocks must pay off badly in particularly unfortunate states of the world. Again, the trouble with predictability is that stocks pay off well in particularly bad states of the world – states with low future returns. This makes stocks even more desirable, requiring even higher risk aversion to explain the equity premium. The alternative, not yet found, is to find some measure of the state of the world that is particularly bad when stocks pay off badly, enough to explain not only the standard equity premium, but the long run equity premium resulting from the fact that stocks are less risky at longer horizons. I write this not to say that such a model is impossible. The point is to show the hurdle that must be overcome, in the hope that someone will overcome it. 21.2.3

Heterogeneous agents and idiosyncratic risks

A long, increasing, and important literature in the equity premium attacks the problem instead with relatively standard preferences, but instead adds uninsured idiosyncratic risk. As with the preference literature, this literature is interesting beyond the equity premium. We are learning a lot about who holds stocks and why, what risks they face. We are challenged to think of new assets and creative ways of using existing assets to share risks better. Constantinides and Duffie (1996) provide a very clever and simple model in which idiosyncratic risk can be tailored to generate any pattern of aggregate consumption and asset prices. It can generate the equity premium, predictability, relatively constant interest rates, smooth and unpredictable aggregate consumption growth and so forth. Furthermore, it requires no transactions costs, borrowing constraints or other frictions, and the individual con430

S ECTION 21.2

N EW MODELS

sumers can have any nonzero value of risk aversion. Of course, we still have to evaluate whether the idiosyncratic risk process we construct to explain asset pricing phenomena are reasonable and consistent with microeconomic data. A simple version of the model

I start with a very simplified version of the Constantinides-Duffie model. Each consumer i has power utility, U =E

X

1−γ e−δt Cit

t

Individual consumption growth Cit+1 is determined by an independent, idiosyncratic normal (0,1) shock ηit , ln

µ

Cit+1 Ci,t

¶

1 2 = ηit+1 yt+1 − yt+1 2

(353)

where yt+1 is, by construction since it multiplies the shock ηit , the cross-sectional standard deviation of consumption growth. yt+1 is dated t + 1 since it is the cross-sectional standard deviation given aggregates at t+1. The aggregates are determined first, and then the shocks ηit are handed out. Now, yt+1 is specified so that people suffer a high cross-sectional variance of consumption growth on dates of a low market return Rt+1 , ·

yt+1 = σ ln

µ

¶¯ ¸ s p Cit+1 ¯¯ 2 Rt+1 = δ − ln Rt+1 . ¯ Cit γ(γ + 1)

(354)

Given this structure, the individual is exactly happy to consume {Cit } without further trading in the stock. (We can call Cit income Iit , and prove the optimal decision rule is to consume income Cit = Iit .) His first-order condition for an optimal consumption-portfolio decision " # µ ¶−γ Cit+1 −δ 1 = Et e Rt+1 Cit holds, exactly. To prove this assertion, just substitute in for Cit+1 /Cit and take the expectation: · ¸ 1 2 1 = Et exp −δ − γη it+1 yt+1 + γyt+1 + ln Rt+1 2

Since η is independent of everything else, we can use E [f (ηy)] = E [E(f (ηy|y)] . Now, 431

C HAPTER 21

E QUITY PREMIUM PUZZLE AND CONSUMPTION - BASED MODELS

with η normal (0,1),

Therefore, we have

· ¸ ¤ ¢ ¡ £ 1 2 2 E exp −γηit+1 yt+1 | yt+1 = exp γ yt+1 . 2 · ¸ 1 2 2 1 2 1 = Et exp −δ + γ yt+1 + γyt+1 + ln Rt+1 . 2 2

Substituting in from (21.354), µ · ¸ ¶ 2 1 1 = Et exp −δ + γ(γ + 1) (δ − ln Rt+1 ) + ln Rt+1 2 γ(γ + 1) 1 = Et 1!

The general model

In the general model, Constantinides and Duffie define s r Ct+1 2 yt+1 = ln mt+1 + δ + γ ln γ(γ + 1) Ct

(355)

where mt is a strictly positive discount factor that prices all assets under consideration, pt = Et [mt+1 xt+1 ] for all xt+1 ∈ X.

(356)

By starting with a discount factor that can price a large collection of assets, where I used the −1 to price the single return Rt+1 in (21.354), idiosyncratic risk can be discount factor Rt+1 constructed to price exactly a large collection of assets. We can exactly match the Sharpe ratio, return forecastability, and other features of the data. Then, they let ln

µ

vit+1 vit

¶

1 2 = ηit+1 yt+1 − yt+1 2

Cit+1 = vit+1 Ct+1 . yt+1 is still the conditional standard deviation of consumption growth, given aggregates – returns and aggregate consumption. This variation allows uncertainty in aggregate consumption. We can tailor the idiosyncratic risk to and consumption-interest rate facts as well.

432

S ECTION 21.2

N EW MODELS

Following exactly the same argument as before, we can now show that " # µ ¶−γ Cit+1 −δ 1 = Et e Rt+1 Cit for all the assets priced by m. A technical assumption

Astute readers will notice the possibility that the square root term in (21.354) and (21.355) might be negative. Constantinides and Duffie rule out this possibility by assuming that the discount factor m satisfies ln mt+1 ≥ δ + γ ln

Ct+1 Ct

(357)

in every state of nature, so that the square root term is positive. We ³ construct ´γ i such discount factors by picking parameters a, b in mt+1 = h can sometimes 0 δ Ct+1 max a + b xt+1 , e to satisfy (21.356). However, neither this construction nor a Ct discount factor satisfying (21.357) is guaranteed to exist for any set of assets. The restriction (21.357) is a tighter form of the familiar restriction that mt+1 ≥ 0 that is equivalent to the absence of arbitrage in the assets under consideration. Ledoit and Bernardo (1997) show that the restriction m > a is equivalent to restrictions on the maximum gain/loss ratio available from the set of assets under consideration. Thus, the theorem really does not apply to any set of arbitrage-free payoffs. The example m = 1/R is a positive discount factor that prices a single asset return 1 = E(R−1 R), but does not necessarily satisfy restriction (21.357). For high R, we can have very negative ln 1/R. This example only works if the distribution of R is limited to R ≤ eδ . How the model works

As the Campbell-Cochrane model is blatantly (and proudly) reverse-engineered to surmount (and here, to illustrate) the known pitfalls of representative consumer models, the Constantinides-Duffie model is reverse engineered to surmount the known pitfalls of idiosyncratic risk models. Idiosyncratic risk stories face two severe challenges, as explained in section 1.2. First, the basic pricing equation applies to each individual. If we are to have low risk aversion and power utility, the required huge volatility of consumption is implausible for any individual. Second, if you add idiosyncratic risk uncorrelated with asset returns, it has no effect on pricing implications. Constantinides and Duffie’s central contribution is to very cleverly solve the second problem. In idiosyncratic risk models, we cannot specify individual consumption directly as we do in representative agent endowment economies, and go straight to finding prices. The endowment economy structure says that aggregate consumption is fixed, and prices have to adjust so that consumers are happy consuming the given aggregate consumption stream. 433

C HAPTER 21

E QUITY PREMIUM PUZZLE AND CONSUMPTION - BASED MODELS

However, individuals can always trade consumption with each other. The whole point of assets is that one individual can sell another some consumption, in exchange for the promise of some consumption in return in the next period. We have to give individuals idiosyncratic income shocks, and then either check that they do not want to trade away the idiosyncratic shock, or find the equilibrium consumption after they do so. Early idiosyncratic risk papers found quickly how clever the consumers could be in getting rid of the idiosyncratic risks by trading the existing set of assets. Telmer (1993) and Lucas (1994) found that if you give people transitory but uninsured income shocks, they respond borrowing and lending or by building up a stock of savings. As in the classic permanent income model, consumption only responds by the interest rate times the change in permanent income, and at low enough interest rates, not at all. “Self-insurance through storage” removes the extra income volatility and we are back to smooth individual consumption and an equity premium puzzle. Constantinides and Duffie get around this problem by making the idiosyncratic shocks permanent. The normal ηit shocks determine consumption growth. In an evaluation in microeconomic data, this makes us look for sources of permanent shocks. This, at a deeper level, is why idiosyncratic consumption shocks have to be uncorrelated with the market. We can give individuals idiosyncratic income shocks that are correlated with the market. Say, agent A gets more income when the market is high, and agent B gets more income when it is low. But then A will short the market, B will go long, and they will trade away any component of the shock that is correlated with the returns on available assets. I argued above that this effect made idiosyncratic shocks hopeless as candidates to explain the equity premium puzzle. Shocks uncorrelated with asset returns have no effect on asset pricing, and shocks correlated with asset returns are quickly traded away. The only way out is to exploit the nonlinearity of marginal utility. We can give people income shocks that are uncorrelated with returns, so they can’t be traded away. Then we have a nonlinear marginal utility function turn these shocks into marginal utility shocks that are correlated with asset returns, and hence can affect pricing implications. This is why Constantinides and Duffie specify that the variance of idiosyncratic risk rises when the market declines. If marginal utility were linear, an increase in variance would have no effect on the average level of marginal utility. Therefore, Constantinides and Duffie specify power utility, and the interaction of nonlinear marginal utility and changing conditional variance produces an equity premium. As a simple calculation that shows the basic idea, start with individuals i with power utility so

0=E

"µ

i Ct+1 Cti

¶−γ

434

Ret+1

#

S ECTION 21.2

N EW MODELS

Now aggregate across people by summing over i, with EN = "

0 = E EN

Ãµ

i Ct+1 Cti

¶−γ !

1 N

#

PN

i=1

Ret+1 .

If the cross-sectional variation of consumption growth is lognormally distributed, ·µ ¶ ¸ 2 −γEN ∆cit+1 + γ2 σ 2N ∆cit+1 e 0=E e Rt+1 As you see, the economy displays more risk aversion than would a “representative agent” with aggregate consumption ∆cat+1 = EN ∆cit+1 . That risk aversion can also vary over time if σ N varies over time, and this variation can generate risk premia. Microeconomic evaluation and risk aversion

Like the Campbell-Cochrane model, this could be either a new view of stock market (and macroeconomic) risk, or just a clever existence proof for a heretofore troubling class of models. The first question is whether the microeconomic picture painted by this model is correct, or even plausible. Is idiosyncratic risk large enough? Does idiosyncratic risk really rise when the market falls, and enough to account for the equity premium? Are there enough permanent idiosyncratic shocks? Do people really shy away from stocks because of stock returns are low at times of high labor market risk? This model does not change the first puzzle. To get power utility consumers to shun stocks, they still must have tremendously volatile consumption growth or high risk aversion. The point of this model is to show how consumers can get stuck with high consumption volatility in equilibrium, already a difficult task. More seriously than volatility itself, consumption growth variance also represents the amount by which the distribution of individual consumption and income spreads out over time, since the shocks must be permanent and independent across people. The 50% or larger consumption growth volatility that we require to reconcile the Sharpe ratio with risk aversion of one means that the distribution of consumption (and income) must also spread out by 50% per year. The distribution of consumption does spread out, but not this much. For example, Deaton and Paxson (1994) report that the cross-sectional variance of log consumption within an age cohort rises from about 0.2 at age 20 to 0.6 at age 60. √ This 0.2 = estimate means that the cross sectional standard deviation of consumption rises from √ . 45 or 45% at age 20 to 0.6 = . 77 or 77% at age 60. (77% means that an individual one standard deviation better off than the mean consumes 77% more than the mean consumer.) We are back to about 1% per year. Finally, and most crucially, the cross-sectional uncertainty about individual income must not only be large, it must be higher when the market is lower. This risk-factor is after all the central element of Constantinides and Duffie’s explanation for the market premium. Figure 51 shows how the cross-sectional standard deviation of consumption growth varies with the 435

C HAPTER 21

E QUITY PREMIUM PUZZLE AND CONSUMPTION - BASED MODELS

market return and risk aversion in my simple version of Constantinides and Duffie’s model. If we insist on low (γ = 1 to 2) risk aversion, the cross-sectional standard deviation of consumption growth must be extremely sensitive to the level of the market return. Looking at the γ = 2 line for example, is it plausible that a year with 5% market return would show a 10% cross-sectional variation in consumption growth, while a mild 5% decline in the market is associated with a 25% cross-sectional variation?

Figure 51. Cross-sectional standard deviation of individual consumption growth as a function of the market return in q the simple q version of the Constantinides-Duffie model. 2 t ln R1t + δ + γ ln CCt−1 . Parameter values are The plot is the variable yt = γ(γ+1) ρ = 0.05, ln Ct /Ct−1 = 0.01.

All of these empirical problems are avoided if we allow high risk aversion rather than a large risk to drive the equity premium. The γ = 25 line in Figure 51 looks possible; a γ = 50 line would look even better. With high risk aversion we do not need to specify highly volatile individual consumption growth, spreading out of the income distribution, or dramatic sensitivity of the cross-sectional variance to the market return. As in any model, a high equity premium must come from a large risk, or from large risk aversion. Labor market risk correlated with the stock market does not seem large enough to account for the equity premium without high risk aversion.

436

S ECTION 21.3

B IBLIOGRAPHY

The larger set of asset pricing facts has not yet been studied in this model. It is clearly able to generate return predictability, but that requires a pattern of variation in idiosyncratic risk that remains to be characterized and evaluated. It can generate cross-sectional patterns such as value premia if value stocks decline at times of higher cross-sectional volatility; that too remains to be studied. Summary

In the end, the Constantinides-Duffie model and the Campbell-Cochrane model are quite similar in spirit. First, both models make a similar, fundamental change in the description of stock market risk. Consumers do not fear much the loss of wealth of a bad market return per se. They fear that loss of wealth because it tends to come in recessions, in one case defined as times of heightened labor market risk, and in the other case defined as a fall of consumption relative to its recent past. This recession state-variable or risk-factor drives most variation in expected returns. Second, both models require high risk aversion. While Constantinides and Duffie’s proof shows that one can dream up a labor income process to rationalize the equity premium for any risk aversion coefficient, we see that even vaguely plausible characterizations of actual labor income uncertainty require high risk aversion to explain the historical equity premium. Third, both models provide long-sought demonstrations that it is possible to rationalize the equity premium in their respective class of models. This existence proof is particularly stunning in Constantinides and Duffie’s case. Many authors (myself included) had come to the conclusion that the effort to generate an equity premium from idiosyncratic risk was hopeless because any idiosyncratic risk that would affect asset prices would be traded away.

21.3

Bibliography

Shiller (1982) made the first calculation that showed either a large risk aversion coefficient or counterfactually large consumption variability was required to explain means and variances of asset returns. Mehra and Prescott (1985) labeled this fact the “equity premium puzzle.” However, they described these puzzles in the context of a two-state Markov model for consumption growth, identifying a stock as a claim to consumption and a risk free bond. Weil (1989) emphasized the interaction between equity premium and risk-free rate puzzles. Hansen and Jagannathan (1991) sparked the kind of calculations I report here in a simplified manner. Cochrane and Hansen (1992) derived many of the extra discount factor moment restrictions I surveyed here, calculating bounds in each case. Luttmer (1996), (1999) tackled the important extension to transactions costs. Kocherlakota (1996) is a nice summary of equity premium facts and models. Much of the material in this Chapter is adapted from a survey in Cochrane (1997). Campbell (1999) and (2000) are two excellent recent surveys. Ferson (1995) is a nice survey of consumption-based model variations as well as some of the beta pricing models discussed in the last chapter. The general picture of all solutions based on changing preferences is that they introduce 437

C HAPTER 21

E QUITY PREMIUM PUZZLE AND CONSUMPTION - BASED MODELS

non-separabilities. If the marginal utility of consumption depends on z as well as c, uc (c, z), then expected returns depend on covariances with z as well. In turn, this happens if we cannot write the utility function as u(c, z) = v(c) + w(z), the separable form. Habit persistence introduces a non-time-separable utility function, since P u(c, x) and x depends on past c–you can’t write a utility function with habits as a sum t vt (ct ), so one period’s consumption affects another period’s marginal utility. The Campbell-Cochrane model I presented here is a tip of an iceberg of habit research, including prominent contributions by Constantinides (1990), Ferson and Constantinides (1991), Heaton (1995), Abel (1990). Models can be nonseparable across goods as well. Leisure is the most natural extra variable to add to a utility function. It’s not clear a priori whether more leisure enhances the marginal utility of consumption (why bother buying a boat if you’re at the office all day and can’t use it) or vice versa (if you have to work all day, it’s more important to come home to a really nice big TV). However, we can let the data speak on this matter. Explicit versions of this approach have not been very successful to date. (Eichenbaum, Hansen and Singleton 1989). On the other hand, recent research has found that adding labor income as an extra adhoc “factor” can be useful in explaining the cross section of average stock returns, especially if it is scaled by a conditioning variable (Jagannathan and Wang 1996, Reyfman 1997, Lettau and Ludvigson 2000). The non-state separable utility functions following Epstein and Zin (1989) are a major omission of this presentation. The expectation E in the standard utility function sums over states of nature, e.g. U = prob(rain) × u(C if it rains) + prob(shine) × u(C if it shines).

“Separability” means one adds across states, so the marginal utility of consumption in one state is unaffected by what happens in another state. But perhaps the marginal utility of a little more consumption in the sunny state of the world is affected by the level of consumption in the rainy state of the world. Epstein and Zin and Hansen, Sargent and Tallarini (1997) propose recursive utility functions of the form £ ¤ Ut = Ct1−γ + βf Et f −1 (Ut+1 ) .

If f (x) = x this expression reduces to power utility. These utility functions are not stateseparable. As with habits, these utility functions distinguish risk aversion from intertemporal substitution–one coefficient can be set to capture the consumption-interest rate facts, and a completely separate coefficient can be set to capture the equity premium. So far, this style of model as in Epstein and Zin (1989), Weil (1989), Kandel and Stambaugh (1991), and Campbell (1996) does not generate time-varying risk aversion, but that modification should not be too difficult, and could lead to a model that works very much like the habit model I surveyed here. Habit persistence is the opposite of durability. If you buy a durable good yesterday, that lowers your marginal utility of an additional purchase today, while buying a habit-forming good raises your marginal utility of an additional purchase today. Thus the durability of goods 438

S ECTION 21.3

B IBLIOGRAPHY

should introduce a non-time-separability of the form u(ct + θxt ), xt = f(ct−1 , ct−2 , ...) rather than the habit persistence form u(ct − θxt ). Since goods are durable, and we have a lot of data on durables purchases, it would be good to include both durability and habit persistence in our models. (In fact, even “nondurables” contain items like clothing; the truly nondurable purchases are such a small fraction of total consumption that we rely on very little data.) One must be careful with the time horizon in such a specification. At a sufficiently small time horizon, all goods are durable. A pizza eaten at noon lowers marginal utility of more pizza at 12:05. Thus, our common continuous time, time separable assumption really cannot be taken literally. Hindy and Huang (1992) argue that consumption should be “locally substitutable” in continuous time models. Heaton (1993) found that at monthly horizons, consumption growth displays the negative autocorrelation suggestive of durability with constant interest rates, while at longer horizons consumption is nearly unforecastable after accounting for time-aggregation. There is also a production first order condition that must be solved, relating asset prices to marginal rates of transformation. The standard here is the q theory of investment, which is based on an adjustment cost. If the stock market is really high, you issue stock and make new investments. The trouble with this view is that f 0 (K) declines very slowly, so the observed price volatility implies huge investment volatility. The q theory adds adjustment costs to damp the investment volatility. The q theory has had as much trouble fitting the data as the consumption-based model. Cochrane (1991d) reports one success when you transform the data to returns – high stock returns are associated with high investment growth. The more recent investment literature has focused on specifying the adjustment cost problem with asymmetries and irreversibilities, for example Abel and Eberly (1996). There is an important literature that puts new utility functions together with production functions, to construct complete explicit economic models that replicate the asset pricing facts. Such efforts should also at least preserve if not enhance our ability to understand the broad range of dynamic microeconomic, macroeconomic, international and growth facts that the standard models were constructed around. Jermann (1998) tried putting habit persistence consumers in a model with a standard technology Y = θf (K, L) from real business cycle models. The easy opportunities for intertemporal transformation provided by that technology meant that the consumers used it to dramatically smooth consumption, destroying the prediction of a high equity premium. To generate the equity premium, Jermann added an adjustment cost technology, as the production-side literature had found necessary. This modification resulted in a high equity premium, but also large variation in riskfree rates. Boldrin, Christiano and Fisher (1997) also added habit-persistence preferences to real business cycle models with frictions in the allocation of resources to two sectors. They generate about 1/2 the historical Sharpe ratio. They find some quantity dynamics are improved over the standard model. However, they still predict highly volatile interest rates and persistent consumption growth. To avoid the implications of highly volatile interest rates, I suspect we will need representations of technology that allow easy transformation across time but not across states of 439

C HAPTER 21

E QUITY PREMIUM PUZZLE AND CONSUMPTION - BASED MODELS

nature, analogous to the need for easy intertemporal substitution but high risk aversion in preferences. Alternatively, the Campbell-Cochrane model above already produces the equity premium with constant interest rates, which can be interpreted as a linear production function f(K). Models with this kind of precautionary savings motive may not be as severely affected. Tallarini (1999) uses non-state separable preferences similar to those of Epstein and Zin in a general equilibrium model with production. He shows a beautiful observational equivalence result: A model with standard preferences and a model with non-state-separable preferences can predict the same path of quantity variables (output, investment, consumption, etc.) but differ dramatically on asset prices. This result offers one explanation of how the real business cycle and growth literature could go on for 25 years examining quantity data in detail and miss all the modifications to preferences that we seem to need to explain asset pricing data. It also means that asset price information is crucial to identifying preferences and calculating welfare costs of policy experiments. Finally, it offers hope that adding the deep modifications necessary to explain asset pricing phenomena will not demolish the success of standard models at describing the movements of quantities. The Constantinides and Duffie model has roots in a calculation by Mankiw (1986) that idiosyncratic risk could make the representative consumer seem more risk averse than the individuals. Work on evaluating the mechanisms in this model in microeconomic data is starting. Heaton and Lucas (1996) calibrate idiosyncratic risk from the PSID, but their model explains at best 1/2 of the sample average stock return, and less still if they allow a net supply of bonds with which people can smooth transitory shocks. More direct tests of these features in microeconomic consumption data are underway, for example Brav, Constantinides and Geczy (1999), Storesletten, Telmer and Yaron (1999). Kandel and Stambaugh (1986) present a model in which a small amount of time-varying consumption volatility and a high risk aversion coefficient generate the large time-varying discount factor volatility we need to generate returns predictability. Aiyagari and Gertler (1991), though aimed at the point that the equity premium might be explained by a “too low” riskless rate, nonetheless was an important paper in specifying and solving models with uninsured individual risks and transactions costs to keep people from trading them away.

21.4 1.

2.

Problems

Derive the analogue to the Hansen-Jagannathan bound in continuous time for an “excess return,” i.e. considering a self-financing portfolio, rather than a single return less the risk free rate. Suppose habit accumulation is linear, and there is a constant riskfree rate or linear

440

S ECTION 21.4

P ROBLEMS

technology equal to the discount rate, Rf = 1/δ. The consumer’s problem is then max

∞ X t=0

3.

δt

(Ct − Xt )1−γ 1−γ

s.t.

X t

δ t Ct =

X

δ t et + W0 ; Xt = θ

t

∞ X

φj Ct−j

j=1

where et is a stochastic endowment. In an internal habit specification, the consumer considers all the effects that current consumption has on future utility through Xt+j . In an external habit specification, the consumer ignores such terms. Show that the two specifications give identical asset pricing predictions in this simple model, by showing that internal-habit marginal utility is proportional to external-habit marginal utility, state by state. Suppose a consumer has quadratic utility with a constant interest rate equal to the subjective discount rate, but a habit or durable consumption good, so that utility is 1 u(ct − θct−1 ) = − (c∗ − ct + θct−1 ). 2

4.

Show that external habit persistence θ > 0 implies positive serial correlation in consumption changes. Show that the same solution holds for internal habits, or durability. Show that durability leads to negative serial correlation in consumption changes. Many models predict too much variation in the conditional mean discount factor, or too much interest rate variation. This problem guides you through a simple example. Introduce a simple form of external habit formation, u = (Ct − θCt−1 )1−γ

5.

and suppose consumption growth Ct+1 /Ct is i.i.d. Show that interest rates still vary despite i.i.d. consumption growth. We showed that if m satisfies the Hansen-Jagannathan bound, then proj(m|X) should also do so. Hansen and Jagannathan also compute bounds with positivity, solutions to min σ(m) s.t. p = E(mx), m ≥ 0, E(m) = µ.

6.

Does proj(m|X) also lie in the same bound? One most often compares consumption-based models to Hansen-Jagannathan bounds. Can you compare the CAPM discount factor m = a − bRem to the bound? To the bound with positivity?

441

Chapter 22.

References

Abel, Andrew B. 1988, “Stock Prices under Time-Varying Dividend Risk: An Exact Solution in an Infinite-Horizon General Equilibrium Model,” Journal of Monetary Economics 22, 375-93. Abel, Andrew B., 1990, “Asset Prices Under Habit Formation and Catching Up With the Jones,” American Economic Review 80, 38-42. Abel, Andrew B., 1994, “Exact Solutions for Expected Rates of Return under Markov Regime Switching: Implications for the Equity Premium Puzzle,” Journal of Money, Credit, and Banking 26, 345-61. Abel, Andrew B., 1999, “Risk Premia and Term Premia in General Equilibrium,” Journal of Monetary Economics 43, 3-33. Abel, Andrew B. and Janice C. Eberly, 1996, “Optimal Investment with Costly Reversibility.” Review of Economic Studies 63, 581-593. Abel, Andrew B. and Janice C. Eberly, 1999, “The Effects of Irreversibility and Uncertainty on Capital Accumulation,” Journal of Monetary Economics 44, 339-77. Aiyagari, S. Rao and Mark Gertler, 1991, “Asset Returns with Transactions Costs and Uninsured Individual Risk: A Stage III exercise,” Journal of Monetary Economics 27, 309-331. Andrews, Donald W. K., 1991, “Heteroskedasticity and Autocorrelation Consistent Covariance Matrix Estimation,” Econometrica 59, 817-58. Atkeson, Andrew, Fernando Alvarez, and Patrick Kehoe, 1999, “Volatile Exchange Rates and the Forward Premium Anomaly: A Segmented Asset Market View,” Working paper, University of Chicago. Bachelier, L. 1900, “Theory of Speculation,” in Cootner, P. (ed), The Random Character of Stock Prices, Cambridge, MA: MIT press 1964. Balduzzi, Pierluigi., Giuseppe Bertola and Silverio Foresi, 1996, “A Model of Target Changes and the Term Structure of Interest Rates,” Journal of Monetary Economics 39, 223, 49. Banz, Rolf W. 1981, “The Relationship Between Return and Market Value of Common Stocks,” Journal of Financial Economics, 9, 3-18. Barsky, Robert and Bradford J. DeLong 1993, “Why Does the Stock Market Fluctuate?” Quarterly Journal of Economics 108, 291-311. Bekaert, Geert and Robert J. Hodrick, 1992, “Characterizing Predictable Components in Excess Returns on Equity and Foreign Exchange Markets,”Journal of Finance 47, 467-509. Becker, Connie, Wayne E. Ferson, Michael Schill, and David Myers 1999, “Conditional Market Timing with Benchmark Investors,” Journal of Financial Economics 52, 119-148. Berk, Jonathan, 1997, “Does Size Really Matter?” Financial Analysts Journal, September/October 1997, 12-18. 442

Bernardo, Antonio and Olivier Ledoit. “Gain Loss and Asset Pricing.” (1999) Journal of Political Economy 108, 144-172. Black, Fischer, Michael Jensen and Myron Scholes, 1972, “The Capital Asset Pricing Model: Some empirical Tests,” in Michael Jensen, Ed., Studies in the Theory of Capital Markets (Praeger, New York NY) Black, Fischer and Myron Scholes, 1973, “The Valuation of Options and Corporate Liabilities.” Journal of Political Economy 81, 637-654. Bollerslev, Tim, R. Chou and K. Kroner 1992, “ARCH Modeling in Finance: A Review of Theory and empirical Evidence,” Journal of Econometrics,52, 5-59. Boudoukh, Jacob, Matthew Richardson 1994 “The Statistics of Long-Horizon Regressions Revisited,” Mathematical Finance 4, 103-119. Boudoukh, Jacob, Matthew Richardson, Robert Stanton and Robert Whitelaw, 1998, “The Stochastic Behavior of Interest Rates: Implications from a Nonlinear, Continuous-time, Multifactor Model,” Manuscript, University of California at Berkeley. Brav, Alon, George Constantinides and Christopher Geczy, 1999, “Asset Pricing with Heterogeneous Consumers and Limited Participation: empirical Evidence,” Manuscript Duke University. Breeden, Douglas T., 1979, “An Intertemporal Asset Pricing Model with Stochastic Consumption and Investment Opportunities,” Journal of Financial Economics 7, 265-296. Breeden, Douglas T., Michael R. Gibbons and Robert H. Litzenberger, 1989, “empirical Tests of the Consumption-Oriented CAPM,” Journal of Finance 44, 231-262. Brown, Stephen, William Goetzmann and Stephen A. Ross, 1995, “Survival” Journal of Finance 50, 853-873. Buraschi, Andrea and Alexei Jiltsov, 1999, “How Large is the Inflation Risk Premium in the U.S. Nominal Term Structure,” Manuscript London Business School. Burnside, Craig, Martin Eichenbaum, and Sergio Rebelo, 1993, “Labor Hoarding and the Business Cycle,” Journal of Political Economy 101, 245-73. Campbell, John Y. 1991, “A Variance Decomposition for Stock Returns,” Economic Journal 101, 157-179. Campbell, John Y. 1995, “Some Lessons from the Yield Curve,” Journal of Economic Perspectives, 9, 129-152. Campbell, John Y. 1996. “Understanding Risk and Return.” Journal of Political Economy 104, 298-345. Campbell, John Y. 1999. “Asset Prices, Consumption, and the Business Cycle.” in John B. Taylor and Michael Woodford eds. Handbook of Macroeconomics. Amsterdam: North-Holland. Campbell, John Y., 2000, “Asset Pricing at the Millennium,” Journal of Finance August Campbell, John Y., and John H. Cochrane, 1999, “By Force of Habit: A Consumption-Based Explanation of Aggregate Stock Market Behavior” Journal of 443

CHAPTER 22

REFERENCES

Political Economy, 107, 205-251. Campbell, John Y., Andrew W. Lo and A. Craig MacKinlay 1997, The Econometrics of Financial Markets Princeton NJ: Princeton University Press. Campbell, John Y. and Robert J. Shiller. 1988a. “The Dividend-Price Ratio and Expectations of Future Dividends and Discount Factors,” Review of Financial Studies 1, 195-227. Campbell, John Y. and Robert J. Shiller. 1988b. “Stock Prices, Earnings, and Expected Dividends.” Journal of Finance 43: 661-676. Campbell, John Y. and Robert J. Shiller 1991, “Yield Spreads and Interest Rates: A Bird’s Eye View” Review of Economic Studies 58, 495-514. Carhart, Mark M. 1997, “On Persistence in Mutual Fund Performance,” Journal of Finance 52, 57-82. Chamberlain, Gary and Michael Rothschild, 1983, “Arbitrage, Factor Structure, and Mean-Variance Analysis on Large Asset Markets,” Econometrica 51, 1281-1304 Chen, Nai-Fu, Richard Roll and Steven Stephen A. Ross, 1986, “Economic Forces and the Stock Market,” Journal of Business 59, 383-403. Christiano, Lawrence, Martin Eichenbaum, and Charles Evans, 1999, “Monetary Policy Shocks: What Have we Learned and to What End?” Forthcoming in John Taylor, ed., Handbook of Monetary Economics Cochrane, John H 1988.,“How Big is the Random Walk in GNP?”, Journal of Political Economy 96, 893-920. Cochrane, John H., 1991a, ”Explaining the Variance of Price-Dividend Ratios” 5, 243-280. Cochrane, John H., 1991b, “A Simple Test of Consumption Insurance,” Journal of Political Economy 99, 957-976. Cochrane, John H., 1991c “Volatility Tests and Efficient Markets: A Review Essay” Journal of Monetary Economics 27, 463-485. Cochrane, John H., 1991d “Production-Based Asset Pricing and the Link Between Stock Returns and Economic Fluctuations,” Journal of Finance 46, 207-234. Cochrane, John H., 1994, “Permanent and Transitory Components of GNP and Stock Prices” Quarterly Journal of Economics 109, 241-266. Cochrane John H. 1994, “Shocks,” Carnegie-Rochester Conference Series on Public Policy 41, 295-364. Cochrane, John H., 1996, “A Cross-Sectional Test of an Investment-Based Asset Pricing Model,” Journal of Political Economy 104, 572-621. Cochrane, John H., 1997, “Where is the Market Going? Uncertain Facts and Novel Theories,” Economic Perspectives 21: 6 (November/December 1997) Federal Reserve Bank of Chicago. Cochrane, John H., 1999a, “New Facts in Finance,” Economic Perspectives Federal Reserve Bank of Chicago 23 (3) 36-58. Cochrane, John H., 1999b, “Portfolio Advice for a Multifactor World” Economic

444

Perspectives Federal Reserve Bank of Chicago 23 (3) 59-78. Cochrane, John H., 2000, “A Resurrection of the Stochastic Discount Factor/GMM Methodology,” Manuscript, University of Chicago. Cochrane, John H. and Lars Peter Hansen, 1992, “Asset Pricing Explorations for Macroeconomics” In Olivier Blanchard and Stanley Fisher, Eds.,1992 NBER Macroeconomics Annual 115-165. Cochrane, John H. and Jesús Saá-Requejo 2000, “Beyond Arbitrage: Good Deal Asset Price Bounds in Incomplete Markets” Journal of Political Economy 108, 79-119. Cochrane, John H. and Argia M. Sbordone, 1988, “Multivariate Estimates of the Permanent Components in GNP and Stock Prices” Journal of Economic Dynamics and Control, 12, 255-296. Constantinides, George M., 1989, “Theory of Valuation: Overview and Recent Developments,” in Sudipto Bhattacharya and George M. Constantinides, eds., Theory of Valuation Totwa NJ: Rowman & Littlefield Constantinides, George M. 1990. “Habit Formation: A Resolution of the Equity Premium Puzzle.” Journal of Political Economy 98: 519-543. Constantinides, George M., 1992, “A Theory of the Nominal Term Structure of Interest Rates,” Review of Financial Studies 5, 531-52. Constantinides, George M. 1998, “Transactions Costs and the Volatility Implied by Option Prices.” Manuscript, Graduate School of Business, University of Chicago. Constantinides, George M. and Darrell Duffie. 1996, “Asset Pricing with Heterogeneous Consumers.” Journal of Political Economy 104, 219–240.. Constantinides, George M. and Thaleia Zariphopoulou, 1997, “Bounds on Option Prices in an Intertemporal Setting with Proportional Transaction Costs and Multiple Securities,” Manuscript, Graduate School of Business, University of Chicago. Cox, John C. and Chi-fu Huang, 1989, “Optimal Consumption and Portfolio Policies when Asset Prices Follow a Diffusion Process,” Journal of Economic Theory 39, 33-83. Cox, John C., Jonathan E. Ingersoll, and Stephen A. Ross. 1985. “A Theory of the Term Structure of Interest Rates.” Econometrica 53: 385-408. Cox, John C., Stephen A. Ross, and Mark Rubinstein, 1979, “Option Pricing: A Simplified Approach,” Journal of Financial Economics 7, 229-63. Cox, John C., and Mark Rubinstein, 1985, Options Markets Englewood Cliffs, NJ: Prentice Hall. Dai, Qiang and Kenneth J. Singleton, 1999, “Specification Analysis of Affine Term Structure Models,” forthcoming, Journal of Finance Daniel, Kent, David Hirshleifer and Avanidhar Subrahmanyam, 1998, “Investor Psychology and Security Market Under- and Overreactions,” Journal of Finance 53, 1839-1885. Debreu, Gerard 1959, The Theory of Value New York: Wiley and Sons. DeBondt, Werner F.M. and Thaler, Richard H., 1985, “Does the Stock Market

445

CHAPTER 22

REFERENCES

Overreact?” Journal of Finance 40, 793-805. Dixit, Avinash and R. Pindyck. Investment Under Uncertainty. Princeton NJ: Princeton University Press, 1994. Duarte, Jefferson 2000 “The Relevance of the Price of Risk in Affine Term-Structure models” Manuscript, University of Chicago Duffee, Gregory, 1999, “Forecasting future interest rates: Are affine models failures?” Manuscript, University of California at Berkeley. Duffie, J. Darrel and Rui Kan, 1996, “A Yield Factor Model of the Term structure of interest rates” Mathematical Finance 6, 379-406 Dybvig, Philip H. and Jonathan E. Ingersoll Jr., 1982, “Mean-Variance Theory in Complete Markets,” Journal of Business 55, 233-51. Dybvig, P., and Stephen Ross, 1985, “Yes, the APT is testable,” Journal of Finance 40, 1173-1188. Dybvig, P., J. Ingersoll Jr. and Stephen Ross 1996, “Long Forward and Zero-Coupon Rates Can Never Fall,” Journal of Business, 69, 1-25. Eichenbaum, Martin, Lars Peter Hansen and Kenneth Singleton, 1988, “A Time-Series Analysis of Representative Agent Models of Consumption and Leisure Choice under Uncertainty,” Quarterly Journal of Economics 103, 51-78. Engel, Charles, 1996, “The Forward Discount Anomaly and the Risk Premium: a Survey of Recent Evidence,” Journal of empirical Finance 3, 123-192. Engle, Robert F. and Clive W. J. Granger, 1987, “Cointegration and Error Correction: Representation, Estimation, and Testing,” Econometrica 55, 251-276. Epstein, Larry G. and Stanley E. Zin. 1989. “Substitution, Risk Aversion and the Temporal Behavior of Asset Returns.” Journal of Political Economy 99: 263-286. Fama, Eugene F., 1965, “The Behavior of Stock Market Prices, Journal of Business 38, 34-105. Fama, Eugene F., 1970, “Efficient Capital Markets: A review of Theory and empirical Work,” Journal of Finance 25, 383-417. Fama, Eugene F., 1984, “Forward and Spot Exchange Rates,” Journal of Monetary Economics 14, 319-338. Fama, Eugene F. 1991, “Efficient Markets II,” Journal of Finance 46, 1575-1618. Fama, Eugene F. and Robert R. Bliss, 1987, “The information in Long-Maturity Forward Rates,” American Economic Review, 77, 680-92. Fama, Eugene F. and Kenneth R. French. 1988a. “Permanent and Temporary Components of Stock Prices.” Journal of Political Economy 96: 246-273. Fama, Eugene F. and Kenneth R. French. 1988b. “Dividend Yields and Expected Stock Returns.” Journal of Financial Economics 22: 3-27. Fama, Eugene F. and Kenneth R. French, 1989, “Business Conditions and Expected Returns on Stocks and Bonds,” Journal of Financial Economics 25, 23-49. Fama, Eugene F. and Kenneth R. French, 1993, “Common Risk Factors in the Returns on Stocks and Bonds,” Journal of Financial Economics 33, 3-56. 446

Fama, Eugene F., and Kenneth R. French, 1996, “Multifactor Explanations of Asset-Pricing Anomalies,” Journal of Finance 47, 426-465. Fama, Eugene F., and Kenneth R. French, 1997a, “Size and Book-to-Market Factors in Earnings and Returns,”Journal of Finance 50, 131-55. Fama, Eugene F., and Kenneth R. French, 1997b, “Industry Costs of Equity,” Journal of Financial Economics 43, 153-193. Fama, Eugene F., and Kenneth R. French, 2000, “The Equity Premium,” Working paper, University of Chicago. Fama, Eugene F., and James D. MacBeth, 1973, “Risk Return and Equilibrium: empirical Tests,” Journal of Financial Political Economy 71, 607-636. Ferson, Wayne E., 1995, “Theory and Empirical Testing of Asset Pricing Models,” in R. A. Jarrow, V. Maksimovic, and W. T. Ziemba, eds., Handbooks in OR & MS, Volume 9, Finance Amsterdam: Elsevier Science B.V. Ferson, Wayne E. and George Constantinides. 1991. “Habit Persistence and Durability in Aggregate Consumption: empirical Tests.” Journal of

John H. Cochrane

June 12, 2000

1

Acknowledgments This book owes an enormous intellectual debt to Lars Hansen and Gene Fama. Most of the ideas in the book developed from long discussions with each of them, and trying to make sense of what each was saying in the language of the other. I am also grateful to all my colleagues in Finance and Economics at the University of Chicago, and to George Constantinides especially, for many discussions about the ideas in this book. I thank George Constantinides, Andrea Eisfeldt, Gene Fama, Wayne Ferson, Owen Lamont, Anthony Lynch, Dan Nelson, Alberto Pozzolo, Michael Roberts, Juha Seppala, Mike Stutzer, Pietro Veronesi, an anonymous reviewer, and several generations of Ph.D. students at the University of Chicago for many useful comments. I thank the NSF and the Graduate School of Business for research support. Additional material and both substantive and typographical corrections will be maintained at http://www-gsb.uchicago.edu/fac/john.cochrane/research/papers c John H. Comments and suggestions are most welcome This book draft is copyright ° Cochrane 1997, 1998, 1999, 2000 John H. Cochrane Graduate School of Business University of Chicago 1101 E. 58th St. Chicago IL 60637 773 702 3059 [email protected] June 12, 2000

2

Contents Acknowledgments

2

Preface

8

Part I.

Asset pricing theory

12

1 Consumption-based model and overview

13

1.1

Basic pricing equation

14

1.2

Marginal rate of substitution/stochastic discount factor

16

1.3

Prices, payoffs and notation

17

1.4

Classic issues in finance

20

1.5

Discount factors in continuous time

33

1.6

Problems

38

2 Applying the basic model

41

2.1

Assumptions and applicability

41

2.2

General Equilibrium

43

2.3

Consumption-based model in practice

47

2.4

Alternative asset pricing models: Overview

49

2.5

Problems

51

3 Contingent Claims Markets

54

3.1

Contingent claims

54

3.2

Risk neutral probabilities

55

3.3

Investors again

57

3.4

Risk sharing

59

3.5

State diagram and price function

60

4 The discount factor

64

4.1

Law of one price and existence of a discount factor

64

4.2

No-Arbitrage and positive discount factors

69

3

4.3

An alternative formula, and x∗ in continuous time

74

4.4

Problems

76

5 Mean-variance frontier and beta representations

77

5.1

Expected return - Beta representations

77

5.2

Mean-variance frontier: Intuition and Lagrangian characterization

80

5.3

An orthogonal characterization of the mean-variance frontier

83

5.4

Spanning the mean-variance frontier

88

5.5

A compilation of properties of R , R

5.6

Mean-variance frontiers for m: the Hansen-Jagannathan bounds

92

5.7

Problems

97

∗

e∗

and x

∗

6 Relation between discount factors, betas, and mean-variance frontiers

89

98

6.1

From discount factors to beta representations

98

6.2

From mean-variance frontier to a discount factor and beta representation

101

6.3

Factor models and discount factors

104

6.4

Discount factors and beta models to mean - variance frontier

108

6.5

Three riskfree rate analogues

109

6.6

Mean-variance special cases with no riskfree rate

115

6.7

Problems

118

7 Implications of existence and equivalence theorems

120

8 Conditioning information

128

8.1

Scaled payoffs

129

8.2

Sufficiency of adding scaled returns

131

8.3

Conditional and unconditional models

133

8.4

Scaled factors: a partial solution

140

8.5

Summary

141

8.6

Problems

142

9 Factor pricing models 9.1

143

Capital Asset Pricing Model (CAPM) 4

145

9.2

Intertemporal Capital Asset Pricing Model (ICAPM)

156

9.3

Comments on the CAPM and ICAPM

158

9.4

Arbitrage Pricing Theory (APT)

162

9.5

APT vs. ICAPM

171

9.6

Problems

172

Part II.

Estimating and evaluating asset pricing models

10 GMM in explicit discount factor models

174 177

10.1 The Recipe

177

10.2 Interpreting the GMM procedure

180

10.3 Applying GMM

184

11 GMM: general formulas and applications

188

11.1 General GMM formulas

188

11.2 Testing moments

192

11.3 Standard errors of anything by delta method

193

11.4 Using GMM for regressions

194

11.5 Prespecified weighting matrices and moment conditions

196

11.6 Estimating on one group of moments, testing on another.

205

11.7 Estimating the spectral density matrix

205

11.8 Problems

212

12 Regression-based tests of linear factor models

214

12.1 Time-series regressions

214

12.2 Cross-sectional regressions

219

12.3 Fama-MacBeth Procedure

228

12.4 Problems

234

13 GMM for linear factor models in discount factor form

235

13.1 GMM on the pricing errors gives a cross-sectional regression

235

13.2 The case of excess returns

237

13.3 Horse Races

239 5

13.4 Testing for characteristics

240

13.5 Testing for priced factors: lambdas or b’s?

241

13.6 Problems

245

14 Maximum likelihood

247

14.1 Maximum likelihood

247

14.2 ML is GMM on the scores

249

14.3 When factors are returns, ML prescribes a time-series regression

251

14.4 When factors are not excess returns, ML prescribes a cross-sectional regression

255

14.5 Problems

256

15 Time series, cross-section, and GMM/DF tests of linear factor models

258

15.1 Three approaches to the CAPM in size portfolios

259

15.2 Monte Carlo and Bootstrap

265

16 Which method?

271

Part III.

284

Bonds and options

17 Option pricing

286

17.1 Background

286

17.2 Black-Scholes formula

293

17.3 Problems

299

18 Option pricing without perfect replication

300

18.1 On the edges of arbitrage

300

18.2 One-period good deal bounds

301

18.3 Multiple periods and continuous time

309

18.4 Extensions, other approaches, and bibliography

317

18.5 Problems

319

19 Term structure of interest rates

320

19.1 Definitions and notation

320 6

19.2 Yield curve and expectations hypothesis

325

19.3 Term structure models – a discrete-time introduction

327

19.4 Continuous time term structure models

332

19.5 Three linear term structure models

337

19.6 Bibliography and comments

348

19.7 Problems

351

Part IV.

Empirical survey

352

20 Expected returns in the time-series and cross-section

354

20.1 Time-series predictability

356

20.2 The Cross-section: CAPM and Multifactor Models

396

20.3 Summary and interpretation

409

20.4 Problems

413

21 Equity premium puzzle and consumption-based models

414

21.1 Equity premium puzzles

414

21.2 New models

423

21.3 Bibliography

437

21.4 Problems

440

22 References

Part V.

442

Appendix

455

23 Continuous time

456

23.1 Brownian Motion

456

23.2 Diffusion model

457

23.3 Ito’s lemma

460

23.4 Problems

462

7

Preface Asset pricing theory tries to understand the prices or values of claims to uncertain payments. A low price implies a high rate of return, so one can also think of the theory as explaining why some assets pay higher average returns than others. To value an asset, we have to account for the delay and for the risk of its payments. The effects of time are not too difficult to work out. However, corrections for risk are much more important determinants of an many assets’ values. For example, over the last 50 years U.S. stocks have given a real return of about 9% on average. Of this, only about 1% is due to interest rates; the remaining 8% is a premium earned for holding risk. Uncertainty, or corrections for risk make asset pricing interesting and challenging. Asset pricing theory shares the positive vs. normative tension present in the rest of economics. Does it describe the way the world does work or the way the world should work? We observe the prices or returns of many assets. We can use the theory positively, to try to understand why prices or returns are what they are. If the world does not obey a model’s predictions, we can decide that the model needs improvement. However, we can also decide that the world is wrong, that some assets are “mis-priced” and present trading opportunities for the shrewd investor. This latter use of asset pricing theory accounts for much of its popularity and practical application. Also, and perhaps most importantly, the prices of many assets or claims to uncertain cash flows are not observed, such as potential public or private investment projects, new financial securities, buyout prospects, and complex derivatives. We can apply the theory to establish what the prices of these claims should be as well; the answers are important guides to public and private decisions. Asset pricing theory all stems from one simple concept, derived in the first page of the first Chapter of this book: price equals expected discounted payoff. The rest is elaboration, special cases, and a closet full of tricks that make the central equation useful for one or another application. There are two polar approaches to this elaboration. I will call them absolute pricing and relative pricing. In absolute pricing, we price each asset by reference to its exposure to fundamental sources of macroeconomic risk. The consumption-based and general equilibrium models described below are the purest examples of this approach. The absolute approach is most common in academic settings, in which we use asset pricing theory positively to give an economic explanation for why prices are what they are, or in order to predict how prices might change if policy or economic structure changed. In relative pricing, we ask a less ambitious question. We ask what we can learn about an asset’s value given the prices of some other assets. We do not ask where the price of the other set of assets came from, and we use as little information about fundamental risk factors as possible. Black-Scholes option pricing is the classic example of this approach. While limited in scope, this approach offers precision in many applications. 8

Asset pricing problems are solved by judiciously choosing how much absolute and how much relative pricing one will do, depending on the assets in question and the purpose of the calculation. Almost no problems are solved by the pure extremes. For example, the CAPM and its successor factor models are paradigms of the absolute approach. Yet in applications, they price assets “relative” to the market or other risk factors, without answering what determines the market or factor risk premia and betas. The latter are treated as free parameters. On the other end of the spectrum, most practical financial engineering questions involve assumptions beyond pure lack of arbitrage, assumptions about equilibrium “market prices of risk.” The central and unfinished task of absolute asset pricing is to understand and measure the sources of aggregate or macroeconomic risk that drive asset prices. Of course, this is also the central question of macroeconomics, and this is a particularly exciting time for researchers who want to answer these fundamental questions in macroeconomics and finance. A lot of empirical work has documented tantalizing stylized facts and links between macroeconomics and finance. For example, expected returns vary across time and across assets in ways that are linked to macroeconomic variables, or variables that also forecast macroeconomic events; a wide class of models suggests that a “recession” or “financial distress” factor lies behind many asset prices. Yet theory lags behind; we do not yet have a well-described model that explains these interesting correlations. In turn, I think that what we are learning about finance must feed back on macroeconomics. To take a simple example, we have learned that the risk premium on stocks – the expected stock return less interest rates – is much larger than the interest rate, and varies a good deal more than interest rates. This means that attempts to line investment up with interest rates are pretty hopeless – most variation in the cost of capital comes from the varying risk premium. Similarly, we have learned that some measure of risk aversion must be quite high, or people would all borrow like crazy to buy stocks. Most macroeconomics pursues small deviations about perfect foresight equilibria, but the large equity premium means that volatility is a first-order effect, not a second-order effect. Standard macroeconomic models predict that people really don’t care much about business cycles (Lucas 1987). Asset prices are beginning to reveal that they do – that they forego substantial return premia to avoid assets that fall in recessions. This fact ought to tell us something about recessions!

This book advocates a discount factor / generalized method of moments view of asset pricing theory and associated empirical procedures. I summarize asset pricing by two equations: pt = E(mt+1 xt+1 ) mt+1 = f(data, parameters). where pt = asset price, xt+1 = asset payoff, mt+1 = stochastic discount factor. 9

The major advantage of the discount factor / moment condition approach are its simplicity and universality. Where once there were three apparently different theories for stocks, bonds, and options, now we see each as just special cases of the same theory. The common language also allows us to use insights from each field of application in other fields. This approach also allows us to conveniently separate the step of specifying economic assumptions of the model (second equation) from the step of deciding which kind of empirical representation to pursue or understand. For a given model – choice of f (·) – we will see how the first equation can lead to predictions stated in terms of returns, price-dividend ratios, expected return-beta representations, moment conditions, continuous vs. discrete time implications and so forth. The ability to translate between such representations is also very helpful in digesting the results of empirical work, which uses a number of apparently distinct but fundamentally connected representations. Thinking in terms of discount factors often turns out to be much simpler than thinking in terms of portfolios. For example, it is easier to insist that there is a positive discount factor than to check that every possible portfolio that dominates every other portfolio has a larger price, and the long arguments over the APT stated in terms of portfolios are easy to digest when stated in terms of discount factors. The discount factor approach is also associated with a state-space geometry in place of the usual mean-variance geometry, and this book emphasizes the state-space intuition behind many classic results. For these reasons, the discount factor language and the associated state-space geometry is common in academic research and high-tech practice. It is not yet common in textbooks, and that is the niche that this book tries to fill. I also diverge from the usual order of presentation. Most books are structured following the history of thought: portfolio theory, mean-variance frontiers, spanning theorems, CAPM, ICAPM, APT, option pricing, and finally consumption-based model. Contingent claims are an esoteric extension of option-pricing theory. I go the other way around: contingent claims and the consumption-based model are the basic and simplest models around; the others are specializations. Just because they were discovered in the opposite order is no reason to present them that way. I also try to unify the treatment of empirical methods. A wide variety of methods are popular, including time-series and cross-sectional regressions, and methods based on generalized method of moments (GMM) and maximum likelihood. However, in the end all of these apparently different approaches do the same thing: they pick free parameters of the model to make it fit best, which usually means to minimize pricing errors; and they evaluate the model by examining how big those pricing errors are. As with the theory, I do not attempt an encyclopedic compilation of empirical procedures. The literature on econometric methods contains lots of methods and special cases (likelihood ratio analogues of common Wald tests; cases with and without riskfree assets and when factors do and don’t span the mean variance frontier, etc.) that are seldom used in practice. I 10

try to focus on the basic ideas and on methods that are actually used in practice. The accent in this book is on understanding statements of theory, and working with that theory to applications, rather than rigorous or general proofs. Also, I skip very lightly over many parts of asset pricing theory that have faded from current applications, although they occupied large amounts of the attention in the past. Some examples are portfolio separation theorems, properties of various distributions, or asymptotic APT. While portfolio theory is still interesting and useful, it is no longer a cornerstone of pricing. Rather than use portfolio theory to find a demand curve for assets, which intersected with a supply curve gives prices, we now go to prices directly. One can then find optimal portfolios, but it is a side issue for the asset pricing question. My presentation is consciously informal. I like to see an idea in its simplest form and learn to use it before going back and understanding all the foundations of the ideas. I have organized the book for similarly minded readers. If you are hungry for more formal definitions and background, keep going, they usually show up later on in the chapter. Again, my organizing principle is that everything can be traced back to specializations of the basic pricing equation p = E(mx). Therefore, after reading the first chapter, one can pretty much skip around and read topics in as much depth or order as one likes. Each major subject always starts back at the same pricing equation. The target audience for this book is economics and finance Ph.D. students, advanced MBA students or professionals with similar background. I hope the book will also be useful to fellow researchers and finance professionals, by clarifying, relating and simplifying the set of tools we have all learned in a hodgepodge manner. I presume some exposure to undergraduate economics and statistics. A reader should have seen a utility function, a random variable, a standard error and a time series, should have some basic linear algebra and calculus and should have solved a maximum problem by setting derivatives to zero. The hurdles in asset pricing are really conceptual rather than mathematical.

11

PART I Asset pricing theory

12

Chapter 1. Consumption-based model and overview I start by thinking of an investor who thinks about how much to save and consume, and what portfolio of assets to hold. The most basic pricing equation comes from the first-order conditions to that problem, and say that price should be the expected discounted payoff, using the investor’s marginal utility to discount the payoff. The marginal utility loss of consuming a little less today and investing the result should equal the marginal utility gain of selling the investment at some point in the future and eating the proceeds. If the price does not satisfy this relation, the investor should buy more of the asset. From this simple idea, I can discuss the classic issues in finance. The interest rate is related to the average future marginal utility, and hence to the expected path of consumption. High real interest rates should be associated with an expectation of growing consumption. In a time of high real interest rates, it makes sense to save, buy bonds, and then consume more tomorrow. Most importantly, risk corrections to asset prices should be driven by the covariance of asset payoffs with consumption or marginal utility. For a given expected payoff of an asset, an asset that does badly in states like a recession, in which the investor feels poor and is consuming little, is less desirable than an asset that does badly in states of nature like a boom when the investor feels wealthy and is consuming a great deal. The former assets will sell for lower prices; their prices will reflect a discount for their riskiness, and this riskiness depends on a co-variance. This is the fundamental point of the whole book. Of course, the fundamental measure of how you feel is marginal utility; given that assets must pay off well in some states and poorly in others, you want assets that pay off poorly in states of low marginal utility, when an extra dollar doesn’t really seem all that important, and you’d rather that they pay off well in states of high marginal utility, when you’re hungry and really anxious to have an extra dollar. Most of the book is about how to go from marginal utility to observable indicators. Consumption is low when marginal utility is high, of course, so consumption may be a useful indicator. Consumption is also low and marginal utility is high when the investor’s other assets have done poorly; thus we may expect that prices are

13

C HAPTER 1

C ONSUMPTION - BASED MODEL AND OVERVIEW

low for assets that covary positively with a large index such as the market portfolio. This is the Capital Asset Pricing Model. The rest of the book comes down to useful indicators for marginal utility, things against which to compute a covariance in order to predict the risk-adjustment for prices.

1.1

Basic pricing equation

An investor’s first order conditions give the basic consumption-based model, · 0 ¸ u (ct+1 ) pt = Et β 0 xt+1 . u (ct ) Our basic objective is to figure out the value of any stream of uncertain cash flows. I start with an apparently simple case, which turns out to capture very general situations. Let us find the value at time t of a payoff xt+1 . For example, if one buys a stock today, the payoff next period is the stock price plus dividend, xt+1 = pt+1 +dt+1 . xt+1 is a random variable: an investor does not know exactly how much he will get from his investment, but he can assess the probability of various possible outcomes. Don’t confuse the payoff xt+1 with the profit or return; xt+1 is the value of the investment at time t + 1, without subtracting or dividing by the cost of the investment. We find the value of this payoff by asking what it is worth to a typical investor. To do this, we need a convenient mathematical formalism to capture what an investor wants. We model investors by a utility function defined over current and future values of consumption, U(ct , ct+1 ) = u(ct ) + βEt [u(ct+1 )] , where ct denotes consumption at date t. We will often use a convenient power utility form, u(ct ) = The limit as γ → 1 is

1 1−γ c . 1−γ t

u(c) = ln(c). The utility function captures the fundamental desire for more consumption, rather than posit a desire for intermediate objectives such as means and variance of portfolio returns. Consumption ct+1 is also random; the investor does not know his wealth tomorrow, and hence how much he will decide to consume. The period utility function u(·) is increasing, reflecting a desire for more consumption, and concave, reflecting the declining marginal value of additional consumption. The last bite is never as satisfying as the first. 14

S ECTION 1.1

BASIC PRICING EQUATION

This formalism captures investors’ impatience and their aversion to risk, so we can quantitatively correct for the risk and delay of cash flows. Discounting the future by β captures impatience, and β is called the subjective discount factor. The curvature of the utility function also generates aversion to risk and to intertemporal substitution: The consumer prefers a consumption stream that is steady over time and across states of nature. Now, assume that the investor can freely buy or sell as much of the payoff xt+1 as he wishes, at a price pt . How much will he buy or sell? To find the answer, denote by e the original consumption level (if the investor bought none of the asset), and denote by ξ the amount of the asset he chooses to buy. Then, his problem is, max u(ct ) + Et βu(ct+1 ) s.t. {ξ}

ct ct+1

= et − pt ξ = et+1 + xt+1 ξ

Substituting the constraints into the objective, and setting the derivative with respect to ξ equal to zero, we obtain the first-order condition for an optimal consumption and portfolio choice, pt u0 (ct ) = Et [βu0 (ct+1 )xt+1 ]

(1)

· 0 ¸ u (ct+1 ) pt = Et β 0 xt+1 . u (ct )

(2)

or,

The investor buys more or less of the asset until this first order condition holds. Equation (1.1) expresses the standard marginal condition for an optimum: pt u0 (ct ) is the loss in utility if the investor buys another unit of the asset; Et [βu0 (ct+1 )xt+1 ] is the increase in (discounted, expected) utility he obtains from the extra payoff at t+1. The investor continues to buy or sell the asset until the marginal loss equals the marginal gain. Equation (1.2) is the central asset-pricing formula. Given the payoff xt+1 and given the investor’s consumption choice ct , ct+1 , it tells you what market price pt to expect. Its economic content is simply the first order conditions for optimal consumption and portfolio formation. Most of the theory of asset pricing just consists of specializations and manipulations of this formula. Notice that we have stopped short of a complete solution to the model, i.e. an expression with exogenous items on the right hand side. We relate one endogenous variable, price, to two other endogenous variables, consumption and payoffs. One can continue to solve this model and derive the optimal consumption choice ct , ct+1 in terms of the givens of the model. In the model I have sketched so far, those givens are the income sequence et , et+1 and a specification of the full set of assets that the investor may buy and sell. We will in fact study 15

C HAPTER 1

C ONSUMPTION - BASED MODEL AND OVERVIEW

such fuller solutions below. However, for many purposes one can stop short of specifying (possibly wrongly) all this extra structure, and obtain very useful predictions about asset prices from (1.2), even though consumption is an endogenous variable.

1.2

Marginal rate of substitution/stochastic discount factor

We break up the basic consumption-based pricing equation into p = E(mx)

m=β

u0 (ct+1 ) u0 (ct )

where mt+1 is the stochastic discount factor. A convenient way to break up the basic pricing equation (1.2) is to define the stochastic discount factor mt+1 mt+1 ≡ β

u0 (ct+1 ) u0 (ct )

(3)

Then, the basic pricing formula (1.2) can simply be expressed as pt = Et (mt+1 xt+1 ).

(4)

When it isn’t necessary to be explicit about time subscripts or the difference between conditional and unconditional expectation, I’ll suppress the subscripts and just write p = E(mx). The price always comes at t, the payoff at t + 1, and the expectation is conditional on time t information. The term stochastic discount factor refers to the way m generalizes standard discount factor ideas. If there is no uncertainty, we can express prices via the standard present value formula pt =

1 xt+1 Rf

(5)

where Rf is the gross risk-free rate. 1/Rf is the discount factor. Since gross interest rates are typically greater than one, the payoff xt+1 sells “at a discount.” Riskier assets have lower prices than equivalent risk-free assets, so they are often valued by using risk-adjusted 16

S ECTION 1.3

P RICES , PAYOFFS AND NOTATION

discount factors,

pit =

1 Et (xit+1 ). Ri

Here, I have added the i superscript to emphasize that each risky asset i must be discounted by an asset-specific risk-adjusted discount factor 1/Ri . In this context, equation (1.4) is obviously a generalization, and it says something deep: one can incorporate all risk-corrections by defining a single stochastic discount factor – the same one for each asset – and putting it inside the expectation. mt+1 is stochastic or random because it is not known with certainty at time t. As we will see, the correlation between the random components of m and xi generate asset-specific risk corrections. mt+1 is also often called the marginal rate of substitution after (1.3). In that equation, mt+1 is the rate at which the investor is willing to substitute consumption at time t + 1 for consumption at time t. mt+1 is sometimes also called the pricing kernel. If you know what a kernel is and express the expectation as an integral, you can see where the name comes from. It is sometimes called a change of measure or a state-price density for reasons that we will see below. For the moment, introducing the discount factor m and breaking the basic pricing equation (1.2) into (1.3) and (1.4) is just a notational convenience. As we will see, however, it represents a much deeper and more useful separation. For example, notice that p = E(mx) would still be valid if we changed the utility function, but we would have a different function connecting m to data. As we will see, all asset pricing models amount to alternative models connecting the stochastic discount factor to data, while p = E(mx) is a convenient accounting identity with almost no content. At the same time, we will study lots of alternative expressions of p = E(mx), and we can summarize many empirical approaches to p = E(mx). By separating our models into these two components, we don’t have to redo all that elaboration for each asset pricing model.

1.3

Prices, payoffs and notation

The price pt gives rights to a payoff xt+1 . In practice, this notation covers a variety of cases, including the following: 17

C HAPTER 1

C ONSUMPTION - BASED MODEL AND OVERVIEW

Stock Return Price-dividend ratio Excess return Managed portfolio Moment condition One-period bond Risk free rate Option

Price pt pt 1 pt dt

0 zt E(pt zt ) pt 1 C

Payoff xt+1 pt+1 + dt+1 Rt+1´ ³ pt+1 dt+1 dt+1 + 1 dt b Ret+1 = Rat+1 − Rt+1 zt Rt+1 xt+1 zt 1 Rf max(ST − K, 0)

The price pt and payoff xt+1 seem like a very restrictive kind of security. In fact, this notation is quite general and allows us easily to accommodate many different asset pricing questions. In particular, we can cover stocks, bonds and options and make clear that there is one theory for all asset pricing. For stocks, the one period payoff is of course the next price plus dividend, xt+1 = pt+1 + dt+1 . We frequently divide the payoff xt+1 by the price pt to obtain a gross return Rt+1 ≡

xt+1 pt

We can think of a return as a payoff with price one. If you pay one dollar today, the return is how many dollars or units of consumption you get tomorrow. Thus, returns obey 1 = E(mR) which is by far the most important special case of the basic formula p = E(mx). I use capital letters to denote gross returns R, which have a numerical value like 1.05. I use lowercase letters to denote net returns r = R − 1 or log (continuously compounded) returns ln(R), both of which have numerical values like 0.05. One may also quote percent returns 100 × r.

Returns are often used in empirical work because they are typically stationary over time. (Stationary in the statistical sense; they don’t have trends and you can meaningfully take an average. “Stationary” does not mean constant.) However, thinking in terms of returns takes us away from the central task of finding asset prices. Dividing by dividends and creating a payoff µ ¶ pt+1 dt+1 xt+1 = 1 + dt+1 dt

corresponding to a price pt /dt is a way to look at prices but still to examine stationary variables. Not everything can be reduced to a return. If you borrow a dollar at the interest rate Rf and invest it in an asset with return R, you pay no money out-of-pocket today, and get the 18

S ECTION 1.4

C LASSIC ISSUES IN FINANCE

payoff R − Rf . This is a payoff with a zero price, so you obviously can’t divide payoff by price to get a return. Zero price does not imply zero payoff. It is a bet in which the chance of losing exactly balances its chance of winning, so that it is not worth paying extra to take the bet. It is common to study equity strategies in which one short sells one stock or portfolio and invests the proceeds in another stock or portfolio, generating an excess return. I denote any such difference between returns as an excess return, Re . It is also called a zero-cost portfolio or a self-financing portfolio. In fact, much asset pricing focuses on excess returns. Our economic understanding of interest rate variation turns out to have little to do with our understanding of risk premia, so it is convenient to separate the two exercises by looking at interest rates and excess returns separately. We also want to think about the managed portfolios, in which one invests more or less in an asset according to some signal. The “price” of such a strategy is the amount invested at time t, say zt , and the payoff is zt Rt+1 . For example a market timing strategy might put a weight in stocks proportional to the price-dividend ratio, investing less when prices are higher. We could represent such a strategy as a payoff using zt = a − b(pt /dt ).

When we think about conditioning information below, we will think of objects like zt as instruments. Then we take an unconditional expectation of pt zt = Et (mt+1 xt+1 )zt , yielding E(pt zt ) = E(mt+1 xt+1 zt ). We can think of this operation as creating a “security” with payoff xt+1 zt+1 , and “price” E(pt zt ) represented with unconditional expectations.

A one period bond is of course a claim to a unit payoff. Bonds, options, investment projects are all examples in which it is often more useful to think of prices and payoffs rather than returns. Prices and returns can be real (denominated in goods) or nominal (denominated in dollars); p = E(mx) can refer to either case. The only difference is whether we use a real or nominal discount factor. If prices, returns and payoffs are nominal, we should use a nominal discount factor. For example, if p and x denote nominal values, then we can create real prices and payoffs to write pt = Et Πt

·µ 0 ¸ ¶ u (ct+1 ) xt+1 β 0 u (ct ) Πt+1

where Π denotes the price level (cpi). Obviously, this is the same as defining a nominal discount factor by ·µ 0 ¶ ¸ u (ct+1 ) Πt pt = Et β 0 xt+1 u (ct ) Πt+1 To accommodate all these cases, I will simply use the notation price pt and payoff xt+1 . These symbols can denote 0, 1, or zt and Rte , rt+1 , or zt Rt+1 respectively, according to the case. Lots of other definitions of p and x are useful as well. 19

C HAPTER 1

1.4

C ONSUMPTION - BASED MODEL AND OVERVIEW

Classic issues in finance

I use simple manipulations of the basic pricing equation to introduce classic issues in finance: the economics of interest rates, risk adjustments, systematic vs. idiosyncratic risk, expected return-beta representations, the mean-variance frontier, the slope of the mean-variance frontier, time-varying expected returns, and present value relations. A few simple rearrangements and manipulations of the basic pricing equation p = E(mx) give a lot of intuition and introduce some classic issues in finance, including determinants of the interest rate, risk corrections, idiosyncratic vs. systematic risk, beta pricing models, and mean variance frontiers. 1.4.1

Risk free rate

Rf = 1/E(m). With lognormal consumption growth, rtf = δ + γEt ∆ ln ct+1 −

γ2 2 σ (∆ ln ct+1 ) 2 t

Real interest rates are high when people are impatient (δ), when expected consumption growth is high (intertemporal substitution), or when risk is low (precautionary saving). A more curved utility function (γ) or a lower elasticity of intertemporal substitution (1/γ) means that interest rates are more sensitive to changes in expected consumption growth. The risk free rate is given by Rf = 1/E(m).

(6)

The risk free rate is known ahead of time, so p = E(mx) becomes 1 = E(mRf ) = E(m)Rf . If a risk free security is not traded, we can define Rf = 1/E(m) as the “shadow” risk-free rate. (In some models it is called the “zero-beta” rate.) If one introduced a risk free security with return Rf = 1/E(m), investors would be just indifferent to buying or selling it. I use Rf to simplify formulas below with this understanding. To think about the economics behind real interest rates in a simple setup, use power utility 20

S ECTION 1.4

C LASSIC ISSUES IN FINANCE

u0 (c) = c−γ . Start by turning off uncertainty, in which case Rf =

1 β

µ

ct+1 ct

¶γ

.

We can see three effects right away: 1. 2.

3.

Real interest rates are high when people are impatient, when β is low. If everyone wants to consume now, it takes a high interest rate to convince them to save. Real interest rates are high when consumption growth is high. In times of high interest rates, it pays investors to consume less now, invest more, and consume more in the future. Thus, high interest rates lower the level of consumption today, while raising its growth rate from today to tomorrow. Real interest rates are more sensitive to consumption growth if the power parameter γ is large. If utility is highly curved, the investor cares more about maintaining a consumption profile that is smooth over time, and is less willing to rearrange consumption over time in response to interest rate incentives. Thus it takes a larger interest rate change to induce him to a given consumption growth. To understand how interest rates behave when there is some uncertainty, I specify that consumption growth is lognormally distributed. In this case, the real riskfree rate equation becomes rtf = δ + γEt ∆ ln ct+1 −

γ2 2 σ (∆ ln ct+1 ) 2 t

where I have defined the log riskfree rate rtf and subjective discount rate δ by rtf = ln Rtf ; β = e−δ , and ∆ denotes the first difference operator, ∆ ln ct+1 = ln ct+1 − ln ct . To derive expression (1.7) for the riskfree rate, start with " µ ¶−γ # ct+1 = 1/Et β . ct

Rtf Using the fact that normal z means

1

E (ez ) = eE(z)+ 2 σ 21

2

(z)

(7)

C HAPTER 1

C ONSUMPTION - BASED MODEL AND OVERVIEW

(you can check this by writing out the integral that defines the expectation), we have · ¸−1 γ2 2 . Rft = e−δ e−γEt (∆ ln ct+1 )+ 2 σt (∆ ln ct+1 ) Then take logarithms. The combination of lognormal distributions and power utility is one of the basic tricks to getting analytical solutions in this kind of model. Section 1.5 shows how to get the same result in continuous time. Looking at (1.7), we see the same results as we had with the deterministic case. Real interest rates are high when impatience δ is high and when consumption growth is high; higher γ makes interest rates more sensitive to consumption growth. The new σ2 term captures precautionary savings. When consumption is more volatile, people with this utility function are more worried about the low consumption states than they are pleased by the high consumption states. Therefore, people want to save more, driving down interest rates. We can also read the same terms backwards: consumption growth is high when real interest rates are high, since people save more now and spend it in the future, and consumption is less sensitive to interest rates as the desire for a smooth consumption stream, captured by γ, rises. . Section 2.2 below takes up the question of which way we should read this equation – as consumption determining interest rates, or as interest rates determining consumption. For the power utility function, the curvature parameter γ simultaneously controls intertemporal substitution – aversion to a consumption stream that varies over time, risk aversion – aversion to a consumption stream that varies across states of nature, and precautionary savings, which turns out to depend on the third derivative of the utility function. This link is particular to the power utility function. We will study utility functions below that loosen the links between these three quantities. 1.4.2

Risk corrections

p=

E(x) + cov(m, x) Rf

¢ ¡ E(Ri ) − Rf = −Rf cov m, Ri .

Payoffs that are positively correlated with consumption growth have lower prices, to compensate investors for risk. Expected returns are proportional to the covariance of returns with discount factors. Using the definition of covariance cov(m, x) = E(mx) − E(m)E(x), we can write 22

S ECTION 1.4

C LASSIC ISSUES IN FINANCE

equation (1.2) as p = E(m)E(x) + cov(m, x).

(8)

Substituting the riskfree rate equation (1.6), we obtain p=

E(x) + cov(m, x) Rf

(9)

The first term in (1.9) is the standard discounted present value formula. This is the asset’s price in a risk-neutral world – if consumption is constant or if utility is linear. The second term is a risk adjustment. An asset whose payoff covaries positively with the discount factor has its price raised and vice-versa. To understand the risk adjustment, substitute back for m in terms of consumption, to obtain p=

E(x) cov [βu0 (ct+1 ), xt+1 ] + Rf u0 (ct )

(10)

Marginal utility u0 (c) declines as c rises. Thus, an asset’s price is lowered if its payoff covaries positively with consumption. Conversely, an asset’s price is raised if it covaries negatively with consumption. Why? Investors do not like uncertainty about consumption. If you buy an asset whose payoff covaries positively with consumption, one that pays off well when you are already feeling wealthy, and pays off badly when you are already feeling poor, that asset will make your consumption stream more volatile. You will require a low price to induce you to buy such an asset. If you buy an asset whose payoff covaries negatively with consumption, it helps to smooth consumption and so is more valuable than its expected payoff might indicate. Insurance is an extreme example. Insurance pays off exactly when wealth and consumption would otherwise be low–you get a check when your house burns down. For this reason, you are happy to hold insurance, even though you expect to lose money—even though the price of insurance is greater than its expected payoff discounted at the risk free rate. To emphasize why the covariance of a payoff with the discount factor rather than its variance determines its riskiness, keep in mind that the investor cares about the volatility of consumption. He does not care about the volatility of his individual assets or of his portfolio, if he can keep a steady consumption. Consider what happens to the volatility of consumption if the investor buys a little more ξ of payoff x: σ2 (c) becomes σ 2 (c + ξx) = σ 2 (c) + 2ξcov(c, x) + ξ 2 σ2 (x) For small (marginal) portfolio changes, the covariance between consumption and payoff determines the effect of adding a bit more of each payoff on the volatility of consumption. We use returns so often that it is worth restating the same intuition for the special case that 23

C HAPTER 1

C ONSUMPTION - BASED MODEL AND OVERVIEW

the price is one and the payoff is a return. Start with the basic pricing equation for returns, 1 = E(mRi ). I denote the return Ri to emphasize that the point of the theory is to distinguish the behavior of one asset Ri from another Rj . The asset pricing model says that, although expected returns can vary across time and assets, expected discounted returns should always be the same, 1. Applying the covariance decomposition, 1 = E(m)E(Ri ) + cov(m, Ri )

(11)

E(Ri ) − Rf = −Rf cov(m, Ri )

(12)

and, using Rf = 1/E(m),

or E(Ri ) − Rf = −

cov[u0 (ct+1 ), Rit+1 ] . E[u0 (ct+1 )]

(13)

All assets have an expected return equal to the risk-free rate, plus a risk adjustment. Assets whose returns covary positively with consumption make consumption more volatile, and so must promise higher expected returns to induce investors to hold them. Conversely, assets that covary negatively with consumption, such as insurance, can offer expected rates of return that are lower than the risk-free rate, or even negative (net) expected returns. Much of finance focuses on expected returns. We think of expected returns increasing or decreasing to clear markets; we offer intuition that “riskier” securities must offer higher expected returns to get investors to hold them, rather than saying “riskier” securities trade for lower prices so that investors will hold them. Of course, a low initial price for a given payoff corresponds to a high expected return, so this is no more than a different language for the same phenomenon. 1.4.3

Idiosyncratic risk does not affect prices

Only the component of a payoff perfectly correlated with the discount factor generates an extra return. Idiosyncratic risk, uncorrelated with the discount factor, generates no premium. You might think that an asset with a high payoff variance is “risky” and thus should have a large risk correction. However, if the payoff is uncorrelated with the discount factor m, the asset receives no risk-correction to its price, and pays an expected return equal to the risk-free 24

S ECTION 1.4

C LASSIC ISSUES IN FINANCE

rate! In equations, if cov(m, x) = 0 then p=

E(x) . Rf

This prediction holds even if the payoff x is highly volatile and investors are highly risk averse. The reason is simple: if you buy a little bit more of such an asset, it has no first-order effect on the variance of your consumption stream. More generally, one gets no compensation or risk adjustment for holding idiosyncratic risk. Only systematic risk generates a risk correction. To give meaning to these words, we can decompose any payoff x into a part correlated with the discount factor and an idiosyncratic part uncorrelated with the discount factor by running a regression, x = proj(x|m) + ε. Then, the price of the residual or idiosyncratic risk ε is zero, and the price of x is the same as the price of its projection on m. The projection of x on m is of course that part of x which is perfectly correlated with m. The idiosyncratic component of any payoff is that part uncorrelated with m. Thus only the systematic part of a payoff accounts for its price. Projection means linear regression without a constant, proj(x|m) =

E(mx) m. E(m2 )

You can verify that regression residuals are orthogonal to right hand variables E(mε) = 0 from this definition. E(mε) = 0 of course means that the price of ε is zero. µ ¶ µ ¶ E(mx) 2 E(mx) p (proj(x|m)) = p m = E m = E(mx) = p(x). E(m2 ) E(m2 ) The words “systematic” and “idiosyncratic” are defined differently in different contexts, which can lead to some confusion. In this decomposition, the residuals ε can be correlated with each other, though they are not correlated with the discount factor. The APT starts with a factor-analytic decomposition of the covariance of payoffs, and the word “idiosyncratic” there is reserved for the component of payoffs uncorrelated with all of the other payoffs. 1.4.4

Expected return-beta representation

25

C HAPTER 1

C ONSUMPTION - BASED MODEL AND OVERVIEW

We can write p = E(mx) as E(Ri ) = Rf + β i,m λm

or

We can express the expected return equation (1.12), for a return Ri , as µ ¶µ ¶ cov(Ri , m) var(m) i f E(R ) = R + − var(m) E(m) E(Ri ) = Rf + β i,m λm

(14)

(15)

where β im is the regression coefficient of the return Ri on m. This is a beta pricing model. It says that expected returns on assets i = 1, 2, ...N should be proportional to their betas in a regression of returns on the discount factor. Notice that the coefficient λm is the same for all assets i,while the β i,m varies from asset to asset. The λm is often interpreted as the price of risk and the β as the quantity of risk in each asset. Obviously, there is nothing deep about saying that expected returns are proportional to betas rather than to covariances. There is a long historical tradition and some minor convenience in favor of betas. The betas refer to the projection of R on m that we studied above, so you see again a sense in which only the systematic component of risk matters. −γ With m = β (ct+1 /ct ) , we can take a Taylor approximation of equation (1.14) to express betas in terms of a more concrete variable, consumption growth, rather than marginal utility. The result, which I derive more explicitly and conveniently in the continuous time limit below, is

E(Ri ) = Rf + β i,∆c λ∆c λ∆c = γvar(∆c).

(1.16)

Expected returns should increase linearly with their betas on consumption growth itself. In addition, though it is treated as a free parameter in many applications, the factor risk premium λ∆c is determined by risk aversion and the volatility of consumption. The more risk averse people are, or the riskier their environment, the larger an expected return premium one must pay to get investors to hold risky (high beta) assets. 1.4.5

Mean-variance frontier

All asset returns lie inside a mean-variance frontier. Assets on the frontier are perfectly correlated with each other and the discount factor. Returns on the frontier can be generated as portfolios of any two frontier returns. We can construct a discount factor from any frontier 26

S ECTION 1.4

C LASSIC ISSUES IN FINANCE

return (except Rf ), and an expected return-beta representation holds using any frontier return (except Rf ) as the factor. Asset pricing theory has focused a lot on the means and variances of asset returns. Interestingly, the set of means and variances of returns is limited. All assets priced by the discount factor m must obey ¯ ¯ ¯E(Ri ) − Rf ¯ ≤ σ(m) σ(Ri ). E(m)

(17)

To derive (1.17) write for a given asset return Ri

1 = E(mRi ) = E(m)E(Ri ) + ρm,Ri σ(Ri )σ(m) and hence E(Ri ) = Rf − ρm,Ri

σ(m) σ(Ri ). E(m)

(18)

Correlation coefficients can’t be greater than one in magnitude, leading to (1.17). This simple calculation has many interesting and classic implications. 1. Means and variances of asset returns must lie in the wedge-shaped region illustrated in Figure 1. The boundary of the mean-variance region in which assets can lie is called the mean-variance frontier. It answers a naturally interesting question, “how much mean return can you get for a given level of variance?” 2. All returns ¯ frontier are perfectly correlated with the discount factor: the frontier ¯ on the is generated by ¯ρm,Ri ¯ = 1. Returns on the upper part of the frontier are perfectly negatively correlated with the discount factor and hence positively correlated with consumption. They are “maximally risky” and thus get the highest expected returns. Returns on the lower part of the frontier are perfectly positively correlated with the discount factor and hence negatively correlated with consumption. They thus provide the best insurance against consumption fluctuations. 3. All frontier returns are also perfectly correlated with each other, since they are all perfectly correlated with the discount factor. This fact implies that we can span or synthesize any frontier return from two such returns. For example if you pick any single frontier return Rm then all frontier returns Rmv must be expressible as

for some number a.

¢ ¡ Rmv = Rf + a Rm − Rf

4. Since each point on the mean-variance frontier is perfectly correlated with the discount 27

C HAPTER 1

C ONSUMPTION - BASED MODEL AND OVERVIEW

E(R)

Mean-variance frontier Slope σ(m)/E(m) Idiosyncratic risk Ri Rf Some asset returns

σ(R)

Figure 1. Mean-variance frontier. The mean and standard deviation of all assets priced by a discount factor m must line in the wedge-shaped region factor, we must be able to pick constants a, b, d, e such that m = a + bRmv Rmv = d + em. Thus, any mean-variance efficient return carries all pricing information. Given a meanvariance efficient return and the risk free rate, we can find a discount factor that prices all assets and vice versa. 5. Given a discount factor, we can also construct a single-beta representation, so expected returns can be described in a single - beta representation using any mean-variance efficient return (except the riskfree rate), £ ¤ E(Ri ) = Rf + β i,mv E(Rmv ) − Rf .

The essence of the β pricing model is that, even though the means and standard deviations of returns fill out the space inside the mean-variance frontier, a graph of mean returns versus betas should yield a straight line. Since the beta model applies to every return including Rmv itself, and Rmv has a beta of one on itself, we can identify the factor risk premium as λ = E(Rmv − Rf ). The last two points suggest an intimate relationship between discount factors, beta models and mean-variance frontiers. I explore this relation in detail in Chapter 6. A problem at the end of this chapter guides you through the algebra to demonstrate points 4 and 5 explicitly. 28

S ECTION 1.4

C LASSIC ISSUES IN FINANCE

6. We can plot the decomposition of a return into a “priced” or “systematic” component and a “residual,” or “idiosyncratic” component as shown in Figure 1. The priced part is perfectly correlated with the discount factor, and hence perfectly correlated with any frontier asset. The residual or idiosyncratic part generates no expected return, so it lies flat as shown in the figure, and it is uncorrelated with the discount factor or any frontier asset.. 1.4.6

Slope of the mean-standard deviation frontier and equity premium puzzle

The Sharpe ratio is limited by the volatility of the discount factor. The maximal risk-return tradeoff is steeper if there is more risk or more risk aversion ¯ ¯ ¯ E(R) − Rf ¯ σ(m) ¯ ¯ ¯ σ(R) ¯ ≤ E(m) ≈ γσ(∆ ln c) This formula captures the equity premium puzzle, which suggests that either people are very risk averse, or the stock returns of the last 50 years were good luck which will not continue. The ratio of mean excess return to standard deviation E(Ri ) − Rf = Sharpe ratio σ(Ri ) is known as the Sharpe ratio. It is a more interesting characterization of any security than the mean return alone. If you borrow and put more money into a security, you can increase the mean return of your position, but you do not increase the Sharpe ratio, since the standard deviation increases at the same rate as the mean. The slope of the mean-standard deviation frontier is the largest available Sharpe ratio, and thus is naturally interesting. It answers “how much more mean return can I get by shouldering a bit more volatility in my portfolio?” Let Rmv denote the return of a portfolio on the frontier. From equation (1.17), the slope of the frontier is ¯ ¯ ¯ E(Rmv ) − Rf ¯ σ(m) f ¯ ¯ ¯ σ(Rmv ) ¯ = E(m) = σ(m)R . Thus, the slope of the frontier is governed by the volatility of the discount factor.

For an economic interpretation, again consider the power utility function, u0 (c) = c−γ , ¯ ¯ ¯ E(Rmv ) − Rf ¯ σ [(ct+1 /ct )−γ ] ¯ ¯ i. (19) ¯ σ(Rmv ) ¯ = h E (ct+1 /ct )−γ

The standard deviation is large if consumption is volatile or if γ is large. We can state this 29

C HAPTER 1

C ONSUMPTION - BASED MODEL AND OVERVIEW

approximation again using the lognormal assumption. If consumption growth is lognormal, ¯ ¯ ¯ E(Rmv ) − Rf ¯ p 2 2 γ σ (∆ ln ct+1 ) − 1 ≈ γσ(∆ ln c). ¯ ¯ ¯ σ(Rmv ) ¯ = e

(20)

(A problem at the end of the chapter guides you though the algebra of the first equality. The relation is exact in continuous time, and thus the approximation is easiest to derive by reference to the continuous time result; see section 1.5.) Reading the equation, the slope of the mean-standard deviation frontier is higher if the economy is riskier – if consumption is more volatile – or if investors are more risk averse. Both situations naturally make investors more reluctant to take on the extra risk of holding risky assets. This expression is also the slope of the expected return beta line of the consumption beta model, (1.16). (Or, conversely, in an economy with a high Sharpe ratio, low risk aversion investors should take on so much risk that their consumption becomes volatile.) In postwar US data, the slope of the mean-standard deviation frontier, or of expected return-beta lines is much higher than reasonable risk aversion and consumption volatility estimates suggest. This is the “equity premium puzzle.” Over the last 50 years in the U.S., real stock returns have averaged 9% with a standard deviation of about 16%, while the real return on treasury bills has been about 1%. Thus, the historical annual market Sharpe ratio has been about 0.5. Aggregate consumption growth has been about 1%. Thus, we can only reconcile these facts with (1.20) if investors have a risk aversion coefficient of 50! Obvious ways of generalizing the calculation just make matters worse. Equation (1.20) relates consumption growth to the mean-variance frontier of all contingent claims. The market indices with 0.5 Sharpe ratios are if anything inside that frontier, so recognizing market incompleteness will only make matters worse. Aggregate consumption has about 0.2 correlation with the market return, while the equality (1.20) takes the worst possible case that consumption growth and asset returns are perfectly correlated. If you add this fact, you need risk aversion of 250 to explain the market Sharpe ratio in the face of 1% consumption volatility! Individuals have riskier consumption streams than aggregate, but as their risk goes up their correlation with any aggregate must decrease proportionally, so to first order recognizing individual risk will not help either. Clearly, either 1) people are a lot more risk averse than we might have thought 2) the stock returns of the last 50 years were largely good luck rather than an equilibrium compensation for risk, or 3) something is deeply wrong with the model, including the utility function and use of aggregate consumption data. This “equity premium puzzle” has attracted the attention of a lot of research in finance, especially on the last item. I return to the equity premium in more detail in Chapter 21. 1.4.7

Random walks and time-varying expected returns

30

S ECTION 1.4

C LASSIC ISSUES IN FINANCE

If investors are risk neutral, returns are unpredictable, and prices follow martingales. In general, prices scaled by marginal utility are martingales, and returns can be predictable if investors are risk averse and if the conditional second moments of returns and discount factors vary over time. This is more plausible at long horizons. So far, we have concentrated on the behavior of prices or expected returns across assets. We should also consider the behavior of the price or return of a given asset over time. Going back to the basic first order condition, pt u0 (ct ) = Et [βu0 (ct+1 )(pt+1 + dt+1 )].

(21)

If investors are risk neutral, i.e. if u(c) is linear or there is no variation in consumption, if the security pays no dividends between t and t + 1, and for short time horizons where β is close to one, this equation reduces to pt = Et (pt+1 ). Equivalently, prices follow a time-series process of the form pt+1 = pt + εt+1 . If the variance σ2t (εt+1 ) is constant, prices follow a random walk. More generally, prices follow a martingale. Intuitively, if the price today is a lot lower than investor’s expectations of the price tomorrow, then people will try to buy the security. But this action will drive up the price of the security until the price today does equal the expected price tomorrow. Another way of saying the same thing is that returns should not be predictable; dividing by pt , expected returns Et (pt+1 /pt ) = 1 should be constant; returns should be like coin flips. The more general equation (1.21) says that prices should follow a martingale after adjusting for dividends and scaling by marginal utility. Since martingales have useful mathematical properties, and since risk-neutrality is such a simple economic environment, many asset pricing results are easily derived by scaling prices and dividends by marginal utility first, and then using “risk-neutral” formulas and economic arguments. Since consumption and risk aversion don’t change much day to day, we might expect the random walk view to hold pretty well on a day-to-day basis. This idea contradicts the still popular notion that there are “systems” or “technical analysis” by which one can predict where stock prices are going on any given day. It has been remarkably successful. Despite decades of dredging the data, and the popularity of television and radio reports that purport to explain where markets are going, trading rules that reliably survive transactions costs and do not implicitly expose the investor to risk have not yet been reliably demonstrated. However, more recently, evidence has accumulated that long-horizon excess returns are quite predictable, and to some this indicates that the whole enterprise of economic explanation of asset returns is flawed. To think about this issue, write our basic equation for expected 31

C HAPTER 1

C ONSUMPTION - BASED MODEL AND OVERVIEW

returns as Et (Rt+1 ) − Rtf

covt (mt+1 , Rt+1 ) Et (mt+1 ) σt (mt+1 ) σt (Rt+1 )ρt (mt+1 , Rt+1 ) = Et (mt+1 ) ≈ γ t σ t (∆ct+1 )σt (Rt+1 )ρt (mt+1 , Rt+1 ). = −

(1.22)

I include the t subscripts to emphasize that the relation applies to conditional moments. Sometimes, the conditional mean or other moment of a random variable is different from its unconditional moment. Conditional on tonight’s weather forecast, you can better predict rain tomorrow than just knowing the average rain for that date. In the special case that random variables are i.i.d. (independent and identically distributed), like coin flips, the conditional and unconditional moments are the same, but that is a special case and not likely to be true of asset prices, returns, and macroeconomic variables. In the theory so far, we have thought of an investor, today, forming expectations of payoffs, consumption, and other variables tomorrow. Thus, the moments are really all conditional, and if we want to be precise we should include some notation to express this fact. I use subscripts Et (xt+1 ) to denote conditional expectation; the notation E(xt+1 |It ) where It is the information set at time t is more precise but a little more cumbersome. Examining equation (1.22), we see that returns can be somewhat predictable. First, if the conditional variance of returns changes over time, we might expect the conditional mean return to vary as well – the return can just move in and out a line of constant Sharpe ratio. This explanation does not seem to help much in the data; variables that forecast means do not seem to forecast variances and vice versa. Unless we want to probe the conditional correlation, predictable excess returns have to be explained by changing risk – σt (∆ct+1 ) – or changing risk aversion γ. It is not plausible that risk or risk aversion change at daily frequencies, but fortunately returns are not predictable at daily frequencies. It is much more plausible that risk and risk aversion change over the business cycle, and this is exactly the horizon at which we see predictable excess returns. Models that make this connection precise are a very active area of current research. 1.4.8

Present value statement

pt = Et

∞ X

mt,t+j dt+j .

j=0

It is convenient to use only the two period valuation, thinking of a price pt and a payoff 32

S ECTION 1.5

D ISCOUNT FACTORS IN CONTINUOUS TIME

xt+1 . But there are times when we want to relate a price to the entire cash flow stream, rather than just to one dividend and next period’s price. The most straightforward way to do this is to write out a longer term objective, Et

∞ X

β j u(ct+j ).

j=0

Now suppose an investor can purchase a stream {dt+j } at price pt . As with the two-period model, his first order condition gives us the pricing formula directly, pt = Et

∞ X

βj

j=0

∞ X u0 (ct+j ) d = E mt,t+j dt+j . t+j t u0 (ct ) j=0

(23)

You can see that if this equation holds at time t and time t + 1, then we can derive the two-period version pt = Et [mt+1 (pt+1 + dt+1 )]

(24)

Thus, the infinite period and two period models are equivalent. (Going in the other direction is a little tougher. If you chain together (1.24), you get (1.23) plus an extra term. To get (1.23) you also need the “transversality condition” limt→∞ Et mt,t+j pt+j = 0. This is an extra first order condition of the infinite period investor, which is not present with overlapping generations of two-period investors. It rules out “bubbles” in which prices grow so fast that people will buy now just to resell at higher prices later, even if there are no dividends.) From (1.23) we can write a risk-adjustment to prices, as we did with one period payoffs, pt =

∞ X Et dt+j j=1

Rft,t+j

+

∞ X

covt (dt+j , mt,t+j )

j=1

f where Rt,t+j ≡ Et (mt,t+j )−1 is the j period interest rate. Again, assets whose dividend streams covary negatively with marginal utility, and positively with consumption, have lower prices, since holding those assets gives the investor a more volatile consumption stream. (It is commonP instead to write prices as a discounted value using a risk adjusted discount factor, ∞ e.g. pt = j=1 Et dt+j /Rt,t+j but this approach is difficult to use correctly for multiperiod problems, especially when expected returns can vary over time.)

1.5

Discount factors in continuous time

Continuous time versions of the basic pricing equations. 33

C HAPTER 1

C ONSUMPTION - BASED MODEL AND OVERVIEW

Discrete 0 P j u (ct+j ) pt = Et ∞ j=1 β u0 (ct ) Dt+j 0

t+1 ) mt+1 = β uu(c 0 (c ) t p = E(mx)

E(R) = Rf − Rf cov(m, R)

Continuous R∞ pt u0 (ct ) = Et s=0 e−δs u0 (ct+s )Dt+s ds Λt = e−δt u0 (ct ) 0 =³ΛD´ dt + Et [d(Λp)] h i f D dΛ dp + Et dp dt = r dt − E t t p p Λ p

It is often convenient to express asset pricing ideas in the language of continuous time stochastic differential equations rather than discrete time stochastic difference equations as I have done so far. The appendix contains a brief introduction to continuous time processes that covers what you need to know for this book. Even if you want to end up with a discrete time representation, manipulations are often easier in continuous time. For example, relating interest rates and Sharpe ratios to consumption growth in the last section required a clumsy lognormal approximation; you’ll see the same sort of thing done much more cleanly in this section. The choice of discrete vs. continuous time is one of modeling convenience. The richness of the theory of continuous time processes often allows one to obtain analytical results that would be unavailable in discrete time. On the other hand, in the complexity of most practical situations, one often ends up resorting to numerical simulation of a discretized model anyway. In those cases, it might be clearer to start with a discrete model. But I emphasize this is all a choice of language. One should become familiar enough with discrete as well as continuous time representations of the same ideas to pick the representation that is most convenient for a particular application. First, we need to think about how to model securities, in place of price pt and one-period payoff xt+1 . Let a generic security have price pt at any moment in time, and let it pay dividends at the rate Dt dt. (I will continue to denote functions of time as pt rather than p(t) to maintain continuity with the discrete-time treatment, and I will drop the time subscripts where they are obvious, e.g. dp in place of dpt . In an interval dt, the security pays dividends Dt dt. I use capital D for dividends to distinguish them from the differential operator d. ) The instantaneous total return is dpt Dt + dt. pt pt We model the price of risky assets as diffusions, for example dpt = µ(·)dt + σ(·)dz. pt (I will reserve the notation dz for increments to a standard Brownian motion, e.g. zt+∆ − zt ∼ N (0, ∆). I use the notation (·) to indicate that the drift and diffusions can be functions of state variables. I limit the discussion to diffusion processes – no jumps.) What’s nice about this diffusion model is that the increments dz are normal; the dependence of µ and σ on state 34

S ECTION 1.5

D ISCOUNT FACTORS IN CONTINUOUS TIME

variables means that the finite time distribution of prices f (pt+∆ |It ) need not be normal.

We can think of a riskfree security as one that has a constant price equal to one and pays the riskfree rate as a dividend, p = 1; Dt = rtf ,

(25)

or as a security that pays no dividend but whose price climbs deterministically at a rate dpt = rtf dt. pt

(26)

Next, we need to express the first order conditions in continuous time. The utility function is U ({ct }) = E

Z

∞

e−δt u(ct )dt.

t=0

Suppose the investor can buy a security whose price is pt and that pays a dividend stream Dt . As we did in deriving the present value price relation in discrete time, the first order condition for this problem gives us the infinite period version of the basic pricing equation right away1 , Z ∞ e−δs u0 (ct+s )Dt+s ds (27) pt u0 (ct ) = Et s=0

This equation is an obvious continuous time analogue to pt = Et

∞ X j=0

βt

u0 (ct+j ) Dt+j . u0 (ct )

It turns out that dividing by u0 (ct ) is not a good idea in continuous time, since the ratio u (ct+∆ )/u0 (ct ) isn’t well behaved for small time intervals. Instead, we keep track of the level of marginal utility. Therefore, define the “discount factor” in continuous time as 0

Λt ≡ e−δt u0 (ct ). Then we can write the pricing equation as pt Λt = Et

Z

∞

Λt+s Dt+s ds.

(28)

s=0

1 One unit of the security pays the dividend stream D , i.e. D dt units of the numeraire consumption good in a t t time interval dt. The security costs pt units of the consumption good. The investor can finance the purchase of ξ units of the security by reducing consumption from et to ct = et − ξpt /dt during time interval dt. The loss in utility from doing so is u0 (ct )(et − ct )dt = u0 (ct )ξpt . The gain is the right hand side of (1.27)

35

C HAPTER 1

C ONSUMPTION - BASED MODEL AND OVERVIEW

(Some people like to define Λt = u0 (ct ), in which case you keep the e−δt in the equation, or R f − τs=0 rt+τ dτ in the equation. to scale Λt by the riskfree rate, in which case you get an extra e The latter procedure makes it look like a risk-neutral or present-value formula valuation.) The analogue to the one period pricing equation p = E(mx) is 0 = ΛD dt + Et [d(Λp)] .

(29)

To derive this fundamental equation, take the difference of equation (1.28) at t and t + ∆. (Or, start directly with the first order condition for buying the security at t and selling it at t + ∆.) Z ∆ Λt+s Dt+s ds + Et [Λt+∆ pt+∆ ] pt Λt = Et s=0

For ∆ small the term in the integral can be approximated pt Λt ≈ Λt Dt ∆ + Et [Λt+∆ pt+∆ ] .

(30)

We want to get to d something, so introduce differences by writing pt Λt ≈ Λt Dt ∆ + Et [Λt pt + (Λt+∆ pt+∆ − Λt pt )] .

(31)

Canceling pt Λt , 0 ≈ Λt Dt ∆ + Et (Λt+∆ pt+∆ − Λt pt ). Taking the limit as ∆ → 0, 0 = Λt Dt dt + Et [d(Λt pt )] or, dropping time subscripts, equation (1.29). Equation (1.29) looks different than p = E(mx) because there is no price on the left hand side; we are used to thinking of the one period pricing equation as determining price at t given other things, including price at t + 1. But price at t is really here, of course, as you can see from equation (1.30) or (1.31). It is just easier to express the difference in price over time rather than price today on the left and payoff (including price tomorrow) on the right. With no dividends and constant Λ, 0 = Et (dpt ) = Et (pt+∆ − pt ) says that price should follow a martingale. Thus, Et [d(Λp)] = 0 means that marginal utility-weighted price should follow a martingale, and (1.29) adjusts for dividends. Thus, it’s the same as the equation (1.21), pt u0 (ct ) = Et (mt+1 (pt+1 + dt+1 )) that we derived in discrete time. Since we will write down price processes for dp and discount factor processes for dΛ, and to interpret (1.29) in terms of expected returns, it is often convenient to break up the d(Λt pt ) term using Ito’s lemma: d(Λp) = pdΛ + Λdp + dpdΛ. 36

(32)

S ECTION 1.5

D ISCOUNT FACTORS IN CONTINUOUS TIME

Using the expanded version (1.32) in the basic equation (1.29), and dividing by pΛ to make it pretty, we obtain an equivalent, slightly less compact but slightly more intuitive version, · ¸ D dΛ dp dΛ dp dt + Et + + 0= . (33) p Λ p Λ p (This formula only works when both Λ and p can never be zero. That is often enough the case that this formula is useful. If not, multiply through by Λ and p and keep them in numerators.) Applying the basic pricing equations (1.29) or (1.33) to a riskfree rate, defined as (1.25) or (1.26), we obtain µ ¶ dΛt (34) rtf dt = −Et Λt This equation is the obvious continuous time equivalent to Rtf =

1 . Et (mt+1 )

If a riskfree rate is not traded, we can use (1.34) to define a shadow riskfree rate or zero-beta rate. With this interpretation, we can rearrange equation (1.33) as µ ¶ · ¸ dpt Dt dΛt dpt f dt = rt dt − Et + . Et pt pt Λt pt

(35)

This is the obvious continuous-time analogue to E(R) = Rf − Rf cov(m, R).

(36)

The last term in (1.35) is the covariance of the return with the discount factor or marginal utility. Since means are order dt, there is no difference between covariance and second moment in the last term of (1.35). The interest rate component of the last term of (1.36) naturally vanishes as the time interval gets short. Ito’s lemma makes many transformations simple in continuous time. For example, the nonlinear transformation between consumption and the discount factor led us to some tricky approximations in discrete time. This transformation is easy in continuous time (diffusions are locally normal, so it’s really the same trick). With Λt = e−δt u0 (ct ) we have 1 dΛt = −δe−δt u0 (ct )dt + e−δt u00 (ct )dct + e−δt u000 (ct )dc2t 2 ct u00 (ct ) dct 1 c2t u000 (ct ) dc2t dΛt = −δdt + 0 + Λt u (ct ) ct 2 u0 (ct ) c2t 37

(37)

C HAPTER 1

C ONSUMPTION - BASED MODEL AND OVERVIEW

Denote the local curvature and third derivative of the utility function as ct u00 (ct ) u0 (ct ) c2t u000 (ct ) . u0 (ct )

γt

= −

ηt

=

(For power utility, the former is the power coefficient γ and the latter is ηt = γ(γ + 1).) Using this formula we can quickly redo the relationship between interest rates and consumption growth, equation (1.7), µ ¶ µ ¶ µ 2¶ 1 1 dΛt dct 1 1 dct f = δ + γ t Et − η t Et . rt = − Et dt Λt dt ct 2 dt c2t We can also easily express asset prices in terms of consumption risk rather than discount factor risk, as in equation (1.16). Using (1.37) in (1.35), µ ¶ µ ¶ dpt Dt dct dpt f dt − rt dt = γEt Et + (38) pt pt ct pt Thus, assets whose returns covary more strongly with consumption get higher mean returns, and the constant relating covariance to mean return is the utility curvature coefficient γ. Since correlations are less than one, equation (1.38) implies that Sharpe ratios are related to utility curvature and consumption volatility directly; we don’t need the ugly lognormal facts and an approximation that we needed in (1.20). Using µp ≡ Et (dpt /pt ) ; h i h i σ2p = Et (dpt /pt )2 ; σ 2c = Et (dct /ct )2 , µp +

Dt pt

dt − rtf dt

σp

1.6

≤ γσ c .

Problems

1. (a)

The absolute risk aversion coefficient is u00 (c) . u0 (c) We scale by u0 (c) because expected utility is only defined up to linear transformations – a + bu(c) gives the same predictions as u(c) – and this measure of the second derivative is invariant to linear transformations. Show that the utility 38

S ECTION 1.6

P ROBLEMS

function with constant absolute risk aversion is u(c) = −e−αc . (b)

The coefficient of relative risk aversion in a one-period model (i.e. when consumption equals wealth) is defined as rra =

(c)

For power utility u0 (c) = c−γ , show that the risk aversion coefficient equals the power. The elasticity of intertemporal substitution is defined as ξI ≡ −

2. 3.

c2 /c1 d(c1 /c2 ) . dR/R

Show that with power utility u0 (c) = c−γ , the intertemporal substitution elasticity is equal to 1/γ. Show that the “idiosyncratic risk” line in Figure 1 is horizontal. (a)

(b)

4.

cu00 (c) . u0 (c)

Suppose you have a mean-variance efficient return Rmv and the risk free rate. using the fact that Rmv is perfectly correlated with the discount factor, construct a discount factor m in terms of Rf and Rmv , with no free parameters. (the constants in m = a + bRmv will be functions of things like E(Rmv )) Using this result, and the beta model in terms of m, show that expected returns can be described in a single - beta representation using any mean-variance efficient return (except the riskfree rate). £ ¤ E(Ri ) = Rf + β i,mv E(Rmv ) − Rf .

Can the “Sharpe ratio” between two risky assets exceed the slope of the mean-variance frontier? I.e. if Rmv is on the frontier, is it possible that E(Rmv ) − Rf E(Ri ) − E(Rj ) > ? σ(Ri − Rj ) σ(Rmv )

5.

6.

Show that if consumption growth is lognormal, then ¯ ¯ −γ p ¯ E(Rmv ) − Rf ¯ ¯ ¯ = σ h[(ct+1 /ct ) ]i = eγ 2 σ2 (∆ ln ct+1 ) − 1 ≈ γσ(∆ ln c). ¯ σ(Rmv ) ¯ E (ct+1 /ct )−γ

1

2

(Start with σ 2 (x) = E(x2 ) − E(x)2 and the lognormal property E(ez ) = eEz+ 2 σ (z) .) There are assets with mean return equal to the riskfree rate, but substantial standard deviation of returns. Long term bonds are pretty close examples. Why would anyone 39

C HAPTER 1

7.

C ONSUMPTION - BASED MODEL AND OVERVIEW

hold such an asset? The first order conditions for an infinitely lived consumer who can buy an asset with dividend stream {dt } are pt = Et

∞ X j=1

βj

u0 (ct+j ) dt+j . u0 (ct )

(39)

The first order conditions for buying a security with price pt and payoff xt+1 = dt+1 + pt+1 are ¸ · 0 u (ct+1 ) (pt+1 + dt+1 ) pt = Et β 0 (40) u (ct )

(a) (b)

8.

9.

Derive (1.40) from (1.39) Derive (1.39) from (1.40). You need an extra condition. Show that this extra condition is a first order condition for maximization. To do this, think about what strategy the consumer could follow to improve utility if the condition did not hold. Suppose a consumer has a utility function that includes leisure. (This could also be a second good, or a good produced in another country.) Using the continuous time setup, show that expected returns will now depend on two covariances, the covariance of returns with leisure and the covariance of returns with consumption, so long as leisure enters non-separably, i.e. u(c, l) cannot be written v(c) + w(l). (This is a three line problem, but you need to apply Ito’s lemma to Λ.) From 1 = E(mR) show that the negative of the mean log discount factor must be larger than any mean return, −E(ln m) > E(ln R). How is it possible that E(ln R) is bounded – what about returns of the form R = (1 − α)Rf + αRm for arbitrarily large α? (Hint: start by assuming m and R are lognormal. Then see if you can generalize the results using Jensen’s inequality, E(f (x)) > f(E(x)) for f convex. The return that solves maxR E(ln R) is known as the growth optimal portfolio.)

40

Chapter 2. 2.1

Applying the basic model

Assumptions and applicability

Writing p = E(mx), we do not assume 1. 2. 3. 4. 5.

Markets are complete, or there is a representative investor Asset returns or payoffs are normally distributed (no options), or independent over time. Two period investors, quadratic utility, or separable utility Investors have no human capital or labor income The market has reached equilibrium, or individuals have bought all the securities they want to. All of these assumptions come later, in various special cases, but we haven’t made them yet. We do assume that the investor can consider a small marginal investment or disinvestment. The theory of asset pricing contains lots of assumptions to derive analytically convenient special cases and empirically useful representations. In writing p = E(mx) or pu0 (ct ) = Et [βu0 (ct+1 )xt+1 ] we have not made most of these assumptions. We have not assumed complete markets or a representative investor. These equations apply to each individual investor, for each asset to which he has access, independently of the presence or absence of other investors or other assets. Complete markets/representative agent assumptions are used if one wants to use aggregate consumption data in u0 (ct ), or other specializations and simplifications of the model. We have not said anything about payoff or return distributions. In particular, we have not assumed that returns are normally distributed or that utility is quadratic. The basic pricing equation should hold for any asset, stock, bond, option, real investment opportunity, etc., and any monotone and concave utility function. In particular, it is often thought that mean-variance analysis and beta pricing models require these kind of limiting assumptions or quadratic utility, but that is not the case. A mean-variance efficient return carries all pricing information no matter what the distribution of payoffs, utility function, etc. This is not a “two-period model.” The fundamental pricing equation holds for any two periods of a multi-period model, as we have seen. Really, everything involves conditional moments, so we have not assumed i.i.d. returns over time. I have written things down in terms of a time- and state-separable utility function and I have extensively used the convenient power utility example. Nothing important lies in either 41

C HAPTER 2

A PPLYING THE BASIC MODEL

choice. Just interpret u0 (ct ) as the partial derivative of a general utility function with respect to consumption at time t. State- or time-nonseparable utility (habit persistence, durability) complicates the relation between the discount factor and real variables, but does not change p = E(mx) or any of the basic structure. We will look at several examples below. We do not assume that investors have no non-marketable human capital, or no outside sources of income. The first order conditions for purchase of an asset relative to consumption hold no matter what else is in the budget constraint. By contrast, the portfolio approach to asset pricing as in the CAPM and ICAPM relies heavily on the assumption that the investor has no non-asset income, and we will study these special cases below. For example, leisure in the utility function just means that u0 (c, l) may depend on l as well as c. We don’t even really need the assumption (yet) that the market is “in equilibrium,” that investor has bought all of the asset that he wants to, or even that he can buy the asset at all. We can interpret p = E(mx) as giving us the value, or willingness to pay for, a small amount of a payoff xt+1 that the investor does not yet have. Here’s why: If the investor had a little ξ more of the payoff xt+1 at time t + 1, his utility u(ct ) + βEt u(ct+1 ) would increase by ¸ · 1 βEt [u(ct+1 + ξxt+1 ) − u(ct+1 )] = βEt u0 (ct+1 )xt+1 ξ + u00 (ct+1 ) (xt+1 ξ)2 + ... 2 If ξ is small, only the first term on the right matters. If the investor has to give up a small amount of money vt ξ at time t, that loss lowers his utility by 1 2 u(ct − vt ξ) = u0 (ct )vt ξ + u00 (ct ) (vt ξ) + .... 2 Again, for small ξ, only the first term matters. Therefore, in order to receive the small extra payoff ξxt+1 , the investor is willing to pay the small amount vt ξ where · 0 ¸ u (ct+1 ) vt = Et β 0 xt+1 . u (ct ) If this private valuation is higher than the market value pt , and if the investor can buy some more of the asset, he will. As he buys more, his consumption will change; it will be higher in states where xt+1 is higher, driving down u0 (ct+1 ) in those states, until the value to the investor has declined to equal the market value. Thus, after an investor has reached his optimal portfolio, the market value should obey the basic pricing equation as well, using post-trade or equilibrium consumption. But the formula can also be applied to generate the marginal private valuation, using pre-trade consumption, or to value a potential, not yet traded security. We have calculated the value of a “small” or marginal portfolio change for the investor. For some investment projects, an investor cannot take a small (“diversified”) position. For example, a venture capitalist or entrepreneur must usually take all or nothing P of a project with payoff stream {xt }. Then the value of a project not already taken, E j β j u(ct+j + xt+j ) 42

S ECTION 2.2

G ENERAL E QUILIBRIUM

P might be substantially different from its marginal counterpart, E β j u0 (ct+j )xt+j . Once the project is taken of course, ct+j + xt+j becomes ct+j , so the marginal valuation still applies to the ex-post consumption stream. Analysts often forget this point and apply marginal (diversified) valuation models such as the CAPM to projects that must be bought in discrete chunks. Also, we have abstracted from short sales and bid/ask spreads; this modification changes p = E(mx) from an equality to a set of inequalities.

2.2

General Equilibrium

Asset returns and consumption: which is the chicken and which is the egg? The exogenous return model, the endowment economy model, and the argument that it doesn’t matter for studying p = E(mx). So far, we have not said where the joint statistical properties of the payoff xt+1 and marginal utility mt+1 or consumption ct+1 come from. We have also not said anything about the fundamental exogenous shocks that drive the economy. The basic pricing equation p = E(mx) tells us only what the price should be, given the joint distribution of consumption (marginal utility, discount factor) and the asset payoff. There is nothing that stops us from writing the basic pricing equation as u0 (ct ) = Et [βu0 (ct+1 )xt+1 /pt ] . We can think of this equation as determining today’s consumption given asset prices and payoffs, rather than determining today’s asset price in terms of consumption and payoffs. Thinking about the basic first order condition in this way gives the permanent income model of consumption. Which is the chicken and which is the egg? Which variable is exogenous and which is endogenous? The answer is, neither, and for many purposes, it doesn’t matter. The first order conditions characterize any equilibrium; if you happen to know E(mx), you can use them to determine p; if you happen to know p, you can use them to determine consumption and savings decisions. For most asset pricing applications we are interested in understanding a wide cross-section of assets. Thus, it is interesting to contrast the cross-sectional variation in their prices (expected returns) with cross-sectional variation in their second moments (betas) with a single discount factor. In most applications, the discount factor is a function of aggregate variables (market return, aggregate consumption), so is plausible to hold the properties of the discount factor constant as we study one individual asset after another. Permanent income studies typically dramatically restrict the number of assets under consideration, often to just an interest rate, and study the time-series evolution of aggregate or individual consumption. Nonetheless, it is an obvious next step to complete the solution of our model economy; to 43

C HAPTER 2

A PPLYING THE BASIC MODEL

find c and p in terms of truly exogenous forces. The results will of course depend on what the rest of the economy looks like, in particular the production or intertemporal transformation technology and the set of markets. Figure 2 shows one possibility for a general equilibrium. Suppose that the production technologies are linear: the real, physical rate of return (the rate of intertemporal transformation) is not affected by how much is invested. Now consumption must adjust to these technologically given rates of return. If the rates of return on the intertemporal technologies were to change, the consumption process would have to change as well. This is, implicitly, how the permanent income model works. This is how many finance theories such as the CAPM and ICAPM and the Cox, Ingersoll and Ross (1986) model of the term structure work as well. These models specify the return process, and then solve the consumer’s portfolio and consumption rules.

Ct+1

R Ct Figure 2. Consumption adjusts when the rate of return is determined by a linear technology. Figure 3 shows another extreme possibility for the production technology. This is an “endowment economy.” Nondurable consumption appears (or is produced by labor) every period. There is nothing anyone can do to save, store, invest or otherwise transform consumption goods this period to consumption goods next period. Hence, asset prices must adjust until people are just happy consuming the endowment process. In this case consumption is exogenous and asset prices adjust. Lucas (1978) and Mehra and Prescott (1985) are two very famous applications of this sort of “endowment economy.” Which of these possibilities is correct? Well, neither, of course. The real economy and all serious general equilibrium models look something like figure 4: one can save or transform consumption from one date to the next, but at a decreasing rate. As investment increases, 44

S ECTION 2.2

G ENERAL E QUILIBRIUM

Ct+1

R Ct Figure 3. Asset prices adjust to consumption in an endowment economy. rates of return decline

Ct+1

R Ct Figure 4. General equilibrium. The solid lines represent the indifference curve and production possibility set. The dashed straight line represents the equilibrium rate of return. The dashed box represents an endowment economy that predicts the same consumption-asset return process.

45

C HAPTER 2

A PPLYING THE BASIC MODEL

Does this observation invalidate any modeling we do with the linear technology (CAPM, CIR, permanent income) model, or the endowment economy model? No. Start at the equilibrium in figure 4. Suppose we model this economy as a linear technology, but we happen to choose for the rate of return on the linear technologies exactly the same stochastic process for returns that emerges from the general equilibrium. The resulting joint consumption, asset return process is exactly the same as in the original general equilibrium! Similarly, suppose we model this economy as an endowment economy, but we happen to choose for the endowment process exactly the stochastic process for consumption that emerges from the equilibrium with a concave technology. Again, the joint consumption-asset return process is exactly the same. Therefore, there is nothing wrong in adopting one of the following strategies for empirical work: 1. 2. 3.

Form a statistical model of bond and stock returns, solve the optimal consumptionportfolio decision. Use the equilibrium consumption values in p = E(mx). Form a statistical model of the consumption process, calculate asset prices and returns directly from the basic pricing equation p = E(mx). Form a completely correct general equilibrium model, including the production technology, utility function and specification of the market structure. Derive the equilibrium consumption and asset price process, including p = E(mx) as one of the equilibrium conditions. If the statistical models for consumption and/or asset returns are right, i.e. if they coincide with the equilibrium consumption or return process generated by the true economy, either of the first two approaches will give correct predictions for the joint consumption-asset return process. As we will see, most finance models, developed from the 1950s through the early 1970s, take the return process as given, implicitly assuming linear technologies. The endowment economy approach, introduced by Lucas (1978), is a breakthrough because it turns out to be much easier. It is much easier to evaluate p = E(mx) for fixed m than it is to solve joint consumption-portfolio problems for given asset returns, all to derive the equilibrium consumption process. To solve a consumption-portfolio problem we have to model the investor’s entire environment: we have to specify all the assets to which he has access, what his labor income process looks like (or wage rate process, and include a labor supply decision). Once we model the consumption stream directly, we can look at each asset in isolation, and the actual computation is almost trivial. This breakthrough accounts for the unusual structure of the presentation in this book. It is traditional to start with an extensive study of consumption-portfolio problems. But by modeling consumption directly, we have been able to study pricing directly, and portfolio problems are an interesting side trip which we can defer. Most uses of p = E(mx) do not require us to take any stand on exogeneity or endogeneity, or general equilibrium. This is a condition that must hold for any asset, for any 46

S ECTION 2.3

C ONSUMPTION - BASED MODEL IN PRACTICE

production technology. Having a taste of the extra assumptions required for a general equilibrium model, you can now appreciate why people stop short of full solutions when they can address an application using only the first order conditions, using knowledge of E(mx) to make a prediction about p. It is enormously tempting to slide into an interpretation that E(mx) determines p. We routinely think of betas and factor risk prices – components of E(mx) – as determining expected returns. For example, we routinely say things like “the expected return of a stock increased because the firm took on riskier projects, thereby increasing its β.” But the whole consumption process, discount factor, and factor risk premia change when the production technology changes. Similarly, we are on thin ice if we say anything about the effects of policy interventions, new markets and so on. The equilibrium consumption or asset return process one has modeled statistically may change in response to such changes in structure. For such questions one really needs to start thinking in general equilibrium terms. It may help to remember that there is an army of permanent-income macroeconomists who make precisely the opposite assumption, taking our asset return processes as exogenous and studying (endogenous) consumption and savings decisions.

2.3

Consumption-based model in practice

The consumption-based model is, in principle, a complete answer to all asset pricing questions, but works poorly in practice. This observation motivates other asset pricing models. The model I have sketched so far can, in principle, give a compete answer to all the questions of the theory of valuation. It can be applied to any security—bonds, stocks, options, futures, etc.—or to any uncertain cash flow. All we need is a functional form for utility, numerical values for the parameters, and a statistical model for the conditional distribution of consumption and payoffs. To be specific, consider the standard power utility function u0 (c) = c−γ .

(41)

" µ # ¶−γ ct+1 e 0 = Et β Rt+1 ct

(42)

Then, excess returns should obey

Taking unconditional expectations and applying the covariance decomposition, expected ex47

C HAPTER 2

A PPLYING THE BASIC MODEL

cess returns should follow e E(Rt+1 )

f

= −R cov

"µ

ct+1 ct

¶−γ

,

e Rt+1

#

.

(43)

Given a value for γ, and data on consumption and returns, one can easily estimate the mean and covariance on the right hand side, and check whether actual expected returns are, in fact, in accordance with the formula. Similarly, the present value formula is pt = Et

∞ X

βj

j=1

µ

ct+j ct

¶−γ

(44)

dt+j .

Given data on consumption and dividends or another stream of payoffs, we can estimate the right hand side and check it against prices on the left. Bonds and options do not require separate valuation theories. For example, an N-period default-free nominal discount bond (a U.S. Treasury strip) is a claim to one dollar at time t + N . Its price should be Ã

pt = Et β

N

µ

ct+N ct

¶−γ

! Πt 1 Πt+N

where Π = price level ($/good). A European option is a claim to the payoff max(St+T − K, 0), where St+T = stock price at time t + T, K = strike price. The option price should be "

pt = Et β

T

µ

ct+T ct

¶−γ

#

max(St+T − K, 0)

again, we can use data on consumption, prices and payoffs to check these predictions. Unfortunately, the above specification of the consumption-based model does not work very well. To give a flavor of some of the problems, Figure 5 presents the mean excess returns on the ten size-ranked portfolios of NYSE stocks vs. the predictions – the right hand side of (2.43) – of the consumption-based model. I picked the utility curvature parameter γ = 241 to make the picture look as good as possible (The section on GMM estimation below goes into detail on how to do this. The Figure presents the first-stage GMM estimate.) As you can see, the model isn’t hopeless–there is some correlation between sample average returns and the consumption-based model predictions. But the model does not do very well. The pricing error (actual expected return - predicted expected return) for each portfolio is of the same order of magnitude as the spread in expected returns across the portfolios. 48

S ECTION 2.4

A LTERNATIVE ASSET PRICING MODELS : OVERVIEW

Figure 5. Mean excess returns of 10 CRSP size portfolios vs. predictions of the power utility consumption-based model. The predictions are generated by −Rf cov(m, Ri ) with m = β(ct+1 /ct )−γ . β = 0.98 and γ = 241 are picked by first-stage GMM to minimize the sum of squared pricing errors (deviation from 45◦ line). Source: Cochrane (1996).

2.4

Alternative asset pricing models: Overview

I motivate exploration of different utility functions, general equilibrium models, and linear factor models such as the CAPM, APT and ICAPM as approaches to circumvent the empirical difficulties of the consumption-based model. The poor empirical performance of the consumption-based model motivates a search for alternative asset pricing models – alternative functions m = f (data). All asset pricing models amount to different functions for m. I give here a bare sketch of some of the different approaches; we study each in detail in later chapters. 1) Different utility functions. Perhaps the problem with the consumption-based model is simply the functional form we chose for utility. The natural response is to try different utility functions. Which variables determine marginal utility is a far more important question than the functional form. Perhaps the stock of durable goods influences the marginal utility of nondurable goods; perhaps leisure or yesterday’s consumption affect today’s marginal utility. These possibilities are all instances of nonseparabilities. One can also try to use micro data on individual consumption of stockholders rather than aggregate consumption. Aggregation of heterogenous investors can make variables such as the cross-sectional variance of income 49

C HAPTER 2

A PPLYING THE BASIC MODEL

appear in aggregate marginal utility. 2) General equilibrium models. Perhaps the problem is simply with the consumption data. General equilibrium models deliver equilibrium decision rules linking consumption to other variables, such as income, investment, etc. Substituting the decision rules ct = f(yt , it , . . . ) in the consumption-based model, we can link asset prices to other, hopefully better-measured macroeconomic aggregates. In addition, true general equilibrium models completely describe the economy, including the stochastic process followed by all variables. They can answer questions such as why is the covariance (beta) of an asset payoff x with the discount factor m the value that it is, rather than take this covariance as a primitive. They can in principle answer structural questions, such as how asset prices might be affected by different government policies. Neither kind of question can be answered by just manipulating investor first order conditions. 3) Factor pricing models. Another sensible response to bad consumption data is to model marginal utility in terms of other variables directly. Factor pricing models follow this approach. They just specify that the discount factor is a linear function of a set of proxies, A B + bB ft+1 +... . mt+1 = a + bA ft+1

(45)

where f i are factors and a, bi are parameters. (This is a different sense of the use of the word “factor” than “discount factor.” I didn’t invent the confusing terminology.) By and large, the factors are just selected as plausible proxies for marginal utility; events that describe whether typical investors are happy or unhappy. Among others, the Capital Asset Pricing Model (CAPM) is the model mt+1 = a + bRW t+1 where RW is the rate of return on a claim to total wealth, often proxied by a broad-based portfolio such as the value-weighted NYSE portfolio. The Arbitrage Pricing Theory (APT) uses returns on broad-based portfolios derived from a factor analysis of the return covariance matrix. The Intertemporal Capital Asset Pricing Model (ICAPM) suggests macroeconomic variables such as GNP and inflation and variables that forecast macroeconomic variables or asset returns as factors. Term structure models such as the Cox-Ingersoll-Ross model specify that the discount factor is a function of a few term structure variables, for example the short rate of interest and a few interest rate spreads. Many factor pricing models are derived as general equilibrium models with linear technologies and no labor income; thus they also fall into the general idea of using general equilibrium relations (from, admittedly, very stylized general equilibrium models) to substitute out for consumption. 4) Arbitrage or near-arbitrage pricing. The mere existence of a representation p = E(mx) and the fact that marginal utility is positive m ≥ 0 (these facts are discussed in the next chapter) can often be used to deduce prices of one payoff in terms of the prices of other payoffs. The Black-Scholes option pricing model is the paradigm of this approach: 50

S ECTION 2.5

P ROBLEMS

Since the option payoff can be replicated by a portfolio of stock and bond, any m that prices the stock and bond gives the price for the option. Recently, there have been several suggestions on how to use this idea in more general circumstances by using very weak further restrictions on m, and we will study these suggestions in Chapter 17. We return to a more detailed derivation and discussion of these alternative models of the discount factor m below. First, and with this brief overview in mind, we look at p = E(mx) and what the discount factor m represents in a little more detail.

2.5 1.

Problems

The representative consumer maximizes a CRRA utility function. X j 1−γ β ct+j . Et Consumption is given by an endowment stream.

(a)

2.

Show that with log utility, the price/consumption ratio of the consumption stream is constant, no matter what the distribution of consumption growth. (b) Suppose there is news at time t that future consumption will be higher. For γ < 1, γ = 1,and γ > 1, evaluate the effect of this news on the price. Make sense of your results. (Note: there is a real-world interpretation here. It’s often regarded as a puzzle that the market declines on good economic news. This is attributed to an expectation by the market that the Fed will respond to such news by raising interest rates. Note that γ > 0 in this problem gives a completely real and frictionless interpretation to this phenomenon! I thank Pete Hecht for this nice problem.) The linear quadratic permanent income model is a very useful general equilibrium model that we can solve in closed form. It specifies a production technology rather than fixed endowments, and it easily allows aggregation of disparate consumers. (Hansen 1987 is a wonderful exposition of what one can do with this setup.) The consumer maximizes µ ¶ ∞ X 1 βt − (ct − c∗ )2 E 2 t=0 subject to a linear technology kt+1 = (1 + r)kt + it it = et − ct et is an exogenous endowment or labor income stream. Assume β = 1/(1 + r); the discount rate equals the interest rate or marginal productivity of capital. 51

C HAPTER 2 (a)

A PPLYING THE BASIC MODEL

Show that optimal consumption follows ct = rkt + rβ

∞ X

β j Et et+j

j=0 ∞ X

ct = ct−1 + (Et − Et−1 ) rβ

(b)

β j et+j

(2.46) (2.47)

j=0

i.e., consumption equals permanent income, precisely defined, and consumption follows a random walk whose innovations are equal to innovations in permanent income. Assume that the endowment et follows an AR(1) et = ρet−1 + εt

3.

and specialize (2.46) and (2.47). Calculate and interpret the result for ρ = 1 and ρ = 0. (The result looks like a “consumption function” relating consumption to capital and current income, except that the slope of that function depends on the persistence of income shocks. Transitory shocks will have little effect on consumption, and permanent shocks a larger effect.) (c) Calculate the one period interest rate (it should come out to r of course) and the price of a claim to the consumption stream. e and k are the only state variables, so the price should be a function of e and k. Interpret the time-variation in the price of the consumption stream. (This consumer gets more risk averse as consumption rises to c∗ . c∗ is the bliss point, so at the bliss point there is no average return that can compensate the consumer for greater risk.) Consider again CRRA utility, Et

X

β j c1−γ t+j .

Consumption growth follows a two-state Markov process. The states are ∆ct = ct /ct−1 = h, l, and a 2×2 matrix π governs the set of transition probabilities, i.e. pr(∆ct+1 = h|∆ct = l) = πl→h . (This is the Mehra-Prescott 1986 model, but it will be faster to do it than to look it up. It is a useful and simple endowment economy.) (a)

Find the riskfree rate (price of a certain real payoff of one) in this economy. This price is generated by pbt = Et (mt,t+1 1).

(b)

You are looking for two values, the price in the l state and the price in the h state. Find the price of the consumption stream (the price at t of {ct+1 , ct+2 , ...}). To do this, guess that the price/consumption ratio must be a function of state (h,l), and find 52

S ECTION 2.5

P ROBLEMS

that function. From ¡ ¢ pct = Et mt,t+1 (pct+1 + ct+1 )

(c)

(d)

find a recursive relation for pct /ct , and hence find the two values of pct /ct , one for the h state and one for the l state. Pick β = 0.99 and try γ = 0.5, 5 (Try more if you feel like it). Calibrate the consumption process to have a 1% mean and 1% standard deviation, and consumption growth uncorrelated over time. Calculate prices and returns in each state. Now introduce serial correlation in consumption growth with γ = 5. (You can do this by adding weight to the diagonal entries of the transition matrix π.) What effect does this have on the model?

53

Chapter 3.

Contingent Claims Markets

Our first task is to understand the p = E(mx) representation a little more deeply. In this chapter I introduce a very simple market structure, contingent claims. This leads us to an inner product interpretation of p = E(mx) which allows an intuitive visual representation of most of the theorems. We see that discount factors exist, are positive, and the pricing function is linear, just starting from prices and payoffs in a complete market, without any utility functions. The next chapter shows that these properties can be built up in incomplete markets as well.

3.1

Contingent claims

I describe contingent claims. I interpret the stochastic discount factor m as contingent claims prices divided by probabilities, and p = E(mx) as a bundling of contingent claims. Suppose that one of S possible states of nature can occur tomorrow, i.e. specialize to a finite-dimensional state space. Denote the individual states by s. For example, we might have S = 2 and s = rain or s = shine. A contingent claim is a security that pays one dollar (or one unit of the consumption good) in one state s only tomorrow. pc(s) is the price today of the contingent claim. I write pc to specify that it is the price of a contingent claim and (s) to denote in which state s the claim pays off. In a complete market investors can buy any contingent claim. They don’t necessarily have to be faced with explicit contingent claims; they just need enough other securities to span or synthesize all contingent claims. For example, if the possible states of nature are (rain, shine), one can span or synthesize any contingent claim or portfolio achieved by combining contingent claims by forming portfolios of a security that pays 2 dollars if it rains and one if it shines, or x1 = (2, 1), and a riskfree security whose payoff pattern is x2 = (1, 1). Now, we are on a hunt for discount factors, and the central point is: If there are complete contingent claims, a discount factor exists, and it is equal to the contingent claim price divided by probabilities. Let x(s) denote an asset’s payoff in state of nature s. We can think of the asset as a bundle of contingent claims—x(1) contingent claims to state 1, x(2) claims to state 2, etc. The asset’s price must then equal the value of the contingent claims of which it is a bundle,

p(x) =

X

pc(s)x(s).

s

54

(48)

S ECTION 3.2

R ISK NEUTRAL PROBABILITIES

I denote the price p(x) to emphasize that it is the price of the payoff x. Where the payoff in question is clear, I suppress the (x). I like to think of equation (3.48) as a happy-meal theorem: the price of a happy meal (in a frictionless market) should be the same as the price of one hamburger, one small fries, one small drink and the toy. It is easier to take expectations rather than sum over states. To this end, multiply and divide the bundling equation (3.48) by probabilities, µ ¶ X pc(s) π(s) x(s) p(x) = π(s) s where π(s) is the probability that state s occurs. Then define m as the ratio of contingent claim price to probability m(s) =

pc(s) . π(s)

Now we can write the bundling equation as an expectation, X p= π(s)m(s)x(s) = E(mx). s

Thus, in a complete market, the stochastic discount factor m in p = E(mx) exists, and it is just a set of contingent claims prices, scaled by probabilities. As a result of this interpretation, the combination of discount factor and probability is sometimes called a state-price density. The multiplication and division by probabilities seems very artificial in this finite-state context. In general, we posit states of nature ω that can take continuous (uncountably infinite) values in a space Ω. In this case, the sums become integrals, and we have to use some measure to integrate over Ω. Thus, scaling contingent claims prices by some probability-like object is unavoidable.

3.2

Risk neutral probabilities

I interpret the discount factor m as a transformation to risk-neutral probabilities such that p = E ∗ (x)/Rf . Another common transformation of p = E(mx) results in “risk-neutral” probabilities. Define π∗ (s) ≡ Rf m(s)π(s) = Rf pc(s) 55

C HAPTER 3

C ONTINGENT C LAIMS M ARKETS

where Rf ≡ 1/

X

pc(s) = 1/E(m).

The π∗ (s) are positive, less than or equal to one and sum to one, so they are a legitimate set of probabilities. Then we can rewrite the asset pricing formula as p(x) =

X

pc(s)x(s) =

s

1 X ∗ E ∗ (x) (s)x(s) = . π Rf Rf

I use the notation E ∗ to remind us that the expectation uses the risk neutral probabilities π∗ instead of the real probabilities π. Thus, we can think of asset pricing as if agents are all risk neutral, but with probabilities π∗ in the place of the true probabilities π. The probabilities π∗ gives greater weight to states with higher than average marginal utility m. There is something very deep in this idea: risk aversion is equivalent to paying more attention to unpleasant states, relative to their actual probability of occurrence. People who report high subjective probabilities of unpleasant events like plane crashes may not have irrational expectations, they may simply be reporting the risk neutral probabilities or the product m × π. This product is after all the most important piece of information for many decisions: pay a lot of attention to contingencies that are either highly probable or that are improbable but have disastrous consequences. The transformation from actual to risk-neutral probabilities is given by π∗ (s) =

m(s) π(s). E(m)

We can also think of the discount factor m as the derivative or change of measure from the real probabilities π to the subjective probabilities π ∗ . The risk-neutral probability representation of asset pricing is quite common, especially in derivative pricing where the results are independent of risk adjustments. The risk-neutral representation is particularly popular in continuous time diffusion processes, because we can adjust only the means, leaving the covariances alone. In discrete time, changing the probabilities typically changes first and second moments. Suppose we start with a process for prices and discount factor dp p dΛ Λ

= µp dt + σ p dz = µΛ dt + σΛ dz. 56

S ECTION 3.3

I NVESTORS AGAIN

The discount factor prices the assets, µ ¶ µ ¶ dp dΛ dp D Et + dt − rf dt = −Et = −σp σ Λ dt p p Λ p In the “risk-neutral measure” we just increase the drift of each price process by its covariance with the discount factor, and write a risk-neutral discount factor, dp p dΛ Λ

=

¡ p ¢ µ + σp σΛ dt + σ p dz = µp∗ dt + σ p dz

= µΛ dt.

Under this new set of probabilities, we can just write, µ ¶ dp D Et∗ + dt − rf dt = 0 p p with Et∗ (dp/p) = µp∗ dt.

3.3

Investors again

We look at investor’s first order conditions in a contingent claims market. The marginal rate of substitution equals the discount factor and the contingent claim price ratio. Though the focus of this chapter is on how to do without utility functions, It’s worth looking at the investor’s first order conditions again in the contingent claim context. The investor starts with a pile of initial wealth y and a state-contingent income y(s). He purchases contingent claims to each possible state in the second period. His problem is then X X X βπ(s)u[c(s)] s.t. c + pc(s)c(s) = y + pc(s)y(s). max u(c) + {c,c(s)}

s

s

s

Introducing a Lagrange multiplier λ on the budget constraint, the first order conditions are u0 (c) = λ βπ(s)u0 [c(s)] = λpc(s).

Eliminating the Lagrange multiplier λ, pc(s) = βπ(s) 57

u0 [c(s)] u0 (c)

C HAPTER 3

C ONTINGENT C LAIMS M ARKETS

or m(s) =

pc(s) u0 [c(s)] =β 0 π(s) u (c)

Coupled with p = E(mx), we obtain the consumption-based model again. The investor’s first order conditions say that marginal rates of substitution between states tomorrow equals the relevant price ratio, u0 [c(s1 )] m(s1 ) = 0 . m(s2 ) u [c(s2 )] m(s1 )/m(s2 ) gives the rate at which the investor can give up consumption in state 2 in return for consumption in state 1 through purchase and sales of contingent claims. u0 [c(s1 )]/u0 [c(s2 )] gives the rate at which the investor is willing to make this substitution. At an optimum, the marginal rate of substitution should equal the price ratio, as usual in economics. We learn that the discount factor m is the marginal rate of substitution between date and state contingent commodities. That’s why it, like c(s), is a random variable. Also, scaling contingent claims prices by probabilities gives marginal utility, and so is not so artificial as it may have seemed above. Figure 6 gives the economics behind this approach to asset pricing. We observe the investor’s choice of date or state-contingent consumption. Once we know his utility function, we can calculate the contingent claim prices that must have led to the observed consumption choice, from the derivatives of the utility function. State 2 or date 2

(c1, c2) Indifference curve

State 1, or date 1

Figure 6. Indifference curve and contingent claim prices The relevant probabilities are the investor’s subjective probabilities over the various states. 58

S ECTION 3.4

R ISK SHARING

Asset prices are set, after all, by investor’s demands for assets, and those demands are set by investor’s subjective evaluations of the probabilities of various events. We often assume rational expectations, namely that subjective probabilities are equal to objective frequencies. But this is an additional assumption that we may not always want to make.

3.4

Risk sharing

Risk sharing: In complete markets, consumption moves together. Only aggregate risk matters for security markets. We deduced that the marginal rate of substitution for any individual investor equals the contingent claim price ratio. But the prices are the same for all investors. Therefore, marginal utility growth should be the same for all investors βi

0 j u0 (cit+1 ) j u (ct+1 ) = β u0 (cit ) u0 (cjt )

(49)

where i and j refer to different investors. If investors have the same homothetic utility function (for example, power utility), then consumption itself should move in lockstep, cj cit+1 = t+1 . i ct cjt More generally, shocks to consumption are perfectly correlated across individuals. This is so radical, it’s easy to misread it at first glance. It doesn’t say that expected consumption growth should be equal; it says that consumption growth should be equal expost. If my consumption goes up 10%, yours goes up exactly 10% as well, and so does everyone else’s. In a complete contingent claims market, all investors share all risks, so when any shock hits, it hits us all equally (after insurance payments). It doesn’t say the consumption level is the same – this is risk-sharing, not socialism. The rich have higher levels of consumption, but rich and poor share the shocks equally. This risk sharing is Pareto-optimal. Suppose a social planner wished to maximize everyone’s utility given the available resources. For example, with two investors i and j, he would maximize max λi

X

β t u(cit ) + λj

X

β t u(cjt ) s.t. cit + cjt = cat

where ca is the total amount available and λi and λj are i and j’s relative weights in the 59

C HAPTER 3

C ONTINGENT C LAIMS M ARKETS

planner’s objective. The first order condition to this problem is λi u0 (cit ) = λj u0 (cjt ) and hence the same risk sharing that we see in a complete market, equation (3.49). This simple fact has profound implications. First, it shows you why only aggregate shocks should matter for risk prices. Any idiosyncratic income risk will be equally shared, and so 1/N of it becomes an aggregate shock. Then the stochastic discount factors m that determine asset prices are no longer affected by truly idiosyncratic risks. Much of this sense that only aggregate shocks matter stays with us in incomplete markets as well. Obviously, the real economy does not yet have complete markets or full risk sharing – individual consumptions do not move in lockstep. However, this observation tells us much about the function of securities markets. Security markets – state-contingent claims – bring individual consumptions closer together by allowing people to share some risks. In addition, better risk sharing is much of the force behind financial innovation. Many successful new securities can be understood as devices to more widely share risks.

3.5

State diagram and price function

I introduce the state space diagram and inner product representation for prices, p(x) = E(mx) = m · x. p(x) = E(mx) implies p(x) is a linear function.

Think of the contingent claims price pc and asset payoffs x as vectors in RS , where each element gives the price or payoff to the corresponding state, £ ¤0 pc = pc(1) pc(2) · · · pc(S) , x=

£

x(1) x(2) · · ·

x(S)

¤0

.

Figure 7 is a graph of these vectors in RS . Next, I deduce the geometry of Figure 7.

The contingent claims price vector pc points in to the positive orthant. We saw in section 3.3 that m(s) = u0 [c(s)]/u0 (c). Now, marginal utility should always be positive (people always want more), so the marginal rate of substitution and discount factor are always nonnegative, m > 0 and pc > 0. Don’t forget, m and pc are vectors, or random variables. Thus, m > 0 means the realization of m is positive in every state of nature, or, equivalently every element of the vector m is positive. The set of payoffs with any given price lie on a (hyper)plane perpendicular to the contin60

S ECTION 3.5

S TATE DIAGRAM AND PRICE FUNCTION

State 2 Payoff

Price = 2 Riskfree rate pc

Price = 1 (returns)

State 1 contingent claim State 1 Payoff Price = 0 (excess returns)

Figure 7. Contingent claims prices (pc) and payoffs. gent claim price vector. We reasoned above that the price of the payoff x must be given by its contingent claim value (3.48), X p(x) = pc(s)x(s). (50) s

Interpreting pc and x as vectors, this means that the price is given by the inner product of the contingent claim price and the payoff. If two vectors are orthogonal – if they point out from the origin at right angles to each other – then their inner product is zero. Therefore, the set of all zero price payoffs must lie on a plane orthogonal to the contingent claims price vector, as shown in figure 7. More generally, the inner product of two vectors x and pc equals the product of the magnitude of the projection of x onto pc times the magnitude of pc. Using a dot to denote inner product, X pc(s)x(s) = pc · x = |pc| × |proj(x|pc)| = |pc| × |x| × cos(θ) p(x) = s

where |x| means the length of the vector x and θ is the angle between the vectors pc and x. Since all payoffs on planes (such as the price planes in figure 7) that are perpendicular to pc have the same projection onto pc, they must have the same price. (Only the price = 0 plane is, strictly speaking, orthogonal to pc. Lacking a better term, I’ve called the nonzero 61

C HAPTER 3

C ONTINGENT C LAIMS M ARKETS

price planes “perpendicular” to pc.) When vectors are finite-dimensional, the prime notation is commonly used for inner products, pc0 x. This notation does not extend well to infinitedimensional spaces. The notation hpc|xi is also often used for inner products. Planes of constant price move out linearly, and the origin x = 0 must have a price of zero. If payoff y = 2x, then its price is twice the price of x, X X pc(s)y(s) = pc(s)2x(s) = 2 p(x). p(y) = s

s

Similarly, a payoff of zero must have a price of zero. We can think of p(x) as a pricing function, a map from the state space or payoff space in which x lies (RS in this case) to the real line. We have just deduced from the definition (3.50) that p(x) is a linear function, i.e. that p(ax + by) = ap(x) + bp(y). The constant price lines in Figure 7 are of course exactly what one expects from a linear function from RS to R. (One might draw the price on the z axis coming out of the page. Then the price function would be a plane going through the origin and sloping up with isoprice lines as given in Figure 7.) Figure 7 also includes the payoffs to a contingent claim to the first state. This payoff is one in the first state and zero in other states and thus located on the axis. The plane of price = 1 payoffs is the plane of asset returns; the plane of price = 0 payoffs is the plane of excess returns. A riskfree unit payoff (the payoff to a risk-free pure discount bond) lies on the (1, 1) point in Figure 7; the riskfree return lies on the intersection of the 45o line (same payoff in both states) and the price = 1 plane (the set of all returns). Geometry with m in place of pc. The geometric interpretation of Figure 7 goes through with the discount factor m in the place of pc. We can define an inner product between the random variables x and y by x · y ≡ E(xy), and retain all the mathematical properties of an inner product. For this reason, random variables for which E(xy) = 0 are often called “orthogonal.” This language may be familiar from linear regressions. When we run a regression of y on x, y = b0 x + ε we find the linear combination of x that is “closest” to y, by minimizing the variance or “size” of the residual ε. We do this by forcing the residual to be “orthogonal” to the right hand variable E(xε) = 0. The projection of y on x is defined as the fitted value, proj(y|x) = 62

S ECTION 3.5

S TATE DIAGRAM AND PRICE FUNCTION

b0 x =E(xx0 )−1 E(yx0 )x. This ideal is often illustrated by a residual vector ε that is perpendicular to a plane defined by the right hand variables x. Thus, when the inner product is defined by a second moment, the operation “project y onto x” is a regression. (If x does not include a constant, you don’t add one.) The geometric interpretation of Figure 7 also is valid if we generalize the setup to an infinite-dimensional state space, i.e. if we think of continuously-valued random variables. Instead of vectors, which are functions from RS to R, random variables are (measurable) functions from Ω to R. Nonetheless, we can still think of them as vectors. The equivalent of Rs is now a Hilbert space L2 , which denotes spaces generated by linear combinations of square integrable functions from Ω to the real line, or the space of random variables with finite second moments. We can still define an “inner product” between two such elements by x · y = E(xy), and p(x) = E(mx) can still be interpreted as “m is perpendicular to (hyper)planes of constant price.” Proving theorems in this context is a bit harder. You can’t just say things like “we can take a line perpendicular to any plane,” such things have to be proved. Sometimes, finite-dimensional thinking can lead you to errors, so it’s important to prove things the right way, keeping the finite dimensional pictures in mind for interpretation. Hansen and Richard (1987) is a very good reference for the Hilbert space machinery.

63

Chapter 4.

The discount factor

Now we look more closely at the discount factor. Rather than derive a specific discount factor as with the consumption-based model in the last chapter, I work backwards. A discount factor is just some random variable that generates prices from payoffs, p = E(mx). What does this expression mean? Can one always find such a discount factor? Can we use this convenient representation without implicitly assuming all the structure of the investors, utility functions, complete markets, and so forth? The chapter focuses on two famous theorems. The law of one price states that if two portfolios have the same payoffs (in every state of nature), then they must have the same price. A violation of this law would give rise to an immediate kind of arbitrage profit, as you could sell the expensive version and buy the cheap version of the same portfolio. The first theorem states that there is a discount factor that prices all the payoffs by p = E(mx) if and only if this law of one price holds. In finance, we reserve the term absence of arbitrage for a stronger idea, that if payoff A is always at least as good as payoff B, and sometimes A is better, then the price of A must be greater than the price of B. The second theorem is that there is a positive discount factor that prices all the payoffs by p = E(mx) if and only if there are no arbitrage opportunities, so defined. These theorems are useful to show that we can use stochastic discount factors without implicitly assuming anything about utility functions, aggregation, complete markets and so on. All we need to know about investors in order to represent prices and payoffs via a discount factor is that they won’t leave law of one price violations or arbitrage opportunities on the table. These theorems can be used to describe aspects of a payoff space (such as law of one price, absence of arbitrage) by restrictions on the discount factor (such as it exists and it is positive). Chapter 18 shows how it can be more convenient to impose and check restrictions on a single discount factor than it is to check the corresponding restrictions on all possible portfolios. Chapter 7 discusses these and other implications of the existence theorems. The theorems are credited to Ross (1978), and Harrison and Kreps (1979). My presentation follows Hansen and Richard (1987).

4.1

Law of one price and existence of a discount factor

Definition of law of one price; price is a linear function. p = E(mx) implies law of one price. The law of one price implies that a discount factor exists: There exists a unique x∗ in X such that p = E(x∗ x) for all x ∈ X = space of all available payoffs. 64

S ECTION 4.1

L AW OF ONE PRICE AND EXISTENCE OF A DISCOUNT FACTOR

Furthermore, for any valid discount factor m, x∗ = proj(m | X).

So far we have derived the basic pricing relation p = E(mx) from environments with a lot of structure: either the consumption-based model or complete markets. Suppose we observe a set of prices p and payoffs x, and that markets — either the markets faced by investors or the markets under study in a particular application — are incomplete, meaning they do not span the entire set of contingencies. In what minimal set of circumstances does some discount factor exists which represents the observed prices by p = E(mx)? This section and the following answer this important question. This treatment is a simplified version of Hansen and Richard (1987), which contains rigorous proofs and some technical assumptions. Payoff space The payoff space X is the set (or a subset) of all the payoffs that investors can purchase, or it is a subset of the tradeable payoffs that is used in a particular study. For example, if there are complete contingent claims to S states of nature, then X = RS . But the whole point is that markets are (as in real life) incomplete, so we will generally think of X as a proper subset of complete markets RS .

The payoff space will include some set of primitive assets, but investors can also form new payoffs by forming portfolios. I assume that investors can form any portfolio of traded assets: A1: (Portfolio formation) x1 , x2 ∈ X ⇒ ax1 + bx2 ∈ X for any real a, b.

Of course, X = RS for complete markets satisfies the portfolio formation assumption. If there is a single basic payoff x, then the payoff space must be at least the ray from the origin through x. If there are two basic payoffs in R3 , then the payoff space X must include the plane defined by these two payoffs and the origin. Figure 8 illustrates these possibilities. The payoff space is not the space of returns. The return space is a subset of the payoff space; if a return R is in the payoff space, then you can pay a price $2 to get a payoff 2R, so the payoff 2R with price 2 is also in the payoff space. Also, −R is in the payoff space.

Free portfolio formation is in fact an important and restrictive simplifying assumption. It rules out short sales constraints, bid/ask spreads, leverage limitations and so on. The theory can be modified to incorporate these frictions, but it is a substantial modification. If investors can form portfolios of a vector of basic payoffs x (say, the returns on the NYSE stocks), then the payoff space consists of all portfolios or linear combinations of these original payoffs X = {c0 x} where c is a vector of portfolio weights. We also can allow truly infinite-dimensional payoff spaces. For example, investors might be able to trade nonlinear 65

C HAPTER 4

T HE DISCOUNT FACTOR

State 3 (into page) State 2

State 2

x2

X

x1

x

Single Payoff in R2

State 1

Two Payoffs in R3

State 1

Figure 8. Payoff spaces X generated by one (left) and two (right) basis payoffs. functions of a basis payoff x, such as call options on x with strike price K, which have payoff max [x(s) − K, 0] . The law of one price. A2: (Law of one price, linearity) p(ax1 + bx2 ) = ap(x1 ) + bp(x2 ) It doesn’t matter how one forms the payoff x. The price of a burger, shake and fries must be the same as the price of a happy meal. Graphically, if the iso-price curves were not planes, then one could buy two payoffs on the same iso-price curve, form a portfolio whose payoff is on the straight line connecting the two original payoffs, and sell the portfolio for a higher price than it cost to assemble it. The law of one price basically says that investors can’t make instantaneous profits by repackaging portfolios. If investors can sell securities, this is a very weak characterization of preferences. It says there is at least one investor for whom marketing doesn’t matter, who values a package by its contents. The law is meant to describe a market that has already reached equilibrium. If there are any violations of the law of one price, traders will quickly eliminate them so they can’t survive in equilibrium. A1 and A2 also mean that the 0 payoff must be available, and must have price 0. The Theorem The existence of a discount factor implies the law of one price. This is obvious to the point of triviality: if x = y + z then E(mx) = E[m(y + z)]. The hard, and interesting part of the theorem reverses this logic. We show that the law of one price implies the existence of a discount factor.

66

S ECTION 4.1

L AW OF ONE PRICE AND EXISTENCE OF A DISCOUNT FACTOR

Theorem: Given free portfolio formation A1, and the law of one price A2, there exists a unique payoff x∗ ∈ X such that p(x) = E(x∗ x) for all x ∈ X. x∗ is a discount factor. A1 and A2 imply that the price function on X looks like Figure 7: parallel hyperplanes marching out from the origin. The only difference is that X may be a subspace of the original state space, as shown in Figure 8. The essence of the proof, then, is that any linear function on a space X can be represented by inner products with a vector that lies in X. Proof 1: (Geometric.) We have established that the price is a linear function as shown in Figure 9. (Figure 9 can be interpreted as the plane X of a larger dimensional space as in the right hand panel of Figure 8, laid flat on the page for clarity.) Now we can draw a line from the origin perpendicular to the price planes. Choose a vector x∗ on this line. Since the line is orthogonal to the price zero plane we have 0 = p(x) = E(x∗ x) for price zero payoffs x immediately. The inner product between any payoff x on the price = 1 plane and x∗ is |proj(x|x∗ )| × |x∗ | Thus, every payoff on the price = 1 plane has the same inner product with x∗ . All we have to do is pick x∗ to have the right length, and we obtain p(x) = 1 = E(x∗ x) for every x on the price = 1 plane. Then, of course we have p(x) = E(x∗ x) for payoffs x on the other planes as well. Thus, the linear pricing function implied by the Law ¤ of One Price can be represented by inner products with x∗ .

Price = 2 x*

Price = 1 (returns)

Price = 0 (excess returns)

Figure 9. Existence of a discount factor x∗ . The basic mathematical point is just that any linear function can be represented by an 67

C HAPTER 4

T HE DISCOUNT FACTOR

inner product. The the Riesz representation theorem extends the proof to infinite-dimensional payoff spaces. See Hansen and Richard (1987). Proof 2: (Algebraic.) We can prove the theorem by construction when the payoff space X is generated by portfolios of a N basis payoffs (for example, N stocks). This is a common situation, so the formulas are ¤0 also useful in practice. Organize the basis payoffs into a vector £ x = x1 x2 ... xN and similarly their prices p. The payoff space is then X = {c0 x}. We want a discount factor that is in the payoff space, as the theorem requires. Thus, it must be of the form x∗ = c0 x. Construct c so that x∗ prices the basis assets. We want p =E(x∗ x) = E(xx0 c). Thus we need c = E(xx0 )−1 p. If E(xx0 ) is nonsingular, this c exists and is unique. A2 implies that E(xx0 ) is nonsingular (after pruning redundant rows of x). Thus, x∗ = p0 E(xx0 )−1 x

(51)

is our discount factor. It is a linear combination of x so it is in X. It prices the basis assets x by construction. It prices every x ∈ X : E[x∗ (x0 c)] = E[p0 E(xx0 )−1 xx0 c] = p0 c. By 0 linearity, p(c0 x) = c p. What the theorem does and does not say The theorem says there is a unique x∗ in X. There may be many other discount factors m not in X. In fact, unless markets are complete, there are an infinite number of random variables that satisfy p = E(mx). If p = E(mx) then p = E [(m + ε)x] for any ε orthogonal to x, E(εx) = 0. Not only does this construction generate some additional discount factors, it generates all of them: Any discount factor m (any random variable that satisfies p = E(mx)) can be represented as m = x∗ +ε with E(εx) = 0. Figure 10 gives an example of a one-dimensional X in a two-dimensional state space, in which case there is a whole line of possible discount factors m. If markets are complete, there is nowhere to go orthogonal to the payoff space X, so x∗ is the only possible discount factor. Reversing the argument, x∗ is the projection of any stochastic discount factor m on the space X of payoffs. This is a very important fact: the pricing implications of any discount factor m for a set of payoffs X are the same as those of the projection of m on X. This discount factor is known as the mimicking portfolio for m. Algebraically, p = E(mx) = E [(proj(m|X) + ε)x] = E [proj(m|X) x] Let me repeat and emphasize the logic. Above, we started with investors or a contingent claim market, and derived a discount factor. p = E(mx) implies the linearity of the pricing function and hence the law of one price, a pretty obvious statement in those contexts. Here we work backwards. Markets are incomplete in that contingent claims to lots of states of nature are not available. We found that the law of one price implies a linear pricing function, and a linear pricing function implies that there exists at least one and usually many discount factors. 68

S ECTION 4.2

N O -A RBITRAGE AND POSITIVE DISCOUNT FACTORS

Payoff space X

x*

m = x* + ε space of discount factors

Figure 10. Many discount facotors m can price a given set of assets in incomplete markets. We do allow arbitrary portfolio formation, and that sort of “completeness” is important to the result. If investors cannot form a portfolio ax + by, they cannot force the price of this portfolio to equal the price of its constituents. The law of one price is not innocuous; it is an assumption about preferences albeit a weak one. The point of the theorem is that this is just enough information about preferences to deduce the existence of a discount factor.

4.2

No-Arbitrage and positive discount factors

The definition of arbitrage: positive payoff implies positive price. There is a strictly positive discount factor m such that p = E(mx) if and only if there are no arbitrage opportunities. No arbitrage is another, slightly stronger, implication of marginal utility, that can be reversed to show that there is a positive discount factor. We need to start with the definition of arbitrage: Definition (Absence of arbitrage): A payoff space X and pricing function p(x) leave no arbitrage opportunities if every payoff x that is always non-negative, x ≥ 0 (almost surely), and positive, x > 0, with some positive probability, has positive

69

C HAPTER 4

T HE DISCOUNT FACTOR

price, p(x) > 0. No-arbitrage says that you can’t get for free a portfolio that might pay off positively, but will certainly never cost you anything. This definition is different from the colloquial use of the word “arbitrage.” Most people use “arbitrage” to mean a violation of the law of one price – a riskless way of buying something cheap and selling it for a higher price. “Arbitrages” here might pay off, but then again they might not. The word “arbitrage” is also widely abused. “Risk arbitrage” is a Wall Street oxymoron that means making specific kinds of bets. An equivalent statement is that if one payoff dominates another, then its price must be higher – if x ≥ y, then p(x) ≥ p(y) (Or, a bit more carefully but more long-windedly, if x ≥ y almost surely and x > y with positive probability, then p(x) > p(y). You can’t forget that x and y are random variables.) m > 0 ⇒No-arbitrage The absence of arbitrage opportunities is clearly a consequence of a positive discount factor, and a positive discount factor naturally results from any sort of utility maximization. Recall, m(s) = β

u0 [c(s)] > 0. u0 (c)

It is a sensible characterization of preferences that marginal utility is always positive. Few people are so satiated that they will throw away money. Therefore, the marginal rate of substitution is positive. The marginal rate of substitution is a random variable, so “positive” means “positive in every state of nature” or “in every possible realization.” Now, if contingent claims prices are all positive, a bundle of positive amounts of contingent claims must also have a positive price, even in incomplete markets. A little more formally, Theorem: p = E(mx) and m(s) > 0 imply no-arbitrage. Proof: m > 0; x ≥ 0 and there are some states where x > 0. Thus, in some states mx > 0 and in other states mx = 0. Therefore E(mx) > 0. ¤ No arbitrage ⇒ m > 0 Now we turn the observation around, which is again the hard and interesting part. As the law of one price property guaranteed the existence of a discount factor m, no-arbitrage guarantees the existence of a positive m. The basic idea is pretty simple. No-arbitrage means that the prices of any payoff in the positive orthant (except zero, but including the axes) must be strictly positive. The price = 70

S ECTION 4.2

N O -A RBITRAGE AND POSITIVE DISCOUNT FACTORS

0 plane divides the region of positive prices from the region of negative prices. Thus, if the region of negative prices is not to intersect the positive orthant, the iso-price lines must march up and to the right, and the discount factor m, must point up and to the right. This is how we have graphed it all along, most recently in figure 9. Figure 11 illustrates the case that is ruled out: a whole region of negative price payoffs lies in the positive orthant. For example, the payoff x is strictly positive, but has a negative price. As a result, the (unique, since this market is complete) discount factor m is negative in the y-axis state. p = -1

p=0

x

p = +1

x*, m

Figure 11. Counter-example for no-arbitrage ⇒ m > 0 theorem. The payoff x is positive, but has negative price. The discount factor is not strictly positive The theorem is easy to prove in complete markets. There is only one m, x∗ . If it isn’t positive in some state, then the contingent claim in that state has a positive payoff and a negative price, which violates no arbitrage. More formally, Theorem: In complete markets, no-arbitrage implies that there exists a unique m > 0 such that p = E(mx).

Proof: No-arbitrage implies the law of one price, so there is an x∗ such that p = E(x∗ x), and in a complete market this is the unique discount factor. Suppose that in those states, and zero x∗ ≤ 0 for some states. Then, form a payoff x that is 1 P elsewhere. This payoff is strictly positive, but its price, s:x∗ (s) 0, p = E(mx) ∀ x ∈ X. Proof : Join (−p(x), x) together to form vectors in RS+1 . Call M the set of all (−p(x), x) pairs, M = {(−p(x), x); x ∈ X} M is still a linear space: m1 ∈ M, m2 ∈ M ⇒ am1 + bm2 ∈ M. No-arbitrage means that elements of M can’t have all positive elements. If x is positive, −p(x) must be negative. Thus, M is a hyperplane that only intersects the positive orthant at the point 0. We can then create a linear function F : RS+1 ⇒ R such that RS+1 + F (−p, x) = 0 for (−p, x) ∈ M, and F (−p, x) > 0 for (−p, x) ∈ RS+1 except + the origin. Since we can represent any linear function by a perpendicular vector, there is a vector (1, m) such that F (−p, x) = (1, m) · (−p, x) = −p + m · x or −p + E(mx) using the second moment inner product. Finally, since F (−p, x) is positive for (−p, x) > 0, m must be positive. ¤ In a larger space than RS+1 + , as generated by continuously valued random variables, the separating hyperplane theorem assures us that there is a linear function that separates the two convex sets M and the equivalent of RS+1 + , and the Riesz representation theorem tells us that we can represent F as an inner product with some vector by F (−p, x) = −p + m · x. What the theorem does and does not say The theorem says that a discount factor m > 0 exists, but it does not say that m > 0 is unique. The left hand panel of Figure 12 illustrates the situation. Any m on the line through x∗ perpendicular to X also prices assets. Again, p = E[(m + ε)x] if E(εx) = 0. All of these discount factors that lie in the positive orthant are positive, and thus satisfy the theorem. There are lots of them! In a complete market, m is unique, but not otherwise. 72

S ECTION 4.2

N O -A RBITRAGE AND POSITIVE DISCOUNT FACTORS

The theorem says that a positive m exists, but it also does not say that every discount factor m must be positive. The discount factors in the left hand panel of Figure 12 outside the positive orthant are perfectly valid – they satisfy p = E(mx), and the prices they generate on X are arbitrage free, but they are not positive in every state of nature. In particular, the discount factor x∗ in the payoff space is still perfectly valid — p(x) = E(x∗ x) — but it need not be positive, again as illustrated in the left hand panel of Figure 12. p=2 p=1

m m>0

x*

p=1 p=2 X

x* X

Figure 12. Existence of a discount factor and extensions. The left graph shows that the positive discount factor is not unique, and that discount factors may also exist that are not strictly positive. In particular, x∗ need not be positive. The right hand graph shows that each particular choice of m > 0 induces an arbitrage free extension of the prices on X to all contingent claims. This theorem shows that we can extend the pricing function defined on X to all possible payoffs RS , and not imply any arbitrage opportunities on that larger space of payoffs. It says that there is a pricing function p(x) defined over all of RS , that assigns the same (correct, or observed) prices on X and that displays no arbitrage on all of RS . Graphically, it says that we can draw parallel planes to represent prices on all of RS in such a way that the planes intersect X in the right places, and the price planes march up and to the right so the positive orthant always has positive prices. Any positive m generates such a no-arbitrage extension, as illustrated in the right hand panel of Figure 12. In fact, there are many ways to do this. Each different choice of m > 0 generates a different extension of the pricing function. We can think of strictly positive discount factors as possible contingent claims prices. We can think of the theorem as answering the question: is it possible that an observed and incomplete set of prices and payoffs is generated by some complete markets, contingent claim economy? The answer is, yes, if there is no arbitrage on the observed prices and payoffs. In fact, since there are typically many positive m’s consistent with a {X, p(x)}, there exist 73

C HAPTER 4

T HE DISCOUNT FACTOR

many contingent claims economies consistent with our observations. Finally, the absence of arbitrage is another very weak characterization of preferences. The theorem tells us that this is enough to allow us to use the p = E(mx) formalism with m > 0. As usual, this theorem and proof do not require that the state space is RS . State spaces generated by continuous random variables work just as well.

4.3

An alternative formula, and x∗ in continuous time

In terms of the covariance matrix of payoffs, x∗ = E(x∗ ) + [p−E(x∗ )E(x)]0 Σ−1 (x−E(x)). Just like x∗ in discrete time, ¶0 µ D dΛ∗ f − r = −r dt − µ + Σ−1 dz. Λ∗ p prices assets by construction in continuous time. Being able to compute x∗ is useful in many circumstances. This section gives an alternative formula in discrete time, and the continuous time counterpart. A formula that uses covariance matrices E(xx0 ) in our previous formula (4.51) is a second moment matrix. We typically summarize data in terms of covariance matrices instead. Therefore, a convenient alternative formula is x∗ = E(x∗ ) + [p−E(x∗ )E(x)]0 Σ−1 (x−E(x))

(52)

where ¡ ¢ Σ ≡ E [x−E(x)] [x−E(x)]0

denotes the covariance matrix of the x payoffs. (We could just substitute E(xx0 ) = Σ + E(x)E(x0 ), but the inverse of the sum is not very useful.) We can derive this formula by postulating a discount factor that is a linear function of the shocks to the payoffs, x∗ = E(x∗ ) + (x−E(x))0 b, and then finding b to ensure that x∗ prices the assets x : ¤ £ p = E(x∗ )E(x) + E (x−Ex)x0 b 74

S ECTION 4.3

A N ALTERNATIVE FORMULA , AND X∗ IN CONTINUOUS TIME

so b = Σ−1 [p−E(x∗ )E(x)] . If a riskfree rate is traded, then we know E(x∗ ) = 1/Rf . If a riskfree rate is not traded – if 1 is not in X – then this formula does not necessarily produce a discount factor x∗ that is in X. In many applications, however, all that matters is producing some discount factor, and the arbitrariness of the risk-free or zero beta rate is not a problem. This formula is particularly useful when the payoff space consists solely of excess returns or price-zero payoffs. In that case, x∗ = p0 E(xx0 )−1 x gives x∗ = 0. x∗ = 0 is in fact the only discount factor in X that prices all the assets, but in this case it’s more interesting (and avoids 1/0 difficulties when we want to transform to expected return/beta or other representations) to pick a discount factor not in X by picking a zero-beta rate or price of the riskfree payoff. In the case of excess returns, for arbitrarily chosen Rf , then, (4.52) gives us x∗ =

1 1 − f E(Re )0 Σ−1 (Re −E(Re )); Σ ≡ cov(Re ) f R R

Continuous time The law of one price implies the existence of a discount factor process, and absence of arbitrage a positive discount factor process in continuous time as well as discrete time. At one level, this statement requires no new mathematics. If we reinvest dividends for simplicity, then a discount factor must satisfy pt Λt = Et (Λt+s pt+s ) . Calling pt+s = xt+s , this is precisely the discrete time p = E(mx) that we have studied all along. Thus, the law of one price or absence of arbitrage are equivalent to the existence of a or a positive Λt+s . The same conditions at all horizons s are thus equivalent to the existence of a discount factor process, or a positive discount factor process Λt for all time t. For calculations it is useful to find explicit formulas for a discount factors. Suppose a set of securities pays dividends Dt dt and their prices follow dpt = µt dt + σt dzt pt where p and z are N × 1 vectors, µt and σ t may vary over time, µ(pt , t,other variables), E (dzt dzt0 ) = I and the division on the left hand side is element-by element. (As usual, I’ll drop the t subscripts when not necessary for clarity, but everything can vary over time.) We can form a discount factor that prices these assets from a linear combination of the 75

C HAPTER 4

T HE DISCOUNT FACTOR

shocks that drive the original assets, µ ¶0 D dΛ∗ f f −r = −r dt − µ + Σ−1 σdz. Λ∗ p

(53)

where Σ ≡ σσ0 again is the covariance matrix of returns. You can easily check that this equation solves µ ¶ µ ∗ ¶ dp dΛ dp D Et + dt − rf dt = −Et (54) p p Λ∗ p and Et

µ

dΛ∗ Λ∗

¶

= −rf dt,

or you can show that this is the only diffusion driven by dz, dt with these properties. If there is a risk free rate rtf (also potentially time-varying), then that rate determines rtf . If there is no risk free rate, (4.53) will price the risky assets for any arbitrary (or convenient) choice of rtf . As usual, this discount factor is not unique; Λ∗ plus orthogonal noise will also act as a discount factor: dΛ∗ dΛ = ∗ + dw; E(dw) = 0; E(dzdw) = 0. Λ Λ You can see that (4.53) is exactly analogous to the discrete time formula (4.52). (If you don’t like answers popping out of hats like this, guess a solution of the form dΛ = µΛ dt + σΛ dz. Λ Then find µΛ and σΛ to satisfy (4.54) for the riskfree and risky assets.)

4.4 1. 2.

3.

Problems

Show that the law of one price loop implies that price is a linear function of payoff and vice versa Does the absence of arbitrage imply the law of one price? Does the law of one price imply the absence of arbitrage? Answer directly using portfolio arguments, and indirectly using the corresponding discount factors. If the law of one price or absence of arbitrage hold in population, must they hold in a sample drawn from that population?

76

Chapter 5. Mean-variance frontier and beta representations Much empirical work in asset pricing is couched in terms of expected return - beta representations and mean-variance frontiers. This chapter introduces expected return - beta representations and mean-variance frontiers. I discuss here the beta representation, most commonly applied to factor pricing models. Chapter 6 shows how an expected return/beta model is equivalent to a linear model for the discount factor, i.e. m = b0 f where f are the right hand variables in the time-series regressions that define betas. Chapter 9 then discusses the derivation of popular factor models such as the CAPM, ICAPM and APT, i.e. under what assumptions the discount factor is a linear function of other variables f such as the market return. I summarize the classic Lagrangian approach to the mean-variance frontier. I then introduce a powerful and useful representation of the mean-variance frontier due to Hansen and Richard (1987). This representation uses the state-space geometry familiar from the existence theorems. It is also useful because it is valid and useful in infinite-dimensional payoff spaces, which we shall soon encounter when we add conditioning information, dynamic trading or options.

5.1

Expected return - Beta representations

The expected return-beta expression of a factor pricing model is E(Ri ) = α + β i,a λa + β i,b λb + . . . The model is equivalent to a restriction that the intercept is the same for all assets in time-series regressions. When the factors are returns excess returns, then λa = E(f a ). If the test assets are also excess returns, then the intercept should be zero, α = 0. Much empirical work in finance is cast in terms of expected return - beta representations of linear factor pricing models, of the form E(Ri ) = α + β i,a λa + β i,b λb + . . . , i = 1, 2, ...N.

77

(55)

C HAPTER 5

M EAN - VARIANCE FRONTIER AND BETA REPRESENTATIONS

The β terms are defined as the coefficients in a multiple regression of returns on factors, Rit = ai + β i,a fta + β i,b ftb + . . . + εit ; t = 1, 2, ...T.

(56)

This is often called a time-series regression, since one runs a regression across time for each security i. The “factors” f are proxies for marginal utility growth. I discuss the stories used to select factors at some length in chapter 9. For the moment keep in mind the canonical examples, f = consumption growth, or f = the return on the market portfolio (CAPM). Notice that we run returns Rti on contemporaneous factors ftj . This regression is not about predicting returns from variables seen ahead of time. Its objective is to measure contemporaneous relations or risk exposure; whether returns are typically high in “good times” or “bad times” as measured by the factors. The point of the beta model(5.55) is to explain the variation in average returns across assets. I write i = 1, 2, ...N in (5.55) to emphasize this fact. The model says that assets with higher betas should get higher average returns. Thus the betas in (5.55) are the explanatory (x) variables, which vary asset by asset. The α and λ – common for all assets – are the intercept and slope in this cross-sectional relation. For example, equation (5.55) says that if we plot expected returns versus betas in a one-factor model, we should expect all (E(Ri ), β i,a ) pairs to line up on a straight line with slope λa and intercept α. β i,a is interpreted as the amount of exposure of asset i to factor a risks, and λa is interpreted as the price of such risk-exposure. Read the beta pricing model to say: “for each unit of exposure β to risk factor a, you must provide investors with an expected return premium λa .” Assets must give investors higher average returns (low prices) if they pay off well in times that are already good, and pay off poorly in times that are already bad, as measured by the factors. One way to estimate the free parameters (α, λ) and to test the model (5.55) is to run a cross sectional regression of average returns on betas, E(Ri ) = α + β i,a λa + β i,b λb + . . . + αi , i = 1, 2, ...N.

(57)

Again, the β i are the right hand variables, and the α and λ are the intercept and slope coefficients that we estimate in this cross-sectional regression. The errors αi are pricing errors. The model predicts αi = 0, and they should be statistically insignificant in a test. (I intentionally use the same symbol for the intercept, or mean of the pricing errors, and the individual pricing errors αi .) In the chapters on empirical technique, we will see test statistics based on the sum of squared pricing errors. The fact that the betas are regression coefficients is crucially important. If the betas are also free parameters then there is no content to the equation. More importantly (and this is an easier mistake to make), the betas cannot be asset-specific or firm-specific characteristics, such as the size of the firm, book to market ratio, or (to take an extreme example) the letter of the alphabet of its ticker symbol. It is true that expected returns are associated with or correlated with many such characteristics. Stocks of small companies or of companies with high 78

S ECTION 5.1

E XPECTED RETURN - B ETA REPRESENTATIONS

book/market ratios do have higher average returns. But this correlation must be explained by some beta. The proper betas should drive out any characteristics in cross-sectional regressions. If, for example, expected returns were truly related to size, one could buy many small companies to form a large holding company. It would be a “large” company, and hence pay low average returns to the shareholders, while earning a large average return on its holdings. The managers could enjoy the difference. What ruins this promising idea? . The “large” holding company will still behave like a portfolio of small stocks. Thus, only if asset returns depend on how you behave, not who you are – on betas rather than characteristics – can a market equilibrium survive such simple repackaging schemes. Some common special cases If there is a risk free rate, its betas in (5.55) are all zero,2 so the intercept is equal to the risk free rate, Rf = α. We can impose this condition rather than estimate α in the cross-sectional regression (5.57). If there is no risk-free rate, then α must be estimated in the cross-sectional regression. Since it is the expected return of a portfolio with zero betas on all factors, α is called the (expected) zero-beta rate in this circumstance. We often examine factor pricing models using excess returns directly. (There is an implicit, though not necessarily justified, division of labor between models of interest rates and models of equity risk premia.) Differencing (5.55) between any two returns Rei = Ri − Rj (Rj does not have to be risk free), we obtain E(Rei ) = β i,a λa + β i,b λb + . . . , i = 1, 2, ...N.

(58)

Here, β ia represents the regression coefficient of the excess return Rei on the factors. It is often the case that the factors are also returns or excess returns. For example, the CAPM uses the return on the market portfolio as the single factor. In this case, the model should apply to the factors as well, and this fact allows us to directly measure the λ coefficients. Each factor has beta of one on itself and zero on all the other factors, of course. Therefore, if the factors are excess returns, we have E(f a ) = λa , and so forth. We can then write the factor model as E(Rei ) = β i,a E(f a ) + β i,b E(f b ) + . . . , i = 1, 2, ...N. The cross-sectional beta pricing model (5.55)-(5.58) and the time-series regression definition of the betas in (5.56) look very similar. It seems that one can take expectations of 2 The betas are zero because the risk free rate is known ahead of time. When we consider the effects of conditioning information, i.e. that the interest rate could vary over time, we have to interpret the means and betas as conditional moments. Thus, if you are worried about time-varying risk free rates, betas, and so forth, either assume all variables are i.i.d. (and thus that the risk free rate is constant), or interpret all moments as conditional on time t information.

79

C HAPTER 5

M EAN - VARIANCE FRONTIER AND BETA REPRESENTATIONS

the time-series regression (5.56) and arrive at the beta model (5.55), in which case the latter would be vacuous since one can always run a regression of anything on anything. The difference is subtle but crucial: the time-series regressions (5.56) will in general have a different intercept ai for each return i, while the intercept α is the same for all assets in the beta pricing equation (5.55). The beta pricing equation is a restriction on expected returns, and thus imposes a restriction on intercepts in the time-series regression. In the special case that the factors are themselves excess returns, the restriction is particularly simple: the time-series regression intercepts should all be zero. In this case, we can avoid the cross-sectional regression entirely, since there are no free parameters left.

5.2

Mean-variance frontier: Intuition and Lagrangian characterization

The mean-variance frontier of a given set of assets is the boundary of the set of means and variances of the returns on all portfolios of the given assets. One can find or define this boundary by minimizing return variance for a given mean return. Many asset pricing propositions and test statistics have interpretations in terms of the mean-variance frontier. Figure 13 displays a typical mean-variance frontier. As displayed in Figure 13, it is common to distinguish the mean-variance frontier of all risky assets, graphed as the hyperbolic region, and the mean-variance frontier of all assets, i.e. including a risk free rate if there is one, which is the larger wedge-shaped region. Some authors reserve the terminology “meanvariance frontier” for the upper portion, calling the whole thing the minimum variance frontier. The risky asset frontier is a hyperbola, which means it lies between two asymptotes, shown as dotted lines. The risk free rate is typically drawn below the intersection of the asymptotes and the vertical axis, or the point of minimum variance on the risky frontier. If it were above this point, investors with a mean-variance objective would try to short the risky assets, which cannot represent an equilibrium. In general, portfolios of two assets or portfolios fill out a hyperbolic curve through the two assets. The curve is sharper the less correlated are the two assets, because the portfolio variance benefits from increasing diversification. Portfolios of a risky asset and risk free rate give rise to straight lines in mean-standard deviation space. In Chapter 1, we derived a similar wedge-shaped region as the set of means and variances of all assets that are priced by a given discount factor. This chapter is about incomplete markets, so we think of a mean-variance frontier generated by a given set of assets, typically less than complete. When does the mean-variance frontier exist? I.e., when is the set of portfolio means and variances less than the whole {E, σ} space? We basically have to rule out a special case: two returns are perfectly correlated but yield different means. In that case one could short one, long the other, and achieve infinite expected returns with no risk. More formally, eliminate purely redundant securities from consideration, then 80

S ECTION 5.2

M EAN - VARIANCE FRONTIER : I NTUITION AND L AGRANGIAN CHARACTERIZATION

Theorem: So long as the variance-covariance matrix of returns is non-singular, there is a mean-variance frontier. To prove this theorem, just follow the construction below. This theorem should sound very familiar: Two perfectly correlated returns with different mean are a violation of the law of one price. Thus, the law of one price implies that there is a mean variance frontier as well as a discount factor. E(R)

Mean-variance frontier Risky asset frontier

Tangency portfolio of risky assets Original assets Rf

σ(R)

Figure 13. Mean-variance frontier

5.2.1

Lagrangian approach to mean-variance frontier

The standard definition and the computation of the mean-variance frontier follows a brute force approach. Problem: Start with a vector of asset returns R. Denote by E the vector £ of mean returns, 0¤ E ≡ E(R), and denote by Σ the variance-covariance matrix Σ = E (R − E)(R − E) . A portfolio is defined by its weights w on the initial securities. The portfolio return is w0 R where the weights sum to one, w0 1 =1. The problem “choose a portfolio to minimize variance for a given mean” is then min{w} w0 Σw s.t. w0 E = µ; w0 1 = 1. Solution: Let A = E 0 Σ−1 E; B = E 0 Σ−1 1; C = 10 Σ−1 1. 81

(59)

C HAPTER 5

M EAN - VARIANCE FRONTIER AND BETA REPRESENTATIONS

Then, for a given mean portfolio return µ, the minimum variance portfolio has variance var (Rp ) = and is formed by portfolio weights w = Σ−1

Cµ2 − 2Bµ + A AC − B 2

(60)

E (Cµ − B) + 1 (A − Bµ) . (AC − B 2 )

Equation (5.60) shows that the variance is a quadratic function of the mean. The square root of a parabola is a hyperbola, which is why we draw hyperbolic regions in mean-standard deviation space. The minimum-variance portfolio is interesting in its own right. It appears as a special case in many theorems and it appears in several test statistics. We can find it by minimizing (5.60) over µ, giving µmin var = B/C. The weights of the minimum variance portfolio are thus w = Σ−1 1/(10 Σ−1 1). We can get to any point on the mean-variance frontier by starting with two returns on the frontier and forming portfolios. The frontier is spanned by any two frontier returns. To see this fact, notice that w is a linear function of µ. Thus, if you take the portfolios corresponding to any two distinct mean returns µ1 and µ2 , the weights on a third portfolio with mean µ3 = λµ1 + (1 − λ)µ2 are given by w3 = λw1 + (1 − λ)w2 . Derivation: To derive the solution, introduce Lagrange multipliers 2λ and 2δ on the constraints. The first order conditions to (5.59) are then Σw−λE − δ1 = 0 w = Σ−1 (λE + δ1).

(61)

We find the Lagrange multipliers from the constraints, E 0 w = E 0 Σ−1 (λE + δ1) = µ 10 w = 10 Σ−1 (λE + δ1) = 1 or

·

E 0 Σ−1 E 10 Σ−1 E ·

A B

E 0 Σ−1 1 10 Σ−1 1

¸·

λ δ

¸

=

¸·

¸

·

µ 1

¸

B C

λ δ

82

=

·

µ 1

¸

S ECTION 5.3

A N ORTHOGONAL CHARACTERIZATION OF THE MEAN - VARIANCE FRONTIER

Hence, λ=

Cµ−B AC−B 2

δ=

A−Bµ AC − B 2

Plugging in to (5.61), we get the portfolio weights and variance.

5.3

An orthogonal characterization of the mean-variance frontier

Every return can be expressed as Ri = R∗ + wi Re∗ + ni . The mean-variance frontier is Rmv = R∗ + wRe∗ Re∗ is defined as Re∗ = proj(1|Re ). It represents mean excess returns, E(Re ) = E(Re∗ Re ) ∀R ∈ Re e

The Lagrangian approach to the mean-variance frontier is straightforward but cumbersome. Our further manipulations will be easier if we follow an alternative approach due to Hansen and Richard (1987). Technically, Hansen and Richard’s approach is also valid when we can’t generate the payoff space by portfolios of a finite set of basis payoffs c0 x. This happens, for example, when we think about conditioning information in Chapter 8. Also, it is the natural geometric way to think about the mean-variance frontier given that we have started to think of payoffs, discount factors and other random variables as vectors in the space of payoffs. Rather than write portfolios as combinations of basis assets, and pose and solve a minimization problem, we first describe any return by a three-way orthogonal decomposition. The mean-variance frontier then pops out easily without any algebra. 5.3.1

Definitions of R∗ , Re∗

I start by defining two special returns. R∗ is the return corresponding to the payoff x∗ that can act as the discount factor. The price of x∗ , is, like any other price, p(x∗ ) = E(x∗ x∗ ). Thus, The definition of R∗ is R∗ ≡

x∗ x∗ = p(x∗ ) E(x∗2 ) 83

(62)

C HAPTER 5

M EAN - VARIANCE FRONTIER AND BETA REPRESENTATIONS

The definition of Re∗ is Re∗ ≡ proj(1 | Re )

(63)

Re ≡ space of excess returns = {x ∈ X s.t. p(x) = 0} Why Re∗ ? We are heading towards a mean-variance frontier, so it is natural to seek a special return that changes means. Re∗ is an excess return that represents means on Re with an inner product in the same way that x∗ is a payoff in X that represents prices with an inner product. As p(x) = E(mx) = E[proj(m|X)x] = E(x∗ x), so E(Re ) = E(1 × Re ) = E [proj(1 | Re ) × Re ] = E(Re∗ Re ). If R∗ and Re∗ are still a bit mysterious at this point, they will make more sense as we use them, and discover their many interesting properties. Now we can state a beautiful orthogonal decomposition. Theorem: Every return Ri can be expressed as Ri = R∗ + wi Re∗ + ni where wi is a number, and ni is an excess return with the property E(ni ) = 0. The three components are orthogonal, E(R∗ Re∗ ) = E(R∗ ni ) = E(Re∗ ni ) = 0. This theorem quickly implies the characterization of the mean variance frontier which we are after: Theorem: Rmv is on the mean-variance frontier if and only if Rmv = R∗ + wRe∗

(64)

for some real number w. As you vary the number w, you sweep out the mean-variance frontier. E(Re∗ ) 6= 0, so adding more w changes the mean and variance of Rmv . You can interpret (5.64) as a “two84

S ECTION 5.3

A N ORTHOGONAL CHARACTERIZATION OF THE MEAN - VARIANCE FRONTIER

fund” theorem for the mean-variance frontier. It expresses every frontier return as a portfolio of R∗ and Re∗ , with varying weights on the latter. As usual, first I’ll argue why the theorems are sensible, then I’ll offer a simple algebraic proof. Hansen and Richard (1987) give a much more careful algebraic proof. 5.3.2

Graphical construction

Figure 14 illustrates the decomposition. Start at the origin (0). Recall that the x∗ vector is perpendicular to planes of constant price; thus the R∗ vector lies perpendicular to the plane of returns as shown. Go to R∗ . Re∗ is the excess return that is closest to the vector 1; it lies at right angles to planes (in R ) of constant mean return, shown in the E = 1, E = 2 lines, just as the return R∗ lies at right angles to planes of constant price. Since Re∗ is an excess return, it is orthogonal to R∗ . Proceed an amount wi in the direction of Re∗ , getting as close to Ri as possible. e

Now move, again in an orthogonal direction, by an amount ni to get to the return Ri . We have thus expressed Ri = R∗ +wi Re∗ +ni in a way that all three components are orthogonal. Returns with n = 0, R∗ + wRe∗ , are the mean-variance frontier. Here’s why. Since E(R2 ) = σ 2 (R) + E(R)2 , we can define the mean-variance frontier by minimizing second moment for a given mean. The length of each vector in Figure 14 is its second moment, so we want the shortest vector that is on the return plane for a given mean. The shortest vectors in the return plane with given mean are on the R∗ + wRe∗ line. The graph also shows how Re∗ represents means in the space of excess returns. Expectation is the inner product with 1. Planes of constant expected value in Figure 14 are perpendicular to the 1 vector, just as planes of constant price are perpendicular to the x∗ or R∗ vectors. I don’t show the full extent of the constant expected payoff planes for clarity; I do show lines of constant expected excess return in Re , which are the intersection of constant expected payoff planes with the Re plane. Therefore, just as we found an x∗ in X to represent prices in X by projecting m onto X, we find Re∗ in Re by projecting of 1 onto Re . Yes, a regression with one on the left hand side. Planes perpendicular to Re∗ in Re are payoffs with constant mean, just as planes perpendicular to x∗ in X are payoffs with the same price. 5.3.3

Algebraic argument

Now, an algebraic proof of the decomposition and characterization of mean variance frontier. The algebra just represents statements about what is at right angles to what with second moments. Proof: Straight from their definitions, (5.62) and (5.63) we know that Re∗ is an

85

C HAPTER 5

M EAN - VARIANCE FRONTIER AND BETA REPRESENTATIONS

R=space of returns (p=1) R*+wiRe* R*

ni Ri=R*+wiRe*+ni

1

Re*

0

E=1

E=0

Re = space of excess returns (p=0) Figure 14. Orthogonal decomposition and mean-variance frontier. excess return (price zero), and hence that R∗ and Re∗ are orthogonal, E(R∗ Re∗ ) =

E(x∗ Re∗ ) = 0. E(x∗2 )

We define ni so that the decomposition adds up to Ri as claimed, and we define wi to make sure that ni is orthogonal to the other two components. Then we prove that E(ni ) = 0. Pick any wi and then define ni ≡ Ri − R∗ − wi Re∗ . ni is an excess return so already orthogonal to R∗ , E(R∗ ni ) = 0. To show E(ni ) = 0 and ni orthogonal to Re∗ , we exploit the fact that since ni is an excess return, E(ni ) = E(Re∗ ni ). Therefore, Re∗ is orthogonal to ni if and only if we pick wi so that E(ni ) = 0. We 86

S ECTION 5.4

S PANNING THE MEAN - VARIANCE FRONTIER

don’t have to explicitly calculate wi for the proof.3 Once we have constructed the decomposition, the frontier drops out. Since E(ni ) = 0 and the three components are orthogonal, E(Ri ) = E(R∗ ) + wi E(Re∗ )

σ2 (Ri ) = σ2 (R∗ + wi Re∗ ) + σ2 (ni ). Thus, for each desired value of the mean return, there is a unique wi . Returns with ¥ ni = 0 minimize variance for each mean. 5.3.4

Decomposition in mean-variance space

Figure 15 illustrates the decomposition in mean-variance space rather than in state-space. First, let’s locate R∗ . R∗ is the minimum second moment return. One can see this fact from the geometry of Figure 14: R∗ is the return closest to the origin, and thus the return with the smallest “length” which is second moment. As with OLS regression, minimizing the length of R∗ and creating an R∗ orthogonal to all excess returns is the same thing. One can also verify this property algebraically. Since any return can be expressed as R = R∗ + wRe∗ + n, E(R2 ) = E(R∗2 ) + w2 E(Re∗2 ) + E(n2 ). n = 0 and w = 0 thus give the minimum second moment return. In mean-standard deviation space, lines of constant second moment are circles. Thus, the minimum second-moment return R∗ is on the smallest circle that intersects the set of all assets, which lie in the mean-variance frontier in the right hand panel of Figure 19. Notice that R∗ is on the lower, or “inefficient” segment of the mean-variance frontier. It is initially surprising that this is the location of the most interesting return on the frontier! R∗ is not the “market portfolio” or “wealth portfolio,” which typically lie on the upper portion of the frontier. Adding more Re∗ moves one along the frontier. Adding n does not change mean but does change variance, so it is an idiosyncratic return that just moves an asset off the frontier as graphed. α is the “zero-beta rate” corresponding to R∗ . It is the expected return of any return that is uncorrelated with R∗ . I demonstrate these properties in section 6.5. 3

Its value wi =

E(Ri ) − E(R∗ ) E(Re∗ )

is not particularly enlightening.

87

C HAPTER 5

M EAN - VARIANCE FRONTIER AND BETA REPRESENTATIONS

E(R)

Ri

R* + wiRe* ni

R*

σ(R) Figure 15. Orthogonal decomposition of a return Ri in mean-standard deviation space.

5.4

Spanning the mean-variance frontier

The characterization of the mean-variance frontier in terms of R∗ and Re∗ is most natural in our setup. However, you can equivalently span the mean-variance frontier with any two portfolios that are on the frontier – any two distinct linear combinations of R∗ and Re∗ . In particular, take any return Rα = R∗ + γRe∗ , γ 6= 0.

(65)

Using this return in place of Re∗ , Rα − R∗ γ

Re∗ =

you can express the mean variance frontier in terms of R∗ and Rα : R∗ + wRe∗

= R∗ + y (Rα − R∗ ) = (1 − y)R∗ + yRα

where I have defined a new weight y = w/γ. 88

(5.66)

S ECTION 5.5

A COMPILATION OF PROPERTIES OF R∗ , Re∗ AND x∗

The most common alternative approach is to use a risk free rate or a risky rate that somehow behaves like the risk free rate in place of Re∗ to span the frontier. When there is a risk free rate, it is on the frontier with representation Rf = R∗ + Rf Re∗ I derive this expression in equation (5.72) below. Therefore, we can use (5.66) with Ra = Rf . When there is no risk free rate, several risky returns that retain some properties of the risk free rate are often used. In section 5.3 below I present a “zero beta” return, which is uncorrelated with R∗ , a “constant-mimicking portfolio” return, which is the return on the traded payoff ˆ = proj(1|X)/p[proj(1|X)] and the minimum variance return. Each of closest to unity, R these returns is on the mean-variance frontier, with form 5.65, though different values of α. Therefore, we can span the mean-variance frontier with R∗ and any of these risk-free rate proxies.

5.5

A compilation of properties of R∗ , Re∗ and x∗

The special returns R∗ , Re∗ that generate the mean variance frontier have lots of interesting and useful properties. Some I derived above, some I will derive and discuss below in more detail, and some will be useful tricks later on. Most properties and derivations are extremely obscure if you don’t look at the picture! 1) 1 . E(x∗2 )

E(R∗2 ) =

(67)

To derive this fact, multiply both sides of (5.62) by R∗ , take expectations, and remember R∗ is a return so 1 = E(x∗ R∗ ). 2) We can reverse the definition and recover x∗ from R∗ via x∗ =

R∗ . E(R∗2 )

(68)

To derive this formula, start with the definition R∗ = x∗ /E(x∗2 ) and substitute from (5.67) for E(x∗2 ) 3) R∗ can be used to represent prices just like x∗ . This is not surprising, since they both point in the same direction, orthogonal to planes of constant price. Most obviously, from 5.68 p(x) = E(x∗ x) =

E (R∗ x) ∀x ∈ X E(R∗2 )

For returns, we can nicely express this result as E(R∗2 ) = E(R∗ R) ∀R ∈ R. 89

(69)

C HAPTER 5

M EAN - VARIANCE FRONTIER AND BETA REPRESENTATIONS

This fact can also serve as an alternative defining property of R∗ . 4) Re∗ represents means on Re via an inner product in the same way that x∗ represents prices on X via an inner product. Re∗ is orthogonal to planes of constant mean in Re as x∗ is orthogonal to planes of constant price. Algebraically, in analogy to p(x) = E(x∗ x) we have E(Re ) = E(Re∗ Re ) ∀Re ∈ Re .

(70)

This fact can serve as an alternative defining property of Re∗ . 5) Re∗ and R∗ are orthogonal, E(R∗ Re∗ ) = 0. More generally, R∗ is orthogonal to any excess return. 6) The mean variance frontier is given by Rmv = R∗ + wRe∗ . ¤ £ To prove this, E(R2 ) = E (R∗ + wRe∗ + n)2 = E(R∗2 ) + w2 E(Re2 ) + E(n2 ), and E(n) = 0, so set n to zero. The conditional mean-variance frontier allows w in the conditioning information set. The unconditional mean variance frontier requires w to equal a constant. 7) R∗ is the minimum second moment return. Graphically, R∗ is the return closest to the origin. To see this, using the decomposition in #6, and set w2 and n to zero to minimize second moment. 8) Re∗ has the same first and second moment, E(Re∗ ) = E(Re∗2 ). Just apply fact (5.70) to Re∗ itself. Therefore var(Re∗ ) = E(Re∗2 ) − E(Re∗ )2 = E(Re∗ ) [1 − E(Re∗ )] . 9) If there is a riskfree rate, then Re∗ can also be defined as the residual in the projection of 1 on R∗ : Re∗ = 1 − proj(1|R∗ ) = 1 −

E(R∗ ) ∗ 1 R = 1 − f R∗ ∗2 E(R ) R

(71)

You’d never have thought of this without looking at Figure 14! Since R∗ and Re are orthogonal and together span X, 1 = proj(1|Re ) + proj(1|R∗ ). You can also verify this statement analytically. Check that Re∗ so defined is an excess return in X – its price is zero–, and E(Re∗ Re ) = E(Re ); E(R∗ Re∗ ) = 0. 90

S ECTION 5.5

A COMPILATION OF PROPERTIES OF R∗ , Re∗ AND x∗

As a result, Rf has the decomposition Rf = R∗ + Rf Re∗ .

(72)

Since Rf > 1 typically, this means that R∗ +Re∗ is located on the lower portion of the meanvariance frontier in mean-variance space, just a bit to the right of Rf . If the risk free rate were one, then the unit vector would lie in the return space, and Rf = R∗ + Re∗ . Typically, the space of returns is a little bit above the unit vector. As you stretch the unit vector by the amount Rf to arrive at the return Rf , so you stretch the amount Re∗ that you add to R∗ to get to Rf . If there is no riskfree rate, then we can use proj(1|X) = proj(proj (1|X) |Re ) + proj(proj (1|X) |R∗ ) = proj(1|Re ) + proj(1|R∗ ) to deduce an analogue to equation (5.71), Re∗ = proj(1|X) − proj(1|R∗ ) = proj(1|X) −

E(R∗ ) ∗ R E(R∗2 )

(73)

10) If a riskfree rate is traded, we can construct Rf from R∗ via Rf =

E(R∗2 ) 1 = . ∗ E(x ) E(R∗ )

(74)

If not, this gives a “zero beta rate” interpretation of the right hand expression. 11) Since we have a formula x∗ = p0 E(xx0 )−1 x for constructing x∗ from basis assets (see section 4.1), we can construct R∗ in this case from R∗ =

p0 E(xx0 )−1 x x∗ = . p(x∗ ) p0 E(xx0 )−1 p

(p(x∗ ) = E(x∗ x∗ ) leading to the denominator.) 12) We can construct Re∗ from a set of basis assets as well. Following the definition to project one on the space of excess returns, Re∗ = E(Re )0 E(Re Re0 )−1 Re where Re is the basis set of excess returns. (You can always use Re = R−R∗ if you want). This construction obviously mirrors the way we constructed x∗ in section 4.1, and you can see the similarity in the result, with E in place of p, since Re∗ represents means rather than prices. . 91

C HAPTER 5

M EAN - VARIANCE FRONTIER AND BETA REPRESENTATIONS

If there is a riskfree rate, we can also use (5.71), Re∗ = 1 −

1 ∗ 1 p0 E(xx0 )−1 x . R = 1 − Rf Rf p0 E(xx0 )−1 p

(75)

If there is no riskfree rate, we can use (5.73) to construct Re∗ . The central ingredient is proj(1|X) = E(x)0 E(xx0 )−1 x.

5.6

Mean-variance frontiers for m: the Hansen-Jagannathan bounds

The mean-variance frontier of all discount factors that price a given set of assets is related to the mean-variance frontier of asset returns by |E(Re )| σ(m) ≥ . E(m) σ(Re ) and hence σ(m) E(Re ) = max e {all m that price x∈X} E(m) {all excess returns R in X} σ(Re ) min

The discount factors on the frontier can be characterized analogously to the mean-variance frontier of asset returns, m = x∗ + we∗ 0 e∗ ≡ 1 − proj(1|X) = proj(1|E) = 1 − E(x) E(xx0 )−1 x E = {m − x∗ } . We derived in Chapter 1 a relation between the Sharpe ratio of an excess return and the volatility of discount factors necessary to price that return, |E(Re )| σ(m) ≥ . E(m) σ(Re ) Quickly, 0 = E(mRe ) = E(m)E(Re ) + ρm,Re σ(m)σ(Re ), 92

S ECTION 5.6

M EAN - VARIANCE FRONTIERS FOR m: THE H ANSEN -JAGANNATHAN BOUNDS

and |ρ| ≤ 1. If we had a riskfree rate, then we know in addition E(m) = 1/Rf . Hansen and Jagannathan (1991) had the brilliant insight to read this equation as a restriction on the set of discount factors that can price a given set of returns, as well as a restriction on the set of returns we will see given a specific discount factor. This calculation teaches us that we need very volatile discount factors with a mean near one to understand stock returns. This and more general related calculations turn out to be a central tool in understanding and surmounting the equity premium puzzle, surveyed in Chapter 21. We would like to derive a bound that uses a large number of assets, and that is valid if there is no riskfree rate. What is the set of {E(m), σ(m)} consistent with a given set of asset prices and payoffs? What is the mean-variance frontier for discount factors? Obviously, the higher the Sharpe ratio, the tighter the bound. This suggests a way to construct the frontier we’re after. For any hypothetical risk-free rate, find the highest Sharpe ratio. That is, of course the tangency portfolio. Then the slope to the tangency portfolio gives the ratio σ(m)/E(m). Figure 16 illustrates. E(R)

σ(m) = Ε(Re)/σ(Re) 1/E(m)

E(Re)/σ(Re)

σ(R)

E(m)

Figure 16. Graphical construction of the Hansen-Jagannathan bound. As we sweep through values of E(m), the slope to the tangency becomes lower, and the Hansen-Jagannathan bound declines. At the mean return corresponding to the minimum variance point, the HJ bound attains its minimum. Continuing, the Sharpe ratio rises again and so does the bound. If there were a riskfree rate, then we know E(m), the return frontier is a V shape, and the HJ bound is purely a bound on variance. This discussion implies a beautiful duality between discount factor volatility and Sharpe ratios. σ(m) E(Re ) = max e . {all m that price x∈X} E(m) {all excess returns R in X} σ(Re ) min

93

(76)

C HAPTER 5

M EAN - VARIANCE FRONTIER AND BETA REPRESENTATIONS

We need formulas for an explicit calculation. In equation (), we found a representation for the set of discount factors that price a given set of asset returns – that satisfy p = E(mx) : m = E(m) + [p − E(m)E(x)] Σ−1 [x − E(x)] + ε

(77)

where Σ ≡ cov(x, x0 ) and E(ε) = 0, E(εx) = 0. You can think of this as a regression or projection of any discount factor on the space of payoffs, plus an error. Since σ2 (ε) > 0, this representation leads immediately to an explicit expression for the Hansen-Jagannathan bound, 0 σ 2 (m) ≥ (p − E(m)E(x)) Σ−1 (p − E(m)E(x)) .

(78)

As all asset returns must lie in a cup-shaped σ(R)} space, all discount ª © region in {E(R), factors must lie in a parabolic region in E(m), σ 2 (m) space, as illustrated in the right hand panel of Figure 16.

We would like an expression for the discount factors on the bound, as we wanted an expression for the returns on the mean variance frontier instead of just a formula for the means and variances. As we found a three way decomposition of all returns, two of which generated the mean-variance frontier of returns, so we can find a three way decomposition of discount factors, two of which generate the mean-variance frontier of discount factors (5.78). I illustrate the construction in Figure 17. Any m must line in the plane marked M, perpendicular to X through x∗ . Any m must be of the form m = x∗ + we∗ + n. Here, I have just broken up the residual ε in the familiar representation m = x∗ + ε into two components. e∗ is defined as the residual from the projection of 1 onto X or, equivalently the projection of 1 on the space E of “excess m’s,” random variables of the form m − x∗ . e∗ ≡ 1 − proj(1|X) = proj(1|E). e∗ generates means of m just as Re∗ did for returns: E(m − x∗ ) = E[1 × (m − x∗ )] = E[proj(1|E)(m − x∗ )]. Finally n, defined as the leftovers, has mean zero since it’s orthogonal to 1 and is orthogonal to X. As with returns, then, the mean-variance frontier of m0 s is given by m∗ = x∗ + we∗ .

(79)

If the unit payoff is in the payoff space, then we know E(m), and the frontier and bound are just m = x∗ , σ 2 (m) ≥ σ2 (x∗ ). This is exactly like the case of risk-neutrality for return mean-variance frontiers. 94

S ECTION 5.6

M EAN - VARIANCE FRONTIERS FOR m: THE H ANSEN -JAGANNATHAN BOUNDS

X = payoff space M = space of discount factors x*+we* x*

n m = x*+we*+n

1

proj(1| X)

e*

0

E(.)=1

E(.)=0 E = space of m-x*

Figure 17. Decomposition of any discount factor m = x∗ + we + n.

95

C HAPTER 5

M EAN - VARIANCE FRONTIER AND BETA REPRESENTATIONS

The construction (5.79) can be used to derive the formula (5.78) for the Hansen-Jagannathan bound for the finite-dimensional cases discussed above. It’s more general, since it can be used in infinite-dimensional payoff spaces as well. Along with the corresponding return formula Rmv = R∗ + wRe∗ , we see in Chapter 8 that it extends more easily to the calculation of conditional vs. unconditional mean-variance frontiers (Gallant, Hansen and Tauchen 1995). It will make construction (5.79) come alive if we find equations for its components. We find x∗ as before, it is the portfolio c0 x in X that prices x: x∗ = p0 E(xx0 )−1 x. Similarly, let’s find e∗ . The projection of 1 on X is proj(1|X) = E(x)0 E(xx0 )−1 x. (After a while you get used to the idea of running regressions with 1 on the left hand side and random variables on the right hand side!) Thus, e∗ = 1 − E(x)0 E(xx0 )−1 x. Again, you can construct time-series of x∗ and e∗ from these definitions.

or

Finally, we now can construct the variance-minimizing discount factors £ ¤ m∗ = x∗ + we∗ = p0 E(xx0 )−1 x + w 1 − E(x)0 E(xx0 )−1 x 0 m∗ = w + [p − wE(x)] E(xx0 )−1 x

(80)

As w varies, we trace out discount factors m∗ on the frontier with varying means and variances. It’s easiest to find mean and second moment: E(m∗ ) = w + [p − wE(x)]0 E(xx0 )−1 E(x) 0 E(m∗2 ) = [p − wE(x)] E(xx0 )−1 [p − wE(x)] ;

variance follows from σ 2 (m) = E(m2 ) − E(m)2 . With a little algebra one can also show that these formulas are equivalent to equation (5.78). As you can see, Hansen-Jagannathan frontiers are equivalent to mean-variance frontiers. For example, an obvious exercise is to see how much the addition of assets raises the HansenJagannathan bound. This is exactly the same as asking how much those assets expand the mean-variance frontier. It was, in fact, this link between Hansen-Jagannathan bounds and mean-variance frontiers rather than the logic I described that inspired Knez and Chen (1996) and DeSantis (1994) to test for mean-variance efficiency using, essentially, HansenJagannathan bounds. 96

S ECTION 5.7

P ROBLEMS

Hansen-Jagannathan bounds have the potential to do more than mean-variance frontiers. Hansen and Jagannathan show how to solve the problem min σ2 (m) s.t. p = E(mx), m > 0. This is the “Hansen-Jagannathan bound with positivity,” and is strictly tighter than the HansenJagannathan bound. It allows you to impose no-arbitrage conditions. In stock applications, this extra bound ended up not being that informative. However, in the option application of this idea of Chapter (18), positivity is really important. That chapter shows how to solve for a bound with positivity. Hansen, Heaton and Luttmer (1995) develop a distribution theory for the bounds. Luttmer (1996) develops bounds with market frictions such as short-sales constraints and bid-ask spreads, to account for ludicrously high apparent Sharpe ratios and bounds in short term bond data. Cochrane and Hansen (1992) survey a variety of bounds, including bounds that incorporate information that discount factors are poorly correlated with stock returns (the HJ bounds use the extreme ρ = 1), bounds on conditional moments that illustrate how many models imply excessive interest rate variation, bounds with short-sales constraints and market frictions, etc. Chapter 21 discusses what the results of Hansen Jagannathan bound calculations and what they mean for discount factors that can price stock and bond return data.

5.7 1. 2. 3. 4.

5. 6. 7.

Problems

Prove that Re∗ lies at right angles to planes (in Re ) of constant mean return, as shown in figure 14. Should we typically draw x∗ above, below or on the plane of returns? Must x∗ always lie in this position? Can you construct Re∗ from knowledge of m, x∗ , or R∗ ? What happens to R∗ , Re∗ and the mean-variance frontier if investors are risk neutral? (a) If a riskfree rate is traded. (b) If no riskfree rate is traded? (Hint: make a drawing or think about the case that payoffs are generated by an N dimensional vector of basis assets x) x∗ = proj(m|X). Is R∗ = proj(m|R)? We showed that all m are of the form x∗ + ε. What about R−1 R? Show that if there is a risk-free rate—if the unit payoff is in the payoff space X—then Re∗ = (Rf − R∗ )/Rf .

97

Chapter 6. Relation between discount factors, betas, and mean-variance frontiers In this chapter, I draw the connection between discount factors, mean-variance frontiers, and beta representations. In the first chapter, I showed how mean-variance and a beta representation follow from p = E(mx) and (in the mean-variance case) complete markets. Here, I discuss the connections in both directions and in incomplete markets, drawing on the representations studied in the last chapter. The central theme of the chapter is that all three representations are equivalent. Figure 18 summarizes the ways one can go from one representation to another. A discount factor, a reference variable for betas – the thing you put on the right hand side in the regressions that give betas – and a return on the mean-variance frontier all carry the same information, and given any one of them, you can find the others. More specifically, 1. 2. 3. 4. 5.

p = E(mx) ⇒ β: Given m such that p = E(mx), m, x∗ , R∗ , or R∗ + wRe∗ all can serve as reference variables for betas. p = E(mx) ⇒ mean-variance frontier. You can construct R∗ from x∗ = proj(m|X), R∗ = x∗ /E(x∗2 ), and then R∗ , R∗ + wRe∗ are on the mean-variance frontier. Mean-variance frontier ⇒ p = E(mx). If Rmv is on the mean-variance frontier, then m = a + bRmv linear in that return is a discount factor; it satisfies p = E(mx). β ⇒ p = E(mx). If we have an expected return/beta model with factors f , then m = b0 f linear in the factors satisfies p = E(mx) (and vice-versa). If a return is on the mean-variance frontier, then there is an expected return/beta model with that return as reference variable. The following subsections discuss the mechanics of going from one representation to the other in detail. The last section of the chapter collects some special cases when there is no riskfree rate. The next chapter discusses some of the implications of these equivalence theorems, and why they are important. Roll (197x) pointed out the connection between mean-variance frontiers and beta pricing. Ross (1978) and Dybvig and Ingersoll (1982) pointed out the connection between linear discount factors and beta pricing. Hansen and Richard (1987) pointed out the connection between a discount factor and the mean-variance frontier.

6.1

From discount factors to beta representations

m, x∗ , and R∗ can all be the single factor in a single beta representation.

98

S ECTION 6.1

F ROM DISCOUNT FACTORS TO BETA REPRESENTATIONS

LOOP Þ m exists

p = E(mx) f = m, x*, R* m = b’f

m = a + bRmv R* is on m.v.f.

E(Ri) = α + βi’λ f = Rmv proj(f|R) on m.v.f.

Rmv on m.v.f.

E(RR’) nonsingular Þ Rmv exists

Figure 18. Relation between three views of asset pricing. 6.1.1

Beta representation using m

p = E(mx) implies E(Ri ) = α + β i,m λm . Start with 1 = E(mRi ) = E(m)E(Ri ) + cov(m, Ri ). Thus, E(Ri ) =

cov(m, Ri ) 1 − . E(m) E(m)

Multiply and divide by var(m), define α ≡ 1/E(m) to get µ ¶µ ¶ cov(m, Ri ) var(m) i E(R ) = α + − = α + β i,m λm . var(m) E(m) As advertised, we have a single beta representation. For example, we can equivalently state the consumption-based model as: mean asset returns should be linear in the regression betas of asset returns on (ct+1 /ct )−γ . Furthermore, the slope of this cross-sectional relationship λm is not a free parameter, though it is usually treated as such in empirical evaluation of factor pricing models. λm should equal the ratio of 99

C HAPTER 6

R ELATION BETWEEN DISCOUNT FACTORS , BETAS , AND MEAN - VARIANCE FRONTIERS

variance to mean of (ct+1 /ct )−γ . The factor risk premium λm for marginal utility growth is negative. Positive expected returns are associated with positive correlation with consumption growth, and hence negative correlation with marginal utility growth and m. Thus, we expect λm < 0. 6.1.2

β representation using x∗ and R∗

It is often useful to express a pricing model in a way that the factor is a payoff rather than a real factor such as consumption growth. In applications, we can then avoid measurement difficulties of real data. We have already seen the idea of “factor mimicking portfolios” formed by projection: project m on to X, and the resulting payoff x∗ also serves as a discount factor. Unsurprisingly, x∗ can also serve as the factor in an expected return-beta representation. It’s even more useful if the reference payoff is a return. Unsurprisingly, the return R∗ = x∗ /E(x∗2 ) can also serve as the factor in a beta pricing model. When the factor is also a return, the model is particularly simple, since the factor risk premium is also the expected excess return. Theorem. 1 = E(mRi ) implies an expected return - beta model with x∗ = proj(m|X) or R∗ ≡ x∗ /E(x∗2 ) as factors, e.g. E(Ri ) = α + β i,x∗ λx∗ and E(Ri ) = α + β i,R∗ [E(R∗ ) − α]. Proof: Recall that p = E(mx) implies p = E [proj(m | X) x], or p = E(x∗ x). Then 1 = E(mRi ) = E(x∗ Ri ) = E(x∗ )E(Ri ) + cov(x∗ , Ri ). Solving for the expected return, E(Ri ) =

cov(x∗ , Ri ) 1 cov(x∗ , Ri ) var(x∗ ) 1 − = − ∗ ∗ ∗ E(x ) E(x ) E(x ) var(x∗ ) E(x∗ )

which we can write as the desired single-beta model, E(Ri ) = α + β i,x∗ λx∗ . Notice that the zero-beta rate 1/E(x∗ ) appears when there is no riskfree rate. To derive a single beta representation with R∗ , recall the definition, R∗ =

x∗ E(x∗2 )

Substituting R∗ for x∗ , equation (6.81) implies that we can in fact construct a return R∗ from m that acts as the single factor in a beta model, µ ¶µ ¶ E(R∗2 ) cov(R∗ , Ri ) E(R∗2 ) cov(R∗ , Ri ) var(R∗ ) i − = + E(R ) = − E(R∗ ) E(R∗ ) E(R∗ ) var(R∗ ) E(R∗ ) 100

(81)

S ECTION 6.2

F ROM MEAN - VARIANCE FRONTIER TO A DISCOUNT FACTOR AND BETA REPRESENTATION

or, defining Greek letters in the obvious way, E(Ri ) = α + β Ri ,R∗ λR∗

(82)

Since the factor R∗ is also a return, its expected excess return over the zero beta rate gives the factor risk premium λR∗ . Applying equation (6.82) to R∗ itself, E(R∗ ) = α −

var(R∗ ) . E(R∗ )

(83)

So we can write the beta model in an even more traditional form E(Ri ) = α + β Ri ,R∗ [E(R∗ ) − α].

(84)

¥ Recall that R∗ is the minimum second moment frontier, on the lower portion of the meanvariance frontier. This is why R∗ has an unusual negative expected excess return or factor risk premium, λR∗ = −var(R∗ )/E(R∗ ) < 0. α is the zero-beta rate on R∗ shown in Figure15. Special cases A footnote to these constructions is that E(m), E(x∗ ), or E(R∗ ) cannot be zero, or you couldn’t divide by them. This is a pathological case: E(m) = 0 implies a zero price for the riskfree asset, and an infinite riskfree rate. If a riskfree rate is traded, we can simply observe that it is not infinite and verify the fact. Also, in a complete market, E(m) cannot be zero since, by absence of arbitrage, m > 0. We will see similar special cases in the remaining theorems: the manipulations only work for discount factor choices that do not imply zero or infinite riskfree rates. I discuss the issue in section 6.6 The manipulation from expected return-covariance to expected return-beta breaks down if var(m), var(x∗ ) or var(R∗ ) are zero. This is the case of pure risk neutrality. In this case, the covariances go to zero faster than the variances, so all betas go to zero and all expected returns become the same as the risk free rate.

6.2

From mean-variance frontier to a discount factor and beta representation

Rmv is on mean-variance frontier ⇒ m = a + bRmv ; E(Ri ) − α = β i [E(Rmv ) − α] We have seen that p = E(mx) implies a single−β model with a mean-variance efficient reference return, namely R∗ . The converse is also true: for (almost) any return on the meanvariance frontier, we can define a discount factor m that prices assets as a linear function of 101

C HAPTER 6

R ELATION BETWEEN DISCOUNT FACTORS , BETAS , AND MEAN - VARIANCE FRONTIERS

the mean-variance efficient return. Also, expected returns mechanically follow a single−β representation using the mean-variance efficient return as reference. I start with the discount factor. Theorem: There is a discount factor of the form m = a + bRmv if and only if Rmv is on the mean-variance frontier, and Rmv is not the riskfree rate. (If there is no riskfree rate, if Rmv is not the constant-mimicking portfolio return.) Graphical argument The basic idea is very simple, and Figure 19 shows the geometry for the complete markets case. The discount factor m = x∗ is proportional to R∗ . The mean-variance frontier is R∗ + wRe∗ . Pick a vector Rmv on the mean-variance frontier as shown in Figure 19. Then stretch it (bRmv ) and then subtract some of the 1 vector (a). Since Re∗ is generated by the unit vector, we can get rid of the Re∗ component and get back to the discount factor x∗ if we pick the right a and b. If the original return vector were not on the mean-variance frontier, then any linear combination a + bRmv with b 6= 0 would point in some of the n direction, which R∗ and x∗ do not. If b = 0, though, just stretching up and down the 1 vector will not get us to x∗ . Thus, we can only get a discount factor of the form a + bRmv if Rmv is on the frontier. You may remember that x∗ is not the only discount factor – all discount factors are of the form m = x∗ + ε with E(εx) = 0. Perhaps a + bR gives one of these discount factors, when R is not on the mean-variance frontier? This doesn’t work, however; n is still in the payoff space X while, by definition, ε is orthogonal to this space. If the mean-variance efficient return Rmv that we start with happens to lie right on the intersection of the stretched unit vector and the frontier, then stretching the Rmv vector and adding some unit vector are the same thing, so we again can’t get back to x∗ by stretching and adding some unit vector. The stretched unit payoff is the riskfree rate, so the theorem rules out the riskfree rate. When there is no riskfree rate, we have to rule out the “constant-mimicking portfolio return.” I treat this case in section 6.1. Algebraic proof Now, an algebraic proof that captures the same ideas. Proof. For an arbitrary R, try the discount factor model m = a + bR = a + b(R∗ + wRe∗ + n). We show that this discount factor prices an arbitrary payoff if and only if n = 0, and except for the w choice that makes R the riskfree rate, or the constant-mimicking portfolio return if there is no riskfree rate. We can determine a and b by forcing m to price any two assets. I find a and b to 102

(85)

S ECTION 6.2

F ROM MEAN - VARIANCE FRONTIER TO A DISCOUNT FACTOR AND BETA REPRESENTATION

bRmv er E - σ fronti

Rmv

Rf

R* 1 x* = a + bRmv Re* 0

Figure 19. There is a discount factor m = a + bRmv if and only if Rmv is on the mean-variance frontier and not the risk free rate. make the model price R∗ and Re∗ . 1 = E(mR∗ ) = aE(R∗ ) + bE(R∗2 ) 0 = E(mRe∗ ) = aE(Re∗ ) + bwE(Re∗2 ) = (a + bw) E(Re∗ ). Solving for a and b, a =

w wE(R∗ ) −

E(R∗2 ) 1 . b = − wE(R∗ ) − E(R∗2 )

Thus, if it is to price R∗ and Re∗ , the discount factor must be m=

w − (R∗ + wRe∗ + n) . wE(R∗ ) − E(R∗2 )

Now, let’s see if m prices an arbitrary payoff xi . Any xi ∈ X can also be decom103

(86)

C HAPTER 6

R ELATION BETWEEN DISCOUNT FACTORS , BETAS , AND MEAN - VARIANCE FRONTIERS

posed as xi = y i R∗ + wi Re∗ + ni . (See Figure 14 if this isn’t obvious.) The price of xi is yi , since both Re∗ and ni are zero-price (excess return) payoffs. Therefore, we want E(mxi ) = y i . Does it? µ ¶ (w − R∗ − wRe∗ − n)(y i R∗ + wi Re∗ + ni ) i E(mx ) = E wE(R∗ ) − E(R∗2 ) Using the orthogonality of R∗ , Re∗ n; E(n) = 0 and E(Re∗2 ) = E(Re∗ ) to simplify the product, E(mxi ) =

wyi E(R∗ ) − y i E(R∗2 ) − E(nni ) E(nni ) i = y . − wE(R∗ ) − E(R∗2 ) wE(R∗ ) − E(R∗2 )

To get p(xi ) = y i = E(mxi ), we need E(nni ) = 0. The only way to guarantee this condition for every payoff xi ∈ X is to insist that n = 0. Obviously, this construction can’t work if the denominator of (6.86) is zero, i.e. if w = E(R∗2 )/E(R∗ ) = 1/E(x∗ ). If there is a riskfree rate, then Rf = 1/E(x∗ ), so we are ruling out the case Rmv = R∗ + Rf Re∗ , which is the risk free rate. If ˆ = R∗ + E(R∗2 )/E(R∗ )Re∗ as a “constant there is no riskfree rate, I interpret R mimicking portfolio return” in section 5.3, and I give a graphical interpretation of this special case in section 6.1 ¥ We can generalize the theorem somewhat. Nothing is special about returns; any payoff of the form yR∗ +wRe∗ or yx∗ +wRe∗ can be used to price assets; such payoffs have minimum variance among all payoffs with given mean and price. Of course, we proved existence not uniqueness: m = a + bRmv + ², E(²x) = 0 also price assets as always. To get from the mean-variance frontier to a beta pricing model, we can just chain this theorem and the theorem of the last section together. There is a slight subtlety about special cases when there is no riskfree rate, but since it is not important for the basic points I relegate the direct connection and the special cases to section 6.2.

6.3

Factor models and discount factors

Beta-pricing models are equivalent to linear models for the discount factor m. E(Ri ) = α + λ0 β i ⇔ m = a + b0 f

104

S ECTION 6.3

FACTOR MODELS AND DISCOUNT FACTORS

We have shown that p = E(mx) implies a single beta representation using m, x∗ or R as factors. Let’s ask the converse question: suppose we have an expected return - beta model such as CAPM, APT, ICAPM, etc. What discount factor model does this imply? I show that an expected return - beta model is equivalent to a model for the discount factor that is a linear function of the factors in the beta model. This is an important and central result. It gives the connection between the discount factor formulation emphasized in this book and the expected return/beta, factor model formulation common in empirical work. ∗

You can write a linear factor model most compactly as m = b0 f , letting one of the factors be a constant. However, since we want a connection to the beta representation based on covariances rather than second moments, it is easiest to fold means of the factors in to the constant, and write m = a + b0 f with E(f) = 0 and hence E(m) = a. The connection is easiest to see in the special case that all the test assets are excess returns. Then 0 = E(mRe ) does not identify the mean of m, and we can normalize a arbitrarily. I find it convenient to normalize to E(m) = 1, or m = 1 + b0 [f − E (f )]. Then, Theorem: Given the model m = 1 + b0 [f − E (f )] ; 0 = E(mRe )

(87)

one can find λ such that E(Re ) = β 0 λ

(88)

where β are the multiple regression coefficients of excess returns Re on the factors. Conversely, given λ in (6.88), we can find b such that (6.87) holds. Proof: From (6.87) 0 = E(mRe ) = E(Re ) + b0 cov(f, Re ) E(Re ) = −b0 cov(f, Re ). From covariance to beta is quick, E(Re ) = −b0 var(f )var(f )−1 cov(f, Re ) = λ0 β Thus, λ and b are related by λ = −var(f )b. ¥ When the test assets are returns, the same idea works just as well, but gets a little more drowned in algebra since we have to keep track of the constant in m and the zero-beta rate in the beta model. 105

C HAPTER 6

R ELATION BETWEEN DISCOUNT FACTORS , BETAS , AND MEAN - VARIANCE FRONTIERS

Theorem: Given the model m = a + b0 f, 1 = E(mRi ),

(89)

one can find α and λ such that E(Ri ) = α + λ0 β i ,

(90)

where β i are the multiple regression coefficients of Ri on f with a constant. Conversely, given α and λ in a factor model of the form (6.90), one can find a, b such that (6.89) holds. Proof: We just have to construct the relation between (α, λ) and (a, b) and show that it works. Start with m = a + b0 f , 1 = E(mR), and hence E(R) =

cov(m, R) 1 E(Rf 0 )b 1 − = − E(m) E(m) a a

(91)

β i is the vector of the appropriate regression coefficients, ¡ ¢−1 β i ≡E ff 0 E(fRi ), so to get β in the formula, continue with E(R) =

1 E(Rf 0 )E(f f 0 )−1 E(f f 0 )b 1 E(f f 0 )b − = − β0 a a a a

Now, define α and λ to make it work, 1 1 = E (m) a 1 λ ≡ − E(ff 0 )b = − αE [mf ] a

α ≡

(6.92)

Using (6.92) we can just as easily go backwards from the expected return-beta representation to m = a + b0 f . As always, we have to worry about a special case of zero or infinite riskfree rates. We rule out E(m) = E(a + b0 f ) = 0 to keep (6.91) from exploding, and we rule ¥ out α = 0 and E(f f 0 ) singular to go from α, β, λ in (6.92) back to m. Given either model there is a model of the other form. They are not unique. We can add to m any random variable orthogonal to returns, and we can add spurious risk factors with zero β and/or λ , leaving pricing implications unchanged. We can also express the multiple beta model as a single beta model with m = a + b0 f as the single factor, or use its corresponding R∗ . Equation (6.92) shows that the factor risk premium λ can be interpreted as the price of the factor; A test of λ 6= 0 is often called a test of whether the “factor is priced.” More precisely, 106

S ECTION 6.3

FACTOR MODELS AND DISCOUNT FACTORS

λ captures the price E(mf ) of the (de-meaned) factors brought forward at the risk free rate. ˜ If we start with underlying factors f˜ such that the demeaned factors are f = f˜−E(f), h i ˜ = −α λ ≡ −α p f˜−E(f)

"

˜ ˜ E(f) p(f)− α

#

λ represents the price of the factors less their risk-neutral valuation, i.e. the factor risk premium. If the factors are not traded, λ is the model’s predicted price rather than a market price. Low prices are high risk premia, resulting in the negative sign. If the factors are returns with price one, then the factor risk premium is the expected return of the factor, less α, λ = E(f ) − α.

Note that the “factors” need not be returns (though they may be); they need not be orthogonal, and they need not be serially uncorrelated or conditionally or unconditionally meanzero. Such properties may occur as natural special cases, or as part of the economic derivation of specific factor models, but they are not required for the existence of a factor pricing representation. For example, if the riskfree rate is constant then Et (mt+1 ) is constant and at least the sum b0f should be uncorrelated over time. But if the riskfree rate is not constant, then Et (mt+1 ) = Et (b0f t+1 ) should vary over time. Factor-mimicking portfolios

It is often convenient to use factor-mimicking payoffs f ∗ = proj(f|X) factor-mimicking returns f∗ =

proj(f|X) p [proj(f |X)]

or factor-mimicking excess returns f ∗ = proj(f|Re ) in place of true factors. These payoffs carry the same pricing information as the original factors, and can serve as reference variables in expected return-beta representations When the factors are not already returns or excess returns, it is convenient to express a beta pricing model in terms of its factor mimicking portfolios rather than the factors themselves. Recall that x∗ = proj(m|X) carries all of m0 s pricing implications on X; p(x) = E(mx) = E(x∗ x). The factor-mimicking portfolios are just the same idea using the individual factors. 107

C HAPTER 6

R ELATION BETWEEN DISCOUNT FACTORS , BETAS , AND MEAN - VARIANCE FRONTIERS

Define the payoffs f ∗ by f ∗ = proj(f|X) Then, m = b0 f ∗ carries the same pricing implications on X as does m = b0 f : p = E(mx) = E(b0 f x) = E [b0 (projf |X) x] = E [b0 f ∗ x] .

(93)

(I include the constant as one of the factors.) The factor-mimicking portfolios also form a beta representation. Just go from (6.93) back to an expected return- beta representation E(Ri ) = α∗ + β ∗0 λ∗ ,

(94)

and find λ∗ , α∗ using (6.92). The β ∗ are the regression coefficients of the returns Ri on the factor-mimicking portfolios, not on the factors, as they should be. It is more traditional to use the returns or excess returns on the factor-mimicking portfolios rather than payoffs as I have done so far. To generate returns, divide the payoff by its price, f∗ =

proj(f |X) . p [proj(f|X)]

The resulting b will be scaled down by the price of the factor-mimicking payoff, and the model is the same. Note you project on the space of payoffs, not of returns. Returns R are not a space, since they don’t contain zero. If the test assets are all excess returns, you can even more easily project the factors on the set of excess returns, which are a space since they do include zero. If we define f ∗ = proj(f|Re ) then of course the excess returns f ∗ carry the same pricing implications as the factors f for a set of excess returns; m = b0 f ∗ satisfies 0 = E(mRei ) and E(Rei ) = β i,f ∗ λ = β i,f ∗ E(f ∗ )

6.4

Discount factors and beta models to mean - variance frontier

From m, we can construct R∗ which is on the mean variance frontier If a beta pricing model holds, then the return R∗ on the mean-variance frontier is a linear combination of the factor-mimicking portfolio returns. 108

S ECTION 6.5

T HREE RISKFREE RATE ANALOGUES

Any frontier return is a combination of R∗ and one other return, a risk free rate or a risk free rate proxy. Thus, any frontier return is a linear function of the factor-mimicking returns plus a risk free rate proxy. It’s easy to show that given m that we can find a return on the mean-variance frontier. Given m construct x∗ = proj(m|X) and R∗ = x∗ /E(x∗2 ). R∗ is the minimum second moment return, and hence on the mean-variance frontier. Similarly, if you have a set of factors f for a beta model, then a linear combination of the factor-mimicking portfolios is on the mean-variance frontier. A beta model is the same as m = b0 f . Since m is linear in f , x∗ is linear in f ∗ = proj(f |X), so R∗ is linear in the factor mimicking payoffs f ∗ or their returns f ∗ /p(f ∗ ). Section 5.4 showed how we can span the mean-variance frontier with R∗ and a risk free rate, if there is one, or the zero-beta, minimum variance, or constant-mimicking portfolio ˆ = proj(1|X)/p[proj(1|X)] if there is no risk free rate. The latter is particularly return R nice in the case of a linear factor model, since we may consider the constant as a factor, so the frontier is entirely generated by factor-mimicking portfolio returns.

6.5

Three riskfree rate analogues

I introduce three counterparts to the risk free rate that show up in asset pricing formulas when there is no risk free rate. The three returns are the zero-beta return, the minimumvariance return and the constant-mimicking portfolio return. Three different generalizations of the riskfree rate are useful when a risk free rate or unit payoff is not in the set of payoffs. These are the zero-beta return, the minimum-variance return and the constant-mimicking portfolio return. I introduce the returns in this section, and I use them in the next section to state some special cases involving the mean-variance frontier. Each of these returns maintains one property of the risk free rate in a market in which there is no risk free rate. The zero-beta return is a mean-variance efficient return that is uncorrelated with another given mean-variance efficient return. The minimum-variance return is just that. The constant-mimicking portfolio return is the return on the payoff “closest” to the unit payoff. Each of these returns one has a representation in the standard form R∗ + wRe∗ with slightly different w. In addition, the expected returns of these risky assets are used in some asset pricing representations. For example, the zero beta rate is often used to refer to the expected value of the zero beta return. Each of these riskfree rate analogues is mean-variance efficient. Thus, I characterize each one by finding its weight w in a representation of the form R∗ + wRe∗ . We derived such a

109

C HAPTER 6

R ELATION BETWEEN DISCOUNT FACTORS , BETAS , AND MEAN - VARIANCE FRONTIERS

representation above for the riskfree rate as equation (5.72), Rf = R∗ + Rf Re∗ .

(95)

In the last subsection, I show how each riskfree rate analogue reduces to the riskfree rate when there is one. 6.5.1

Zero-beta return for R∗

The zero beta return for R∗ , denoted Rα , is the mean-variance efficient return uncorrelated with R∗ . Its expected return is the zero beta rate α = E(Ra ). This zero beta return has representation Ra = R∗ +

var(R∗ ) Re∗ , E(R∗ )E(Re∗ )

and the corresponding zero beta rate is α = E(Rα ) =

1 E(R∗2 ) = . E(R∗ ) E(x∗ )

The zero beta rate is found graphically in mean-standard deviation space by extending the tangency at R∗ to the vertical axis. It is also the inverse of the price that x∗ and R∗ assign to the unit payoff. The riskfree rate Rf is of course uncorrelated with R∗ . Risky returns uncorrelated with R earn the same average return as the risk free rate if there is one, so they might take the place of Rf when the latter does not exist. For any return Rα that is uncorrelated with R∗ we have E(R∗ Rα ) = E(R∗ )E(Rα ), so ∗

α = E(Rα ) =

1 E(R∗2 ) = . E(R∗ ) E(x∗ )

The first equality introduces a popular notation α for this rate. I call α the zero beta rate, and Ra the zero beta return. There is no riskfree rate, so there is no security that just pays α. As you can see from the formula, the zero-beta rate is the inverse of the price that R∗ and x assign to the unit payoff, which is another natural generalization of the riskfree rate. It is called the zero beta rate because cov(R∗ , Rα ) = 0 implies that the regression beta of Rα on R∗ is zero. More precisely, one might call it the zero beta rate on R∗ , since one can calculate zero-beta rates for returns other than R∗ and they are not the same as the zero-beta rate for R∗ In particular, the zero-beta rate on the “market portfolio” will generally be different from the zero beta rate on R∗ . ∗

110

S ECTION 6.5

T HREE RISKFREE RATE ANALOGUES

E(R) α= E(R*2 )/ E(R* ) = 1/E(x*)

Rα R*

σ(R) Figure 20. Zero-beta rate α and zero-beta return Ra for R∗ .

111

C HAPTER 6

R ELATION BETWEEN DISCOUNT FACTORS , BETAS , AND MEAN - VARIANCE FRONTIERS

I draw α in Figure 20 as the intersection of the tangency and the vertical axis. This is a property of any return on the mean variance frontier: The expected return on an asset uncorrelated with the mean-variance efficient asset (a zero-beta asset) lies at the point so constructed. To check this geometry, use similar triangles: p The length ofp R∗ in Figure 20 is p ∗ ∗2 ∗2 E(R ), and its vertical extent is E(R ). Therefore, α/ E(R ) = E(R∗2 )/E(R∗ ), or α = E(R∗2 )/E(R∗ ). Since R∗ is on the lower portion of the mean-variance frontier, this zero beta rate α is above the minimum variance return. Note that in general α 6= 1/E(m). Projecting m on X preserves asset pricing implications on X but not for payoffs not in X. Thus if a risk free rate is not traded, x∗ and m may differ in their predictions for the riskfree rate as for other nontraded assets. The zero beta return is the rate of return on the mean-variance frontier with mean equal to the zero beta rate, as shown in Figure 20. We want to characterize this return in R∗ + wRe∗ form. To do this, we want to find w such that E(Ra ) =

E(R∗2 ) = E(R∗ ) + wE(Re∗ ). E(R∗ )

Solving, the answer is w=

var(R∗ ) E(R∗2 ) − E(R∗ )2 = . ∗ e∗ E(R )E(R ) E(R∗ )E(Re∗ )

Thus, the zero beta return is Ra = R∗ +

var(R∗ ) Re∗ , E(R∗ )E(Re∗ )

expression (6.103). Note that the weight is not E(Ra ) = E(R∗2 )/E(R∗ ). When there is no risk free rate, the weight and the mean return are different. 6.5.2

Minimum variance return

The minimum variance return has the representation Rmin. var. = R∗ +

E(R∗ ) Re∗ . 1 − E(Re∗ )

The riskfree rate obviously is the minimum variance return when it exists. When there is no risk free rate, the minimum variance return is Rmin. var. = R∗ +

E(R∗ ) Re∗ . 1 − E(Re∗ )

112

(96)

S ECTION 6.5

T HREE RISKFREE RATE ANALOGUES

Taking expectations, E(Rmin. var. ) = E(R∗ ) +

E(R∗ ) E(R∗ ) E(Re∗ ) = . e∗ 1 − E(R ) 1 − E(Re∗ )

The minimum variance return retains the nice property of the risk free rate, that its weight on Re∗ is the same as its mean, Rmin. var. = R∗ + E(Rmin. var. )Re∗ just as Rf = R∗ + Rf Re∗ . When there is no risk free rate, the zero-beta and minimum variance returns are not the same. You can see this fact clearly in Figure 20. We can derive expression (6.96) for the minimum variance return by brute force: choose w in R∗ + wRe∗ to minimize variance. min var(R∗ + wRe∗ ) = E[(R∗ + wRe∗ )2 ] − E(R∗ + wRe∗ )2 = w

= E(R∗2 ) + w2 E(Re∗ ) − E(R∗ )2 − 2wE(R∗ )E(Re∗ ) − w2 E(Re∗ )2 . The first order condition is 0 = wE(Re∗ )[1 − E(Re∗ )] − E(R∗ )E(Re∗ ) w=

6.5.3

E(R∗ ) . 1 − E(Re∗ )

Constant-mimicking portfolio return

The constant-mimicking portfolio return is defined as the return on the projection of the unit vector on the payoff space, ˆ = proj(1|X) . R p [proj(1|X)] It has the representation ∗2 ˆ = R∗ + E(R ) Re∗ . R E(R∗ )

When there is a risk free rate, it is the rate of return on a unit payoff, Rf = 1/p(1). When there is no risk free rate, we might define the rate of return on the mimicking portfolio for a 113

C HAPTER 6

R ELATION BETWEEN DISCOUNT FACTORS , BETAS , AND MEAN - VARIANCE FRONTIERS

unit payoff, ˆ = proj(1|X) . R p [proj(1|X)] I call this object the constant-mimicking portfolio return. The mean-variance representation of the constant-mimicking portfolio return is ∗2 ˆ = R∗ + αRe∗ = R∗ + E(R ) Re∗ . R E(R∗ )

(97)

Note that the weight α equal to the zero beta rate creates the constant-mimicking return, not the zero beta return. To show (6.97), start with property (5.73), Re∗ = proj(1|X) −

E(R∗ ) ∗ R . E(R∗2 )

(98)

Take the price of both sides. Since the price of Re∗ is zero and the price of R∗ is one, we establish p [proj(1|X)] =

E(R∗ ) . E(R∗2 )

(99a)

Solving (6.98) for proj(1|X), dividing by (6.99a) we obtain the right hand side of (6.97). 6.5.4

Risk free rate

The risk free rate has the mean-variance representation Rf = R∗ + Rf Re∗ . The zero-beta, minimum variance and constant-mimicking portfolio returns reduce to this formula when there is a risk free rate. Again, we derived in equation (5.72) that the riskfree rate has the representation, Rf = R∗ + Rf Re∗ .

(100)

Obviously, we should expect that the zero-beta return, minimum-variance return, and constantmimicking portfolio return reduce to the riskfree rate when there is one. These other rates are ∗2 ˆ = R∗ + E(R ) Re∗ constant-mimicking: R E(R∗ )

114

(101)

S ECTION 6.6

M EAN - VARIANCE SPECIAL CASES WITH NO RISKFREE RATE

minimum-variance: Rmin. var. = R∗ +

zero-beta: Rα = R∗ +

E(R∗ ) Re∗ 1 − E(Re∗ )

var(R∗ ) Re∗ . E(R∗ )E(Re∗ )

(102)

(103)

To establish that these are all the same when there is a riskfree rate, we need to show that Rf =

E(R∗ ) var(R∗ ) E(R∗2 ) = = E(R∗ ) 1 − E(Re∗ ) E(R∗ )E(Re∗ )

(104)

We derived the first equality above as equation (5.74). To derive the second equality, take expectations of (6.95), Rf = E(R∗ ) + Rf E(Re∗ )

(105)

and solve for Rf . To derive the third equality, use the first equality from (6.104) in (6.105), E(R∗2 ) = E(R∗ ) + Rf E(Re∗ ). E(R∗ ) Solving for Rf , Rf =

6.6

var(R∗ ) E(R∗2 ) − E(R∗ )2 = . E(R∗ )E(Re∗ ) E(R∗ )E(Re∗ )

Mean-variance special cases with no riskfree rate

We can find a discount factor from any mean-variance efficient return except the constantmimicking return. We can find a beta representation from any mean-variance efficient return except the minimum-variance return. I collect in this section the special cases for the equivalence theorems of this chapter. The special cases all revolve around the problem that the expected discount factor, price of a unit payoff or riskfree rate must not be zero or infinity. This is typically an issue of theoretical rather than practical importance. In a complete, arbitrage free market, m > 0 so we know E(m) > 0. If a riskfree rate is traded you can observe ∞ > E(m) = 1/Rf > 0. However, in an incomplete market in which no riskfree rate is traded, there are many discount factors with the same asset pricing implications, and you might have happened to choose one with E(m) = 0 in your manipulations. By and large, this is easy to avoid: choose another of the 115

C HAPTER 6

R ELATION BETWEEN DISCOUNT FACTORS , BETAS , AND MEAN - VARIANCE FRONTIERS

many discount factors with the same pricing implications that does not have E(m) = 0. More generally, when you choose a particular discount factor you are choosing an extension of the current set of prices and payoffs; you are viewing the current prices and payoffs as a subset of a particular contingent-claim economy. Make sure you pick a sensible one. Therefore, we could simply state the special cases as “when a riskfree rate is not traded, make sure you use discount factors with 0 < E(m) < ∞.” However, it is potentially useful and it certainly is traditional to specify the special return on the mean-variance frontier that leads to the infinite or zero implied riskfree rate, and to rule it out directly. This section works out what those returns are and shows why they must be avoided. 6.6.1

The special case for mean variance frontier to discount factor

When there is no riskfree rate, we can find a discount factor that is a linear function of any mean-variance efficient return except the constant-mimicking portfolio return. In section 6.2, we saw that we can form a discount factor a + bRmv from any mean∗2 ) e∗ variance efficient return Rmv except one particular return, of the form R∗ + E(R E(R∗ ) R . This return led to an infinite m. We now recognize this return as the risk-free rate, when there is one, or the constant-mimicking portfolio return, if there is no riskfree rate. Figure 21 shows the geometry of this case. To use no more than three dimensions I had to reduce the return and excess return spaces to lines. The payoff space X is the plane joining the return and excess return sets as shown. The set of all discount factors is m = x∗ + ε, E(εx) = 0, the line through x∗ orthogonal to the payoff space X in the figure. I draw the unit payoff (the dot marked “1” in Figure 21) closer to the viewer than the plane X, and I draw a vector through the unit payoff coming out of the page. Take any return on the mean-variance frontier, Rmv . (Since the return space only has two dimensions, all returns are on the frontier.) For a given Rmv , the space a + bRmv is the plane spanned by Rmv and the unit payoff. This plane lies sideways in the figure. As the figure shows, there is a vector a + bRmv in this plane that lies on the line of discount factors. Next, the special case. This construction would go awry if the plane spanning the unit payoff and the return Rmv were parallel to the plane containing the discount factor. Thus, ˆ in the Figure. This is a return the construction would not work for the return marked R corresponding to a payoff that is the projection of the unit payoff on to X, so that the residual will be orthogonal to X, as is the line of discount factors. With Figure 21 in front of us, we can also see why the constant-mimicking portfolio return is not the same thing as the minimum-variance return. Variance is the size or second moment

116

S ECTION 6.6

M EAN - VARIANCE SPECIAL CASES WITH NO RISKFREE RATE

Discount factors Rmv

^ Constant-mimicking return R

R*

R

a+b Rmv x* 1

X Re 0 Figure 21. One can construct a discount factor m = a + bRmv from any ˆ mean-variance-efficient return except the constant-mimicking return R. of the residual in a projection (regression) on 1. ¤ £ ¤ £ var(x) = E (x − E(x))2 = E (x − proj(x|1))2 = ||x − proj(x|1)||2

Thus, the minimum variance return is the return closest to extensions of the unit vector. It is formed by projecting returns on the unit vector. The constant-mimicking portfolio return is the return on the payoff closest to 1 It is formed by projecting the unit vector on the set of payoffs. 6.6.2

The special case for mean-variance frontier to a beta model

We can use any return on the mean-variance frontier as the reference return for a single beta representation, except the minimum-variance return. We already know mean variance frontiers ⇔ discount factor and discount factor ⇔ single beta representation, so at a superficial level we can string the two theorems together to go 117

C HAPTER 6

R ELATION BETWEEN DISCOUNT FACTORS , BETAS , AND MEAN - VARIANCE FRONTIERS

from a mean-variance efficient return to a beta representation. However it is more elegant to go directly, and the special cases are also a bit simpler this way. Theorem: There is a single beta representation with a return Rmv as factor, E(Ri ) = αRmv + β i,Rmv [E(Rmv ) − α] , if and only if Rmv is mean-variance efficient and not the minimum variance return. This famous theorem is given by Roll (1976) and Hansen and Richard (1987). We rule out minimum variance to rule out the special case E(m) = 0. Graphically, the zero-beta rate is formed from the tangency to the mean-variance frontier as in Figure 20. I use the notation αRmv to emphasize that we use the zero-beta rate corresponding to the particular meanvariance return Rmv that we use as the reference return. If we used the minimum-variance return, that would lead to an infinite zero-beta rate. Proof: The mean-variance frontier is Rmv = R∗ + wRe∗ . Any return is Ri = R∗ + wi Re∗ + ni . Thus, E(Ri ) = E(R∗ ) + wi E(Re∗ )

(106)

Now, £ ¤ cov(Ri , Rmv ) = cov (R∗ + wRe∗ ) , (R∗ + wi Re∗ )

= var(R∗ ) + wwi var(Re∗ ) − (w + wi )E(R∗ )E(Re∗ ) = var(R∗ ) − wE(R∗ )E(Re∗ ) + wi [w var(Re∗ ) − E(R∗ )E(Re∗ )]

Thus, cov(Ri , Rmv ) and E(Ri ) are both linear functions of wi . We can solve cov(Ri , Rmv ) for wi , plug into the expression for E(Ri ) and we’re done. To do this, of course, we must be able to solve cov(Ri , Rmv ) for wi . This requires w 6=

E(R∗ )E(Re∗ ) E(R∗ ) E(R∗ )E(Re∗ ) = = var(Re∗ ) E(Re∗2 ) − E(Re∗ )2 1 − E(Re∗ )

which is the condition for the minimum variance return.

6.7 1.

2.

(107) ¥

Problems

In the argument that Rmv on the mean variance frontier, Rmv = R∗ + wRe∗ , implies a discount factor m = a + bRmv , do we have to rule out the case of risk neutrality? (Hint: What is Re∗ when the economy is risk-neutral?) If you use factor mimicking portfolios as in (6.93), you know that the predictions for expected returns are the same as they are if you use the factors themselves . Are the α∗ , 118

S ECTION 6.7

3. 4.

5.

6.

P ROBLEMS

λ∗ , and β ∗ for the factor mimicking portfolio representation the same as the original α, λ, and β of the factor pricing model? Suppose the CAPM is true, m = a − bRm prices a set of assets, and there is a risk-free rate Rf . Find R∗ in terms of the moments of Rm , Rf . If you express the mean-variance frontier as a linear combination of factor-mimicking portfolios from a factor model, do the relative weights of the various factor portfolios in the mean-variance efficient return change as you sweep out the frontier, or do they stay the same? (Start with the riskfree rate case) For an arbitrary mean-variance efficient return of the form R∗ + wRe∗ , find its zero-beta return and zero-beta rate. Show that your rate reduces to the riskfree rate when there is one. When the economy is risk neutral, and if there is no risk-free rate, show that the zero-beta, minimum-variance, and constant-mimicking portfolio returns are again all equivalent, though not equal to the risk-free rate. (In this case, the mean-variance frontier is just the minimum-variance point.)

119

Chapter 7. Implications of existence and equivalence theorems Existence of a discount factor means p = E(mx) is innocuous, and all content flows from the discount factor model. The theorems apply to sample moments too; the dangers of fishing up ex-post or sample mean-variance efficient portfolios. Sources of discipline in factor fishing expeditions. The joint hypothesis problem. How efficiency tests are the same as tests of economic discount factor models. Factors vs. their mimicking portfolios. Testing the number of factors. Plotting contingent claims on the axis vs. mean and variance. The theorems on the existence of a discount factor, and the equivalence between the p = E(mx), expected return - beta, and mean-variance views of asset pricing have important implications for how we approach and evaluate empirical work. The equivalence theorems are obviously important, especially to the theme of this book, to show that the choice of discount factor language versus expected return-beta language or mean-variance frontier is entirely one of convenience. Nothing in the more traditional statements is lost. p = E(mx) is innocuous Before Roll (1976), expected return – beta representations had been derived in the context of special and explicit economic models, especially the CAPM. In empirical work, the success of any expected return - beta model seemed like a vindication of the whole structure. The fact that, for example, one might use the NYSE value-weighted index portfolio in place of the return on total wealth predicted by the CAPM seemed like a minor issue of empirical implementation. When Roll showed that mean-variance efficiency implies a single beta representation, all that changed. Some single beta representation always exists, since there is some meanvariance efficient return. The asset pricing model only serves to predict that a particular return (say, the “market return”) will be mean-variance efficient. Thus, if one wants to “test the CAPM” it becomes much more important to be choosy about the reference portfolio, to guard against stumbling on something that happens to be mean-variance efficient and hence prices assets by construction. 120

This insight led naturally to the use of broader wealth indices (Stambaugh 1982) in the reference portfolio to provide a more grounded test of the CAPM. However, this approach has not caught on. Stocks are priced with stock factors, bonds with bond factors, and so on. More recently, stocks sorted on size, book/market, and past performance characteristics are priced by portfolios sorted on those characteristics. Part of the reason for this is that the betas are small; stocks and bonds are not highly correlated so risk premia from one source of betas have small impacts on another set of average returns. Larger measures of wealth including human capital and real estate do not come with high frequency price data, so adding them to a wealth portfolio has little effect on betas. The good news in this existence theorem is that you can always start by writing an expected return-beta model, knowing that you have imposed almost no structure in doing so. The bad news is that you haven’t gotten very far. All the economic, statistical and predictive content comes in picking the factors. The theorem that, from the law of one price, there exists some discount factor m such that p = E(mx) is just an updated restatement of Roll’s theorem. The content is all in m = f (data) not in p = E(mx). Again, an asset pricing framework that initially seemed to require a lot of completely unbelievable structure–the representative consumer consumptionbased model in complete frictionless markets–turns out to require (almost) no structure at all. Again, the good news is that you can always start by writing p = E(mx), and need not suffer criticism about hidden contingent claim or representative consumer assumptions in so doing. The bad news is that you haven’t gotten very far by writing p = E(mx) as all the economic, statistical and predictive content comes in picking the discount factor model m = f (data). Ex-ante and ex-post. I have been deliberately vague about the probabilities underlying expectations and other moments in the theorems. The fact is, the theorems hold for any set of probabilities4 . Thus, the existence and equivalence theorems work equally well ex-ante as ex-post: E(mx), β, E(R) and so forth can refer to agent’s subjective probability distributions, objective population probabilities, or to the moments realized in a given sample. Thus, if the law of one price holds in a sample, one may form an x∗ from sample moments that satisfies p(x) = E(x∗ x), exactly, in that sample, where p(x) refers to observed prices and E(x∗ x) refers to the sample average. Equivalently, if the sample covariance matrix of a set of returns is nonsingular, there exists an ex-post mean-variance efficient portfolio for which sample average returns line up exactly with sample regression betas. This observation points to a great danger in the widespread exercise of searching for and statistically evaluating ad-hoc asset pricing models. Such models are guaranteed empirical success in a sample if one places little enough structure on what is included in the discount factor function. The only reason the model doesn’t work perfectly is the restrictions the researcher has imposed on the number or identity of the factors included in m, or the parameters of the function relating the factors to m. Since these restrictions are the entire content of the 4

Precisely, any set of probabilities that agree on impossible (zero-probability) events.

121

C HAPTER 7

I MPLICATIONS OF EXISTENCE AND EQUIVALENCE THEOREMS

model, they had better be interesting, carefully described and well motivated! Obviously, this is typically not the case or I wouldn’t be making such a fuss about it. Most empirical asset pricing research posits an ad-hoc pond of factors, fishes around a bit in that set, and reports statistical measures that show “success,” in that the model is not statistically rejected in pricing an ad-hoc set of portfolios. The set of discount factors is usually not large enough to give the zero pricing errors we know are possible, yet the boundaries are not clearly defined. Discipline What is wrong, you might ask, with finding an ex-post efficient portfolio or x∗ that prices assets by construction? Perhaps the lesson we should learn from the existence theorems is to forget about economics, the CAPM, marginal utility and all that, and simply price assets with ex-post mean variance efficient portfolios that we know set pricing errors to zero! The mistake is that a portfolio that is ex-post efficient in one sample, and hence prices all assets in that sample, is unlikely to be mean-variance efficient, ex-ante or ex-post, in the next sample, and hence is likely to do a poor job of pricing assets in the future. Similarly, the portfolio x∗ = p0 E(xx0 )−1 x (using the sample second moment matrix) that is a discount factor by construction in one sample is unlikely to be a discount factor in the next sample; the required portfolio weights p0 E(xx0 )−1 change, often drastically, from sample to sample. For example, suppose the CAPM is true, the market portfolio is ex-ante mean-variance efficient, and sets pricing errors to zero if you use true or subjective probabilities. Nonetheless, the market portfolio is unlikely to be ex-post mean-variance efficient in any given sample. In any sample, there will be lucky winners and unlucky losers. An ex-post mean variance efficient portfolio will be a Monday-morning quarterback; it will tell you to put large weights on assets that happened to be lucky in a given sample, but are no more likely than indicated by their betas to generate high returns in the future. “Oh, if I had only bought Microsoft in 1982...” is not a useful guide to forming a mean-variance efficient portfolio today. (In fact, mean-reversion in the market and book/market effects in individual stocks suggest that if anything, assets with unusually good returns in the past are likely to do poorly in the future!) The only solution is to impose some kind of discipline in order to avoid dredging up spuriously good in-sample pricing. The situation is the same as in traditional regression analysis. Regressions are used to forecast or to explain a variable y by other variables x in a regression y = x0 β + ε. By blindly including right hand variables, one can produce models with arbitrarily good statistical measures of fit. But this kind of model is typically unstable out of sample or otherwise useless for explanation or forecasting. One has to carefully and thoughtfully limit the search for right hand variables x to produce good models. What makes for an interesting set of restrictions? Econometricians wrestling with y = x0 β + ε have been thinking about this question for about 50 years, and the best answers are 1) use economic theory to carefully specify the right hand side and 2) use a battery of 122

cross-sample and out-of-sample stability checks. Alas, this advice is hard to follow. Economic theory is usually either silent on what variables to put on the right hand side of a regression, or allows a huge range of variables. The same is true in finance. “What are the fundamental risk factors?” is still an unanswered question. At the same time one can appeal to the APT and ICAPM to justify the inclusion of just about any desirable factor (Fama 1991 calls these theories a “fishing license.”) Thus, you will grow old waiting for theorists to provide useful answers to this kind of question. Following the purely statistical advice, the battery of cross-sample and out-of-sample tests often reveals the model is unstable, and needs to be changed. Once it is changed, there is no more out-of-sample left to check it. Furthermore, even if one researcher is pure enough to follow the methodology of classical statistics, and wait 50 years for another fresh sample to be available before contemplating another model, his competitors and journal editors are unlikely to be so patient. In practice, then, out of sample validation is not a strong guard against fishing. Nonetheless, these are the only standards we have to guard against fishing. In my opinion, the best hope for finding pricing factors that are robust out of sample and across different markets, is to try to understand the fundamental macroeconomic sources of risk. By this I mean, tying asset prices to macroeconomic events, in the way the ill-fated consumption based model does via mt+1 = βu0 (ct+1 )/u0 (ct ). The difficulties of the consumption-based model have made this approach lose favor in recent years. However, the alternative approach is also running into trouble that the number and identity of empirically-determined risk factors does not seem stable. Every time a new anomaly or data set pops up, a new set of ad-hoc factors gets created to explain them! In any case, one should always ask of a factor model, “what is the compelling economic story that restricts the range of factors used?” and / or “what statistical restraints are used” to keep from discovering ex-post mean variance efficient portfolios, or to ensure that the results will be robust across samples. The existence theorems tell us that the answers to these questions are the only content of the exercise. If the purpose of the model is not just to predict asset prices but also to explain them, this puts an additional burden on economic motivation of the risk factors. There is a natural resistance to such discipline built in to our current statistical methodology for evaluating models (and papers). When the last author fished around and produced an ad-hoc factor pricing model that generates 1% average pricing errors, it is awfully hard to persuade readers, referees, journal editors, and clients that your economically motivated factor pricing model is better despite 2% average pricing errors. Your model may really be better and will therefore continue to do well out of sample when the fished model falls by the wayside of financial fashion, but it is hard to get past statistical measures of in-sample fit. One hungers for a formal measurement of the number of hurdles imposed on a factor fishing ¯ 2 . Absent a numerical correction, we expedition, like the degrees of freedom correction in R have to use judgment to scale back apparent statistical successes by the amount of economic and statistical fishing that produced them. 123

C HAPTER 7

I MPLICATIONS OF EXISTENCE AND EQUIVALENCE THEOREMS

Mimicking portfolios The theorem x∗ = proj(m|X) also has interesting implications for empirical work. The pricing implications of any model can be equivalently represented by its factor-mimicking portfolio. If there is any measurement error in a set of economic variables driving m, the factor-mimicking portfolios for the true m will price assets better than an estimate of m that uses the measured macroeconomic variables. Thus, it is probably not a good idea to evaluate economically interesting models with statistical horse races against models that use portfolio returns as factors. Economically interesting models, even if true and perfectly measured, will just equal the performance of their own factor-mimicking portfolios, even in large samples. They will always lose in sample against ad-hoc factor models that find nearly ex-post efficient portfolios. This said, there is an important place for models that use returns as factors. After we have found the underlying true macro factors, practitioners will be well advised to look at the factor-mimicking portfolio on a day-by-day basis. Good data on the factor-mimicking portfolios will be available on a minute-by-minute basis. For many purposes, one does not have to understand the economic content of a model. But this fact does not tell us to circumvent the process of understanding the true macroeconomic factors by simply fishing for factor-mimicking portfolios. The experience of practitioners who use factor models seems to bear out this advice. Large commercial factor models resulting from extensive statistical analysis (otherwise known as fishing) perform poorly out of sample, as revealed by the fact that the factors and loadings (β) change all the time. Also models specified with economic fundamentals will always seem to do poorly in a given sample against ad-hoc variables (especially if one fishes an ex-post mean-variance efficient portfolio out of the latter!). But what other source of discipline do we have? Irrationality and Joint Hypothesis Finance contains a long history of fighting about “rationality” vs. “irrationality” and “efficiency” vs. “inefficiency” of asset markets. The results of many empirical asset pricing papers are sold as evidence that markets are “inefficient” or that investors are “irrational.” For example, the crash of October 1987, and various puzzles such as the small-firm, book/market, seasonal effects or long-term predictability have all been sold this way. However, none of these puzzles documents an arbitrage opportunity5 . Therefore, we know that there is a “rational model”–a stochastic discount factor, an efficient portfolio to use in a single-beta representation—that rationalizes them all. And we can confidently predict this situation to continue; real arbitrage opportunities do not last long! Fama (1970) contains a famous statement of the same point. Fama emphasized that any test of “efficiency” is a joint test of efficiency and a “model of market equilibrium.” Translated, an asset pricing model, or a model of m. 5 The closed-end fund puzzle comes closest since it documents an apparent violation of the law of one price. However, you can’t costlessly short closed end funds, and we have ignored short sales constraints so far.

124

But surely markets can be “irrational” or “inefficient” without requiring arbitrage opportunities? Yes, they can, if (and only if) the discount factors that generate asset prices are disconnected from marginal rates of substitution or transformation in the real economy. But now we are right back to specifying and testing economic models of the discount factor! At best, an asset pricing puzzle might be so severe that we can show that the required discount factors are completely “unreasonable” (by some standard) measures of real marginal rates of substitution and/or transformation, but we still have to say something about what a reasonable marginal rate looks like. In sum, the existence theorems mean that there are no quick proofs of “rationality” or “irrationality.” The only game in town for the purpose of explaining asset prices is thinking about economic models of the discount factor. The number of factors. Many asset pricing tests focus on the number of factors required to price a cross-section of assets. The equivalence theorems imply that this is a silly question. A linear factor model m = b0 f or its equivalent expected return / beta model E(Ri ) = α + β 0if λf are not unique representations. In particular, given any multiple-factor or multiple-beta representation we can easily find a single-beta representation. The single factor m = b0 f will price assets just as well as the original factors f, as will x∗ = proj(b0 f | X) or the corresponding R∗ . All three options give rise to single-beta models with exactly the same pricing ability as the multiple factor model. We can also easily find equivalent representations with different numbers (greater than one) of factors. For example, write µ ¶ b3 m = a + b1 f1 + b2 f2 + b3 f3 = a + b1 f1 + b2 f2 + f3 = a + b1 f1 + b2 fˆ2 b2 to reduce a “three factor” model to a “two factor” model. In the ICAPM language, consumption itself could serve as a single state variable, in place of the S state variables presumed to drive it. There are times when one is interested in a multiple factor representation. Sometimes the factors have an economic interpretation that is lost on taking a linear combination. But the pure number of pricing factors is not a meaningful question. Discount factors vs. mean, variance and beta. The point of the previous chapter was to show how the discount factor, mean-variance, and expected return- beta models are all equivalent representations of asset pricing. It seems a good moment to contrast them as well; to understand why the mean-variance and beta language developed first, and to think about why the discount factor language seems to be taking over. Asset pricing started by putting mean and variance of returns on the axes, rather than payoff in state 1 payoff in state 2, etc. as we do now. The early asset pricing theorists posed the question just right: they wanted to treat assets in the apples-and-oranges, indifference 125

C HAPTER 7

I MPLICATIONS OF EXISTENCE AND EQUIVALENCE THEOREMS

curve and budget set framework of macroeconomics. The problem was, what labels to put on the axis? Clearly, “IBM stock” and “GM stock” is not a good idea; investors do not value securities per se, but value some aspects of the stream of random cash flows that those securities give rise to. Their brilliant insight was to put the mean and variance of the portfolio return on the axis; to treat these as “hedonics” by which investors valued their portfolios. Investors plausibly want more mean and less variance. They gave investors “utility functions” defined over this mean and variance, just as standard utility functions are defined over apples and oranges. The mean-variance frontier is the “budget set.” With this focus on portfolio mean and variance, the next step was to realize that each security’s mean return measures its contribution to the portfolio mean, and that regression betas on the overall portfolio give each security’s contribution to the portfolio variance. The mean-return vs. beta description for each security followed naturally. In a deep sense, the transition from mean-variance frontiers and beta models to discount factors represents the realization that putting consumption in state 1 and consumption in state 2 on the axes — specifying preferences and budget constraints over state-contingent consumption — is a much more natural mapping of standard microeconomics into finance than putting mean, variance, etc. on the axes. If for no other reason, the contingent claim budget constraints are linear, while the mean-variance frontier is not. Thus, I think, the focus on means and variance, the mean-variance frontier and expected return/beta models is all due to an accident of history, that the early asset pricing theorists happened to put mean and variance on the axes rather than state contingent consumption. Well, here we are, why prefer one language over another? The discount factor language has an advantage for its simplicity, generality, mathematical convenience, and elegance. These virtues are to some extent in the eye of the beholder, but to this beholder, it is inspiring to be able to start every asset pricing calculation with one equation, p = E(mx). This equation covers all assets, including bonds, options, and real investment opportunities, while the expected return/beta formulation is not useful or very cumbersome in the latter applications. Thus, it has seemed that there are several different asset pricing theories: expected return/beta for stocks, yield-curve models for bonds, arbitrage models for options. In fact all three are just cases of p = E(mx). As a particular example, arbitrage, in the precise sense of positive payoffs with negative prices, has not entered the equivalence discussion at all. I don’t know of any way to cleanly graft absence of arbitrage on to expected return/beta models. You have to tack it on after the fact – “by the way, make sure that every portfolio with positive payoffs has a positive price.” It is trivially easy to graft it on to a discount factor model: just add m > 0. The discount factor and state space language also makes it easier to think about different horizonsP and the present value statement of models. p = E(mx) generalizes quickly to pt = Et j mt,t+j xt+j , while returns have to be chained together to think about multiperiod models. Papers are still written arguing about geometric vs. arithmetic average returns for multiperiod discounting. 126

The choice of language is not about normality or return distributions. There is a lot of confusion about where return distribution assumptions show up in finance. I have made no distributional assumptions in any of the discussion so far. Second moments as in betas and the variance of the mean-variance frontier show up because p = E(mx) involves a second moment. One does not need to assume normality to talk about the mean-variance frontier. Returns on the mean-variance frontier price other assets even when returns are not normally distributed.

127

Chapter 8.

Conditioning information

The asset pricing theory I have sketched so far really describes prices at time t in terms of conditional moments. The investor’s first order conditions are pt u0 (ct ) = βEt [u0 (ct+1 )xt+1 ] where Et means expectation conditional on the investor’s time t information. Sensibly, the price at time t should be higher if there is information at time t that the discounted payoff is likely to be higher than usual at time t + 1. The basic asset pricing equation should be pt = Et (mt+1 xt+1 ). (Conditional expectation can also be written pt = E [mt+1 xt+1 |It ] when it is important to specify the information set It .). If payoffs and discount factors were independent and identically distributed (i.i.d.) over time, then conditional expectations would be the same as unconditional expectations and we would not have to worry about the distinction between the two concepts. But stock price/dividend ratios, bond and option prices all change over time, which must reflect changing conditional moments of something on the right hand side. One approach is to specify and estimate explicit statistical models of conditional distributions of asset payoffs and discount factor variables (e.g. consumption growth). This approach is sometimes used, and is useful in some applications, but it is usually cumbersome. As we make the conditional mean, variance, covariance, and other parameters of the distribution of (say) N returns depend flexibly on M information variables, the number of required parameters can quickly exceed the number of observations. More importantly, this explicit approach typically requires us to assume that investors use the same model of conditioning information that we do. We obviously don’t even observe all the conditioning information used by economic agents, and we can’t include even a fraction of observed conditioning information in our models. The basic feature and beauty of asset prices (like all prices) is that they summarize an enormous amount of information that only individuals see. The events that make the price of IBM stock change by a dollar, like the events that make the price of tomatoes change by 10 cents, are inherently unobservable to economists or would-be social planners (Hayek 1945). Whenever possible, our treatment of conditioning information should allow agents to see more than we do. If we don’t want to model conditional distributions explicitly, and if we want to avoid assuming that investors only see the variables that we include in an empirical investigation, we eventually have to think about unconditional moments, or at least moments conditioned on less information than agents see. Unconditional implications are also interesting in and of themselves. For example, we may be interested in finding out why the unconditional mean 128

S ECTION 8.1

S CALED PAYOFFS

returns on some stock portfolios are higher than others, even if every agent fundamentally seeks high conditional mean returns. Most statistical estimation essentially amounts to characterizing unconditional means, as we will see in the chapter on GMM. Thus, rather than model conditional distributions, this chapter focuses on what implications for unconditional moments we can derive from the conditional theory.

8.1

Scaled payoffs

pt = Et (mt+1 xt+1 ) ⇒ E(pt zt ) = E(mt+1 xt+1 zt ) One can incorporate conditioning information by adding scaled payoffs and doing everything unconditionally. I interpret scaled returns as payoffs to managed portfolios.

8.1.1

Conditioning down

The unconditional implications of any pricing model are pretty easy to state. From pt = Et (mt+1 xt+1 ) we can take unconditional expectations to obtain6 E(pt ) = E(mt+1 xt+1 ).

(108)

Thus, if we just interpret p to stand for E(pt ), everything we have done above applies to unconditional moments. In the same way, we can also condition down from agents’ fine information sets to coarser sets that we observe, pt

= E(mt+1 xt+1 | Ω) ⇒ E(pt |I ⊂ Ω) = E(mt+1 xt+1 | I ⊂ Ω) ⇒ pt = E(mt+1 xt+1 | It ⊂ Ωt ) if pt ∈ It .

In making the above statements I used the law of iterated expectations, which is important enough to highlight it. This law states that if you take an expected value using less information of an expected value that is formed on more information, you get back the expected value using less information. Your best forecast today of your best forecast tomorrow is the same We need a small technical assumption that the unconditional moment³or moment ´ conditioned ³ on ´ a coarser information set exists. For example, if X and Y are normal (0, 1), then E X |Y = 0 but E X is infinite. Y Y 6

129

C HAPTER 8

C ONDITIONING INFORMATION

as your best forecast today. In various useful guises, E(Et (x)) = E(x), Et−1 (Et (xt+1 )) = Et−1 (xt+1 ) E [E(x|Ω) | I ⊂ Ω] = E [x|I] 8.1.2

Instruments and managed portfolios

We can do more than just condition down. Suppose we multiply the payoff and price by an instrument zt observed at time t. Then, zt pt = Et (mt+1 xt+1 zt ) and, taking unconditional expectations, E(pt zt ) = E(mt+1 xt+1 zt ).

(109)

This is an additional implication of the conditional model, not captured by just conditioning down as in (8.108). This trick originates from the GMM method of estimating asset pricing models, discussed below. The word instruments for the z variables comes from the instrumental variables estimation heritage of GMM. To think about equation (8.109), group (xt+1 zt ). Call this product a payoff x = xt+1 zt , with price p = E(pt zt ). Then 8.109 reads p = E(mx) once again. Rather than thinking about (8.109) as a instrumental variables estimate of a conditional model, we can think of it as a price and a payoff, and apply all the asset pricing theory directly. This interpretation is not as artificial as it sounds. zt xt+1 are the payoffs to managed portfolios. An investor who observes zt can, rather than “buy and hold,” invest in an asset according to the value of zt . For example, if a high value of zt forecasts that asset returns are likely to be high the next period, the investor might buy more of the asset when zt is high and vice-versa. If the investor follows a linear rule, he puts zt pt dollars into the asset each period and receives zt xt+1 dollars the next period. This all sounds new and different, but practically every test uses managed portfolios. For example, the size, beta, industry, book/market and so forth portfolios of stocks are all managed portfolios, since their composition changes every year in response to conditioning information – the size, beta, etc. of the individual stocks. This idea is also closely related to the deep idea of dynamic spanning. Markets that are apparently very incomplete can in 130

S ECTION 8.2

S UFFICIENCY OF ADDING SCALED RETURNS

reality provide many more state-contingencies through dynamic (conditioned on information) trading strategies. Equation (8.109) offers a very simple view of how to incorporate the extra information in conditioning information: Add managed portfolio payoffs, and proceed with unconditional moments as if conditioning information didn’t exist! Linearity is not important. If the investor wanted to place, say, 2 + 3z 2 dollars in the asset, we could capture this desire with an instrument z2 = 2 + 3z 2 . Nonlinear (measurable) transformations of time−t random variables are again random variables. We can thus incorporate conditioning information while still looking at unconditional moments instead of conditional moments, without any of the statistical machinery of explicit models with time-varying moments. The only subtleties are 1) The set of asset payoffs expands dramatically, since we can consider all managed portfolios as well as basic assets, potentially multiplying every asset return by every information variable. 2) Expected prices of managed portfolios show up for p instead of just p = 0 and p = 1 if we started with basic asset returns and excess returns.

8.2

Sufficiency of adding scaled returns

Checking the expected price of all managed portfolios is, in principle, sufficient to check all the implications of conditioning information. E(zt ) = E(mt+1 Rt+1 zt ) ∀zt ∈ It ⇒ 1 = E(mt+1 Rt+1 |It ) E(pt ) = E(mt+1 xt+1 ) ∀ xt+1 ∈ X t+1 ⇒ pt = E(mt+1 xt+1 |It ) We have shown that we can derive some extra implications from the presence of conditioning information by adding scaled returns. But does this exhaust the implications of conditioning information? Are we missing something important by relying on this trick? The answer is, in principle no. I rely on the following mathematical fact: The conditional expectation of a variable yt+1 given an information set It , E(yt+1 | It ) is equal to a regression forecast of yt+1 using every variable zt ∈ It . Now, “every random variable” means every variable and every nonlinear (measurable) transformation of every variable, so there are a lot of variables in this regression! (The word projection and proj(yt+1 |zt ) is used to distinguish the best forecast of yt+1 using only linear combinations of zt from the conditional expectation.) Applying this fact to our case, let yt+1 = mt+1 Rt+1 − 1. Then E [(mt+1 Rt+1 − 1) zt ] = 0 for every zt ∈ It implies 1 = E(mt+1 Rt+1 | It ). Thus, no implications are lost in principle by looking at scaled returns. 131

C HAPTER 8

C ONDITIONING INFORMATION

Another way of looking at the same idea is that Rt+1 zt+1 is the return on a payoff available at time t + 1. Thus, the space of all payoffs X t+1 should be understood to include the time-t + 1 payoff you can generate with a basis set of assets Rt+1 and all dynamic strategies that use information in the set It . With that definition of the space X t+1 we can write the sufficiency of scaled returns with the more general second equality above. “All linear and nonlinear transformations of all variables observed at time t” sounds like a lot of instruments, and it is. But there is a practical limit to the number of instruments zt one needs to scale by, since only variables that forecast returns or m (or their higher moments and co-moments) add any information. Since adding instruments is the same thing as including potential managed portfolios, thoughtfully choosing a few instruments is the same thing as the thoughtful choice of a few assets or portfolios that one makes in any test of an asset pricing model. Even when evaluating completely unconditional asset pricing models, one always forms portfolios and omits many possible assets from analysis. Few studies, in fact, go beyond checking whether a model correctly prices 10-25 stock portfolios and a few bond portfolios. Implicitly, one feels that the chosen payoffs do a pretty good job of spanning the set of available risk-loadings (mean returns) and hence that adding additional assets will not affect the results. Nonetheless, since data are easily available on all 2000 or so NYSE stocks, plus AMEX and NASDAQ stocks, to say nothing of government and corporate bonds, returns of mutual funds, foreign exchange, foreign equities, real investment opportunities, etc., the use of a few portfolios means that a tremendous number of potential asset payoffs are left out in an ad-hoc manner. In a similar manner, if one had a small set of instruments that capture all the predictability of discounted returns mt+1 Rt+1 , then there would be no need to add more instruments. Thus, we carefully but arbitrarily select a few instruments that we think do a good job of characterizing the conditional distribution of returns. Exclusion of potential instruments is exactly the same thing as exclusion of assets. It is no better founded, but the fact that it is a common sin may lead one to worry less about it. There is nothing special about unscaled returns, and no economic reason to place them above scaled returns. A mutual fund might come into being that follows the managed portfolio strategy and then its unscaled returns would be the same as an original scaled return. Models that cannot price scaled returns are no more interesting than models that can only price (say) stocks with first letter A through L. (There may be econometric reasons to trust results for nonscaled returns a bit more, but we haven’t gotten to statistical issues yet.) Of course, the other way to incorporate conditioning information is by constructing explicit parametric models of conditional distributions. With this procedure one can in fact check all of a model’s implications about conditional moments. However, the parametric model may be incorrect, or may not reflect some variable used by investors. Including instruments may not be as efficient, but it is still consistent if the parametric model is incorrect. The wrong parametric model of conditional distributions may lead to inconsistent estimates. In addition, one avoids estimating nuisance parameters of the parametric distribution model.

132

S ECTION 8.3

8.3

C ONDITIONAL AND UNCONDITIONAL MODELS

Conditional and unconditional models

A conditional factor model does not imply a fixed-weight or unconditional factor model: mt+1 = b0t ft+1 , pt = Et (mt+1 xt+1 ) does not imply that ∃b s.t. mt+1 = b0 ft+1 , E(pt ) = E(mt+1 xt+1 ). Et (Rt+1 ) = β 0t λt does not imply E(Rt+1 ) = β 0 λ. Conditional mean-variance efficiency does not imply unconditional mean-variance efficiency. The converse statements are true, if managed portfolios are included. For explicit discount factor models—models whose parameters are constant over time— the fact that one looks at a conditional vs. unconditional implications makes no difference to the statement of the model. pt = Et (mt+1 xt+1 ) ⇒ E(pt ) = E(mt+1 xt+1 ) and that’s it. Examples include the consumption-based model with power utility, mt+1 = W . β(ct+1 /ct )−γ , and the log utility CAPM, mt+1 = 1/Rt+1 However, linear factor models include parameters that may vary over time and as functions of conditioning information. In these cases the transition from conditional to unconditional moments is much more subtle. We cannot easily condition down the model at the same time as the prices and payoffs. 8.3.1

Conditional vs. unconditional factor models in discount factor language

As an example, consider the CAPM m = a − bRW where RW is the return on the market or wealth portfolio. We can find a and b from the condition that this model correctly price any two returns, for example RW itself and a riskfree rate: W ½ ) a = R1f + bEt (Rt+1 ) 1 = Et (mt+1 RW t t+1 f ⇒ . (110) Et (RW )−R f t 1 = Et (mt+1 )Rt b = f t+1 R σ2 (RW ) t

t

t+1

As you can see, b > 0 and a > 0: to make a payoff proportional to the minimum secondmoment return (on the inefficient part of the mean-variance frontier) we need a portfolio long the risk free rate and short the market RW . 133

C HAPTER 8

C ONDITIONING INFORMATION

2 W More importantly for our current purposes, a and b vary over time, as Et (RW t+1 ), σ t (Rt+1 ), f and Rt vary over time. If it is to price assets conditionally, the CAPM must be a linear factor model with time-varying weights, of the form

mt+1 = at − bt RW t+1 . This fact means that we can no longer transparently condition down. The statement that £ ¤ 1 = Et (at + bt RW t+1 )Rt+1

does not imply that we can find constants a and b so that ¤ £ 1 = E (a + bRW t+1 )Rt+1 .

Just try it. Taking unconditional expectations, ¤ £ ¤ £ W W )Rt+1 = E at Rt+1 + bt Rt+1 Rt+1 1 = E (at + bt Rt+1 W = E(at )E(Rt+1 ) + E(bt )E(RW t+1 Rt+1 ) + cov(at , Rt+1 ) + cov(bt , Rt+1 Rt+1 )

Thus, the unconditional model 1=E

£¡

¢ ¤ W Rt+1 E(at ) + E(bt )Rt+1

only holds if the covariance terms above happen to be zero. Since at and bt are formed from conditional moments of returns, the covariances will not, in general be zero. On the other hand, suppose it is true that at and bt are constant over time. Then £ ¤ 1 = Et (a + bRW t+1 )Rt+1

does imply

¤ £ W )Rt+1 , 1 = E (a + bRt+1

just like any other constant-parameter factor pricing model. Furthermore, the latter unconditional model implies the former conditional model, if the latter holds for all managed portfolios. 8.3.2

Conditional vs. unconditional in an expected return / beta model

To put the same observation in beta-pricing language, Et (Ri ) = Rtf + β t λt 134

(111)

S ECTION 8.3

C ONDITIONAL AND UNCONDITIONAL MODELS

does not imply that E(Ri ) = α + βλ

(112)

The reason is that β t and β represent conditional and unconditional regression coefficients respectively. Again, if returns and factors are i.i.d., the unconditional model can go through. In that case, cov(·) = covt (·), var(·) = vart (·), so the unconditional regression beta is the same as the conditional regression beta, β = β t . Then, we can take expectations of (8.111) to get (8.112), with λ = E(λt ). But to condition down in this way, the covariance and variance must each be constant over time. It is not enough that their ratio, or conditional betas are constant. If covt and vart change over time, then the unconditional regression beta, β = cov/var is not equal to the average conditional regression beta, E(β t ) or E(covt /vart ). Some models specify that covt and vart vary over time, but covt /vart is a constant. This specification still does not imply that the unconditional regression beta β ≡ cov/var is equal to the constant covt /vart . Similarly, it is not enough that λ be constant, since E(β t ) 6= β. The betas must be regression coefficients, not just numbers. If the betas do not vary over time, the λt may still vary and λ = E(λt ). 8.3.3

A precise statement

Let’s formalize these observations somewhat. Let X denote the space of all portfolios of the primitive assets, including managed portfolios in which the weights may depend on conditioning information, i.e. scaled returns. A conditional factor pricing model is a model mt+1 = at + b0t ft+1 that satisfies pt = Et+1 (mt+1 xt+1 ) for all xt+1 ∈ X. An unconditional factor pricing model is model mt+1 = a + b0 ft+1 satisfies E(pt ) = E(mt+1 xt+1 ) for all xt+1 ∈ X. It might be more appropriately called a fixed-weight factor pricing model. Given these definitions it’s almost trivial that the unconditional model is just a special case of the conditional model, one that happens to have fixed weights. Thus, a conditional factor model does not imply an unconditional factor model (because the weights may vary) but an unconditional factor model does imply a conditional factor model. There is one important subtlety. The payoff space X is common, and contains all managed portfolios in both cases. The payoff space for the unconditional factor pricing model is not just fixed combinations of a set of basis assets. For example, we might simply check that the static (constant a, b) CAPM captures the unconditional mean returns of a set of assets. If this model does not also price those assets scaled by instruments, then it is not a conditional model, or, as I argued above, really a valid factor pricing model at all. Of course, everything applies for the relation between a conditional factor pricing model 135

C HAPTER 8

C ONDITIONING INFORMATION

using a fine information set (like investors’ information sets) and conditional factor pricing models using coarser information sets (like ours). If you think a set of factors prices assets with respect to investors’ information, that does not mean the same set of factors prices assets with respect to our, coarser, information sets. 8.3.4

Mean-variance frontiers

Define the conditional mean-variance frontier as the set of returns that minimize vart (Rt+1 ) given Et (Rt+1 ). (This definition includes the lower segment as usual.) Define the unconditional mean-variance frontier as the set of returns including managed portfolio returns that minimize var(Rt+1 ) given E(Rt+1 ). These two frontiers are related by: If a return is on the unconditional mean-variance frontier, it is on the conditional mean-variance frontier. However, If a return is on the conditional mean-variance frontier, it need not be on the unconditional mean-variance frontier. These statements are exactly the opposite of what you first expect from the language. The law of iterated expectations E(Et (x)) = E(x) leads you to expect that “conditional” should imply “unconditional.” But we are studying the conditional vs. unconditional mean-variance frontier, not raw conditional and unconditional expectations, and it turns out that exactly the opposite words apply. Of course “unconditional” can also mean “conditional on a coarser information set.” Again, keep in mind that the unconditional mean variance frontier includes returns on managed portfolios. This definition is eminently reasonable. If you’re trying to minimize variance for given mean, why tie your hands to fixed weight portfolios? Equivalently, why not allow yourself to include in your portfolio the returns of mutual funds whose advisers promise the ability to adjust portfolios based on conditioning information? You could form a mean-variance frontier of fixed-weight portfolios of a basis set of assets, and this is what many people often mean by “unconditional mean-variance frontier.” The return on the true unconditional mean-variance frontier will, in general, include some managed portfolio returns, and so will lie outside this mean-variance frontier of fixed-weight portfolios. Conversely, a return on the fixed-weight portfolio MVF is, in general, not on the unconditional or conditional mean-variance frontier. All we know is that the fixed-weight frontier lies inside the other two. It may touch, but it need not. This is not to say the fixed-weight unconditional frontier is uninteresting. For example, returns on this frontier will price fixed-weight portfolios of the basis assets. The point is that this frontier has no connection to the other two frontiers. In particular, a conditionally mean-variance efficient return (conditional CAPM) need not unconditionally price the fixed weight portfolios. 136

S ECTION 8.3

C ONDITIONAL AND UNCONDITIONAL MODELS

I offer several ways to see this important statement. Using the connection to factor models We have seen that the conditional CAPM mt+1 = at − bt RW t+1 does not imply an unconW . We have seen that the existence of such a conditional ditional CAPM mt+1 = a − bRt+1 W lies on the conditional meanfactor model is equivalent to the statement that the return Rt+1 variance frontier, and the existence of an unconditional factor model mt+1 = a − bRW t+1 is equivalent to the statement that RW is on the unconditional mean-variance frontier. Then, from the “trivial” fact that an unconditional factor model is a special case of a conditional one, we know that RW on the unconditional frontier implies RW on the conditional frontier but not vice-versa. Using the orthogonal decomposition We can see the relation between conditional and unconditional mean-variance frontiers using the orthogonal decomposition characterization of mean-variance efficiency given above. This beautiful proof is the main point of Hansen and Richard (1987). By the law of iterated expectations, x∗ and R∗ generate expected prices and Re∗ generates unconditional means as well as conditional means: E [p = Et (x∗ x)] ⇒ E(p) = E(x∗ x) ¤ £ E Et (R∗2 ) = Et (R∗ R) ⇒ E(R∗2 ) = E(R∗ R)

E [Et (Re∗ Re ) = Et (Re )] ⇒ E(Re∗ Re ) = E(Re ) This fact is subtle and important. For example, starting with x∗ = p0t Et (xt+1 x0t+1 )−1 xt+1 , you might think we need a different x∗ , R∗ , Re∗ to represent expected prices and unconditional means, using unconditional probabilities to define inner products. The three lines above show that this is not the case. The same old x∗ , R∗ , Re∗ represent conditional as well as unconditional prices and means. Recall that a return is mean-variance efficient if and only if it is of the form Rmv = R∗ + wRe∗ . Thus, Rmv is conditionally mean-variance efficient if w is any number in the time t information set. mv ∗ = Rt+1 + wt Re∗ conditional frontier: Rt+1 t+1 ,

and Rmv is unconditionally mean-variance efficient if w is any constant. mv ∗ unconditional frontier: Rt+1 = Rt+1 + wRe∗ t+1 .

137

C HAPTER 8

C ONDITIONING INFORMATION

Constants are in the t information set; time t random variables are not necessarily constant. Thus unconditional efficiency (including managed portfolios) implies conditional efficiency but not vice versa. As with the factor models, once you see the decomposition, it is a trivial argument about whether a weight is constant or time-varying. Brute force and examples. If you’re still puzzled, an additional argument by brute force may be helpful. If a return is on the unconditional MVF it must be on the conditional MVF at each date. If not, you could improve the unconditional mean-variance trade-off by moving to the conditional MVF at each date. Minimizing unconditional variance given mean is the same as minimizing unconditional second moment given mean, min E(R2 ) s.t. E(R) = µ Writing the unconditional moment in terms of conditional moments, the problem is £ ¤ min E Et (R2 ) s.t. E [Et (R)] = µ

Now, suppose you could lower Et (R2 ) at one date t without affecting Et (R) at that date. This change would lower the objective, without changing the constraint. Thus, you should have done it: you should have picked returns on the conditional mean variance frontiers. It almost seems that reversing the argument we can show that conditional efficiency implies unconditional efficiency, but it doesn’t. Just because you have minimized Et (R2 ) for given value of Et (R) at each date t does not imply that you have minimized E(R2 ) for a given value of E(R). In showing that unconditional efficiency implies conditional efficiency we held fixed Et (R) at each date at µ, and showed it is a good idea to minimize σ t (R). In trying to go backwards, the problem is that a given value of E(R) does not specify what Et (R) should be at each date. We can increase Et (R) in one conditioning information set and decrease it in another, leaving the return on the conditional MVF. Figure 22 presents an example. Return B is conditionally mean-variance efficient. It also has zero unconditional variance, so it is the unconditionally mean-variance efficient return at the expected return shown. Return A is on the conditional mean-variance frontiers, and has the same unconditional expected return as B. But return A has some unconditional variance, and so is inside the unconditional mean-variance frontier. As a second example,the riskfree rate is only on the unconditional mean-variance frontier if it is a constant. Remember the expression (6.95) for the risk free rate, Rf = R∗ + Rf Re∗ . The unconditional mean-variance frontier is R∗ + wRe∗ with w a constant. Thus, the riskfree rate is only unconditionally mean-variance efficient if it is a constant. 138

S ECTION 8.4

S CALED FACTORS : A PARTIAL SOLUTION

Et(R)

Info. set 1

A Info. set 2 B A

σt(R) Figure 22. Return A is on the conditional mean-variance frontiers but not on the unconditional mean variance frontier. 8.3.5

Implications: Hansen-Richard Critique.

Many models, such as the CAPM, imply a conditional linear factor model mt+1 = at + b0t ft+1 . These theorems show that such a model does not imply an unconditional model. Equivalently, if the model predicts that the market portfolio is conditionally mean-variance efficient, this does not imply that the market is unconditionally mean-variance efficient. We often test the CAPM by seeing if it explains the average returns of some portfolios or (equivalently) if the market is on the unconditional mean-variance frontier. The CAPM may quite well be true (conditionally) and fail these tests; many assets may do better in terms of unconditional mean vs. unconditional variance. The situation is even worse than these comments seem, and is not repaired by simple inclusion of some conditioning information. Models such as the CAPM imply a conditional linear factor model with respect to investors’ information sets. However, the best we can hope to do is to test implications conditioned down on variables that we can observe and include in a test. Thus, a conditional linear factor model is not testable! I like to call this observation the “Hansen-Richard critique” by analogy to the “Roll Critique.” Roll pointed out, among other things, that the wealth portfolio might not be observable, making tests of the CAPM impossible. Hansen and Richard point out that the conditioning information of agents might not be observable, and that one cannot omit it in testing a conditional model. Thus, even if the wealth portfolio were observable, the fact that we cannot observe agents’ information sets dooms tests of the CAPM.

139

C HAPTER 8

8.4

C ONDITIONING INFORMATION

Scaled factors: a partial solution

You can expand the set of factors to test conditional factor pricing models factors = ft+1 ⊗ zt The problem is that the parameters of the factor pricing model mt+1 = at + bt ft+1 may vary over time. A partial solution is to model the dependence of parameters at and bt on variables in the time−t information set; let at = a(zt ), bt = b(zt ) where zt is a vector of variables observed at time t (including a constant). In particular, why not try linear models at = a0 z t , bt = b0 z t Linearity is not restrictive: zt2 is just another instrument. The only criticism one can make is that some instrument zjt is important for capturing the variation in at and bt , and was omitted. For instruments on which we have data, we can meet this objection by trying zjt and seeing whether it does, in fact, enter significantly. However, for instruments zt that are observed by agents but not by us, this criticism remains valid. Linear discount factor models lead to a nice interpretation as scaled factors, in the same way that linearly managed portfolios are scaled returns. With a single factor and instrument, write mt = a(zt ) + b(zt )ft+1

(113)

= a0 + a1 zt + (b0 + b1 zt )ft+1 = a0 + a1 zt + b0 ft+1 + b1 (zt ft+1 ) .

(114)

Thus, in place of the one-factor model with time-varying coefficients (8.113), we have a three-factor model (zt , ft+1 , zt ft+1 ) with fixed coefficients, (8.114). Since the coefficients are now fixed, we can use the scaled-factor model with unconditional moments. pt = Et [(a0 + a1 zt + b0 ft+1 + b1 (zt ft+1 )) xt+1 ] ⇒ E(pt ) = E [(a0 + a1 zt + b0 ft+1 + b1 (zt ft+1 )) xt+1 ] For example, in standard derivations of CAPM, the market (wealth portfolio) return is conditionally mean-variance efficient; investors want to hold portfolios on the conditional 140

S ECTION 8.5

S UMMARY

mean-variance frontier; conditionally expected returns follow a conditional single-beta representation, or the discount factor m follows a conditional linear factor model W mt+1 = at − bt Rt+1

as we saw above. But none of these statements mean that we can use the CAPM unconditionally. Rather than throw up our hands, we can add some scaled factors. Thus, if, say, the dividend/price ratio and term premium do a pretty good job of summarizing variation in conditional moments, the conditional CAPM implies an unconditional, five-factor (plus constant) model. The factors are a constant, the market return, the dividend/price ratio, the term premium, and the market return times the dividend-price ratio and the term premium. The unconditional pricing implications of such a five-factor model could, of course, be summarized by a single−β representation. (See the caustic comments in the section on implications and equivalence.) The reference portfolio would not be the market portfolio, of course, but a mimicking portfolio of the five factors. However, the single mimicking portfolio would not be easily interpretable in terms of a single factor conditional model and two instruments. In this case, it might be more interesting to look at a multiple −β or multiplefactor representation. If we have many factors f and many instruments z, we should in principle multiply every factor by every instrument, m = b1 f1 + b2 f1 z1 + b3 f1 z2 + ... + bN+1 f2 + bN+2 f2 z1 + bN+3 f2 z2 + ... This operation can be compactly summarized with the Kronecker product notation, a ⊗ b, which means “multiply every element in vector a by every element in vector b, or mt+1 = b0 (ft+1 ⊗ zt ).

8.5

Summary

When you first think about it, conditioning information sounds scary – how do we account for time-varying expected returns, betas, factor risk premia, variances, covariances, etc. However, the methods outlined in this chapter allow a very simple and beautiful solution to the problems raised by conditioning information. To express the conditional implications of a given model, all you have to do is include some scaled or managed portfolio returns, and then pretend you never heard about conditioning information. Some factor models are conditional models, and have coefficients that are functions of investors’ information sets. In general, there is no way to test such models, but if you are willing to assume that the relevant conditioning information is well summarized by a few variables, then you can just add new factors, equal to the old factors scaled by the conditioning 141

C HAPTER 8

C ONDITIONING INFORMATION

variables, and again forget that you ever heard about conditioning information. You may want to remember conditioning information as a diagnostic and in economic interpretation of the results. It may be interesting to take estimates of a many factor model, mt = a0 + a1 zt + b0 ft+1 + b1 zt ft+1 , and see what they say about the implied conditional model, mt = (a0 + a1 zt ) + (b0 + b1 zt )ft+1 . You may want to make plots of conditional bs, betas, factor risk premia, expected returns,etc. But you don’t have to worry about it in estimation and testing.

8.6 1. 2.

3.

Problems

If there is a risk free asset, is it on the a) conditional b) unconditional c) both mean-variance frontier? If there is a conditionally riskfree asset – a claim to 1 is traded at each date, does this mean that there is an unconditionally risk free asset? (Define the latter first!) How about vice versa? Suppose you took the unconditional population moments E(R), E(RR0 ) of assets returns and constructed the mean-variance frontier. Does this frontier correspond to the conditional or the unconditional MV frontier, or neither? What is the key assumption underlying your answer?

142

Chapter 9.

Factor pricing models

In Chapter 2, I noted that the consumption-based model, while a complete answer to most asset pricing questions in principle, does not (yet) work well in practice. This observation motivates efforts to tie the discount factor m to other data. Linear factor pricing models are the most popular models of this sort in finance. They dominate discrete time empirical work. Factor pricing models replace the consumption-based expression for marginal utility growth with a linear model of the form mt+1 = a + b0 f t+1 a and b are free parameters. This specification is equivalent to a multiple-beta model E(Rt+1 ) = α + β 0 λ where β are multiple regression coefficients of returns R on the factors f . Here, α and λ are the free parameters. The big question is, what should one use for factors ft+1 ? Factor pricing models look for variables that are good proxies for aggregate marginal utility growth, i.e., variables for which β

u0 (ct+1 ) ≈ a + b0 f t+1 u0 (ct )

(115)

is a sensible and economically interpretable approximation. More directly and interpretably, the essence of asset pricing is that there are special states of the world in which investors are especially concerned that their portfolios not do badly. They are willing to trade off some overall performance – average return – to make sure that portfolios do not do badly in these particular states of nature. The factors are variables that indicate that these “bad states” have occurred. The factors that result from this search are and should be intuitively sensible. In any sensible economic model, as well as in the data, consumption is related to returns on broadbased portfolios, to interest rates, to growth in GNP, investment, or other macroeconomic variables, and to returns on production processes. All of these variables measure “wealth” or the state of the economy. Consumption is and should be high in “good times” and low in “bad times.” Furthermore, consumption and marginal utility respond to news: if a change in some variable today signals high income in the future, then consumption rises now, by permanent income logic. This fact opens the door to forecasting variables: any variable that forecasts asset returns (“changes in the investment opportunity set”) or macroeconomic variables is a candidate factor. Variables such as the term premium, dividend/price ratio, stock returns, etc. can be defended as pricing factors on this logic. Though they themselves are not measures of aggregate good or bad times, they forecast such times. 143

C HAPTER 9

FACTOR PRICING MODELS

Should factors be independent over time? The answer is, sort of. If there is a constant real interest rate, then marginal utility growth should be unpredictable. (“Consumption is a random walk” in the quadratic utility permanent income model.) To see this, just look at the first order condition with a constant interest rate, u0 (ct ) = βRf Et [u0 (ct+1 )] or in a more time-series notation, 1 u0 (ct+1 ) = + εt+1 ; Et (εt+1 ) = 0. 0 u (ct ) βRf The real risk free rate is not constant, but it does not vary a lot, especially compared to asset returns. Measured consumption growth is not exactly unpredictable but it is the least predictable macroeconomic time series, especially if one accounts properly for temporal aggregation (consumption data are quarterly averages). Thus, factors that proxy for marginal utility growth, though they don’t have to be totally unpredictable, should not be highly predictable. If one chooses highly predictable factors, the model will counterfactually predict large interest rate variation. In practice, this consideration means that one should choose the right units: Use GNP growth rather than level, portfolio returns rather than prices or price/dividend ratios, etc. However, unless one wants to impose an exactly constant risk free rate, one does not have to filter or prewhiten factors to make them exactly unpredictable. This view of factors as intuitively motivated proxies for marginal utility growth is sufficient to carry the reader through current empirical tests of factor models. The extra constraints of a formal exposition of theory in this part have not yet constrained the factor-fishing expedition. The precise derivations all proceed in the way I have motivated factor models: One writes down a general equilibrium model, in particular a specification of the production technology by which real investment today results in real output tomorrow. This general equilibrium produces relations that express the determinants of consumption from exogenous variables, and relations linking consumption and other endogenous variables; equations of the form ct = g(ft ). One then uses this kind of equation to substitute out for consumption in the basic first order conditions. The formal derivations accomplish two things: they determine one particular list of factors that can proxy for marginal utility growth, and they prove that the relation should be linear. Some assumptions can often be substituted for others in the quest for these two features of a factor pricing model. This is a point worth remembering: all factor models are derived as specializations of the consumption-based model. Many authors of factor model papers disparage the consumptionbased model, forgetting that their factor model is the consumption-based model plus extra assumptions that allow one to proxy for marginal utility growth from some other variables. 144

S ECTION 9.1

C APITAL A SSET P RICING MODEL (CAPM)

My presentation follows Constantinides’ (1989) derivation of traditional models as instances of the consumption-based model in this regard. Above, I argued that clear economic foundation was important for factor models, since it is the only guard against fishing. Alas, we discover here that the current state of factor pricing models is not a particularly good guard against fishing. One can call for better theories or derivations, more carefully aimed at limiting the list of potential factors and describing the fundamental macroeconomic sources of risk, and thus providing more discipline for empirical work. The best minds in finance have been working on this problem for 40 years though, so a ready solution is not immediately in sight. On the other hand, we will see that even current theory can provide much more discipline than is commonly imposed in empirical work. For example, the derivations of the CAPM and ICAPM do leave predictions for the risk free rate and for factor risk premia that are often ignored. The ICAPM gives tighter restrictions on state variables than are commonly checked: “State variables” do have to forecast something! We also see how special and unrealistic are the general equilibrium setups necessary to derive popular specifications such as CAPM and ICAPM. This observation motivates a more serious look at real general equilibrium models below.

9.1

Capital Asset Pricing Model (CAPM)

The CAPM is the model m = a + bRw ; Rw = wealth portfolio return. I derive it from the consumption based model by 1) Two period quadratic utility; 2) Two periods, exponential utility and normal returns; 3) Infinite horizon, quadratic utility and i.i.d. returns; 4) Log utility and normally distributed returns. The CAPM is the first, most famous and (so far) most widely used model in asset pricing. It ties the discount factor m to the return on the “wealth portfolio.” The function is linear, W . mt+1 = a + bRt+1

a and b are free parameters. One can find theoretical values for the parameters a and b by requiring the discount factor m to price any two assets, such as the wealth portfolio return and risk-free rate, 1 = E(mRW ) and 1 = E(m)Rf . (As an example, we did this in equation (8.110) above.) In empirical applications, we can also pick a and b to “best” price larger cross-sections of assets. We do not have good data on, or even a good empirical definition for, the return on total wealth. It is conventional to proxy RW by the return on a broad-based stock portfolio such as the value- or equally-weighted NYSE, S&P500, etc. The CAPM is of course most frequently stated in equivalent expected return / beta language, E(Ri ) = α + β i,RW [E(Rw ) − α] . 145

C HAPTER 9

FACTOR PRICING MODELS

This section briefly describes some classic derivations of the CAPM. Again, we need to find assumptions that defend which factors proxy for marginal utility (RW here), and assumptions to defend the linearity between m and the factor. I present several derivations of the same model. Many of these derivations use classic modeling assumptions which are important in their own sake. This is also an interesting place in which to see that various sets of assumptions can often be used to get to the same place. The CAPM is often criticized for one or another assumption. By seeing several derivations, we can see how one assumption can be traded for another. For example, the CAPM does not in fact require normal distributions, if one is willing to swallow quadratic utility instead. 9.1.1

Two-period quadratic utility

Two period investors with no labor income and quadratic utility imply the CAPM. Investors have quadratic preferences and only live two periods, 1 1 U(ct , ct+1 ) = − (ct − c∗ )2 − βE[(ct+1 − c∗ )2 ]. 2 2

(116)

Their marginal rate of substitution is thus mt+1 = β

(ct+1 − c∗ ) u0 (ct+1 ) =β . 0 u (ct ) (ct − c∗ )

The quadratic utility assumption means marginal utility is linear in consumption. Thus, the first target of the derivation, linearity. Investors are born with wealth Wt in the first period and earn no labor income. They can invest in lots of assets with prices pit and payoffs xit+1 , or, to keep the notation simple, i . They choose how much to consume at the two dates, ct and ct+1 , and the returns Rt+1 portfolio weights αi for their investment portfolio. Thus, the budget constraint is (117)

ct+1 = Wt+1 Wt+1 = RW t+1 (Wt − ct ) RW =

N X

αi Ri ;

i=1

N X i=1

RW is the rate of return on total wealth. 146

αi = 1.

S ECTION 9.1

C APITAL A SSET P RICING MODEL (CAPM)

The two-period assumption means that investors consume everything in the second period, by constraint (9.117). This fact allows us to substitute wealth and the return on wealth for consumption, achieving the second goal of the derivation, naming the factor that proxies for consumption or marginal utility: mt+1 = β i.e.

W Rt+1 (Wt − ct ) − c∗ −βc∗ β(Wt − ct ) W = + Rt+1 ct − c∗ ct − c∗ ct − c∗

mt+1 = at + bt RW t+1 . 9.1.2

Exponential utility, normal distributions

u(c) = −e−αc and a normally distributed set of returns also produces the CAPM. The combination of exponential utility and normal distributions is another set of assumptions that deliver the CAPM in a one or two period model. This structure has a particularly convenient analytical form. Since it gives rise to linear demand curves, it is very widely used in models that complicate the trading structure, by introducing incomplete markets or asymmetric information. I present a model with consumption only in the last period. (You can do the quadratic utility model of the last section this way as well.) Utility is ¤ £ E [u(c)] = E −e−αc . α is known as the coefficient of absolute risk aversion. If consumption is normally distributed, we have Eu(c) = −e−αE(c)+

α2 2

σ 2 (c)

.

Suppose this investor has initial wealth W which can be split between a riskfree asset paying Rf and a set of risky assets paying return R. Let y denote the amount of this wealth W (amount, not fraction) invested in each security. Then, the budget constraint is c = y f Rf + y 0 R W = y f + y0 1 Plugging the first constraint into the utility function we obtain Eu(c) = −e−α[y

f

2

Rf +y 0 E(R)]+ α2 y 0 Σy

147

.

(118)

C HAPTER 9

FACTOR PRICING MODELS

As with quadratic utility, the two-period model is what allows us to set consumption to wealth and then substitute the return on the wealth portfolio for consumption growth in the discount factor. Maximizing (9.118) with respect to y, y f , we obtain the first order condition describing the optimal amount to be invested in the risky asset, y = Σ−1

E(R) − Rf α

Sensibly, the investor invests more in risky assets if their expected return is higher, less if his risk aversion coefficient is higher, and less if the assets are riskier. Notice that total wealth does not appear in this expression. With this setup, the amount invested in risky assets is independent of the level of wealth. This is why we say that this investor has an aversion to absolute rather than relative (to wealth) risk aversion. Note also that these “demands” for the risky assets are linear in expected returns, which is a very convenient property. Inverting the first order conditions, we obtain E(R) − Rf = αΣy = α cov(R, Rm ).

(119)

The investor’s total risky portfolio is y 0 R. Hence, Σy gives the covariance of each return with y0 R, and also with the investor’s overall portfolio y f Rf + y0 R. If all investors are identical, then the market portfolio is the same as the individual’s portfolio so Σy also gives the correlation of each return with Rm = yf Rf + y 0 R. (If investors differ in risk aversion α, the same thing goes through but with an aggregate risk aversion coefficient.) Thus, we have the CAPM. This version is especially interesting because it ties the market price of risk to the risk aversion coefficient. Applying (9.119) to the market return itself, we have E(Rm ) − Rf = α. σ2 (Rm ) 9.1.3

Quadratic value function, dynamic programming.

We can let investors live forever in the quadratic utility CAPM so long as we assume that the environment is independent over time. Then the value function is quadratic, taking the place of the quadratic second-period utility function. This case is a nice first introduction to dynamic programming. The two-period structure given above is unpalatable, since (most) investors do in fact live longer than two periods. It is natural to try to make the same basic ideas work with less 148

S ECTION 9.1

C APITAL A SSET P RICING MODEL (CAPM)

restrictive and more palatable assumptions. We can derive the CAPM in a multi-period context by replacing the second-period quadratic utility function with a quadratic value function. However, the quadratic value function requires the additional assumption that returns are i.i.d. (no “shifts in the investment opportunity set”). This observation, due to Fama (1970), is also a nice introduction to dynamic programming, which is a powerful way to handle multiperiod problems by expressing them as two period problems. Finally, I think this derivation makes the CAPM more realistic, transparent and intuitively compelling. Buying stocks amounts to taking bets over wealth; really the fundamental assumption driving the CAPM is that marginal utility of wealth is linear in wealth and does not depend on other state variables. Let’s start in a simple ad-hoc manner by just writing down a “utility function” defined over this period’s consumption and next period’s wealth, U = u(ct ) + βEt V (Wt+1 ). This is a reasonable objective for an investor, and does not require us to make the very artificial assumption that he will die tomorrow. If an investor with this “utility function” can buy an asset at price pt with payoff xt+1 , his first order condition (buy a little more, then x contributes to wealth next period) is pt u0 (ct ) = βEt [V 0 (Wt+1 )xt+1 ] . Thus, the discount factor uses next period’s marginal value of wealth in place of the more familiar marginal utility of consumption mt+1 = β

V 0 (Wt+1 ) u0 (ct )

(The envelope condition states that, at the optimum, a penny saved has the same value as a penny consumed u0 (ct ) = V 0 (Wt ). We could use this condition to express the denominator in terms of wealth also.) Now, suppose the value function were quadratic, η V (Wt+1 ) = − (Wt+1 − W ∗ )2 . 2 Then, we would have mt+1

W Rt+1 (Wt − ct ) − W ∗ Wt+1 − W ∗ = −βη = −βη u0 (ct ) u0 (ct ) · ¸ · ¸ ∗ βηW βη(Wt − ct ) W = , + − Rt+1 u0 (ct ) u0 (ct )

149

C HAPTER 9

FACTOR PRICING MODELS

or, once again, mt+1 = at + bt RW t+1 , the CAPM! Let’s be clear about the assumptions and what they do. 1) The value function only depends on wealth. If other variables entered the value function, then ∂V /∂W would depend on those other variables, and so would m. This assumption bought us the first objective of any derivation: the identity of the factors. The ICAPM, below, allows other variables in the value function, and obtains more factors. (Actually, other variables could enter so long as they don’t affect the marginal value of wealth. The weather is an example: You like me might be happier on sunny days, but you do not value additional wealth more on sunny than on rainy days. Hence, covariance with weather does not affect how you value stocks.) 2) The value function is quadratic. We wanted the marginal value function V 0 (W ) be linear, to buy us the second objective, showing m is linear in the factor. Quadratic utility and value functions deliver a globally linear marginal value function V 0 (W ). By the usual Taylor series logic, linearity of V 0 (W ) is probably not a bad assumption for small perturbations, and not a good one for large perturbations. Why is the value function quadratic? You might think we are done. But economists are unhappy about a utility function that has wealth in it. Few of us are like Disney’s Uncle Scrooge, who got pure enjoyment out of a daily swim in the coins in his vault. Wealth is valuable because it gives us access to more consumption. Utility functions should always be written over consumption. One of the few real rules in economics that keep our theories from being vacuous is that ad-hoc “utility functions” over other objects like wealth (or means and variances of portfolio returns, or “status” or “political power”) should be defended as arising from a more fundamental desire for consumption. More practically, being careful about the derivation makes clear that the superficially plausible assumption that the value function is only a function of wealth derives from the much less plausible, in fact certainly false, assumption that interest rates are constant, the distribution of returns is i.i.d., and that the investor has no risky labor income. So, let us see what it takes to defend the quadratic value function in terms of some utility function. Suppose investors last forever, and have the standard sort of utility function ∞ 1 X j β u(ct+j ). U = − Et 2 j=0

Again, investors start with wealth W0 which earns a random return RW and they have no other source of income. In addition, suppose that interest rates are constant, and stock returns 150

S ECTION 9.1

C APITAL A SSET P RICING MODEL (CAPM)

are i.i.d. over time. Define the value function as the maximized value of the utility function in this environment. Thus, define V (W ) as7 V (Wt ) ≡ max{ct ,ct+1 ,ct+2 ...αt ,αt+1 ,...} Et s.t. Wt+1

∞ X

β j u(ct+j )

(9.120)

j=0

W 0 0 = RW t+1 (Wt − ct ); Rt = αt Rt ; αt 1 = 1

¤0 £ (I used vector notation to simplify the statement of the portfolio problem; R ≡ R1 R2 ... RN , etc.) The value function is the total level of utility the investor can achieve, given how much wealth he has, and any other variables constraining him. This is where the assumptions of no labor income, a constant interest rate and i.i.d. returns come in. Without these assumptions, the value function as defined above might depend on these other characteristics of the investor’s environment. For example, if there were some variable, say, “D/P” that indicated returns would be high or low for a while, then the investor would be happier, and have a high value, when D/P is high, for a given level of wealth. Thus, we would have to write V (Wt , D/Pt ) Value functions allow you to express an infinite period problem as a two period problem. Break up the maximization into the first period and all the remaining periods, as follows ∞ X Et+1 β j u(ct+1+j ) s. t. .. max V (Wt ) = max{ct ,αt } u(ct ) + βEt {ct+1 ,ct+2 ..,αt+1 ,αt+2 ....} j=0

or

V (Wt ) = max{ct ,αt } {u(ct ) + βEt V (Wt+1 )} s.t. ...

(121)

Thus, we have defended the existence of a value function. Writing down a two period “utility function” over this period’s consumption and next period’s wealth is not as crazy as it might seem. The value function is also an attractive view of how people actually make decisions. You don’t think “If I buy a sandwich today, I won’t be able to go out to dinner one night 20 years from now” – trading off goods directly as expressed by the utility function. You think “I can’t afford a new car” meaning that the decline in the value of wealth is not worth the increase in the marginal utility of consumption. Thus, the maximization in (9.121) describes your psychological approach to utility maximization. 7 There is also a transversality condition or a lower limit on wealth in the budget constraints. This keeps the consumer from consuming a bit more and rolling over more and more debt, and it means we can write the budget constraint in present value form.

151

C HAPTER 9

FACTOR PRICING MODELS

The remaining question is, can the value function be quadratic? What utility function assumption leads to a quadratic value function? Here is the fun fact: A quadratic utility function leads to a quadratic value function in this environment. This is not a law of nature; it is not true that for any u(c), V (W ) has the same functional form. But it is true here and a few other special cases. The “in this environment” clause is not innocuous. The value function – the achieved level of expected utility – is a result of the utility function and the constraints. How could we show this fact? One way would be to try to calculate the value function by brute force from its definition, equation (9.120). This approach is not fun, and it does not exploit the beauty of dynamic programming, which is the reduction of an infinite period problem to a two period problem. Instead solve (9.121) as a functional equation. Guess that the value function V (Wt+1 ) is quadratic, with some unknown parameters. Then use the recursive definition of V (Wt ) in (9.121), and solve a two period problem–find the optimal consumption choice, plug it into (9.121) and calculate the value function V (Wt ). If the guess was right, you obtain a quadratic function for V (Wt ), and determine any free parameters. Let’s do it. Specify 1 u(ct ) = − (ct − c∗ )2 . 2 Guess γ V (Wt+1 ) = − (Wt+1 − W ∗ )2 2 with γ and W ∗ parameters to be determined later. Then the problem (9.121) is (I don’t write the portfolio choice α part for simplicity; it doesn’t change anything) · ¸ γ 1 V (Wt ) = max − (ct − c∗ )2 − β E(Wt+1 − W ∗ )2 s. t. Wt+1 = RW t+1 (Wt − ct ). {ct } 2 2 (Et is now E since I assumed i.i.d.) Substituting the constraint into the objective, · ¸ ¤ γ £ W 1 ∗ 2 ∗ 2 V (Wt ) = max − (ct − c ) − β E Rt+1 (Wt − ct ) − W . 2 2 {ct }

The first order condition with respect to ct , using cˆ to denote the optimal value, is ¤ W ª ©£ ˆt ) − W ∗ Rt+1 cˆt − c∗ = βγE RW t+1 (Wt − c

Solving for cˆt ,

cˆt = c∗ + βγE

¤ª ©£ W 2 W2 W Rt+1 Wt − cˆt Rt+1 − W ∗ Rt+1 152

(122)

S ECTION 9.1

C APITAL A SSET P RICING MODEL (CAPM)

£ ¤ 2 ∗ W2 ∗ W cˆt 1 + βγE(RW t+1 ) = c + βγE(Rt+1 )Wt − βγW E(Rt+1 ) cˆt =

∗ W2 c∗ − βγE(RW t+1 )W + βγE(Rt+1 )Wt 2 1 + βγE(RW t+1 )

(123)

This is a linear function of Wt . Writing (9.122) in terms of the optimal value of c, we get ¤2 1 γ £ W V (Wt ) = − (ˆ ct − c∗ )2 − β E Rt+1 (Wt − cˆt ) − W ∗ 2 2

(124)

This is a quadratic function of Wt and cˆ. A quadratic function of a linear function is a quadratic function, so the value function is a quadratic function of Wt . If you want to spend a pleasant few hours doing algebra, plug (9.123) into (9.124), check that the result really is quadratic in Wt , and determine the coefficients γ, W ∗ in terms of fundamental parameters β, c∗ , E(RW ), E(RW 2 ) (or σ2 (RW )). The expressions for γ, W ∗ do not give much insight, so I don’t do the algebra here. 9.1.4

Log utility

Log utility rather than quadratic utility also implies a CAPM. Log utility implies that consumption is proportional to wealth, allowing us to substitute the wealth return for consumption data. The point of the CAPM is to avoid the use of consumption data, and so to use wealth or the rate of return on wealth instead. Log utility is another special case that allows this substitution. Log utility is much more plausible than quadratic utility. Suppose that the investor has log utility u(c) = ln(c). Define the wealth portfolio as a claim to all future consumption. Then, with log utility, the price of the wealth portfolio is proportional to consumption itself. pW t = Et

∞ X j=1

βj

∞ X β u0 (ct+j ) ct c ct = E βj ct+j = t+j t u0 (ct ) c 1 − β t+j j=1

The return on the wealth portfolio is proportional to consumption growth, W Rt+1

pW + ct+1 = t+1 W = pt

β 1−β + 1 ct+1 β ct 1−β

153

=

1 ct+1 1 u0 (ct ) . = β ct β u0 (ct+1 )

C HAPTER 9

FACTOR PRICING MODELS

Thus, the log utility discount factor equals the inverse of the wealth portfolio return, mt+1 =

1 . RW t+1

(125)

Equation (9.125) could be used by itself: it attains the goal of replacing consumption data by some other variable. (Brown and Gibbons 1982 test a CAPM in this form.) Note that log utility is the only assumption so far. We do not assume constant interest rates, i.i.d. returns or the absence of labor income. Log utility has a special property that “income effects offset substitution effects,” or in an asset pricing context that “discount rate effects offset cashflow effects.” News of higher consumption = dividend should make the claim to consumption more valuable. However, through u0 (c) it also raises the discount rate, lowering the value of the claim to consumption. For log utility, these two effects exactly offset. 9.1.5

Linearizing any model: Taylor approximations and normal distributions.

Any nonlinear model m = f(z) can be turned into a linear model m = a + bz in discrete time by assuming normal returns. It is traditional in the CAPM literature to try to derive a linear relation between m and the wealth portfolio return. We could always do this by a Taylor approximation, mt+1 ∼ = at + bt RW t+1 . We can make this approximation exact in a special case, that the factors and all asset returns are normally distributed. (We can also take the continuous time limit, which is really the same thing. However, this discrete-time trick is common and useful.) First, I quote without proof the central mathematical trick as a lemma Lemma 1 (Stein’s lemma) If f, R are bivariate normal, g(f ) is differentiable and E | g0 (f) |< ∞, then cov [g(f), R] = E[g 0 (f )] cov(f, R).

(126)

Now we can use the lemma to state the theorem. Theorem 2 If m = g(f ), if f and a set of the payoffs priced by m are normally distributed returns, and if |E[g0 (f )]| < ∞, then there is a linear model m = a + bf that prices the normally distributed returns. 154

S ECTION 9.1

C APITAL A SSET P RICING MODEL (CAPM)

Proof: First, the definition of covariance means that the pricing equation can be rewritten as a restriction between mean returns and the covariance of returns with m: 1 = E(mR) ⇔ 1 = E(m)E(R) + cov(m, R).

(127)

Now, given m = g(f), f and R jointly normal, apply Stein’s lemma (9.126) and (9.127), 1 = E[g(f)]E(R) + E[g0 (f)]cov(f, R) 1 = E[g(f)]E(R) + cov(E[g 0 (f)]f, R) Exploiting the ⇐ part of (9.127), we know that an m with mean E(g(f )) and that depends on f via E(g 0 (f ))f will price assets, m = E[g(f )] + E[g0 (f )][f − E(f )]. ¥ Using this trick, and recalling that we have not assumed i.i.d. so all these moments are conditional, the log utility CAPM implies the linear model "µ µ ¶ ¶2 # ¤ £ W 1 1 mt+1 = Et (128) Rt+1 − Et (RW − Et t+1 ) W Rt+1 RW t+1 if RW t+1 and all asset returns to be priced are normally distributed. From here it is a short step to an expected return-beta representation using the wealth portfolio return as the factor.

In the same way, we can trade the quadratic utility function for normal distributions in the dynamic programming derivation of the CAPM. Starting from £ ¤ V 0 RW V 0 (Wt+1 ) t+1 (Wt − ct ) =β mt+1 = β u0 (ct ) u0 (ct ) we can derive an expression that links m linearly to RW t+1 by assuming normality. Using the same trick, the consumption-based model can be written in linear fashion, i.e. expected returns can be expressed as a linear function of betas on consumption growth rather than betas on consumption growth raised to a power. However, for large risk aversion coefficients (more than about 10 in postwar consumption data) or other transformations, the inaccuracies due to the normal or lognormal approximation can be very significant in discrete data. The normal distribution assumption seems rather restrictive, and it is. However, the most popular class of continuous-time models specify instantaneously normal distributions even for things like options that have very non-normal distributions for discrete time intervals. 155

C HAPTER 9

FACTOR PRICING MODELS

Therefore, one can think of the Stein’s lemma tricks as a way to get to continuous time approximations without doing it in continuous time. I demonstrate the explicit continuous time approach with the ICAPM, in the next section. 9.1.6

Portfolio intuition

The classic derivation of the CAPM contains some useful intuition. The classic derivation starts with a mean-variance objective for portfolio wealth, max Eu(W ). Beta drives average returns because beta measures how much adding a bit of the asset to a diversified portfolio increases the volatility of the portfolio. The central insight that started it all is that investors care about portfolio returns, not about the behavior of specific assets. Once the characteristics of portfolios replaced demand curves for individual stocks, modern finance was born.

9.2

Intertemporal Capital Asset Pricing Model (ICAPM)

Any “state variable” zt can be a factor. The ICAPM is a linear factor model with wealth and state variables that forecast changes in the distribution of future returns or income. The ICAPM generates linear discount factor models mt+1 = a + b0 ft+1 in which the factors are “state variables” for the investor’s consumption-portfolio decision. The “state variables” are the variables that determine how well the investor can do in his maximization. Current wealth is obviously a state variable. Additional state variables describe the conditional distribution of income and asset returns the agent will face in the future or “shifts in the investment opportunity set.” In multiple good or international models, relative price changes are also state variables. Optimal consumption is a function of the state variables, ct = g(zt ). We can use this fact once again to substitute out consumption, and write mt+1 = β

u0 [g(zt+1 )] . u0 [g(zt )]

From here, it is a simple linearization to deduce that the state variables zt+1 will be factors. Alternatively, the value function depends on the state variables V (Wt+1 , zt+1 ), 156

S ECTION 9.2

I NTERTEMPORAL C APITAL A SSET P RICING M ODEL (ICAPM)

so we can write mt+1 = β

VW (Wt+1 , zt+1 ) VW (Wt , zt )

(The marginal value of a dollar must be the same in any use, so I made the denominator pretty by writing u0 (ct ) = VW (Wt , zt ). This fact is known as the envelope condition.) This completes the first step, naming the proxies. To obtain a linear relation, we can take a Taylor approximation, assume normality and use Stein’s lemma, or, most conveniently, move to continuous time (which is really just a more convenient way of making the normal approximation.) We saw above that we can write the basic pricing equation in continuous time as µ ¶ dΛ dp dp − rf dt = −E . E p Λ p (for simplicity of the formulas, I’m folding any dividends into the price process). The discount factor is marginal utility, which is the same as the marginal value of wealth, dVW (Wt , zt ) du0 (ct ) dΛt = = 0 Λt u (ct ) VW Our objective is to express the model in terms of factors z rather than marginal utility or value, and Ito’s lemma makes this easy VW z W VW W dW 1 dVW + = dz + (second derivative terms) VW VW W VW 2 (We don’t have to grind out the second derivative terms if we are going to take rf dt = Et (dΛ/Λ) , though this approach removes a potentially interesting and testable implication of the model). The elasticity of marginal value with respect to wealth is often called the coefficient of relative risk aversion, rra ≡ −

W VW W . VW

Substituting, we obtain the ICAPM, which relates expected returns to the covariance of returns with wealth, and also with the other state variables, µ µ ¶ ¶ dW dp dp dp VW z − rf dt = rra E E dz E − . p W p VW p From here, it is fairly straightforward to express the ICAPM in terms of betas rather than covariances, or as a linear discount factor model. Most empirical work occurs in discrete time; we often simply approximate the continuous time result as E(R) − Rf ≈ rra cov(R, ∆W ) + λz cov(R, ∆z). 157

C HAPTER 9

FACTOR PRICING MODELS

One often substitutes covariance with the wealth portfolio for covariance with wealth, and one uses factor-mimicking portfolios for the other factors dz as well. The factor-mimicking portfolios are interesting for portfolio advice as well, as they give the purest way of hedging against or profiting from state variable risk exposure. This short derivation does not do justice to the beauty of Merton’s portfolio theory and ICAPM. What remains is to actually state the consumer’s problem and prove that the value function depends on W and z, the state variables for future investment opportunities, and that the optimal portfolio holds the market and hedge portfolios for the investment opportunity variables.

9.3

Comments on the CAPM and ICAPM

Conditional vs. unconditional models. Do they price options? Why bother linearizing? The wealth portfolio. Ex-post returns. The implicit consumption-based model. What are the ICAPM state variables? CAPM and ICAPM as general equilibrium models Is the CAPM conditional or unconditional? Is the CAPM a conditional or an unconditional factor model? I.e., are the parameters a and b in m = a − bRW constants, or do they change at each time period, as conditioning information changes? We saw above that a conditional CAPM does not imply an unconditional CAPM, so additional steps must be taken to say anything about observed average returns. The two period quadratic utility based derivation results in a conditional CAPM, since the parameters at and bt depend on consumption which changes over time. Also we know that a and b must vary over time if the conditional moments of RW , Rf vary over time. This twoperiod investor chooses a portfolio on the conditional mean variance frontier, which is not on the unconditional frontier. The multiperiod quadratic utility CAPM only holds if returns are i.i.d. so it only holds if there is no difference between conditional and unconditional models. The log utility CAPM expressed with the inverse market return is a beautiful model, since it holds both conditionally and unconditionally. There are no free parameters that can change

158

S ECTION 9.3

C OMMENTS ON THE CAPM AND ICAPM

with conditioning information: 1 = Et

µ

1 RW t+1

Rt+1

¶

⇔1=E

µ

1 RW t+1

¶

Rt+1 .

In fact there are no free parameters at all! Furthermore, the model makes no distributional assumptions, so it can apply to any asset, including options. Finally it requires no specification of the investment opportunity set, or (macro language) no specification of technology. Linearizing the log utility CAPM comes at enormous price. The expectations in the linearized log utility CAPM (9.128) are conditional. Thus, the apparent simplification of linearity destroys the nice unconditional feature of the log utility CAPM. Should the CAPM price options? As I have derived them, the quadratic utility CAPM and the nonlinear log utility CAPM should apply to all payoffs: stocks, bonds, options, contingent claims, etc. However, if we assume normal return distributions to obtain a linear CAPM from log utility, we can no longer hope to price options, since option returns are non-normally distributed (that’s the point of options!) Even the normal distribution for regular returns is a questionable assumption. You may hear the statement “the CAPM is not designed to price derivative securities”; the statement refers to the log utility plus normal-distribution derivation of the linear CAPM. Why linearize? Why bother linearizing a model? Why take the log utility model m = 1/RW which W that loses the clean conditioningshould price any asset, and turn it into mt+1 = at + bt Rt+1 down property and cannot price non-normally distributed payoffs? These tricks were developed before the p = E(mx) expression of asset pricing models, when (linear) expected return-beta models were the only thing around. You need a linear model of m to get an expected return - beta model. More importantly, the tricks were developed when it was hard to estimate nonlinear models. It’s clear how to estimate a β and a λ by regressions, but estimating nonlinear models used to be a big headache. Now, GMM has made it easy to estimate and evaluate nonlinear models. Thus, in my opinion, linearization is mostly intellectual baggage. The desire for linear representations and this normality trick is one of the central reasons why many asset pricing models are written in continuous time. In most continuous time models, everything is locally normal. Unfortunately for empiricists, this approach adds timeaggregation and another layer of unobservable conditioning information into the predictions of the model. For this reason, most empirical work is still based on discrete-time models. However, the local normal distributions in continuous time, even for option returns, is a good reminder that normal approximations probably aren’t that bad, so long as the time interval is kept reasonably short. What about the wealth portfolio? The log utility derivation makes clear just how expansive is the concept of the wealth portfolio. To own a (share of) the consumption stream, you have to own not only all stocks, 159

C HAPTER 9

FACTOR PRICING MODELS

but all bonds, real estate, privately held capital, publicly held capital (roads, parks, etc.), and human capital – a nice word for “people.” Clearly, the CAPM is a poor defense of common proxies such as the value-weighted NYSE portfolio. And keep in mind that since it is easy to find ex-post mean-variance efficient portfolios of any subset of assets (like stocks) out there, taking the theory seriously is our only guard against fishing. Implicit consumption-based models Many users of alternative models clearly are motivated by a belief that the consumptionbased model doesn’t work, no matter how well measured consumption might be. This view is not totally unreasonable; as above, perhaps transactions costs de-link consumption and asset returns at high frequencies, and some diagnostic evidence suggests that the consumption behavior necessary to save the consumption model is too wild to be believed. However, the derivations make clear that the CAPM and ICAPM are not alternatives to the consumption-based model, they are special cases of that model. In each case mt+1 = βu0 (ct+1 )/u0 (ct ) still operates. We just added assumptions that allowed us to substitute other variables in place of ct . One cannot adopt the CAPM on the belief that the consumption based model is wrong. If you think the consumption-based model is wrong, the economic justification for the alternative factor models evaporates. The only plausible excuse for factor models is a belief that consumption data are unsatisfactory. However, while asset return data are well measured, it is not obvious that the S&P500 or other portfolio returns are terrific measures of the return to total wealth. “Macro factors” used by Chen, Roll and Ross (1986) and others are distant proxies for the quantities they want to measure, and macro factors based on other NIPA aggregates (investment, output, etc.) suffer from the same measurement problems as aggregate consumption. In large part, the “better performance” of the CAPM and ICAPM relative to consumptionbased models comes from throwing away content. Again mt+1 = δu0 (ct+1 )/u0 (ct ) is there in any CAPM or ICAPM. The CAPM and ICAPM make predictions concerning consumption data that are wildly implausible, not only of admittedly poorly measured aggregate consumption data but any imaginable perfectly measured individual consumption data as well. For example, equation (9.129) says that the standard deviation of the wealth portfolio return equals the standard deviation of consumption growth. The latter is about 1% per year. All the miserable failures of the log-utility consumption-based model apply equally to the log utility CAPM. Finally, most models take the market price of risk as a free parameter. Of course it isn’t; it is related to risk aversion and consumption volatility and is very hard to justify as such. Ex-post returns The log utility model also allows us for the first time to look at what moves returns ex-post as well as ex-ante. Recall that, in the log utility model, we have RW t+1 =

1 ct+1 . β ct

160

(129)

S ECTION 9.3

C OMMENTS ON THE CAPM AND ICAPM

Thus, the wealth portfolio return is high, ex-post, when consumption is high. This holds at every frequency: If stocks go up between 12:00 and 1:00, it must be because (on average) we all decided to have a big lunch. This seems silly. Aggregate consumption and asset returns are likely to be de-linked at high frequencies, but how high (quarterly?) and by what mechanism are important questions to be answered. In any case, this is another implication of the log utility CAPM that is just thrown out. In sum, the poor performance of the consumption-based model is an important nut to chew on, not just a blind alley or failed attempt that we can safely disregard and go on about our business. Identity of state variables The ICAPM does not tell us the identity of the state variables zt , and many authors use the ICAPM as an obligatory citation to theory on the way to using factors composed of ad-hoc portfolios, leading Fama (1991) to characterize the ICAPM as a “fishing license.” The ICAPM really isn’t quite such an expansive license. One could do a lot to insist that the factor-mimicking portfolios actually are the projections of some identifiable state variables on to the space of returns, and one could do a lot to make sure the candidate state variables really are plausible state variables for an explicitly stated optimization problem. For example, one could check that investment-opportunity set state variables actually do forecast something. The fishing license comes as much from habits of applying the theory as from the theory itself. General equilibrium models The CAPM and other models are really general equilibrium models. Looking at the derivation through general-equilibrium glasses, we have specified a set of linear technologies with returns Ri that do not depend on the amount invested. Some derivations make further assumptions, such as an initial capital stock, and no labor or labor income. The CAPM is obviously very artificial. Its central place really comes from its long string of empirical successes rather than its theoretical purity. The theory was extended and multiple factors anticipated long before they became empirically popular. Portfolio intuition I have derived all the models as instances of the consumption-based model. The more traditional portfolio intuition for multifactor models is also useful. The intuition (and historical development) comes from looking past consumption to its determinants in sources of income or news. The CAPM simplifies matters by assuming that the average investor only cares the performance of his investment portfolio. Most of us have jobs, so events like recessions hurt the majority of investors. People with jobs will prefer stocks that don’t fall in recessions, even if their market betas, mean returns, and standard deviations are the same as stocks that do fall in recessions. Demanding such stocks, they drive down the corresponding expected returns. Thus, we expect expected returns to depend on additional betas that capture labor market 161

C HAPTER 9

FACTOR PRICING MODELS

conditions. The traditional ICAPM intuition works the same way. Even jobless investors have long horizons. Thus, they will prefer stocks that do well when news comes that future returns are lower. Demanding more of such stocks, they depress expected returns. Thus, expected returns come to depend on covariation with news of future returns, not just covariation with the current market return. The ICAPM remained on the theoretical shelf for 20 years mostly because it took that long to accumulate empirical evidence that returns are, in fact, predictable. It is vitally important that the extra factors affect the average investor. If an event makes investor A worse off and investor B better off, then investor A buys assets that do well when the event happens, and investor B sells them. They transfer the risk of the event, but the price or expected return of the asset is unaffected. For a factor to affect prices or expected returns, the average investor must be affected by it, so investors collectively bid up or down the price and expected return of assets that covary with the event rather than just transfer the risk without affecting equilibrium prices. As you can see, this traditional intuition is encompassed by consumption. Bad labor market outcomes or bad news about future returns are bad news that raise the marginal utility of wealth, which equals the marginal utility of consumption.

9.4

Arbitrage Pricing Theory (APT)

The APT: If a set of asset returns are generated by a linear factor model Ri = E(Ri ) +

N X

β ij f˜j + εi

j=1

E(εi ) = E(εi f˜j ) = 0. Then (with additional assumptions) there is a discount factor m linear in the factors m = a + b0 f that prices the returns. The APT starts from a statistical characterization. There is a big common component to stock returns: when the market goes up, most individual stocks also go up. Beyond the market, groups of stocks move together such as computer stocks, utilities, small stocks, value stocks and so forth. Finally, each stock’s return has some completely idiosyncratic movement. This is a characterization of realized returns, outcomes or payoffs. The point of the APT is to start with this statistical characterization of outcomes, and derive something about expected returns or prices. The intuition behind the APT is that the completely idiosyncratic movements in asset 162

S ECTION 9.4

A RBITRAGE P RICING T HEORY (APT)

returns should not carry any risk prices, since investors can diversify them away by holding portfolios. Therefore, risk prices or expected returns on a security should be related to the security’s covariance with the common components or “factors” only. The job of this section is then 1) to describe a mathematical model of the tendency for stocks to move together, and thus to define the “factors” and residual idiosyncratic components, and 2) to think carefully about what it takes for the idiosyncratic components to have zero (or small) risk prices, so that only the common components matter to asset pricing. There are two lines of attack for the second item. 1) If there were no residual, then we could price securities from the factors by arbitrage (really, by the law of one price, but the current distinction between law of one price and arbitrage came after the APT was named.) Perhaps we can extend this logic and show that if the residuals are small, they must have small risk prices. 2) If investors all hold well-diversified portfolios, then only variations in the factors drive consumption and hence marginal utility. Much of the original appeal and marketing of the APT came from the first line of attack, the idea that we could derive pricing implications without the economic structure required of the CAPM, ICAPM, or any other model derived as a specialization of the consumptionbased model. In this section, I will first try to see how far we can in fact get with purely law of one price arguments. I will conclude that the answer is, “not very far,” and that the most satisfactory argument for the APT is in fact just another specialization of the consumptionbased model. 9.4.1

Factor structure in covariance matrices

I define and examine the factor decomposition xi = αi + β 0i f + εi ; E(εi ) = 0, E(fεi ) = 0 The factor decomposition is equivalent to a restriction on the payoff covariance matrix. The APT models the tendency of asset payoffs (returns) to move together via a statistical factor decomposition xi = αi +

M X

β ij fj + εi = αi + β 0i f + εi .

(130)

j=1

The fj are the factors, the β ij are the betas or factor loadings and the εi are residuals. letter without subscripts to denote a vector, for example f = £As usual, I use the ¤same 0 f1 f2 ... fK . A discount factor m, pricing factors f in m = b0 f and this factor decomposition (or factor structure) for returns are totally unrelated uses of the word “factor.” 163

C HAPTER 9

FACTOR PRICING MODELS

I didn’t invent the terminology! The APT is conventionally written with xi = returns, but it ends up being much less confusing to use prices and payoffs. It is a convenient and conventional simplification to fold the factor means into the first, constant, factor and write the factor decomposition with zero-mean factors f˜ ≡ f − E(f). xi = E(xi ) +

M X

β ij f˜j + εi .

(131)

j=1

Remember that E(xi ) is still just a statistical characterization, not a prediction of a model. We can construct the factor decomposition as a regression equation. Define the β ij as regression coefficients, and then the εi are uncorrelated with the factors by construction, E(εi f˜j ) = 0. The content — the assumption that keeps (9.131) from describing any arbitrary set of returns — is an assumption that the εi are uncorrelated with each other. E(εi εj ) = 0. (More general versions of the model allow some limited correlation across the residuals but the basic story is the same.) The factor structure is thus a restriction on the covariance matrix of payoffs. For example, if there is only one factor, then ½ 2 σεi if i = j i j i j 2 ˜ ˜ cov(x , x ) = E[(β i f + ε )(β j f + ε )] = β i β j σ (f) + . 0 if i 6= j Thus, with N = number of securities, the N (N − 1)/2 elements of a variance-covariance matrix are described by N betas, and N + 1 variances. A vector version of the same thing is 2 0 σ1 0 2 cov(x, x0 ) = ββ 0 σ 2 (f ) + 0 σ 2 0 . .. . 0 0

With multiple (orthogonalized) factors, we obtain

cov(x, x0 ) = β 1 β 01 σ2 (f1 ) + β 2 β 02 σ2 (f2 ) + . . . + (diagonal matrix) In all these cases, we describe the covariance matrix a singular matrix ββ 0 (or a sum of a few such singular matrices) plus a diagonal matrix. If we know the factors we want to use ahead of time, say the market (value-weighted portfolio) and industry portfolios, or size and book to market portfolios, we can estimate a factor structure by running regressions. Often, however, we don’t know the identities of the factor portfolios ahead of time. In this case we have to use one of several statistical 164

S ECTION 9.4

A RBITRAGE P RICING T HEORY (APT)

techniques under the broad heading of factor analysis (that’s where the word “factor” came from in this context) to estimate the factor model. One can estimate a factor structure quickly by simply taking an eigenvalue decomposition of the covariance matrix, and then setting small eigenvalues to zero. 9.4.2

Exact factor pricing

With no error term, xi = E(xi )1 + β 0i f˜. implies p(xi ) = E(xi )p(1) + β 0i p(f˜) and thus m = a + b0 f ; p(xi ) = E(mxi ) E(Ri ) = Rf + β 0i λ. using only the law of one price. Suppose that there are no idiosyncratic terms εi . This is called an exact factor model. Now look again at the factor decomposition, xi = E(xi )1 + β 0i f˜.

(132)

It started as a statistical decomposition. But it also says that the payoff xi can be synthesized as a portfolio of the factors and a constant (risk-free payoff). Thus, the price of xi can only depend on the prices of the factors f, ˜ p(xi ) = E(xi )p(1) + β 0i p(f).

(133)

The law of one price assumption lets you take prices of right and left sides. If the factors are returns, their prices are 1. If the factors are not returns, their prices are free parameters which can be picked to make the model fit as well as possible. Since there are fewer factors than payoffs, this procedure is not vacuous. (Recall that the prices of the factors are related to the λ in expected return beta representations. λ is determined by the expected return of a return factor, and is a free parameter for non-return factor models.) We are really done, but the APT is usually stated as “there is a discount factor linear in f that prices returns Ri ,” or “there is an expected return-beta representation with f as 165

C HAPTER 9

FACTOR PRICING MODELS

factors.” Therefore, we should take a minute to show that the rather obvious relationship (9.133) between prices is equivalent to discount factor and expected return statements. Assuming only the law of one price, we know there is a discount factor m linear in factors that prices the factors. We usually call it x∗ , but call it f ∗ here to remind us that it £ ¤0 prices the factors. Denote fˆ = 1 f˜ the factors including the constant. As with x∗ , f ∗ = p(fˆ)0 E(fˆfˆ0 )−1 fˆ = a + b0 f satisfies p(fˆ) = E(f ∗ fˆ) and p(1) = E(f ∗ ). If the discount factor prices the factors, it must price any portfolio of the factors; hence f ∗ prices all payoffs xi that follow the factor structure (9.132). We could now go from m linear in the factors to an expected return-beta model using the above theorems that connect the two representations. But there is a more direct and elegant connection. Start with (9.133), specialized to returns xi = Ri and of course p(Ri ) = 1. Use p(1) = 1/Rf and solve for expected return as h i ˜ = Rf + β 0i λ. E(Ri ) = Rf + β 0i −Rf p(f)

The last equality defines λ. Expected returns are linear in the betas, and the constants (λ) are related to the prices of the factors. In fact, this is the same definition of λ that we arrived at above connecting m = b0 f to expected return-beta models. 9.4.3

Approximate APT using the law of one price

Attempts to extend the exact factor model to an approximate factor pricing model when errors are “small,” or markets are “large,” still only using law of one price. For fixed m, the APT gets better and better as R2 or the number of assets increases. However, for any fixed R2 or size of market, the APT can be arbitrarily bad. These observations mean that we must go beyond the law of one price to derive factor pricing models. Actual returns do not display an exact factor structure. There is some idiosyncratic or residual risk; we cannot exactly replicate the return of a given stock with a portfolio of a few large factor portfolios. However, the idiosyncratic risks are often small. For example, factor model regressions of the form (9.130) often have very high R2 , especially when portfolios rather than individual securities are on the left hand side. And the residual risks are still idiosyncratic: Even if they are a large part of an individual security’s variance, they should be a small contributor to the variance of well diversified portfolios. Thus, there is reason to hope that the APT holds approximately, especially for reasonably large portfolios. Surely, if the residuals are “small” and/or “idiosyncratic,” the price of an asset can’t be “too different” from the price predicted from its factor content? 166

S ECTION 9.4

A RBITRAGE P RICING T HEORY (APT)

To think about these issues, start again from a factor structure, but this time put in a residual, xi = E(xi )1 + β 0i f˜ + εi Again take prices of both sides, p(xi ) = E(xi )p(1) + β 0i p(f˜) + E(mεi ) Now, what can we say about the price of the residual p(εi ) = E(mεi )? Figure 23 illustrates the situation. Portfolios of the factors span a payoff space, the line from the origin through β 0i f in the Figure. The payoff we want to price, xi is not in that space, since the residual εi is not zero. A discount factor f ∗ that is in the f payoff space prices the factors. The set of all discount factors that price the factors is the line m perpendicular to f ∗ . The residual εi is orthogonal to the factor space, since it is a regression residual, and to f ∗ in particular, E(f ∗ εi ) = 0. This means that f ∗ assigns zero price to the residual. But the other discount factors on the m line are not orthogonal to εi , so generate non-zero price for the residual εi . As we sweep along the line of discount factors m that price the f, in fact, we generate every price from −∞ to ∞ for the residual. Thus, the law of one price does not nail down the price of the residual εi and hence the price or expected return of xi .

β’if

All m m>0

εi xi

f*

m: σ2(m) < A

m

Figure 23. Approximate arbitrage pricing. 167

C HAPTER 9

FACTOR PRICING MODELS

Limiting arguments We would like to show that the price of xi has to be “close to” the price ofβ 0i f . One notion of “close to” is that in some appropriate limit the price of xi converges to the price of β 0i f. “Limit” means, of course, that you can get arbitrarily good accuracy by going far enough in the direction of the limit (for every ε > 0 there is a δ....). Thus, establishing a limit result is a way to argue for an approximation. Here is one theorem that seems to imply that the APT should be a good approximation for portfolios that have high R2 on the factors. I state the argument for the case that there is a constant factor, so the constant is in the f space and E(εi ) = 0. The same ideas work in the less usual case that there is no constant factor, using second moments in place of variance. Theorem: Fix a discount factor m that prices the factors. Then, as var(εi ) → 0, p(xi ) → p(β 0i f ). This is easiest to see by just looking at the graph. E(εi ) = 0 so var(εi ) = E(εi2 ) = ||ε || . Thus, as the size of the εi vector in Figure 23 gets smaller, xi gets closer and closer to β 0i f. For any fixed m, the induced pricing function (lines perpendicular to the chosen m) is continuous. Thus, as xi gets closer and closer to β 0i f , its price gets closer and closer to β 0i f. i 2

The factor model is defined as a regression, so var(xi ) = var(β 0i f ) + var(εi ) Thus, the variance of the residual is related to the regression R2 . var(εi ) = 1 − R2 var(xi ) The theorem says that as R2 → 1, the price of the residual goes to zero.

We were hoping for some connection between the fact that the risks are idiosyncratic and factor pricing. Even if the idiosyncratic risks are a large part of the payoff at hand, they are a small part of a well-diversified portfolio. The next theorem shows that portfolios with high R2 don’t have to happen by chance; well-diversified portfolios will always have this characteristic. Theorem: As the number of primitive assets increases, the R2 of well-diversified portfolios increases to 1. Proof: Start with an equally weighted portfolio xp =

N 1 X i x. N i=1

Going back to the factor decomposition (9.130) for each individual asset xi , the 168

S ECTION 9.4

A RBITRAGE P RICING T HEORY (APT)

factor decomposition of xp is xp =

N N N N ¢ 0 1 X 1 X¡ 1 X 0 1 X i ai + β 0i f + εi = ai + β if + ε = ap + β p f + εp . N i=1 N i=1 N i=1 N i=1

The last equality defines notation αp , β p , εp . But p

var(ε ) = var

Ã

N 1 X i ε N i=1

!

So long as the variance of εi are bounded, and given the factor assumption E(εi εj ) = 0, lim var(εp ) = 0.

N→∞

Obviously, the same idea goes through so long as the portfolio spreads some weight on all the new assets, i.e. so long as it is “well-diversified.” ¥ These two theorems can be interpreted to say that the APT holds approximately (in the usual limiting sense) for portfolios that either naturally have high R2 , or well-diversified portfolios in large enough markets. We have only used the law of one price. Law of one price arguments fail Now, let me pour some cold water on these results. I fixed m and then let other things take limits. The flip side is that for any nonzero residual εi , no matter how small, we can pick a discount factor m that prices the factors and assigns any price to xi ! As often in mathematics, the order of “for all” and “there exists” matters a lot. Theorem: For any nonzero residual εi there is a discount factor that prices the factors f (consistent with the law of one price) and that assigns any desired price in (−∞, ∞) to the payoff xi . So long as ||εi || > 0, as we sweep the choice of m along the dashed line, the inner product of m with εi and hence xi varies from −∞ to ∞. Thus, for a given size R2 < 1, or a given finite market, the law of one price says absolutely nothing about the prices of payoffs that do not exactly follow the factor structure. The law of one price says that two ways of constructing the same portfolio must give the same price. If the residual is not exactly zero, there is no way of replicating the payoff xi from the factors and no way to infer anything about the price of xi from the price of the factors. I think the contrast between this theorem and those of the last subsection accounts for most of the huge theoretical controversy over the APT. If you fix m and take limits of N or 169

C HAPTER 9

FACTOR PRICING MODELS

ε, the APT gets arbitrarily good. But if you fix N or ε, as one does in any application, the APT can get arbitrarily bad as you search over possible m. The lesson I learn is that the effort to extend prices from an original set of securities (f in this case) to new payoffs that are not exactly spanned by the original set of securities, using only the law of one price, is fundamentally doomed. To extend a pricing function, you need to add some restrictions beyond the law of one price. 9.4.4

Beyond the law of one price: arbitrage and Sharpe ratios

We can find a well-behaved approximate APT if we impose the law of one price and a restriction on the volatility of discount factors, or, equivalently, a bound on the Sharpe ratio achievable by portfolios of the factors and test assets. The approximate APT based on the law of one price fell apart because we could always choose a discount factor sufficiently “far out” to generate an arbitrarily large price for an arbitrarily small residual. But those discount factors are surely “unreasonable.” Surely, we can rule them out, reestablishing an approximate APT, without jumping all the way to fully specified discount factor models such as the CAPM or consumption-based model A natural first idea is to impose the no-arbitrage restriction that m must be positive. Graphically, we are now restricted to the solid m line in Figure 23. Since that line only extends a finite amount, restricting us to strictly positive m0 s gives rise to finite upper and lower arbitrage bounds on the price of εi and hence xi . (The word arbitrage bounds comes from option pricing, and we will see these ideas again in that context. If this idea worked, it would restore the APT to “arbitrage pricing” rather than “law of one-pricing.”) Alas, in applications of the APT (as often in option pricing), the arbitrage bounds are too wide to be of much use. The positive discount factor restriction is equivalent to saying “if portfolio A gives a higher payoff than portfolio B in every state of nature, then the price of A must be higher than the price of B.” Since stock returns and factors are continuously distributed, not two-state distributions as I have graphed for figure 23, there typically are no strictly dominating portfolios, so adding m > 0 does not help. A second restriction does let us derive an approximate APT that is useful in finite markets with R2 < 1. We can restrict the variance and hence the size (||m|| = E(m2 ) = σ 2 (m) + E(m)2 = σ 2 (m) + 1/Rf2 ) of the discount factor. Figure 23 includes a plot of the discount factors with limited variance, size, or length in the geometry of that Figure. The restricted range of discount factors produces a restricted range of prices for xi . The restricted range of discount factors gives us upper and lower price bounds for the price of xi in terms of the

170

S ECTION 9.5

APT VS . ICAPM

factor prices. Precisely, the upper and lower bounds solve the problem min ( or max) p(xi ) = E(mxi ) s.t. E(mf ) = p(f), m ≥ 0, σ2 (m) ≤ A. {m}

{m}

Limiting the variance of the discount factor is of course the same as limiting the maximum Sharpe ratio (mean / standard deviation of excess return) available from portfolios of the factors and xi . Recall that σ(m) E (Re ) ≤ . σ(Re ) E(m) Though a bound on Sharpe ratios or discount factor volatility is not a totally preferencefree concept, it clearly imposes a great deal less structure than the CAPM or ICAPM which are essentially full general equilibrium models. Ross (1976) included this suggestion in his original APT paper, though it seems to have disappeared from the literature since then in the failed effort to derive an APT from the law of one price alone. Ross pointed out that deviations from factor pricing could provide very high Sharpe ratio opportunities, which seem implausible though not violations of the law of one price. Saá-Requejo and I (2000) dub this idea “good-deal” pricing, as an extension of “arbitrage pricing.” Limiting σ(m) rules out “good deals” as well as pure arbitrage opportunities. Having imposed a limit on discount factor volatility or Sharpe ratio A, then the APT limit does work, and does not depend on the order of “for all” and “there exists.” Theorem: As εi → 0 and R2 → 1, the price p(xi ) assigned by any discount factor m that satisfies E(mf) = p(f ), m ≥ 0, σ 2 (m) ≤ A approaches p(β 0i f ).

9.5

APT vs. ICAPM

A factor structure in the covariance of returns or high R2 in regressions of returns on factors can imply factor pricing (APT) but factors can price returns without describing their covariance matrix (ICAPM). Differing inspiration for factors. The disappearance of absolute pricing. The APT and ICAPM stories are often confused. Factor structure can imply factor pricing (APT), but factor pricing does not require a factor structure. In the ICAPM there is no presumption that factors f in a pricing model m = b0 f describe the covariance matrix of returns. The factors don’t have to be orthogonal or i.i.d. either. High R2 in time-series regressions of the returns on the factors may imply factor pricing (APT), but again are not 171

C HAPTER 9

FACTOR PRICING MODELS

necessary (ICAPM). The regressions of returns on factors can have low R2 in the ICAPM. Factors such as industry may describe large parts of returns’ variances but not contribute to the explanation of average returns. The biggest difference between APT and ICAPM for empirical work is in the inspiration for factors. The APT suggests that one start with a statistical analysis of the covariance matrix of returns and find portfolios that characterize common movement. The ICAPM suggests that one start by thinking about state variables that describe the conditional distribution of future asset returns and non-asset income. More generally, the idea of proxying for marginal utility growth suggests macroeconomic indicators, and indicators of shocks to non-asset income in particular. The difference between the derivations of factor pricing models, and in particular an approximate law-of-one-price basis vs. a proxy for marginal utility basis seems not to have had much impact on practice. In practice, we just test models m = b0 f and rarely worry about derivations. The best evidence for this view is the introductions of famous papers. Chen, Roll and Ross (1986) describe one of the earliest popular multifactor models, using industrial production and inflation as some of the main factors. They do not even present a factor decomposition of test asset returns, or the time-series regressions. A reader might well categorize the paper as a macroeconomic factor model or perhaps an ICAPM. Fama and French (1993) describe the currently most popular multifactor model, and their introduction describes it as an ICAPM in which the factors are state variables. But the factors are sorted on size and book/market just like the test assets, the time-series R2 are all above 90%, and much of the explanation involves “common movement” in test assets captured by the factors. A a reader might well categorize the model as much closer to an APT. In the first chapter, I made a distinction between relative pricing and absolute pricing. In the former, we price one security given the prices of others, while in the latter, we price each security by reference to fundamental sources of risk. The factor pricing stories are interesting in that they start with a nice absolute pricing model, the consumption-based model, and throw out enough information to end up with relative models. The CAPM prices Ri given the market, but throws out the consumption-based model’s description of where the market return came from.

9.6 1.

2. 3.

Problems

Suppose the investor only has a one-period horizon. He invests wealth W at date zero, and only consumes with expected utility Eu(c) = Eu(W ) in period one. Derive the quadratic utility CAPM in this case. (This is an even simpler derivation. The Lagrange multiplier on initial wealth W now becomes the denominator of m in place of u0 (c0 )). Express the log utility CAPM in continuous time to derive a discount factor linear in wealth. Figure 23 suggests that m > 0 is enough to establish a well-behaved approximate APT. 172

S ECTION 9.6

4.

P ROBLEMS

The text claims this is not true. Which is right? Can you use any excess return for the market factor in the CAPM, or must it be the market less the riskfree rate?

173

PART II Estimating and evaluating asset pricing models

174

S ECTION 9.6

P ROBLEMS

Our first task in bringing an asset pricing model to data is to estimate the free parameters; the β and γ in m = β(ct+1 /ct )−γ , or the b in m = b0 f. Then we want to evaluate the model. Is it a good model or not? Is another model better? Statistical analysis helps us to evaluate a model by providing a distribution theory for numbers such as parameter estimates that we create from the data. A distribution theory pursues the following idea: Suppose that we generate artificial data over and over again from a statistical model. For example, we could specify that the market return is an i.i.d. normal + εit . After random variable, and a set of stock returns is generated by Rtei = αi + β i Rem t 2 i picking values for the mean and variance of the market return and the αi , β i , σ (ε ), we could ask a computer to simulate many artificial data sets. We can repeat our statistical procedure in each of these artificial data sets, and graph the distribution of any statistic which we have estimated from the real data, i.e. the frequency that it takes on any particular value in our artificial data sets. In particular, we are interested in a distribution theory for the estimated parameters, to give us some sense of how much the data really has to say about their values; and for the pricing errors, which helps us to judge whether pricing errors are just bad luck of one particular historical accident or if they indicate a failure of the model. We also will want to generate distributions for statistics that compare one model to another, or provide other interesting evidence, to judge how much sample luck affects those calculations. All of the statistical methods I discuss in this part achieve these ends. They give methods for estimating free parameters; they provide a distribution theory for those parameters, and they provide distributions for statistics that we can use to evaluate models, most often a ˆ. quadratic form of pricing errors in the form α ˆ 0 V −1 α I start by focusing on the GMM approach. The GMM approach is a natural fit for a discount factor formulation of asset pricing theories, since we just use sample moments in the place of population moments. As you will see, there is no singular “GMM estimate and test.” GMM is a large canvas and a big set of paints and brushes; a flexible tool for doing all kinds of sensible (and, unless you’re careful, not-so-sensible) things to the data. Then I consider traditional regression tests (naturally paired with expected return-beta statements of factor models) and their maximum likelihood formalization. I emphasize the fundamental similarities between these three methods, as I emphasized the similarity between p = E(mx),

175

C HAPTER 9

FACTOR PRICING MODELS

expected return-beta models, and mean-variance frontiers. A concluding chapter highlights some of the differences between the methods, as I contrasted p = E(mx) and beta or meanvariance representations of the models.

176

Chapter 10. GMM in explicit discount factor models The basic idea in the GMM approach is very straightforward. The asset pricing model predicts E(pt ) = E [m(datat+1 , parameters) xt+1 ] .

(134)

The most natural way to check this prediction is to examine sample averages, i.e. to calculate T T 1X 1X pt and [m(datat+1 , parameters) xt+1 ] . T t=1 T t=1

(135)

GMM estimates the parameters by making the sample averages as close to each other as possible. It seems natural, before evaluating a model, to pick parameters that give it its best chance. GMM then works out a distribution theory for the estimates. This distribution theory is a generalization of the simplest exercise in statistics: the distribution of the sample mean. Then, it suggests that we evaluate the model by looking at how close the sample averages of price and discounted payoff are to each other, or equivalently by looking at the pricing errors. It gives a statistical test of the hypothesis that the underlying population means are in fact zero.

10.1

The Recipe

Definitions ut+1 (b) ≡ mt+1 (b)xt+1 − pt gT (b) ≡ ET [ut (b)] ∞ X S ≡ E [ut (b) ut−j (b)0 ] j=−∞

GMM estimate ˆb2 = argminb gT (b)0 Sˆ−1 gT (b). Standard errors 1 ∂gT (b) var(ˆb2 ) = (d0 S −1 d)−1 ; d ≡ T ∂b 177

C HAPTER 10

GMM IN EXPLICIT DISCOUNT FACTOR MODELS

Test of the model (“overidentifying restrictions”) £ ¤ T JT = T min gT (b)0 S −1 gT (b) ∼ χ2 (#moments − #parameters). It’s easiest to start our discussion of GMM in the context of an explicit discount factor model, such as the consumption-based model. I treat the special structure of linear factor models later. I start with the basic classic recipe as given by Hansen and Singleton (1982). Discount factor models involve some unknown parameters as well as data, so I write mt+1 (b) when it’s important to remind ourselves of this dependence. For example, if mt+1 = β(ct+1 /ct )−γ , then b ≡ [β γ]0 . I write ˆb to denote an estimate when it is important to distinguish estimated from other values. Any asset pricing model implies E(pt ) = E [mt+1 (b)xt+1 ] .

(136)

It’s easiest to write this equation in the form E(·) = 0 E [mt+1 (b)xt+1 − pt ] = 0.

(137)

x and p are typically vectors; we typically check whether a model for m can price a number of assets simultaneously. Equations (10.137) are often called the moment conditions. It’s convenient to define the errors ut (b) as the object whose mean should be zero, ut+1 (b) = mt+1 (b)xt+1 − pt Given values for the parameters b, we could construct a time series on ut and look at its mean. Define gT (b) as the sample mean of the ut errors, when the parameter vector is b in a sample of size T : T 1X ut (b) = ET [ut (b)] = ET [mt+1 (b)xt+1 − pt ] . gT (b) ≡ T t=1

The second equality introduces the handy notation ET for sample means, T 1X (·). ET (·) = T t=1

ˆ and gˆ. However, Hansen’s T subscript (It might make more sense to denote these estimates E notation is so widespread that doing so would cause more confusion than it solves.) 178

S ECTION 10.1

T HE R ECIPE

The first stage estimate of b minimizes a quadratic form of the sample mean of the errors, ˆb1 = argmin ˆ gT (ˆb)0 W gT (ˆb) {b} for some arbitrary matrix W (often, W = I). This estimate is consistent and asymptotically normal. You can and often should stop here, as I explain below. Using ˆb1 , form an estimate Sˆ of ∞ X

S≡

E [ut (b) ut−j (b)0 ] .

(138)

j=−∞

(Below I discuss various interpretations of and ways to construct this estimate.) Form a second stage estimate ˆb2 using the matrix Sˆ in the quadratic form, ˆb2 = argmin gT (b)0 Sˆ−1 gT (b). b ˆb2 is a consistent, asymptotically normal, and asymptotically efficient estimate of the parameter vector b. “Efficient” means that it has the smallest variance-covariance matrix among all estimators that set different linear combinations of gT (b) to zero. The variance-covariance matrix of ˆb2 is 1 var(ˆb2 ) = (d0 S −1 d)−1 T where d≡

∂gT (b) ∂b

or, more explicitly, d = ET

µ

¶¯ ¯ ∂ [(mt+1 (b)xt+1 − pt )] ¯¯ ∂b b=ˆ b

(More precisely, d should be written as the object to which ∂gT /∂b converges, and then ∂gT /∂b is an estimate of that object used to form a consistent estimate of the asymptotic variance-covariance matrix.) This variance-covariance matrix can be used to test whether a parameter or group of parameters are equal to zero, via ˆb q i ∼ N (0, 1) var(ˆb)ii 179

C HAPTER 10

GMM IN EXPLICIT DISCOUNT FACTOR MODELS

and h i−1 ˆbj ∼ χ2 (#included b0 s) ˆbj var(ˆb)jj

where bj =subvector, var(b)jj =submatrix.

Finally, the test of overidentifying restrictions is a test of the overall fit of the model. It states that T times the minimized value of the second-stage objective is distributed χ2 with degrees of freedom equal to the number of moments less the number of estimated parameters. £ ¤ T JT = T min gT (b)0 S −1 gT (b) ∼ χ2 (#moments − #parameters). {b}

10.2

Interpreting the GMM procedure

gT (b) is a pricing error. It is proportional to α. GMM picks parameters to minimize a weighted sum of squared pricing errors. The second-stage picks the linear combination of pricing errors that are best measured, by having smallest sampling variation. First and second stage are like OLS and GLS regressions. The standard error formula is a simple application of the delta method. The JT test evaluates the model by looking at the sum of squared pricing errors.

Pricing errors The moment conditions are gT (b) = ET [mt+1 (b)xt+1 ] − ET [pt ] . Thus, each moment is the difference between actual (ET (p)) and predicted (ET (mx)) price, or pricing error. What could be more natural than to pick parameters so that the model’s predicted prices are as close as possible to the actual prices, and then to evaluate the model by how large these pricing errors are? In the language of expected returns, the moments gT (b) are proportional to the difference between actual and predicted returns; Jensen’s alphas, or the vertical distance between the points and the line in Figure 5. To see this fact, recall that 0 = E(mRe ) can be translated to a predicted expected return, E(Re ) = −

cov(m, Re ) . E(m)

180

S ECTION 10.2

I NTERPRETING THE GMM PROCEDURE

Therefore, we can write the pricing error as µ µ ¶¶ cov(m, Re ) e g(b) = E(mR ) = E(m) E(R ) − − E(m) 1 (actual mean return - predicted mean return.) g(b) = Rf e

If we express the model in expected return-beta language, E(Rei ) = αi + β 0i λ then the GMM objective is proportional to the Jensen’s alpha measure of mis-pricing, g(b) =

1 αi . Rf

First-stage estimates If we could, we’d pick b to make every element of gT (b) = 0 — to have the model price assets perfectly in sample. However, there are usually more moment conditions (returns times instruments) than there are parameters. There should be, because theories with as many free parameters as facts (moments) are vacuous. Thus, we choose b to make gT (b) as small as possible, by minimizing a quadratic form, min gT (b)0 W gT (b). {b}

(139)

W is a weighting matrix that tells us how much attention to pay to each moment, or how to trade off doing well in pricing one asset or linear combination of assets vs. doing well in pricing another. In the common case W = I, GMM treats all assets symmetrically, and the objective is to minimize the sum of squared pricing errors. The sample pricing error gT (b) may be a nonlinear function of b. Thus, you may have to use a numerical search to find the value of b that minimizes the objective in (10.139). However, since the objective is locally quadratic, the search is usually straightforward. Second-stage estimates: Why S −1 ? What weighting matrix should you use? The weighting matrix directs GMM to emphasize some moments or linear combinations of moments at the expense of others. You might start with W = I, i.e., try to price all assets equally well. A W that is not the identity matrix can be used to offset differences in units between the moments. You also might also start with different elements on the diagonal of W if you think some assets are more interesting, more informative, or better measured than others. The second-stage estimate picks a weighting matrix based on statistical considerations. 181

C HAPTER 10

GMM IN EXPLICIT DISCOUNT FACTOR MODELS

Some asset returns may have much more variance than other assets. For those assets, the sample mean gT = ET (mt Rt − 1) will be a much less accurate measurement of the population mean E(mR − 1), since the sample mean will vary more from sample to sample. Hence, it seems like a good idea to pay less attention to pricing errors from assets with high variance of mt Rt − 1. One could implement this idea by using a W matrix composed of inverse variances of ET (mt Rt − 1) on the diagonal. More generally, since asset returns are correlated, one might think of using the covariance matrix of ET (mt Rt −1). This weighting matrix pays most attention to linear combinations of moments about which the data set at hand has the most information. This idea is exactly the same as heteroskedasticity and cross-correlation corrections that lead you from OLS to GLS in linear regressions. The covariance matrix of gT = ET (ut+1 ) is the variance of a sample mean. Exploiting the assumption that E(ut ) = 0, and that ut is stationary so E(u1 u2 ) = E(ut ut+1 ) depends only on the time interval between the two us, we have

var(gT ) = var =

Ã

T 1 X ut+1 T t=1

!

¡ ¢ ¤ 1 £ T E(ut u0t ) + (T − 1) E(ut u0t−1 ) + E(ut u0t+1 )) + ... 2 T

As T → ∞, (T − j)/T → 1, so

var(gT ) →

∞ 1 X 1 E(ut u0t−j ) = S. T j=−∞ T

The last equality denotes S, known for other reasons as the spectral density matrix at frequency zero of ut . (Precisely, S so defined is the variance-covariance matrix of the gT for fixed b. The actual variance-covariance matrix of gT must take into account the fact that we chose b to set a linear combination of the gT to zero in each sample. I give that formula below. The point here is heuristic.) This fact suggests that a good weighting matrix might be the inverse of S. In fact, Hansen (1982) shows formally that the choice W = S −1 , S ≡

∞ X

E(ut u0t−j )

j=−∞

is the statistically optimal weighing matrix, meaning that it produces estimates with lowest asymptotic variance. √ You may be more used to the formula σ(u)/ T for the standard deviation of a sample mean. This formula is a special case that holds when the u0t s are uncorrelated over time. If 182

S ECTION 10.2

I NTERPRETING THE GMM PROCEDURE

Et (ut u0t−j ) = 0, j 6= 0, then the previous equation reduces to var

Ã

T 1X ut+1 T t=1

!

=

var(u) 1 E(uu0 ) = . T T

This is probably the first statistical formula you ever saw – the variance of the sample mean. In GMM, it is the last statistical formula you’ll ever see as well. GMM amounts to just generalizing the simple ideas behind the distribution of the sample mean to parameter estimation and general statistical contexts. The first and second stage estimates should remind you of standard linear regression models. You start with an OLS regression. If the errors are not i.i.d., the OLS estimates are consistent, but not efficient. If you want efficient estimates, you can use the OLS estimates to obtain a series of residuals, estimate a variance-covariance matrix of residuals, and then do GLS. GLS is also consistent and more efficient, meaning that the sampling variation in the estimated parameters is lower. Standard errors The formula for the standard error of the estimate, 1 var(ˆb2 ) = (d0 S −1 d)−1 T

(140)

can be understood most simply as an instance of the “delta method” that the asymptotic variance of f (x) is f 0 (x)2 var(x). Suppose there is only one parameter and one moment. S/T is the variance matrix of the moment gT . d−1 is [∂gT /∂b]−1 = ∂b/∂gT . Then the delta method formula gives 1 ∂b ∂b var(ˆb2 ) = var(gT ) . T ∂gT ∂gT The actual formula (10.140) just generalizes this idea to vectors. 10.2.1

JT Test

Once you’ve estimated the parameters that make a model “fit best,” the natural question is, how well does it fit? It’s natural to look at the pricing errors and see if they are “big.” The JT test asks whether they are “big” by statistical standards – if the model is true, how often should we see a (weighted) sum of squared pricing errors this big? If not often, the model is “rejected.” The test is h i T JT = T gT (ˆb)0 S −1 gT (ˆb) ∼ χ2 (#moments − #parameters).

Since S is the variance-covariance matrix of gT , this statistic is the minimized pricing errors 183

C HAPTER 10

GMM IN EXPLICIT DISCOUNT FACTOR MODELS

divided by their variance-covariance matrix. Sample means converge to a normal distribution, so sample means squared divided by variance converges to the square of a normal, or χ2 . The reduction in degrees of freedom corrects for the fact that S is really the covariance matrix of gT for fixed b. We set a linear combination of the gT to zero in each sample, so the actual covariance matrix of gT is singular, with rank #moments - #parameters. More details below.

10.3

Applying GMM

Notation. Forecast errors and instruments. Stationarity and choice of units.

Notation; instruments and returns Most of the effort involved with GMM is simply mapping a given problem into the very general notation. The equation E [mt+1 (b)xt+1 − pt ] = 0 can capture a lot. We often test asset pricing models using returns, in which case the moment conditions are E [mt+1 (b)Rt+1 − 1] = 0. It is common to add instruments as well. Mechanically, you can multiply both sides of 1 = Et [mt+1 (b)Rt+1 ] by any variable zt observed at time t before taking unconditional expectations, resulting in E(zt ) = E [mt+1 (b)Rt+1 zt ] . Expressing the result in E(·) = 0 form, 0 = E {[mt+1 (b)Rt+1 − 1] zt } .

(141)

We can do this for a whole vector of returns and instruments, multiplying each return by each instrument. For example, if we start with two returns R = [Ra Rb ]0 and one instrument z, 184

S ECTION 10.3

A PPLYING GMM

equation (10.141) looks like mt+1 (b) Rat+1 mt+1 (b) Rbt+1 E mt+1 (b) Rat+1 zt mt+1 (b) Rbt+1 zt

1 1 − = zt zt

0 0 . 0 0

Using the Kronecker product ⊗ meaning “multiply every element by every other element” we can denote the same relation compactly by E {[mt+1 (b) Rt+1 − 1] ⊗ zt } = 0,

(142)

or, emphasizing the managed-portfolio interpretation and p = E(mx) notation, E [mt+1 (b)(Rt+1 ⊗ zt ) − (1 ⊗ zt )] = 0. Forecast errors and instruments The asset pricing model says that, although expected returns can vary across time and assets, expected discounted returns should always be the same, 1. The error ut+1 = mt+1 Rt+1 − 1 is the ex-post discounted return. ut+1 = mt+1 Rt+1 − 1 represents a forecast error. Like any forecast error, ut+1 should be conditionally and unconditionally mean zero. In an econometric context, z is an instrument because it is uncorrelated with the error ut+1 . E(zt ut+1 ) is the numerator of a regression coefficient of ut+1 on zt ; thus adding instruments basically checks that the ex-post discounted return is unforecastable by linear regressions. If an asset’s return is higher than predicted when zt is unusually high, but not on average, scaling by zt will pick up this feature of the data. Then, the moment condition checks that the discount rate is unusually low at such times, or that the conditional covariance of the discount rate and asset return moves sufficiently to justify the high conditionally expected return. As I explained in Section 8.1, the addition of instruments is equivalent to adding the returns of managed portfolios to the analysis, and is in principle able to capture all of the model’s predictions. Stationarity and distributions The GMM distribution theory does require some statistical assumption. Hansen (1982) and Ogaki (1993) cover them in depth. The most important assumption is that m, p, and x must be stationary random variables. (“Stationary” of often misused to mean constant, or i.i.d.. The statistical definition of stationarity is that the joint distribution of xt , xt−j depends only on j and not on t.) Sample averages must converge to population means as the sample size grows, and stationarity implies this result. Assuring stationarity usually amounts to a choice of sensible units. For example, though 185

C HAPTER 10

GMM IN EXPLICIT DISCOUNT FACTOR MODELS

we could express the pricing of a stock as pt = Et [mt+1 (dt+1 + pt+1 )] it would not be wise to do so. For stocks, p and d rise over time and so are typically not stationary; their unconditional means are not defined. It is better to divide by pt and express the model as · ¸ dt+1 + pt+1 1 = Et mt+1 = Et (mt+1 Rt+1 ) pt The stock return is plausibly stationary. Dividing by dividends is an alternative and I think underutilized way to achieve stationarity (at least for portfolios, since many individual stocks do not pay regular dividends): · µ ¶ ¸ pt+1 dt+1 pt = Et mt+1 1 + . dt dt+1 dt ´ ³ dt+1 pt Now we map 1 + pdt+1 dt into xt+1 and dt into pt . This formulation allows us to focus t+1 on prices rather than one-period returns. Bonds are a claim to a dollar, so bond prices and yields do not grow over time. Hence, it might be all right to examine pbt = E(mt+1 1) with no transformations. Stationarity is not always a clear-cut question in practice. As variables become “less stationary,” as they experience longer swings in a sample, the asymptotic distribution can becomes a less reliable guide to a finite-sample distribution. For example, the level of nominal interest rates is surely a stationary variable in a fundamental sense: it was 6% in ancient Babylon, about 6% in 14th century Italy, and about 6% again today. Yet it takes very long swings away from this unconditional mean, moving slowly up or down for even 20 years at a time. Therefore, in an estimate and test that uses the level of interest rates, the asymptotic distribution theory might be a bad approximation to the correct finite sample distribution theory. This is true even if the number of data points is large. 10,000 data points measured every minute are a “smaller” data set than 100 data points measured every year. In such a case, it is particularly important to develop a finite-sample distribution by simulation or bootstrap, which is easy to do given today’s computing power. It is also important to choose test assets in a way that is stationary. For example, individual stocks change character over time, increasing or decreasing size, exposure to risk factors, leverage, and even nature of the business. For this reason, it is common to sort stocks into portfolios based on characteristics such as betas, size, book/market ratios, industry and so forth. The statistical characteristics of the portfolio returns may be much more constant than 186

S ECTION 10.3

A PPLYING GMM

the characteristics of individual securities, which float in and out of the various portfolios. (One can alternatively include the characteristics as instruments.) Many econometric techniques require assumptions about distributions. As you can see, the variance formulas used in GMM do not include the usual assumptions that variables are i.i.d., normally distributed, homoskedastic, etc. You can put such assumptions in if you want to – we’ll see how below, and adding such assumptions simplifies the formulas and can improve the small-sample performance when the assumptions are justified – but you don’t have to add these assumptions.

187

Chapter 11. GMM: general formulas and applications Lots of calculations beyond formal parameter estimation and overall model testing are useful in the process of evaluating a model and comparing it to other models. But you want to understand sampling variation in such calculations, and mapping the questions into the GMM framework allows you to do this easily. In addition, alternative estimation and evaluation procedures may be more intuitive or robust to model misspecification than the two (or multi) stage procedure described above. In this chapter I lay out the general GMM framework, and I discuss four applications and variations on the basic GMM method. 1) I show how to derive standard errors of nonlinear functions of sample moments, such as correlation coefficients. 2) I apply GMM to OLS regressions, easily deriving standard error formulas that correct for autocorrelation and conditional heteroskedasticity. 3) I show how to use prespecified weighting matrices W in asset pricing tests in order to overcome the tendency of efficient GMM to focus on spuriously lowvariance portfolios 4) As a good parable for prespecified linear combination of moments a, I show how to mimic “calibration” and “evaluation” phases of real business cycle models. 5) I show how to use the distribution theory for the gT beyond just forming the JT test in order to evaluate the importance of individual pricing errors. The next chapter continues, and collects GMM variations useful for evaluating linear factor models and related mean-variance frontier questions. Many of these calculations amount to creative choices of the aT matrix that selects which linear combination of moments are set to zero, and reading off the resulting formulas for variance covariance matrix of the estimated coefficients, equation (11.146) and variance covariance matrix of the moments gT , equation (11.147).

11.1

General GMM formulas

The general GMM estimate aT gT (ˆb) = 0 Distribution of ˆb : T cov(ˆb) = (ad)−1 aSa0 (ad)−10 Distribution of gT (ˆb) : h i ¡ ¢ ¡ ¢0 T cov gT (ˆb) = I − d(ad)−1 a S I − d(ad)−1 a 188

S ECTION 11.1

G ENERAL GMM FORMULAS

The “optimal” estimate uses a = d0 S −1 . In this case, T cov(ˆb) = (d0 S −1 d)−1 h i T cov gT (ˆb) = S − d(d0 S −1 d)−1 d0

and

T JT = T gT (ˆb)0 S −1 gT (ˆb) → χ2 (#moments − #parameters). An analogue to the likelihood ratio test, T JT (restricted) − T JT (unrestricted) ∼ χ2Number of restrictions GMM procedures can be used to implement a host of estimation and testing exercises. Just about anything you might want to estimate can be written as a special case of GMM. To do so, you just have to remember (or look up) a few very general formulas, and then map them into your case. Express a model as E[f (xt , b)] = 0 Everything is a vector: f can represent a vector of L sample moments, xt can be M data series, b can be N parameters. f(xt , b) is a slightly more explicit statement of the errors ut (b) in the last chapter Definition of the GMM estimate. We estimate parameters ˆb to set some linear combination of sample means of f to zero, ˆb : set aT gT (ˆb) = 0

(143)

where gT (b) ≡

T 1X f(xt , b) T t=1

and aT is a matrix that defines which linear combination of gT (b) will be set to zero. This defines the GMM estimate. If there are as many moments as parameters, you will set each moment to zero; when there are fewer parameters than moments, (11.143) just captures the natural idea that you will set some moments, or some linear combination of moments to zero in order to estimate the parameters. The minimization of the last chapter is a special case. If you estimate b by 189

C HAPTER 11

GMM: GENERAL FORMULAS AND APPLICATIONS

min gT0 (b)0 W gT (b), the first order conditions are ∂gT0 W gT (b) = 0, ∂b which is of the form (11.143) with aT = ∂gT0 /∂bW . The general GMM procedure allows you to pick arbitrary linear combinations of the moments to set to zero in parameter estimation. Standard errors of the estimate. Hansen (1982), Theorem 3.1 tells us that the asymptotic distribution of the GMM estimate is √ ¤ £ T (ˆb − b) → N 0, (ad)−1 aSa0 (ad)−10

where

d≡E

·

(144)

¸ ∂f ∂gT (b) (x , b) = t ∂b0 ∂b0

(i.e., d is defined as the population moment in the first equality, which we estimate in sample by the second equality), where a ≡ plim aT , and where S≡ Don’t forget the

√

∞ X

E [f(xt , b), f (xt−j b)0 ] .

(145)

j=−∞

T in (11.144)! In practical terms, this means to use 1 var(ˆb) = (ad)−1 aSa0 (ad)−10 T

(146)

as the covariance matrix for standard errors and tests. As in the last chapter, you can understand this formula as an application of the delta method. Distribution of the moments. Hansen’s Lemma 4.1 gives the sampling distribution of the moments gT (b) : h ¡ √ ¢ ¡ ¢0 i T gT (ˆb) → N 0, I − d(ad)−1 a S I − d(ad)−1 a .

(147)

As we have seen, S would be the asymptotic variance-covariance matrix of sample means, if we did not estimate any parameters, which sets some linear combinations of the gT to zero. The I − d(ad)−1 a terms account for the fact that in each sample some linear combinations of gT are set to zero. Thus, this variance-covariance matrix is singular. 190

S ECTION 11.1

G ENERAL GMM FORMULAS

χ2 tests. A sum of squared standard normals is distributed χ2 . Therefore, it is natural to use the distribution theory for gT to see if the gT are jointly “too big.” Equation (11.147) suggests that we form the statistic h¡ ¢ ¡ ¢0 i−1 T gT (ˆb)0 I − d(ad)−1 a S I − d(ad)−1 a gT (ˆb) (148)

and that it should have a χ2 distribution. It does, but with a hitch: The variance-covariance matrix is singular, soP you have to pseudo-invert it. For example, you can perform an eigenvalue decomposition = QΛQ0 and then invert only the non-zero eigenvalues. Also, the χ2 distribution has degrees of freedom given by the number non-zero linear combinations of gT , the number of moments less number of estimated parameters. You can similarly use (11.147) to construct tests of individual moments (“are the small stocks mispriced?”) or groups of moments. Efficient estimates

The theory so far allows us to estimate parameters by setting any linear combination of moments to zero. Hansen shows that one particular choice is statistically optimal, a = d0 S −1 .

(149)

This choice is the first order condition to min{b} gT (b)0 S −1 gT (b) that we studied in the last Chapter. With this weighting matrix, the standard error formula (11.146) reduces to √ ¤ £ (150) T (ˆb − b) → N 0, (d0 S −1 d)−1 .

This is Hansen’s Theorem 3.2. The sense in which (11.149) is “efficient” is that the sampling variation of the parameters for arbitrary a matrix, (11.146), equals the sampling variation of the “efficient” estimate in (11.150) plus a positive semidefinite matrix. With the optimal weights (11.149), the variance of the moments (11.147) simplifies to cov(gT ) =

¢ 1¡ S − d(d0 S −1 d)−1 d0 . T

(151)

We can use this matrix in a test of the form (11.148). However, Hansen’s Lemma 4.2 tells us that there is an equivalent and simpler way to construct this test, T gT (ˆb)0 S −1 gT (ˆb) → χ2 (#moments − #parameters).

(152)

This result is nice since we get to use the already-calculated and non-singular S −1 . To derive (11.152) from (11.147), factor S = CC 0 and then find the asymptotic covariance matrix of C −1 gT (ˆb) using (11.147). The result is i h√ var T C −1 gT (ˆb) = I − C −1 d(d0 S −1 d)−1 d0 C −10 . 191

C HAPTER 11

GMM: GENERAL FORMULAS AND APPLICATIONS

This is an idempotent matrix of rank #moments-#parameters, so (11.152) follows. Alternatively, note that S −1 is a pseudo-inverse of the second stage cov(gT ). (A pseudoinverse times cov(gT ) should result in an idempotent matrix of the same rank as cov(gT ).) ¡ ¢ S −1 cov(gT ) = S −1 S − d(d0 S −1 d)−1 d0 = I − S −1 d(d0 S −1 d)−1 d0

Then, check that the result is idempotent. ¡ ¢¡ ¢ I − S −1 d(d0 S −1 d)−1 d0 I − S −1 d(d0 S −1 d)−1 d0 = I − S −1 d(d0 S −1 d)−1 d0 .

This derivation not only verifies that JT has the same distribution as gT0 cov(gT )−1 gT , but that they are numerically the same in every sample. I emphasize that (11.150) and (11.152) only apply to the “optimal” choice of weights, (11.149). If you use another set of weights, as in a first-stage estimate, you must use the general formulas (11.146) and (11.147). Model comparisons You often want to compare one model to another. If one model can be expressed as a special or “restricted” case of the other or “unrestricted” model we can perform a statistical comparison that looks very much like a likelihood ratio test. If we use the same S matrix – usually that of the unrestricted model – the restricted JT must rise. But if the restricted model is really true, it shouldn’t rise “much.” How much? T JT (restricted) − T JT (unrestricted) ∼ χ2 (#of restrictions) This is a “χ2 difference” test, due to Newey and West (1987a), who call it the “D-test.”

11.2

Testing moments

How to test one or a group of pricing errors. 1) Use the formula for var(gT ) 2) A χ2 difference test. You may want to see how well a model does on particular moments or particular pricing errors. For example, the celebrated “small firm effect” states that an unconditional CAPM (m = a+ bRW , no scaled factors) does badly in pricing the returns on a portfolio that always holds the smallest 1/10th or 1/20th of firms in the NYSE. You might want to see whether a new model prices the small returns well. The standard error of pricing errors also allows you to add error bars to a plot of predicted vs. actual mean returns such as Figure 5 or other diagnostics based on pricing errors. We have already seen that individual elements of gT measure the pricing errors or expected return errors. Thus, the sampling variation of gT given by (11.147) provides exactly 192

S ECTION 11.3

S TANDARD ERRORS OF ANYTHING BY DELTA METHOD

the standard error we are looking for. You can use the sampling distribution of gT , to evaluate the significance of individual pricing errors, to construct a t-test (for a single gT , such as small firms) or χ2 test (for groups of gT , such as small firms ⊗ instruments). As usual this is the Wald test. Alternatively, you can use the χ2 difference approach. Start with a general model that includes all the moments, and form an estimate of the spectral density matrix S. Now set to zero the moments you want to test, and denote gsT (b) the vector of moments, including the zeros (s for “smaller”). Choose bs to minimize gsT (bs )0 S −1 gsT (bs ) using the same weighting matrix S. The criterion will be lower than the original criterion gT (b)0 S −1 gT (b), since there are the same number of parameters and fewer moments. But, if the moments we want to test truly are zero, the criterion shouldn’t be that much lower. The χ2 difference test applies, T gT (ˆb)0 S −1 gT (ˆb) − T gsT (ˆbs )S −1 gsT (ˆbs ) ∼ χ2 (#eliminated moments). Of course, don’t fall into the obvious trap of picking the largest of 10 pricing errors and noting it’s more than two standard deviations from zero. The distribution of the largest of 10 pricing errors is much wider than the distribution of a single one. To use this distribution, you have to pick which pricing error you’re going to test before you look at the data.

11.3

Standard errors of anything by delta method

One quick application illustrates the usefulness of the GMM formulas. Often, we want to estimate a quantity that is a nonlinear function of sample means, b = φ [E(xt )] = φ(µ). In this case, the formula (11.144) reduces to · ¸ · ¸0 X ∞ 1 dφ dφ cov(xt , x0t−j ) var(bT ) = . T dµ j=−∞ dµ

(153)

The formula is very intuitive. The variance of the sample mean is the covariance term inside. The derivatives just linearize the function φ near the true b. For example, a correlation coefficient can be written as a function of sample means as

Thus, take

E(xt yt ) − E(xt )E(yt ) p corr(xt , yt ) = p E(x2t ) − E(xt )2 E(yt2 ) − E(yt )2 µ=

£

E(xt ) E(x2t ) E(yt ) E(yt2 ) E(xt yt )

¤0

.

A problem at the end of the chapter asks you to take derivatives and derive the standard error of the correlation coefficient. One can derive standard errors for impulse-response functions, 193

C HAPTER 11

GMM: GENERAL FORMULAS AND APPLICATIONS

variance decompositions, and many other statistics in this way.

11.4

Using GMM for regressions

By mapping OLS regressions in to the GMM framework, we derive formulas for OLS standard errors that correct for autocorrelation and conditional heteroskedasticity of the errors. The general formula is ∞ X 1 ˆ = E(xt x0 )−1 E(ut xt x0t−j ut−j ) E(xt x0t )−1 . var(β) t T j=−∞ and it simplifies in special cases.

Mapping any statistical procedure into GMM makes it easy to develop an asymptotic distribution that corrects for statistical problems such as non-normality, serial correlation and conditional heteroskedasticity. To illustrate, as well as to develop the very useful formulas, I map OLS regressions into GMM. Correcting OLS standard errors for econometric problems is not the same thing as GLS. When errors do not obey the OLS assumptions, OLS is consistent, and often more robust than GLS, but its standard errors need to be corrected. OLS picks parameters β to minimize the variance of the residual: £ ¤ min ET (yt − β 0 xt )2 . {β}

ˆ from the first order condition, which states that the residual is orthogonal to the We find β right hand variable: h i ˆ =0 ˆ = ET xt (yt − x0 β) gT (β) (154) t This condition is exactly identified–the number of moments equals the number of parameters. Thus, we set the sample moments exactly to zero and there is no weighting matrix (a = I). We can solve for the estimate analytically, ˆ = [ET (xt x0t )]−1 ET (xt yt ). β This is the familiar OLS formula. The rest of the ingredients to equation (11.144) are d = E(xt x0t ) 194

S ECTION 11.4

U SING GMM FOR REGRESSIONS

f (xt , β) = xt (yt − x0t β) = xt et where et is the regression residual. Equation (11.144) gives a formula for OLS standard errors, ∞ X 1 ˆ = E(xt x0 )−1 var(β) E(ut xt x0t−j ut−j ) E(xt x0t )−1 . (155) t T j=−∞ This formula reduces to some interesting special cases. Serially uncorrelated, homoskedastic errors These are the usual OLS assumptions, and it’s good the usual formulas emerge. Formally, the OLS assumptions are E(et | xt , xt−1 ...et−1 , et−2 ...) = 0

(156)

E(e2t | xt , xt−1 ...et , et−1 ...) = constant = σ2e .

(157)

To use these assumptions, I use the fact that E(ab) = E(E(a|b)b). The first assumption means that only the j = 0 term enters the sum ∞ X

E(et xt x0t−j et−j ) = E(e2t xt x0t ).

j=−∞

The second assumption means that E(e2t xt x0t ) = E(e2t )E(xt x0t ) = σ2e E(xt x0t ). Hence equation (11.155) reduces to our old friend, 1 2 −1 σe E(xt x0t )−1 = σ 2e (X 0 X) . T £ The last notation is typical of econometrics texts, in which X = x1 resents the data matrix. ˆ = var(β)

x2

... xT

¤0

rep-

Heteroskedastic errors If we delete the conditional homoskedasticity assumption (11.157), we can’t pull the u out 195

C HAPTER 11

GMM: GENERAL FORMULAS AND APPLICATIONS

of the expectation, so the standard errors are ˆ = var(β)

1 E(xt x0t )−1 E(u2t xt x0t )E(xt x0t )−1 . T

These are known as “Heteroskedasticity consistent standard errors” or “White standard errors” after White (1980). Hansen-Hodrick errors Hansen and Hodrick (1982) run forecasting regressions of (say) six month returns, using monthly data. We can write this situation in regression notation as yt+k = β 0 xt + εt+k t = 1, 2, ...T. Fama and French (1988) also use regressions of overlapping long horizon returns on variables such as dividend/price ratio and term premium. Such regressions are an important part of the evidence for predictability in asset returns. Under the null that one-period returns are unforecastable, we will still see correlation in the εt due to overlapping data. Unforecastable returns imply E(εt εt−j ) = 0 for |j| ≥ k but not for |j| < k. Therefore, we can only rule out terms in S lower than k. Since we might as well correct for potential heteroskedasticity while we’re at it, the standard errors are k X 1 E(ut xt x0t−j ut−j ) E(xt x0t )−1 . var(bT ) = E(xt x0t )−1 T j=−k

11.5

Prespecified weighting matrices and moment conditions

Prespecified rather than “optimal” weighting matrices can emphasize economically interesting results, they can avoid the trap of blowing up standard errors rather than improving pricing errors, they can lead to estimates that are more robust to small model misspecifications. This is analogous to the fact that OLS is often preferable to GLS in a regression context. The GMM formulas for a fixed weighting matrix W are 1 var(ˆb) = (d0 W d)−1 d0 W SW d(d0 W d)−1 T var(gT ) =

1 (I − d(d0 W d)−1 d0 W )S(I − W d(d0 W d)−1 d0 ). T 196

S ECTION 11.5

P RESPECIFIED WEIGHTING MATRICES AND MOMENT CONDITIONS

In the basic approach outlined in Chapter 10, our final estimates were based on the “efficient” S −1 weighting matrix. This objective maximizes the asymptotic statistical information in the sample about a model, given the choice of moments gT . However, you may want to use a prespecified weighting matrix W 6= S −1 instead, or at least as a diagnostic accompanying more formal statistical tests. A prespecified weighting matrix lets you, rather than the S matrix, specify which moments or linear combination of moments GMM will value in the minimization min{b} gT (b)0 W gT (b). A higher value of Wii forces GMM to pay more attention to getting the ith moment right in the parameter estimation. For example, you might feel that some assets suffer from measurement error, are small and illiquid and hence should be deemphasized, or you may want to keep GMM from looking at portfolios with strong long and short position. I give some additional motivations below. You can also go one step further and impose which linear combinations aT of moment conditions will be set to zero in estimation rather than use the choice resulting from a minimization, aT = d0 S −1 or aT = d0 W . The fixed W estimate still trades off the accuracy of individual moments according £ to the¤0 sensitivity of each moment with respect to the parameter. For example, if gT = gT1 gT2 , W = I, but ∂gT /∂b = [1 10], so that the second moment is 10 times more sensitive to the parameter value than the first moment, then GMM with fixed weighting matrix sets 1 × gT1 + 10 × gT2 = 0. The second moment condition will be 10 times closer to zero than the first. If you really want GMM to pay equal attention to the two moments, then you can fix the aT matrix directly, for example aT = [1 1] or aT = [1 − 1].

Using a prespecified weighting matrix or using a prespecified set of moments is not the same thing as ignoring correlation of the errors ut in the distribution theory. The S matrix will still show up in all the standard errors and test statistics. 11.5.1

How to use prespecified weighting matrices

Once you have decided to use a prespecified weighting matrix W or a prespecified set of moments aT gT (b) = 0, the general distribution theory outlined in section 11.1 quickly gives standard errors of the estimates and moments, and therefore a χ2 statistic that can be used to test whether all the moments are jointly zero. Section 11.1 gives the formulas for the case that aT is prespecified. If we use weighting matrix W , the first order conditions to min{b} gT0 (b)W gT (b) are ∂gT (b)0 W gT (b) = d0 W gT (b) = 0, ∂b so we map into the general case with aT = d0 W. Plugging this value into (11.146), the 197

C HAPTER 11

GMM: GENERAL FORMULAS AND APPLICATIONS

variance-covariance matrix of the estimated coefficients is 1 var(ˆb) = (d0 W d)−1 d0 W SW d(d0 W d)−1 . T

(158)

(You can check that this formula reduces to 1/T (d0 S −1 d)−1 with W = S −1 .) Plugging a = d0 W into equation (11.147), we find the variance-covariance matrix of the moments gT var(gT ) =

1 (I − d(d0 W d)−1 d0 W )S(I − W d(d0 W d)−1 d0 ) T

(159)

As in the general formula, the terms to the left and right of S account for the fact that some linear combinations of moments are set to zero in each sample. Equation (11.159) can be the basis of χ2 tests for the overidentifying restrictions. If we interpret ()−1 to be a generalized inverse, then gT0 var(gT )−1 gT ∼ χ2 (#moments − #parameters). As in the general case, you have to pseudo-invert the singular var(gT ), for example by inverting only the non-zero eigenvalues. The major danger in using prespecified weighting matrices or moments aT is that the choice of moments, units, and (of course) the prespecified aT or W must be made carefully. For example, if you multiply the second moment by 10 times its original value, the S matrix will undo this transformation and weight them in their original proportions. The identity weighting matrix will not undo such transformations, so the units should be picked right initially. 11.5.2

Motivations for prespecified weighting matrices

Robustness, as with OLS vs. GLS. When errors are autocorrelated or heteroskedastic, every econometrics textbook shows you how to “improve” on OLS by making appropriate GLS corrections. If you correctly model the error covariance matrix and if the regression is perfectly specified, the GLS procedure can improve efficiency, i.e. give estimates with lower asymptotic standard errors. However, GLS is less robust. If you model the error covariance matrix incorrectly, the GLS estimates can be much worse than OLS. Also, the GLS transformations can zero in on slightly misspecified areas of the model, producing garbage. GLS is “best,” but OLS is “pretty darn good.” One often has enough data that wringing every last ounce of statistical precision (low standard errors) from the data is less important than producing estimates that do not depend on questionable statistical assumptions, and that transparently focus on the interesting features of the data. In these cases, it is often a good idea to use OLS estimates. The OLS standard error formulas are wrong, though, so you must correct the standard errors of the 198

S ECTION 11.5

P RESPECIFIED WEIGHTING MATRICES AND MOMENT CONDITIONS

OLS estimates for these features of the error covariance matrices, using the formulas we developed in section 11.4. GMM works the same way. First-stage or otherwise fixed weighting matrix estimates may give up something in asymptotic efficiency, but they are still consistent, and they can be more robust to statistical and economic problems. You still want to use the S matrix in computing standard errors, though, as you want to correct OLS standard errors, and the GMM formulas show you how to do this. Even if in the end you want to produce “efficient” estimates and tests, it is a good idea to calculate standard errors and model fit tests for the first-stage estimates. Ideally, the parameter estimates should not change by much, and the second stage standard errors should be tighter. If the “efficient” parameter estimates do change a great deal, it is a good idea to diagnose why this is so. It must come down to the “efficient” parameter estimates strongly weighting moments or linear combinations of moments that were not important in the first stage, and that the former linear combination of moments disagrees strongly with the latter about which parameters fit well. Then, you can decide whether the difference in results is truly due to efficiency gain, or whether it signals a model misspecification. Chapter 16 argues more at length for judicious use of “inefficient” methods such as OLS to guard against inevitable model misspecifications. Near-singular S. The spectral density matrix is often nearly singular, since asset returns are highly correlated with each other, and since we often include many assets relative to the number of data points. As a result, second stage GMM (and, as we will see below, maximum likelihood or any other efficient technique) tries to minimize differences and differences of differences of asset returns in order to extract statistically orthogonal components with lowest variance. One may feel that this feature leads GMM to place a lot of weight on poorly estimated, economically uninteresting, or otherwise non-robust aspects of the data. In particular, portfolios of the form 100R1 − 99R2 assume that investors can in fact purchase such heavily leveraged portfolios. Short-sale costs often rule out such portfolios or significantly alter their returns, so one may not want to emphasize pricing them correctly in the estimation and evaluation. For example, suppose that S is given by S=

·

1 ρ ρ 1

¸

.

so S −1 =

1 1 − ρ2

·

1 −ρ −ρ 1

¸

.

We can factor S −1 into a “square root” by the Choleski decomposition. This produces a 199

C HAPTER 11

GMM: GENERAL FORMULAS AND APPLICATIONS

triangular matrix C such that C 0 C = S −1 . You can check that the matrix " # √1 √−ρ 2 2 1−ρ 1−ρ C= 0 1

(160)

works. Then, the GMM criterion min gT0 S −1 gT is equivalent to min(gT0 C 0 )(CgT ). CgT gives the linear combination of moments that efficient GMM is trying to minimize. Looking at (11.160), as ρ → 1, the (2,2) element stays at 1, but the (1,1) and (1,2) elements get very large and of opposite signs. For example, if ρ = 0.95, then · ¸ 3.20 −3.04 C= . 0 1 In this example, GMM pays a little attention to the second moment, but places three times as much weight on the difference between the first and second moments. Larger matrices produce even more extreme weights. At a minimum, it is a good idea to look at S −1 and its Choleski decomposition to see what moments GMM is prizing. The same point has a classic interpretation, and is a well-known danger with classic regression-based tests. Efficient GMM wants to focus on well-measured moments. In asset pricing applications, the errors are typically close to uncorrelated over time, so GMM is looking for portfolios with small values of var(mt+1 Ret+1 ). Roughly speaking, those will be asset with small return variance. Thus, GMM will pay most attention to correctly pricing the sample minimum-variance portfolio, and GMM’s evaluation of the model by JT test will focus on its ability to price this portfolio. Now, consider what happens in a sample, as illustrated in Figure 24. The sample meanvariance frontier is typically a good deal wider than the true, or ex-ante mean-variance frontier. In particular, the sample minimum-variance portfolio may have little to do with the true minimum-variance portfolio. Like any portfolio on the sample frontier, its composition largely reflects luck – that’s why we have asset pricing models in the first place rather than just price assets with portfolios on the sample frontier. The sample minimum variance return is also likely to be composed of strong long-short positions. In sum, you may want to force GMM not to pay quite so much attention to correctly pricing the sample minimum variance portfolio, and you may want to give less importance to a statistical measure of model evaluation that almost entirely prizes GMM’s ability to price that portfolio. Economically interesting moments. 200

S ECTION 11.5

P RESPECIFIED WEIGHTING MATRICES AND MOMENT CONDITIONS

Sample minimum-variance portfolio E(R)

Sample, ex-post frontier

True, ex-ante frontier

σ(R)

Figure 24. True or ex ante and sample or ex-post mean-variance frontier. The sample often shows a spurious minimum-variance portfolio.

201

C HAPTER 11

GMM: GENERAL FORMULAS AND APPLICATIONS

The optimal weighting matrix makes GMM pay close attention to linear combinations of moments with small sampling error in both estimation and evaluation. One may want to force the estimation and evaluation to pay attention to economically interesting moments instead. The initial portfolios are usually formed on an economically interesting characteristic such as size, beta, book/market or industry. One typically wants in the end to see how well the model prices these initial portfolios, not how well the model prices potentially strange portfolios of those portfolios. If a model fails, one may want to characterize that failure as “the model doesn’t price small stocks” not “the model doesn’t price a portfolio of 900× small firm returns −600× large firm returns −299× medium firm returns.”

Level playing field.

The S matrix changes as the model and as its parameters change. (See the definition, (10.138) or (11.145).) As the S matrix changes, which assets the GMM estimate tries hard to price well changes as well. For example, the S matrix from one model may value strongly pricing the T bill well, while that of another model may value pricing a stock excess return well. Comparing the results of such estimations is like comparing apples and oranges. By fixing the weighting matrix, you can force GMM to pay attention to the various assets in the same proportion while you vary the model. The fact that S matrices change with the model leads to another subtle trap. One model my may “improve” a JT = gT0 S −1 gT statistic because it blows up the estimates of S, rather than making any progress on lowering the pricing errors gT . No one would formally use a comparison of JT tests across models to compare them, of course. But it has proved nearly irresistible for authors to claim success for a new model over previous ones by noting improved JT statistics, despite different weighting matrices, different moments, and sometimes much larger pricing errors. For example, if you take a model mt and create a new model by simply adding noise, unrelated to asset returns (in sample), m0t = mt + εt , then the moment condition gT = ETh(m0t Rte ) = ET ((mit + εt ) Rte ) is unchanged. However, the spectral den2 sity matrix S = E (mt + εt ) Rte Rte0 can rise dramatically. This can reduce the JT leading to a false sense of “improvement.” Conversely, if the sample contains a nearly riskfree portfolio of the test assets, or a portfolio with apparently small variance of mt+1 Ret+1 , then the JT test essentially evaluates the model by how will it can price this one portfolio. This can lead to a false rejection – even a very small gT will produce a large gT0 S −1 gT if there is an eigenvalue of S that is (spuriously) too small. If you use a common weighting matrix W for all models, and evaluate the models by gT0 W gT , then you can avoid this trap. Beware that the individual χ2 statistics are based on gT0 var(gT )−1 gT , and var(gT ) contains S, even with a prespecified weighting matrix W . You should look at the pricing errors, or at some statistic such as the sum of absolute or squared pricing errors to see if they are bigger or smaller, leaving the distribution aside. The question “are the pricing errors small?” is as interesting as the question “if we drew artificial data over and over again from a null statistical model, how often would we estimate a ratio 202

S ECTION 11.5

P RESPECIFIED WEIGHTING MATRICES AND MOMENT CONDITIONS

of pricing errors to their estimated variance gT0 S −1 gT this big or larger?” 11.5.3

Some prespecified weighting matrices

Two examples of economically interesting weighting matrices are the second-moment matrix of returns, advocated by Hansen and Jagannathan (1997) and the simple identity matrix, which is used implicitly in much empirical asset pricing. Second moment matrix. Hansen and Jagannathan (1997) advocate the use of the second moment matrix of payoffs W = E(xx0 )−1 in place of S. They motivate this weighting matrix as an interesting distance measure between a model for m, say y, and the space of true m’s. Precisely, the minimum distance (second moment) between a candidate discount factor y and the space of true discount factors is the same as the minimum value of the GMM criterion with W = E(xx0 )−1 as weighting matrix.

X proj(y| X)

m x*

y Nearest m

Figure 25. Distance between y and nearest m = distance between proj(y|X) and x∗ . To see why this is true, refer to Figure 25. The distance between y and the nearest valid m is the same as the distance between proj(y | X) and x∗ . As usual, consider the case that X is generated from a vector of payoffs x with price p. From the OLS formula, proj(y | X) = E(yx0 )E(xx0 )−1 x. 203

C HAPTER 11

GMM: GENERAL FORMULAS AND APPLICATIONS

x∗ is the portfolio of x that prices x by construction, x∗ = p0 E(xx0 )−1 x. Then, the distance between y and the nearest valid m is: ky − nearest mk = kproj(y|X) − x∗ k ° ° = °E(yx0 )E(xx0 )−1 x − p0 E(xx0 )−1 x° ° ° = °(E(yx0 ) − p0 ) E(xx0 )−1 x° = [E(yx) − p]0 E(xx0 )−1 [E(yx) − p] = gT0 E(xx0 )−1 gT

You might want to choose parameters of the model to minimize this “economic” measure of model fit, or this economically motivated linear combination of pricing errors, rather than the statistical measure of fit S −1 . You might also use the minimized value of this criterion to compare two models. In that way, you are sure the better model is better because it improves on the pricing errors rather than just blowing up the weighting matrix. Identity matrix. Using the identity matrix weights the initial choice of assets or portfolios equally in estimation and evaluation. This choice has a particular advantage with large systems in which S is nearly singular, as it avoids most of the problems associated with inverting a near-singular S matrix. Many empirical asset pricing studies use OLS cross-sectional regressions, which are the same thing as a first stage GMM estimate with an identity weighting matrix. Comparing the second moment and identity matrices. The second moment matrix gives an objective that is invariant to the initial choice of assets or portfolios. If we form a portfolio Ax of the initial payoffs x, with nonsingular A (i.e. a transformation that doesn’t throw away information) then [E(yAx) − Ap]0 E(Axx0 A0 )−1 [E(yAx) − Ap] = [E(yx) − p]0 E(xx0 )−1 [E(yx) − p]. The optimal weighting matrix S shares this property. It is not true of the identity or other fixed matrices. In those cases, the results will depend on the initial choice of portfolios. Kandel and Stambaugh (1995) have suggested that the results of several important asset pricing model tests are highly sensitive to the choice of portfolio; i.e. that authors inadvertently selected a set of portfolios on which the CAPM does unusually badly in a particular sample. Insisting that weighting matrices have this kind of invariance to portfolio selection might be a good device to ward against this problem. On the other hand, if you want to focus on the model’s predictions for economically interesting portfolios, then it wouldn’t make much sense for the weighting matrix to undo the specification of economically interesting portfolios! For example, many studies want 204

S ECTION 11.6

E STIMATING ON ONE GROUP OF MOMENTS , TESTING ON ANOTHER .

to focus on the ability of a model to describe expected returns that seem to depend on a characteristic such as size, book/market, industry, momentum, etc. Also, the second moment matrix is often even more nearly singular than the spectral density matrix, since E(xx0 ) = cov(x)+E(x)E(x)0 . Therefore, it often emphasizes portfolios with even more extreme short and long positions, and is no help on overcoming the near singularity of the S matrix.

11.6

Estimating on one group of moments, testing on another.

You may want to force the system to use one set of moments for estimation and another for testing. The real business cycle literature in macroeconomics does this extensively, typically using “first moments” for estimation (“calibration”) and “second moments” (i.e. first moments of squares) for evaluation. A statistically minded macroeconomist might like to know whether the departures of model from data “second moments” are large compared to sampling variation, and would like to include sampling uncertainty about the parameter estimates in this evaluation. You might want to choose parameters using one set of asset returns (stocks; domestic assets; size portfolios, first 9 size deciles, well-measured assets) and then see how the model does “out of sample” on another set of assets (bonds; foreign assets; book/market portfolios, small firm portfolio, questionably measured assets, mutual funds). However, you want the distribution theory for evaluation on the second set of moments to incorporate sampling uncertainty about the parameters in their estimation on the first set of moments. You can do all this very simply by using an appropriate weighting matrix or a prespecified moment matrix aT . For example, if the first N moments will be used to estimate N parameters, and the remaining M moments will be used to test the model “out of sample,” use aT = [IN 0N×M ] . If there are more moments N than parameters in the “estimation” block, you can construct a weighting matrix W which is an identity matrix in the N × N estimation block and zero elsewhere. Then aT = ∂gT0 /∂bW will simply contain the first N columns of ∂gT0 /∂b followed by zeros. The test moments will not be used in estimation. You could even use the inverse of the upper N × N block of S (not the upper block of the inverse of S!) to make the estimation a bit more efficient.

11.7

Estimating the spectral density matrix

Hints on estimating the spectral density or long run covariance matrix. 1) Use a sensible first stage estimate 2) Remove means 3) Downweight higher order correlations 4) Consider parametric structures for autocorrelation and heteroskedasticity 5) Use the null to limit the number of correlations or to impose other structure on S? 6) Size problems; consider a factor or other parametric cross-sectional structure for S. 7) Iteration and simultaneous b, S estimation. 205

C HAPTER 11

GMM: GENERAL FORMULAS AND APPLICATIONS

The optimal weighting matrix S depends on population moments, and depends on the parameters b. Work back through the definitions, ∞ X

E(ut u0t−j );

S

=

ut

≡ (mt (b)xt − pt−1 )

j=−∞

How do we estimate this matrix? The big picture is simple: following the usual philosophy, estimate population moments by their sample counterparts. Thus, use the first stage b estimates and the data to construct sample versions of the definition of S. This produces a consistent estimate of the true spectral density matrix, which is all the asymptotic distribution theory requires. The details are important however, and this section gives some hints. Also, you may want a different, and less restrictive, estimate of S for use in standard errors than you do when you are estimating S for use in a weighting matrix. 1) Use a sensible first stage W, or transform the data. In the asymptotic theory, you can use consistent first stage b estimates formed by any nontrivial weighting matrix. In practice, of course, you should use a sensible weighting matrix so that the first stage estimates are not ridiculously inefficient. W = I is often a good choice. Sometimes, some moments will have different units than other moments. For example, the dividend/price ratio is a number like 0.04. Therefore, the moment formed by Rt+1 × d/pt will be about 0.04 as big as large as the moment formed by Rt+1 × 1. If you use W = I, GMM will pay much less attention to the Rt+1 × d/pt moment. It is wise, then, to either use an initial weighting matrix that overweights the Rt+1 × d/pt moment, or to transform the data so the two moments are about the same mean and variance. For example, you could use Rt+1 × (1 + d/pt ). It is also useful to start with moments that are not horrendously correlated with each other, or to remove such correlation with a clever W . For example, you might consider Ra and Rb − Ra rather than Ra and Rb . You can accomplish this directly, or by starting with

W =

·

1 −1 0 1

¸·

1 0 −1 1

¸

=

·

2 −1 −1 1

¸

.

2) Remove means. Under the null, E(ut ) = 0, so it does not matter to the asymptotic distribution theory 206

S ECTION 11.7

E STIMATING THE SPECTRAL DENSITY MATRIX

whether you estimate the covariance matrix by removing means, using T T 1X 1X 0 [(ut − u ¯)(ut − u ¯) ] ; u ¯≡ ut T t=1 T t=1

or whether you estimate the second moment matrix by not removing means. However, Hansen and Singleton (1982) advocate removing the means in sample, and this is generally a good idea. It is already a major obstacle to second-stage estimation that estimated S matrices (and even simple variance-covariance matrices) are often nearly singular, providing an unreliable weighting matrix when inverted. Since second moment matrices E(uu0 ) = cov(u, u0 ) + E(u)E(u0 ) add a singular matrix E(u)E(u0 ) they are often even worse. 3) Downweight higher order correlations. You obviously cannot use a direct sample counterpart to the spectral density matrix. In a sample of size 100, there is no way to estimate E(ut u0t+101 ).Your estimate of E(ut u0t+99 ) is based on one data point, u1 u0100 . Hence, it will be a pretty unreliable estimate. For this reason, the estimator using all possible autocorrelations in a given sample is inconsistent. (Consistency means that as the sample grows, the probability distribution of the estimator converges to the true value. Inconsistent estimates typically have very large sample variation.) Furthermore, even S estimates that use few autocorrelations are not always positive definite in sample. This is embarrassing when one tries to invert the estimated spectral density matrix, which you have to do if you use it as a weighting matrix. Therefore, it is a good idea to construct consistent estimates that are automatically positive definite in every sample. One such estimate is the Bartlett estimate, used in this application by Newey and West (1987b). It is ¶ k µ T X k − |j| 1X (ut u0t−j ). (161) Sˆ = k T t=1 j=−k

As you can see, only autocorrelations up to kth (k < T ) order are included, and higher order autocorrelations are downweighted. (It’s important to use 1/T not 1/(T − k); this is a further downweighting.) The Newey-West estimator is basically the variance of kth sums, which is why it is positive definite in sample: k X ut−j = kE(ut u0t ) + (k − 1)[E(ut u0t−1 ) + E(ut−1 u0t )] + · · · V ar j=1

+[E(ut u0t−k ) + E(ut−k u0t )] = k

k X k − |j| E(ut u0t−k ). k

j=−k

Andrews (1991) gives some additional weighting schemes for spectral density estimates. 207

C HAPTER 11

GMM: GENERAL FORMULAS AND APPLICATIONS

This calculation also gives some intuition the S matrix. We’re looking for the variance Pfor T across samples of the sample mean var( T1 t=1 ut ). We only have one sample mean to look at, so we estimate the variance the sample mean by looking at the variance in a single ´ ³ of Pk sample of shorter sums, var k1 j=1 uj . The S matrix is sometimes called the long-run covariance matrix for this reason. In fact, one could estimate S directly as a variance of kth sums and obtain almost the same estimator, that would also be positive definite in any sample, vt

=

k X

ut−j ; v¯ =

j=1

Sˆ =

1 1 kT −k

T X 1 vt T −k t=k+1

T X

t=k+1

(vt − v¯) (vt − v¯)0 .

This estimator has been used when measurement of S is directly interesting (Cochrane 1998, Lo and MacKinlay 1988). A variety of other weighting schemes have been advocated. What value of k, or how wide a window if of another shape, should you use? Here again, you have to use some judgment. Too short values of k, together with a ut that is significantly autocorrelated, and you don’t correct for correlation that might be there in the errors. Too long a value of k, together with a series that does not have much autocorrelation, and the performance of the estimate and test deteriorates. If k = T /2 for example, you are really using only two data points to estimate the variance of the mean. The optimum value then depends on how much persistence or low-frequency movement there is in a particular application, vs. accuracy of the estimate. There is an extensive statistical literature about optimal window width, or size of k. Alas, this literature mostly characterizes the rate at which k should increase with sample size. You must promise to increase k as sample size increases, but not as quickly as the sample size increases – limT →∞ k = ∞, limT →∞ k/T = 0 – in order to obtain consistent estimates. In practice, promises about what you’d do with more data are pretty meaningless, and usually broken once more data arrives. 4) Consider parametric structures for autocorrelation and heteroskedasticity. “Nonparametric” corrections such as (11.161) often don’t perform very well in typical samples. The problem is that “nonparametric” techniques are really very highly parametric; you have to estimate many correlations in the data. Therefore, the nonparametric estimate varies a good deal from sample to sample, while the asymptotic distribution theory ignores sampling variation in covariance matrix estimates. The asymptotic distribution can therefore be a poor approximation to the finite-sample distribution of statistics like the JT . The S −1 weighting matrix will also be unreliable. One answer is to use a Monte-Carlo or bootstrap to estimate the finite-sample distribution parameters and test statistics rather than to rely on asymptotic theory. Alternatively, you can impose a parametric structure on the S matrix. Just because the 208

S ECTION 11.7

E STIMATING THE SPECTRAL DENSITY MATRIX

formulas are expressed in terms of a sum of covariances does not mean you have to estimate them that way; GMM is not inherently tied to “nonparametric” covariance matrix estimates. For example, if you model a scalar u as an AR(1) with parameter ρ, then you can estimate two numbers ρ and σ 2u rather than a whole list of autocorrelations, and calculate S=

∞ X

E(ut ut−j ) = σ2u

j=−∞

∞ X

ρ|j| = σ 2u

j=−∞

1+ρ 1−ρ

If this structure is not a bad approximation, imposing it can result in more reliable estimates and test statistics since one has to estimate many fewer coefficients. You could transform the data in such a way that there is less correlation to correct for in the first place. (This is a very useful formula, by the way. You are probably used to calculating the standard error of the mean as σ(x) σ(¯ x) = √ . T This formula assumes that the x are uncorrelated over time. If an AR(1) is not a bad model for their correlation, you can quickly adjust for correlation by using r σ(x) 1 + ρ √ σ(¯ x) = 1−ρ T instead.) This sort of parametric correction is very familiar from OLS regression analysis. The textbooks commonly advocate the AR(1) model for serial correlation as well as parametric models for heteroskedasticity corrections. There is no reason not to follow a similar approach for GMM statistics. 5) Use the null to limit correlations? In the typical asset pricing setup, the null hypothesis specifies that Et (ut+1 ) = Et (mt+1 Rt+1 − 1) = 0, as well as E(ut+1 ) = 0. This implies that all the autocorrelation terms of S drop out; E(ut u0t−j ) = 0 for j 6= 0. The lagged u could be an instrument z; the discounted return should be unforecastable, using past discounted returns as well as any other variable. In this situation, one could exploit the null to only include one term, and estimate T 1X ut u0t . Sˆ = T t=1

Similarly, if one runs a regression forecasting returns from some variable zt , Rt+1 = a + bzt + εt+1 , the null hypothesis that returns are not forecastable by any variable at time t means that the 209

C HAPTER 11

GMM: GENERAL FORMULAS AND APPLICATIONS

errors should not be autocorrelated. One can then simplify the standard errors in the OLS regression formulas given in section 11.4, eliminating all the leads and lags. In other situations, the null hypothesis can suggest a functional form for E(ut u0t−j ) or that some but not all are zero. For example, as we saw in section 11.4, regressions of long horizon returns on overlapping data lead to a correlated error term, even under the null hypothesis of no return forecastability. We can impose this null, ruling out terms past the overlap, as suggested by Hansen and Hodrick, k X 1 E(et xt x0t−j et−j ) E(xt x0t )−1 . var(bT ) = E(xt x0t )−1 T

(162)

j=−k

However, the null might not be correct, and the errors might be correlated. If so, you might make a mistake by leaving them out. If the null is correct, the extra terms will converge to zero and you will only have lost a few (finite-sample) degrees of freedom needlessly estimating them. If the null is not correct, you have an inconsistent estimate. With this in mind, you might want to include at least a few extra autocorrelations, even when the null says they don’t belong. Furthermore, there is no guarantee that the unweighted sum in (11.162) is positive definite in sample. If the sum in the middle is not positive definite, you could add a weighting to the sum, possibly increasing the number of lags so that the lags near k are not unusually underweighted. Again, estimating extra lags that should be zero under the null only loses a little bit of power. Monte Carlo evidence (Hodrick 1992) suggests that imposing the null hypothesis to simplify the spectral density matrix helps to get the finite-sample size of test statistics right – the probability of rejection given the null is true. One should not be surprised that if the null is true, imposing as much of it as possible makes estimates and tests work better. On the other hand, adding extra correlations can help with the power of test statistics – the probability of rejection given that an alternative is true – since they converge to the correct spectral density matrix. This trade-off requires some thought. For measurement rather than pure testing, using a spectral density matrix that can accommodate alternatives may be the right choice. For example, in the return forecasting regressions, one is really focused on measuring return forecastability rather than just formally testing the hypothesis that it is zero. On the other hand, the small-sample performance of the nonparametric estimators with many lags is not very good. If you are testing an asset pricing model that predicts u should not be autocorrelated, and there is a lot of correlation – if this issue makes a big difference – then this is an indication that something is wrong with the model; that including u as one of your instruments z would result in a rejection or at least substantially change the results. If the u are close to uncorrelated, then it really doesn’t matter if you add a few extra terms or not. 210

S ECTION 11.7

E STIMATING THE SPECTRAL DENSITY MATRIX

6) Size problems; consider a factor or other parametric cross-sectional structure. If you try to estimate a covariance matrix that is larger than the number of data points (say 2000 NYSE stocks and 800 monthly observations), the estimate of S, like any other covariance matrix, is singular by construction. This fact leads to obvious problems when you try to invert S! More generally, when the number of moments is more than around 1/10 the number of data points, S estimates tend to become unstable and near-singular. Used as a weighting matrix, such an S matrix tells you to pay lots of attention to strange and probably spurious linear combinations of the moments, as I emphasized in section 11.5. For this reason, most second-stage GMM estimations are limited to a few assets and a few instruments. A good, but as yet untried alternative might be to impose a factor structure or other wellbehaved structure on the covariance matrix. The near-universal practice of grouping assets into portfolios before analysis already implies an assumption that the true S of the underlying assets has a factor structure. Grouping in portfolios means that the individual assets have no information not contained in the portfolio, so that a weighting matrix S −1 would treat all assets in the portfolio identically. It might be better to estimate an S imposing a factor structure on all the primitive assets. Another response to the difficulty of estimating S is to stop at first stage estimates, and only use S for standard errors. One might also use a highly structured estimate of S as weighting matrix, while using a less constrained estimate for the standard errors. This problem is of course not unique to GMM. Any estimation technique requires us to calculate a covariance matrix. Many traditional estimates simply assume that ut errors are cross-sectionally independent. This false assumption leads to understatements of the standard errors far worse than the small sample performance of any GMM estimate. Our econometric techniques all are designed for large time series and small cross-sections. Our data has a large cross section and short time series. A large unsolved problem in finance is the development of appropriate large-N small-T tools for evaluating asset pricing models. 7) Alternatives to the two-stage procedure: iteration and one-step. Hansen and Singleton (1982) describe the above two-step procedure, and it has become popular for that reason. Two alternative procedures may perform better in practice, i.e. may result in asymptotically equivalent estimates with better small-sample properties. They can also be simpler to implement, and require less manual adjustment or care in specifying the setup (moments, weighting matrices) which is often just as important. a) Iterate. The second stage estimate ˆb2 will not imply the same spectral density as the first stage. It might seem appropriate that the estimate of b and of the spectral density should be consistent, i.e. to find a fixed point of ˆb = min{b} [gT (b)0 S(ˆb)−1 gT (b)]. One way to search for such a fixed point is to iterate: find b2 from ˆb2 = min gT (b)0 S −1 (b1 )gT (b) {b}

(163)

where b1 is a first stage estimate, held fixed in the minimization over b2 . Then use ˆb2 to find 211

C HAPTER 11

GMM: GENERAL FORMULAS AND APPLICATIONS

S(ˆb2 ), find ˆb3 = min[gT (b)0 S(ˆb2 )−1 gT (b)], {b}

and so on. There is no fixed point theorem that such iterations will converge, but they often do, especially with a little massaging. (I once used S [(bj + bj−1 )/2] in the beginning part of an iteration to keep it from oscillating between two values of b). Ferson and Foerster (1994) find that iteration gives better small sample performance than two-stage GMM in Monte Carlo experiments. This procedure is also likely to produce estimates that do not depend on the initial weighting matrix. b) Pick b and S simultaneously. It is not true that S must be held fixed as one searches for b. Instead, one can use a new S(b) for each value of b. Explicitly, one can estimate b by min [gT (b)0 S −1 (b)gT (b)] {b}

(164)

The estimates produced by this simultaneous search will not be numerically the same in a finite sample as the two-step or iterated estimates. The first order conditions to (11.163) are ¶0 µ ∂gT (b) S −1 (b1 )gT (b) = 0 (165) ∂b while the first order conditions in (11.164) add a term involving the derivatives of S(b) with respect to b. However, the latter terms vanish asymptotically, so the asymptotic distribution theory is not affected. Hansen, Heaton and Yaron (1996) conduct some Monte Carlo experiments and find that this estimate may have small-sample advantages in certain problems. A problem is that the one-step minimization may find regions of the parameter space that blow up the spectral density matrix S(b) rather than lower the pricing errors gT . Often, one choice will be much more convenient than another. For linear models, one can find the minimizing value of b from the first order conditions (11.165) analytically. This fact eliminates the need to search so even an iterated estimate is much faster. For nonlinear 0 models, each step involves a numerical search over gT (b) SgT (b). Rather than perform this 0 search many times, it may be much quicker to minimize once over gT (b) S(b)gT (b). On the other hand, the latter is not a locally quadratic form, so the search may run into greater numerical difficulties.

11.8 1. 2.

Problems

Use the delta method version of the GMM formulas to derive the sampling variance of an autocorrelation coefficient. Write a formula for the standard error of OLS regression coefficients that corrects for autocorrelation but not heteroskedasticity 212

S ECTION 11.8 3. 4.

P ROBLEMS

Write a formula for the standard error of OLS regression coefficients if E(et et−j ) = ρj σ2 . If the GMM errors come from an asset pricing model, ut = mt Rt − 1, can you ignore lags in the spectral density matrix? What if you know that returns are predictable? What if the error is formed from an instrument/managed portfolio ut zt−1 ?

213

Chapter 12. Regression-based tests of linear factor models This and the next three chapters study the question, how should we estimate and evaluate linear factor models; models of the form p = E(mx), m = b0 f or equivalently E(Re ) = βλ? These models are by far the most common in empirical asset pricing, and there is a large literature on econometric techniques to estimate and evaluate them. Each technique focuses on the same questions: how to estimate parameters, how to calculate standard errors of the estimated parameters, how to calculate standard errors of the pricing errors, and how to test ˆ. the model, usually with a test statistic of the form α ˆ 0 V −1 α I start with simple and longstanding time-series and cross-sectional regression tests. Then, I pursue GMM approach to the model expressed in p = E(mx), m = b0 f form. The following chapter summarizes the principle of maximum likelihood estimation and derives maximum likelihood estimates and tests. Finally, a chapter compares the different approaches. As always, the theme is the underlying unity. All of the techniques come down to one of two basic ideas: time-series regression or cross-sectional regression. The GMM, p = E(mx) approach turns out to be almost identical to cross-sectional regressions. Maximum likelihood (with appropriate statistical assumptions) justifies the time-series and cross-sectional regression approaches. The formulas for parameter estimates, standard errors, and test statistics are all strikingly similar.

12.1

Time-series regressions

When the factor is also a return, we can evaluate the model E(Rei ) = β i E(f) by running OLS time series regressions i Rei t = αi + β i ft + εt ; t = 1, 2, ...T

for each asset. The OLS distribution formulas (with corrected standard errors) provide standard errors of α and β. With errors that are i.i.d. over time, homoskedastic and independent of the factors, the asymptotic joint distribution of the intercepts gives the model test statistic, "

T 1+

µ

ET (f) σ ˆ (f )

¶2 #−1 214

ˆ −1 α α ˆ 0Σ ˆ ∼χ2N

S ECTION 12.1

T IME - SERIES REGRESSIONS

The Gibbons-Ross-Shanken test is a multivariate, finite sample counterpart to this statistic, when the errors are also normally distributed, ´−1 T −N −K ³ ˆ −1 ET (f ) ˆ −1 α 1 + ET (f)0 Ω α ˆ 0Σ ˆ ∼FN,T −N−K . N

I show how to construct the same test statistics with heteroskedastic and autocorrelated errors via GMM. I start with the simplest case. We have a factor pricing model with a single factor. The factor is an excess return (for example, the CAPM, with Rem = Rm − Rf ), and the test assets are all excess returns. We express the model in expected return - beta form. The betas are defined by regression coefficients Rtei = αi + β i ft + εit

(166)

and the model states that expected returns are linear in the betas: E(Rei ) = β i E(f ).

(167)

Since the factor is also an excess return, the model applies to the factor as well, so E(f ) = 1 × λ.

Comparing the model (12.167) and the expectation of the time series regression (12.166) we see that the model has one and only one implication for the data: all the regression intercepts αi should be zero. The regression intercepts are equal to the pricing errors.

Given this fact, Black Jensen and Scholes (1972) suggested a natural strategy for estimation and evaluation: Run time-series regressions (12.166) for each test asset. The estimate of the factor risk premium is just the sample mean of the factor, ˆ = ET (f ). λ Then, use standard OLS formulas for a distribution theory of the parameters. In particular you can use t-tests to check whether the pricing errors α are in fact zero. These distributions are usually presented for the case that the regression errors in (12.166) are uncorrelated and homoskedastic, but the formulas in section 11.4 show easily how to calculate standard errors for arbitrary error covariance structures. We also want to know whether all the pricing errors are jointly equal to zero. This requires us to go beyond standard formulas for the regression (12.166) taken alone, as we want to know the joint distribution of α estimates from separate regressions running side by side but with errors correlated across assets (E(εit εjt ) 6= 0). (We can think of 12.166 as a panel regression, and then it’s a test whether the firm dummies are jointly zero.) The classic form of these tests assume no autocorrelation or heteroskedasticity, but allow the errors to be correlated across assets. Dividing the α ˆ regression coefficients by their variance-covariance 215

C HAPTER 12

R EGRESSION - BASED TESTS OF LINEAR FACTOR MODELS

matrix leads to a χ2 test, "

T 1+

µ

ET (f) σ ˆ (f )

¶2 #−1

ˆ −1 α α ˆ 0Σ ˆ ∼χ2N

(168)

where ET (f ) denotes sample mean, σ ˆ 2 (f ) denotes sample variance, α ˆ is a vector of the estimated intercepts, ¤0 £ ˆ 2 ... α ˆN ˆ1 α α ˆ= α ˆ is the residual covariance matrix, i.e. the sample estimate of E(εt ε0t ) = Σ, where Σ ¤0 £ . εt = ε1t ε2t · · · εN t

As usual when testing hypotheses about regression coefficients, this test is valid asymptotically. The asymptotic distribution theory assumes that σ2 (f) (i.e. X 0 X) and Σ have converged to their probability limits; therefore it is asymptotically valid even though the factor is stochastic and Σ is estimated, but it ignores those sources of variation in a finite sample. It does not require that the errors are normal, relying on the central limit theorem so that α ˆ is normal. I derive (12.168) below. Also as usual in a regression context, we can derive a finite-sample F distribution for the hypothesis that a set of parameters are jointly zero, for fixed values of the right hand variable ft ,. " µ ¶2 #−1 T −N −1 ET (f) ˆ −1 α α ˆ 0Σ ˆ ∼FN,T −N−1 (169) 1+ N σ ˆ (f ) This is the Gibbons Ross and Shanken (1989) or “GRS” test statistic. The F distribution ˆ which is not included in (12.168). This distribution recognizes sampling variation in Σ, requires that the errors ε are normal as well as uncorrelated and homoskedastic. With normal ˆ is an independent Wishart (the multivariate version of a χ2 ), errors, the α ˆ are normal and Σ so the ratio is F . This distribution is exact in a finite sample. Tests (12.168) and (12.169) have a very intuitive form. The basic part of the test is a ˆ −1 α ˆ . If there were no βf in the model, then the α ˆ quadratic form in the pricing errors, α ˆ 0Σ would simply be the sample mean of the regression errors εt . Assuming i.i.d. εt , the variance ˆ would be a sum of their sample mean is just 1/T Σ. Thus, if we knew Σ then T α ˆ 0 Σ−1 α of squared sample means divided by their variance-covariance matrix, which would have an asymptotic χ2N distribution, or a finite sample χ2N distribution if the εt are normal. But we have to estimate Σ, which is why the finite-sample distribution is F rather than χ2 . We also estimate the β, and the second term in (12.168) and (12.169) accounts for that fact. Recall that a single beta representation exists if and only if the reference return is on the mean-variance frontier. Thus, the test can also be interpreted as a test whether f is ex216

S ECTION 12.1

T IME - SERIES REGRESSIONS

ante mean-variance efficient – whether it is on the mean-variance frontier using population moments – after accounting for sampling error. Even if f is on the true or ex-ante meanvariance frontier, other returns will outperform it in sample due to luck, so the return f will usually be inside the ex-post mean-variance frontier – i.e. the frontier drawn using sample moments. Still, it should not be too far inside the sample frontier. Gibbons Ross and Shanken show that the test statistic can be expressed in terms of how far inside the ex-post frontier the return f is, ´2 ³ ´2 ³ µq ET (f ) − σ ˆ (f ) T − N − 1 σq (170) ´2 . ³ N ET (f ) 1 + σˆ (f ) ³

µq σq

´2

is the Sharpe ratio of the ex-post tangency portfolio (maximum ex-post Sharpe ratio) formed from the test assets plus the factor f . If there are many factors that are excess returns, the same ideas work, with some cost of algebraic complexity. The regression equation is Rei = αi + β 0i ft + εit . The asset pricing model E(Rei ) = β 0i E(f) again predicts that the intercepts should be zero. We can estimate α and β with OLS timeˆ −1 α ˆ has the distriseries regressions. Assuming normal i.i.d. errors, the quadratic form α ˆ0Σ bution, ´−1 T −N −K ³ ˆ −1 ET (f) ˆ −1 α 1 + ET (f )0 Ω α ˆ0Σ ˆ ∼FN,T −N−K (171) N where

N K

= =

ˆ = Ω

Number of assets Number of factors T 1X [ft − ET (f)] [ft − ET (f)]0 T t=1

The main difference is that the Sharpe ratio of the single factor is replaced by the natural 0 ˆ −1 ET (f). generalization ET (f) Ω 12.1.1

Derivation of the χ2 statistic and distributions with general errors.

I derive (12.168) as an instance of GMM. This approach allows us to generate straightforwardly the required corrections for autocorrelated and heteroskedastic disturbances. (MacKin217

C HAPTER 12

R EGRESSION - BASED TESTS OF LINEAR FACTOR MODELS

lay and Richardson (1991) advocate GMM approaches to regression tests in this way.) It also serves to remind us that GMM and p = E(mx) are not necessarily paired; one can do a GMM estimate of an expected return - beta model too. The mechanics are only slightly different than what we did to generate distributions for OLS regression coefficients in section 11.4, since we keep track of N OLS regressions simultaneously. Write the equations for all N assets together in vector form, Ret = α + βft + εt . We use the usual OLS moments to estimate the coefficients, · ¸ µ· ¸¶ ET (Rte − α − βft ) εt = ET =0 gT (b) = ft εt ET [(Rte − α − βft ) ft ] These moments exactly identify the parameters α, β, so the a matrix in agT (ˆb) = 0 is the identity matrix. Solving, the GMM estimates are of course the OLS estimates, ˆ T (ft ) α ˆ = ET (Rte ) − βE e e e ˆ = ET [(Rt − ET (Rt )) ft ] = covT (Rt , ft ) . β ET [(ft − ET (ft )) ft ] varT (ft ) The d matrix in the general GMM formula is · ¸ · ¸ ∂gT (b) IN 1 E(ft ) IN E(ft ) = − d≡ = − ⊗ IN IN E(ft ) IN E(ft2 ) E(ft ) E(ft2 ) ∂b0 where IN is an N × N identity matrix. The S matrix is

¸ ∞ · X E(εt ε0t−j ) E(εt ε0t−j ft−j ) S= . E(ft εt ε0t−j ) E(ft εt ε0t−j ft−j ) j=−∞

Using the GMM variance formula (11.146) with a = I we have µ· ¸¶ α ˆ 1 var = d−1 Sd−10 . ˆ β T

(172)

At this point, we’re done. The upper left hand corner of var(α β) gives us var(ˆ α) and the −1 test we’re looking for is α ˆ 0 var(ˆ α) α ˆ ∼ χ2N .

The standard formulas make this expression prettier by assuming that the errors are uncorrelated over time and not heteroskedastic to simplify the S matrix, as we derived the standard OLS formulas in section 11.4. If we assume that f and ε are independent as well as orthogonal, E(fεε0 ) = E(f )E(εε0 ) and E(f 2 εε0 ) = E(f 2 )E(εε0 ). If we assume that the errors are independent over time as well, we lose all the lead and lag terms. Then, the S matrix 218

S ECTION 12.2

C ROSS - SECTIONAL REGRESSIONS

simplifies to S=

·

E(εt ε0t ) E(εt ε0t )E(ft ) 0 E(ft )E(εt εt ) E(εt ε0t )E(ft2 )

¸

=

·

1 E(ft ) E(ft ) E(ft2 )

¸

⊗Σ

(173)

Now we can plug into (12.172). Using (A⊗B)−1 = A−1 ⊗B −1 and (A⊗B)(C ⊗D) = AC ⊗ BD, we obtain var

µ·

α ˆ ˆ β

¸¶

1 = T

Ã·

1 E(ft ) E(ft ) E(ft2 )

¸−1

!

⊗Σ .

Evaluating the inverse, var

µ·

α ˆ ˆ β

¸¶

1 1 = T var(f)

·

E(ft2 ) −E(ft ) 1 −E(ft )

¸

⊗Σ

We’re interested in the top left corner. Using E(f 2 ) = E(f )2 + var(f), 1 var (ˆ α) = T

µ ¶ E(f )2 1+ Σ. var(f )

This is the traditional formula (12.168), but there is now no real reason to assume that the errors are i.i.d. or independent of the factors. By simply calculating 12.172, we can easily construct standard errors and test statistics that do not require these assumptions.

12.2

Cross-sectional regressions

We can fit E(Rei ) = β 0i λ + αi by running a cross-sectional regression of average returns on the betas. This technique can be used whether the factor is a return or not. I discuss OLS and GLS cross-sectional regressions, I find formulas for the standard errors of λ, and a χ2 test whether the α are jointly zero. I derive the distributions as an instance of GMM, and I show how to implement the same approach for autocorrelated and heteroskedastic errors. I show that the GLS cross-sectional regression is the same as the time-series regression when the factor is also an excess return, and is included in the set of test assets.

219

C HAPTER 12

R EGRESSION - BASED TESTS OF LINEAR FACTOR MODELS

E ( Rei )

αi

Assets i

Slope λ

βi

Figure 26. Cross-sectional regression Start again with the K factor model, written as E(Rei ) = β 0i λ; i = 1, 2, ...N

The central economic question is why average returns vary across assets; expected returns of an asset should be high if that asset has high betas or risk exposure to factors that carry high risk premia. Figure 26 graphs the case of a single factor such as the CAPM. Each dot represents one asset i. The model says that average returns should be proportional to betas, so plot the sample average returns against the betas. Even if the model is true, this plot will not work out perfectly in each sample, so there will be some spread as shown. Given these facts, a natural idea is to run a cross-sectional regression to fit a line through the scatterplot of Figure 26. First find estimates of the betas from a time series regression, 0 i Rei t = ai + β i ft + εt , t = 1, 2, ...T for each i.

(174)

Then estimate the factor risk premia λ from a regression across assets of average returns on the betas, ET (Rei ) = β 0i λ + αi , i = 1, 2....N.

220

(175)

S ECTION 12.2

CROSS - SECTIONAL REGRESSIONS

As in the figure, β are the right hand variables, λ are the regression coefficients, and the cross-sectional regression residuals αi are the pricing errors. This is also known as a twopass regression estimate, because one estimates first time-series and then cross-sectional regressions. You can run the cross-sectional regression with or without a constant. The theory says that the constant or zero-beta excess return should be zero. You can impose this restriction or estimate a constant and see if it turns out to be small. The usual tradeoff between efficiency (impose the null as much as possible to get efficient estimates) and robustness applies. 12.2.1

OLS cross-sectional regression

It will simplify notation to consider a single factor; the case of multiple factors looks the same with vectors in place of scalars. I denote from 1 to N with¤ missing sub or ¤ vectors £ 1 2 £ 0 N 0 superscripts, i.e. εt = εt εt · · · εt , β = β 1 β 2 · · · β N , and similarly for Ret and α. For simplicity take the case of no intercept in the cross-sectional regression. With this notation OLS cross-sectional estimates are ¡ ¢ ˆ = β 0 β −1 β 0 ET (Re ) λ (12.176) e ˆ α ˆ = ET (R ) − λβ. Next, we need a distribution theory for the estimated parameters. The most natural place to start is with the standard OLS distribution formulas. I start with the traditional assumption that the true errors are i.i.d. over time, and independent of the factors. This will give us some easily interpretable formulas, and we will see most of these terms remain when we do the distribution theory right later on. In an OLS regression Y = Xβ + u and E(uu0 ) = Ω, the standard error of the β estimate is (X 0 X)−1 X 0 ΩX(X 0 X)−1 . The residual covariance matrix is (I − X(X 0 X)−1 X 0 )Ω(I − X(X 0 X)−1 X 0 )0 Denote Σ = E (εt ε0t ). Since the αi are just time series averages of the true εit shocks (the average of the sample residuals is always zero), the errors in the cross-sectional regression have covariance matrix E (αα0 ) = T1 Σ. Thus the conventional OLS formulas for the covariance matrix of OLS estimates and residual with correlated errors give ³ ´ ¡ ¢−1 1 ¡ 0 ¢−1 0 ˆ = ββ σ2 λ β Σβ β 0 β (12.177) T ¡ ¢−1 0 ´ ³ ¡ ¢−1 0 ´ 1³ I − β β0β cov(ˆ α) = β Σ I − β β0β β (12.178) T We could test whether all pricing errors are zero with the statistic α ˆ 0 cov(ˆ α)−1 α ˆ ∼χ2N−1 .

221

(179)

C HAPTER 12

R EGRESSION - BASED TESTS OF LINEAR FACTOR MODELS

The distribution is χ2N−1 not χ2N because the covariance matrix is singular. The singularity and the extra terms in (12.178) result from the fact that the λ coefficient was estimated along the way, and means that we have to use a generalized inverse. (If there are K factors, we obviously end up with χ2N−K .) A test of the residuals is unusual in OLS regressions. We do not usually test whether the residuals are “too large,” since we have no information other than the residuals themselves about how large they should be. In this case, however, the first stage time-series regression gives us some independent information about the size of cov(αα0 ), information that we could not get from looking at the cross-sectional residual α itself. 12.2.2

GLS cross-sectional regression

Since the residuals in the cross-sectional regression (12.175) are correlated with each other, standard textbook advice is to run a GLS cross-sectional regression rather than OLS, using E(αα0 ) = T1 Σ as the error covariance matrix: ¡ 0 −1 ¢−1 0 −1 βΣ β β Σ ET (Re ) ˆ α ˆ = ET (Re ) − λβ. ˆ = λ

The standard regression formulas give the variance of these estimates as ³ ´ 1 ¡ 0 −1 ¢−1 ˆ = βΣ β σ2 λ T ³ ¡ ¢−1 0 ´ 1 Σ − β β 0 Σ−1 β cov(ˆ α) = β T

(12.180)

(12.181) (12.182)

The comments of section 11.5 warning that OLS is sometimes much more robust than GLS apply in this case. The GLS regression should improve efficiency, i.e. give more precise estimates. However, Σ may be hard to estimate and to invert, especially if the cross-section N is large. One may well choose the robustness of OLS over the asymptotic statistical advantages of GLS. A GLS regression can be understood as a transformation of the space of returns, to focus attention on the statistically most informative portfolios. Finding (say, by Choleski decomposition) a matrix C such that CC 0 = Σ−1 , the GLS regression is the same as an OLS regression of CET (Re ) on Cβ , i.e. of testing the model on the portfolios CRe . The statistically most informative portfolios are those with the lowest residual variance Σ. But this asymptotic statistical theory assumes that the covariance matrix has converged to its true value. In most samples, the ex-post or sample mean-variance frontier still seems to indicate lots of luck, and this is especially true if the cross section is large, anything more than 1/10 of the time series. The portfolios CRe are likely to contain many extreme long-short positions. Again, we could test the hypothesis that all the α are equal to zero with (12.179). Though the appearance of the statistic is the same, the covariance matrix is smaller, reflecting the 222

S ECTION 12.2

CROSS - SECTIONAL REGRESSIONS

greater power of the GLS test. As with the JT test, (11.152) we can develop an equivalent test that does not require a generalized inverse; Tα ˆ 0 Σ−1 α ˆ ∼χ2N−1 .

(183)

To derive (12.183), I proceed exactly as in the derivation of the JT test (11.152). Define, say 0 −1 by Choleski √ decomposition, a matrix C such that CC = Σ . Now, find the covariance ˆ: matrix of T C 0 α ³ √ ¡ ¢−1 0 ¡ ¢−1 0 ´ −1 cov( T Cα) = C 0 (CC 0 ) − β β 0 CC 0 β β C = I − δ δ0 δ δ

where

δ = C 0 β. √ √ In sum, α ˆ is asymptotically normal so T C 0 α ˆ is asymptotically normal, cov( T C 0 α ˆ ) is an ˆ 0 CC 0 α ˆ = Tα ˆ 0 Σ−1 α ˆ is χ2N−1 . idempotent matrix with rank N − 1; therefore T α 12.2.3 Correction for the fact that β are estimated, and GMM formulas that don’t need i.i.d. errors.

In applying standard OLS formulas to a cross-sectional regression, we assume that the right hand variables β are fixed. The β in the cross-sectional regression are not fixed, of course, but are estimated in the time series regression. This turns out to matter, even as T → ∞.

In this section, I derive the correct asymptotic standard errors. With the simplifying assumption that the errors ε are i.i.d. over time and independent of the factors, the result is i h ´ ¡ ¢ ³ ˆ OLS ) = 1 (β 0 β)−1 β 0 Σβ β 0 β −1 1 + λ0 Σ−1 λ + Σf σ 2 (λ (12.184) f T i h¡ ´ ¢ ³ ˆ GLS ) = 1 β 0 Σ−1 β −1 1 + λ0 Σ−1 λ + Σf σ 2 (λ f T

where Σf is the variance-covariance matrix of the factors. This correction is due to Shanken (1992). Comparing these³ standard errors ´ to (12.177) and (12.181), we see that there is a 0 −1 multiplicative correction 1 + λ Σf λ and an additive correction Σf . The asymptotic variance-covariance matrix of the pricing errors is ´ ¢³ ¡ ¢−1 0 ´ ¡ 1³ IN − β β 0 β cov(ˆ αOLS ) = β Σ IN − β(β 0 β)−1 β 0 1 + λ0 Σ−1 (12.185) λ f T ´ ¡ 0 −1 ¢−1 0 ´ ³ 1³ Σ−β β Σ β 1 + λ0 Σ−1 cov(ˆ αGLS ) = β (12.186) f λ T

Comparing these results to (12.178) and (12.182) we see the same multiplicative correction applies. 223

C HAPTER 12

R EGRESSION - BASED TESTS OF LINEAR FACTOR MODELS

We can form the asymptotic χ2 test of the pricing errors by dividing pricing errors by their ˆ cov(ˆ α)−1 α ˆ . Following (12.183), we can simplify this result for variance-covariance matrix, α the GLS pricing errors resulting in ³ ´ T 1 + λ0 Σ−1 λ α ˆ 0GLS Σ−1 α ˆ GLS ∼ χ2N−K . (187) f Are the corrections important relative to the simple OLS formulas given above? In the CAPM λ = E(Rem ) so λ2 /σ2 (Rem ) ≈ (0.08/0.16)2 = 0.25 in annual data. In annual data, then, the multiplicative term is too large to ignore. However, the mean and variance both scale with horizon, so the Sharpe ration scales with the square root of horizon. Therefore, for a monthly interval λ2 /σ 2 (Rem ) ≈ 0.25/12 ≈ 0.02 which is quite small and ignoring the multiplicative term makes little difference. ˆ can be very important. Consider a one factor The additive term in the standard error of λ model, suppose all the β are 1.0, all the residuals are uncorrelated so Σ is diagonal, suppose all assets have the same residual covariance σ2 (ε), and ignore the multiplicative term. Now we can write either covariance matrix in (12.184) as ¸ · 1 1 2 2 ˆ 2 σ (ε) + σ (f) σ (λ) = T N Even with N = 1, most factor models have fairly high R2 , so σ2 (ε) < σ2 (f ). Typical CAPM values of R2 = 1 − σ2 (ε)/σ2 (f ) for large portfolios are 0.6-0.7; and multifactor models such as the Fama French 3 factor model have R2 often over 0.9. Typical numbers of assets N = 10 to 50 make the first term vanish compared to the second term. More generally, suppose the factor were in fact a return. Then the factor risk premium is λ = E(f), and we’d use Σf /T as the standard error of λ. This is the “correction” term in (12.184), so we expect it to be, in fact, the most important term. Note that Σf /T is the standard error of the mean of f . Thus, in the case that the return is a factor, so E(f ) = λ, this is the only term you would use. This example suggests that Σf is not just an important correction, it is likely to be the ˆ. dominant consideration in the sampling error of the λ Comparing (12.187) to the GRS tests for a time-series regression, (12.168), (12.169), (12.171) we see the same statistic. The only difference is that by estimating λ from the cross-section rather than imposing λ = E(f ), the cross-sectional regression loses degrees of freedom equal to the number of factors. Though these formulas are standard classics, I emphasize that we don’t have to make the severe assumptions on the error terms that are used to derive them. As with the time-series ˆ and α ˆ , and only at the last moment case, I derive a general formula for the distribution of λ make classic error term assumptions to make the spectral density matrix pretty. Derivation and formulas that don’t require i.i.d. errors.

224

S ECTION 12.2

CROSS - SECTIONAL REGRESSIONS

The easy and elegant way to account for the effects of “generated regressors” such as the β in the cross-sectional regression is to map the whole thing into GMM. Then, we treat the moments that generate the regressors β at the same time as the moments that generate the cross-sectional regression coefficient λ, and the covariance matrix S between the two sets of moments captures the effects of generating the regressors on the standard error of the cross-sectional regression coefficients. Comparing this straightforward derivation with the ˆ and noting difficulty of Shanken’s (1992) paper that originally derived the corrections for λ, that Shanken did not go on to find the formulas (12.185) that allow a test of the pricing errors is a nice argument for the simplicity and power of the GMM framework. To keep the algebra manageable, I treat the case of a single factor. The moments are 0 E(Rte − a − βft ) gT (b) = E [(Rte − a − βft )ft ] = 0 0 E (Re − βλ)

(188)

The top two moment conditions exactly identify a and β as the time-series OLS estimates. (Note a not α. The time-series intercept is not necessarily equal to the pricing error in a cross-sectional regression.) The bottom moment condition is the asset pricing model. It is in general overidentified in a sample, since there is only one extra parameter (λ) and N extra moment conditions. If we use a weighting vector β 0 on this condition, we obtain the OLS cross-sectional estimate of λ. If we use a weighting vector β 0 Σ−1 , we obtain the GLS crosssectional estimate of λ. To accommodate both cases, use a weighting vector γ 0 , and then substitute γ 0 = β 0 , γ 0 = β 0 Σ−1 , etc. at the end. ˆ come straight from the general GMM standard error formula The standard errors for λ ˆ are not parameters, but are the last N moments. Their covariance matrix is (11.146). The α thus given by the GMM formula (11.147) for the sample variation of the gT . All we have to do is map the problem into the GMM notation. The parameter vector is b0 =

£

a0

β0

λ

¤

The a matrix chooses which moment conditions are set to zero in estimation, a=

·

I2N 0

0 γ0

¸

.

The d matrix is the sensitivity of the moment conditions to the parameters, −IN −IN E(f ) 0 ∂gT d= = −IN E(f ) −IN E(f 2 ) 0 ∂b0 −β 0 −λIN

225

C HAPTER 12

R EGRESSION - BASED TESTS OF LINEAR FACTOR MODELS

The S matrix is the long-run covariance matrix of the moments.

S

0 e Rt−j − a − βft−j Ret − a − βft e − a − βft−j )ft−j = E (Ret − a − βft )ft (Rt−j e e j=−∞ Rt − βλ Rt−j − βλ 0 ∞ εt εt−j X εt ft εt−j ft−j = E j=−∞ β(ft − Ef ) + εt β(ft−j − Ef) + εt−j ∞ X

In the second expression, I have used the regression model and the restriction under the null that E (Rte ) = βλ. In calculations, of course, you could simply estimate the first expression. We are done. We have the ingredients to calculate the GMM standard error formula (11.146) and formula for the covariance of moments (11.147). We can recover the classic formulas (12.184), (12.185), (12.186) by adding the assumption that the errors are i.i.d. and independent of the factors, and that the factors are uncorrelated over time as well. The assumption that the errors and factors are uncorrelated over time means we can ignore the lead and lag terms. Thus, the top left corner is E(εt ε0t ) = Σ. The assumption that the errors are independent from the factors ft simplifies the terms in which εt and ft are multiplied: E(εt (ε0t ft )) = E(f )Σ for example. The result is Σ E(f )Σ Σ E(f )Σ S = E(f )Σ E(f 2 )Σ 0 2 Σ E(f )Σ ββ σ (f ) + Σ

Multiplying a, d, S together as specified by the GMM formula for the covariance matrix of parameters (11.146) we obtain the covariance matrix of all the parameters, and its (3,3) ˆ . Multiplying the terms together as specified by (11.147), we element gives the variance of λ obtain the sampling distribution of the α ˆ , (12.185). The formulas (12.184) reported above are derived the same way with a vector of factors ft rather than a scalar; the second moment condition in (12.188) then reads E [(Rte − a − βf t ) ⊗ ft ]. The matrix multiplication is not particularly enlightening. Once again, there is really no need to make the assumption that the errors are i.i.d. and especially that they are conditionally homoskedastic – that the factor f and errors ε are independent. It is quite easy to estimate an S matrix that does not impose these conditions and calculate standard errors. They will not have the pretty analytic form given above, but they will more closely report the true sampling uncertainty of the estimate. Furthermore, if one is really interested in efficiency, the GLS cross-sectional estimate should use the spectral density matrix as weighting matrix rather than Σ−1 . 226

S ECTION 12.2 12.2.4

CROSS - SECTIONAL REGRESSIONS

Time series vs. cross-section

How are the time-series and cross-sectional approaches different? Most importantly, you can run the cross-sectional regression when the factor is not a return. The time-series test requires factors that are also returns, so that you can estimate ˆ = ET (f ). The asset pricing model does predict a restriction on the factor risk premia by λ intercepts in the time-series regression. Why not just test these? If you impose the restriction E(Rei ) = β 0i λ, you can write the time-series regression (12.174) as 0 0 i Rei t = β i λ + β i (ft − E(f)) + εt , t = 1, 2, ...T for each i.

Comparing this with (12.174), you see that the intercept restriction is ai = β 0i (λ − E(f )) .

This restriction makes sense. The model says that mean returns should be proportional to betas, and the intercept in the time-series regression controls the mean return. You can also see how λ = E(f ) results in a zero intercept. Finally, however, you see that without an estimate of λ, you can’t check this intercept restriction. If the factor is not a return, you will be forced to do something like a cross-sectional regression. When the factor is a return, so that we can compare the two methods, they are not necessarily the same. The time-series regression estimates the factor risk premium as the sample mean of the factor. Hence, the factor receives a zero pricing error. Also, the predicted zerobeta excess return is also zero. Thus, the time-series regression describes the cross-section of expected returns by drawing a line as in Figure 26 that runs through the origin and through the factor, ignoring all of the other points. The OLS cross-sectional regression picks the slope and intercept, if you include one, to best fit all the points; to minimize the sum of squares of all the pricing errors. If the factor is a return, the GLS cross-sectional regression, including the factor as a test asset, is identical to the time-series regression. The time-series regression for the factor is, of course, ft = 0 + 1ft + 0

so it has a zero intercept, beta equal to one, and zero residual in every sample. The residual variance covariance matrix of the returns, including the factor, is µ· e ¸ ¸ ¶ · R − a − βf Σ 0 E [·]0 = 0 0 f − 0 − 1f Since the factor has zero residual variance, a GLS regression puts all its weight on that asset. ˆ = ET (f ) just as for the time-series regression. The pricing errors are the same, Therefore, λ as is their distribution and the χ2 test. (You gain a degree of freedom by adding the factor to the cross sectional regression, so the test is a χ2N .) 227

C HAPTER 12

R EGRESSION - BASED TESTS OF LINEAR FACTOR MODELS

Why does the “efficient” technique ignore the pricing errors of all of the other assets in estimating the factor risk premium, and focus only on the mean return? The answer is simple, though subtle. In the regression model Ret = a + βft + εt ,

the average return of each asset in a sample is equal to beta times the average return of the factor in the sample, plus the average residual in the sample. An average return carries no additional information about the mean of the factor. A signal plus noise carries no additional information beyond that in the same signal. Thus, an “efficient” cross-sectional regression wisely ignores all the information in the other asset returns and uses only the information in the factor return to estimate the factor risk premium.

12.3

Fama-MacBeth Procedure

I introduce the Fama-MacBeth procedure for running cross sectional regression and calculating standard errors that correct for cross-sectional correlation in a panel. I show that, when the right hand variables do not vary over time, Fama-MacBeth is numerically equivalent to pooled time-series, cross-section OLS with standard errors corrected for cross-sectional correlation, and also to a single cross-sectional regression on time-series averages with standard errors corrected for cross-sectional correlation. Fama-MacBeth standard errors do not include corrections for the fact that the betas are also estimated. Fama and MacBeth (1973) suggest an alternative procedure for running cross-sectional regressions, and for producing standard errors and test statistics. This is a historically important procedure, it is computationally simple to implement, and is still widely used, so it is important to understand it and relate it to other procedures. First, you find beta estimates with a time-series regression. Fama and MacBeth use rolling 5 year regressions, but one can also use the technique with full-sample betas, and I will consider that simpler case. Second, instead of estimating a single cross-sectional regression with the sample averages, we now run a cross-sectional regression at each time period, i.e. Rtei = β 0i λt + αit i = 1, 2, ...N for each t.

I write the case of a single factor for simplicity, but it’s easy to extend the model to multiple factors. Then, Fama and MacBeth suggest that we estimate λ and αi as the average of the cross sectional regression estimates, T T X 1X ˆ= 1 ˆt; α λ λ ˆi = α ˆ it . T t=1 T t=1

228

S ECTION 12.3

FAMA -MAC B ETH P ROCEDURE

Most importantly, they suggest that we use the standard deviations of the cross-sectional regression estimates to generate the sampling errors for these estimates, ˆ = σ2 (λ)

T T ´2 1 X ³ˆ 1 X 2 ˆ λ − λ ; σ (ˆ α ) = (ˆ αit − α ˆ i )2 . t i T 2 t=1 T 2 t=1

It’s 1/T 2 because we’re finding standard errors of sample means, σ2 /T This is an intuitively appealing procedure once you stop to think about it. Sampling error is, after all, about how a statistic would vary from one sample to the next if we repeated the observations. We can’t do that with only one sample, but why not cut the sample in half, and deduce how a statistic would vary from one full sample to the next from how it varies from the first half of the sample to the next half? Proceeding, why not cut the sample in fourths, eights and so on? The Fama-MacBeth procedure carries this idea to is logical conclusion, ˆ t over time to deduce its sampling variation. using the variation in the statistic λ We are used to deducing the sampling variance of the sample mean of a series xt by x) = σ2 (x)/T = looking at the variation of xt through time in the sample, using σ 2 (¯ P 2 1 ¯) . The Fama-MacBeth technique just applies this idea to the slope and prict (xt − x T2 ing error estimates. The formula assumes that the time series is not autocorrelated, but one ˆ t that are correlated over time by using a long run could easily extend the idea to estimates λ variance matrix, i.e. estimate . ˆ = σ 2 (λ)

∞ 1 X ˆt, λ ˆ t−j ) covT (λ T j=−∞

One should of course use some sort of weighting matrix or a parametric description of the ˆ , as explained in section 11.7. Asset return data are usually not highly autocorrelations of λ correlated, but accounting for such correlation could have a big effect on the application of the Fama-MacBeth technique to corporate finance data or other regressions in which the cross-sectional estimates are highly correlated over time. It is natural to use this sampling theory to test whether all the pricing errors are jointly zero as we have before. Denote by α the vector of pricing errors across assets. We could estimate the covariance matrix of the sample pricing errors by α ˆ =

cov(ˆ α) =

T 1X α ˆt T t=1

T 1 X (ˆ αt − α ˆ ) (ˆ αt − α ˆ )0 T 2 t=1

(or a general version that accounts for correlation over time) and then use the test α ˆ 0 cov(ˆ α)−1 α ˆ ∼ χ2N−1 .

229

C HAPTER 12 12.3.1

R EGRESSION - BASED TESTS OF LINEAR FACTOR MODELS

Fama MacBeth in depth

The GRS procedure and the formulas given above for a single cross-sectional regression are familiar from any course in regression. We will see them justified by maximum likelihood below. The Fama MacBeth procedure seems unlike anything you’ve seen in any econometrics course, and it is obviously a useful and simple technique that can be widely used in panel data in economics and corporate finance as well as asset pricing. Is it truly different? Is there something different about asset pricing data that requires a fundamentally new technique not taught in standard regression courses? Or is it similar to standard techniques? To answer these questions it is worth looking in a little more detail at what it accomplishes and why. It’s easier to do this in a more standard setup, with left hand variable y and right hand variable x. Consider a regression yit = β 0 xit + εit i = 1, 2, ...N; t = 1, 2, ...T.

The data in this regression has a cross-sectional element as well as a time-series element. In corporate finance, for example, one might be interested in the relationship between investment and financial variables, and the data set has many firms (N ) as well as time series observations for each firm (T ). In and expected return-beta asset pricing model, the xit standing for the β i and β stands for λ. An obvious thing to do in this context is simply to stack the i and t observations together and estimate β by OLS. I will call this the pooled time-series cross-section estimate. However, the error terms are not likely to be uncorrelated with each other. In particular, the error terms are likely to be cross-sectionally correlated at a given time. If one stock’s return is unusually high this month, another stock’s return is also likely to be high; if one firm invests an unusually great amount this year, another firm is also likely to do so. When errors are correlated, OLS is still consistent, but the OLS distribution theory is wrong, and typically suggests standard errors that are much too small. In the extreme case that the N errors are perfectly correlated at each time period, there really only one observation for each time period, so one really has T rather than NT observations. Therefore, a real pooled time-series cross-section estimate must include corrected standard errors. People often ignore this fact and report OLS standard errors. Another thing we could do is first take time series averages and then run a pure crosssectional regression of ET (yit ) = β 0 ET (xit ) + ui i = 1, 2, ...N

This procedure would lose any information due to variation of the xit over time, but at least it might be easier to figure out a variance-covariance matrix for ui and correct the standard errors for residual correlation. (You could also average cross-sectionally and than run a single time-series regression. We’ll get to that option later.) In either case, the standard error corrections are just applications of the standard formula 230

S ECTION 12.3

FAMA -MAC B ETH P ROCEDURE

for OLS regressions with correlated error terms. Finally, we could run the Fama-MacBeth procedure: run a cross-sectional regression at each point in time; average the cross-sectional βˆ t estimates to get an estimate βˆ , and use the time-series standard deviation of βˆ t to estimate the standard error of βˆ . It turns out that the Fama MacBeth procedure is just another way of calculating the standard errors, corrected for cross-sectional correlation: Proposition: If the xit variables do not vary over time, and if the errors are cross-sectionally correlated but not correlated over time, then the Fama-MacBeth estimate, the pure crosssectional OLS estimate and the pooled time-series cross-sectional OLS estimates are identical. Also, the Fama-MacBeth standard errors are identical to the cross-sectional regression or stacked OLS standard errors, corrected for residual correlation. None of these relations hold if the x vary through time.

Since they are identical procedures, whether one calculates estimates and standard errors in one way or the other is a matter of taste. I emphasize one procedure that is incorrect: pooled time series and cross section OLS with no correction of the standard errors. The errors are so highly cross-sectionally correlated in most finance applications that the standard errors so computed are often off by a factor of 10. The assumption that the errors are not correlated over time is probably not so bad for asset pricing applications, since returns are close to independent. However, when pooled time-series cross-section regressions are used in corporate finance applications, errors are likely to be as severely correlated over time as across firms, if not more so. The “other factors” (ε) that cause, say, company i to invest more at time t than predicted by a set of right hand variables is surely correlated with the other factors that cause company j to invest more. But such factors are especially likely to cause company i to invest more tomorrow as well. In this case, any standard errors must also correct for serial correlation in the errors; the GMM based formulas in section 11.4 can do this easily. The Fama-MacBeth standard errors also do not correct for the fact that βˆ are generated regressors. If one is going to use them, it is a good idea to at least calculate the Shanken correction factors outlined above, and check that the corrections are not large. Proof: We just have to write out the three approaches and compare them. Having assumed that the x variables do not vary over time, the regression is yit = x0i β + εit .

We can stack up the cross-sections i = 1...N and write the regression as yt = xβ + εt . x is now a matrix with the x0i as rows. The error assumptions mean E(εt ε0t ) = Σ.

231

C HAPTER 12

R EGRESSION - BASED TESTS OF LINEAR FACTOR MODELS

Pooled OLS: To run pooled OLS, we stack the time series and cross sections by writing ε1 x y1 ε2 x y2 Y = . ; X = . ; ² = . .. .. .. yT

x

εT

and then

Y = Xβ + ²

with

E(²²0 ) = Ω =

Σ

..

. Σ

The estimate and its standard error are then

ˆ OLS = (X 0 X)−1 X 0 Y β −1 −1 0 ˆ X 0 ΩX (X 0 X) cov(β OLS ) = (X X)

Writing this out from the definitions of the stacked matrices, with X 0 X =T x0 x, ˆ β OLS

−1

= (x0 x) x0 ET (yt ) 1 0 −1 0 −1 ˆ (x x) (x Σx) (x0 x) . cov(β OLS ) = T

We can estimate this sampling variance with ¡ ¢ ˆ OLS ˆ = ET ˆεtˆε0t ; ˆεt ≡ yt − xβ Σ

Pure cross-section: The pure cross-sectional estimator runs one cross-sectional regression of the time-series averages. So, take those averages, ET (yt ) = xβ + ET (εt )

where x = ET (x ) since x is constant. Having assumed i.i.d. errors over time, the error covariance matrix is E (ET (εt ) ET (ε0t )) =

1 Σ. T

The cross sectional estimate and corrected standard errors are then ˆ XS β

−1

= (x0 x) x0 ET (yt ) ˆ ) = 1 (x0 x)−1 x0 Σx−1 (x0 x)−1 σ 2 (β XS T

232

S ECTION 12.3

FAMA -MAC B ETH P ROCEDURE

Thus, the cross-sectional and pooled OLS estimates and standard errors are exactly the same, in each sample. Fama-MacBeth: The Fama–MacBeth estimator is formed by first running the crosssectional regression at each moment in time, ˆ = (x0 x)−1 x0 yt . β t

Then the estimate is the average of the cross-sectional regression estimates, ³ ´ ˆ = (x0 x)−1 x0 ET (yt ) . ˆ F M = ET β β t

Thus, the Fama-MacBeth estimator is also the same as the OLS estimator, in each sample. The Fama-MacBeth standard error is based on the time-series standard deviation of the βˆ t . Using covT to denote sample covariance, ´ ³ ´ ³ ˆ = 1 (x0 x)−1 x0 covT (yt ) x (x0 x)−1 . ˆ F M = 1 covT β cov β t T T with

yt = xβ F M + ˆεt

we have ˆ covT (yt ) = ET (ˆεtˆε0t ) = Σ

and finally ´ ³ ˆ F M = 1 (x0 x)−1 x0 Σx ˆ (x0 x)−1 . cov β T

Thus, the FM estimator of the standard error is also numerically equivalent to the OLS corrected standard error. Varying x If the xit vary through time, none of the three procedures are equal anymore, since the cross-sectional regressions ignore time-series variation in the xit . As an extreme example, suppose a scalar xit varies over time but not cross-sectionally, yit = α + xt β + εit ; i = 1, 2, ...N ; t = 1, 2, ...T.

The grand OLS regression is P P P ˜t N1 i yit ˜t yit tx it x ˆ P P β = = OLS ˜2t ˜2t it x tx

where x˜ = x − ET (x) denotes the demeaned variables. The estimate is driven by the covariance over time of xt with the cross-sectional average of the yit , which is sensible because all of the information in the sample lies in time variation. It is identical to a regression over time 233

C HAPTER 12

R EGRESSION - BASED TESTS OF LINEAR FACTOR MODELS

of cross-sectional averages. However, you can’t even run a cross-sectional estimate, since the right hand variable is constant across i. As a practical example, you might be interested in a CAPM specification in which the betas vary over time (β t ) but not across test assets. This sample still contains information about the CAPM: the time-variation in betas should be matched by time variation in expected returns. But any method based on cross-sectional ¥ regressions will completely miss it. In historical context, the Fama MacBeth procedure was also important because it allowed changing betas, which a single cross-sectional regression or a time-series regression test cannot easily handle.

12.4 1.

2. 3. 4.

Problems

When we express the CAPM in excess return form, can the test assets be differences between risky assets, Ri −Rj ? Can the market excess return also use a risky ¡ asset, or must ¢ it be relative to a risk free rate? (Hint: start with E(Ri ) − Rf = β i,m E(Rm ) − Rf and see if you can get to the other forms. Betas must be regression coefficients.) Can you run the GRS test on a model that uses industrial production growth as a factor, E(Ri ) − Rf = β i,∆ip λip ? Fama and French (1997b) report that pricing errors are correlated with betas in a test of a factor pricing model on industry portfolios. How is this possible? We saw that a GLS cross-sectional regression of the CAPM passes through the market and riskfree rate by construction. Show that if the market return is an equally weighted portfolio of the test assets, then an OLS cross-sectional regression with an estimated intercept passes through the market return by construction. Does it also pass through the riskfree rate or origin?

234

Chapter 13. GMM for linear factor models in discount factor form 13.1

GMM on the pricing errors gives a cross-sectional regression

The first stage estimate is an OLS cross-sectional regression, and the second stage is a GLS regression, First stage : Second stage :

ˆb1 = (d0 d)−1 d0 ET (p) ˆb2 = (d0 S −1 d)d0 S −1 E(p).

Standard errors are the corresponding regression formulas, and the variance of the pricing errors are the standard regression formula for variance of a residual. Treating the constant a × 1 as a constant factor, the model is m = b0 f E(p) = E(mx).

or simply E(p) = E(xf 0 )b.

(189)

Keep in mind that p and x are N × 1 vectors of asset prices and payoffs respectively; f is a K × 1 vector of factors, and b is a K × 1 vector of parameters. I suppress the time indices mt+1 , ft+1 , xt+1, pt . The payoffs are typically returns or excess returns, including returns scaled by instruments. The prices are typically one (returns) zero (excess returns) or instruments. To implement GMM, we need to choose a set of moments. The obvious set of moments to use are the pricing errors, gT (b) = ET (xf 0 b − p).

This choice is natural but not necessary. You don’t have to use p = E(mx) with GMM, and you don’t have to use GMM with p = E(mx). You can (we will) use GMM on expected return-beta models, and you can use maximum likelihood on p = E(mx). It is a choice, and the results will depend on this choice of moments as well as the specification of the model. 235

C HAPTER 13

GMM FOR LINEAR FACTOR MODELS IN DISCOUNT FACTOR FORM

The GMM estimate is formed from min gT (b)0 W gT (b) b

with first order condition d0 W gT (b) = d0 W ET (xf 0 b − p) = 0

where d0 =

∂gT0 (b) = ET (f x0 ). ∂b

This is the second moment matrix of payoffs and factors. The first stage has W = I , the second stage has W = S −1 . Since this is a linear model, we can solve analytically for the GMM estimate, and it is First stage : Second stage :

ˆb1 = (d0 d)−1 d0 ET (p) ˆb2 = (d0 S −1 d)d0 S −1 ET (p).

The first stage estimate is an OLS cross-sectional regression of average prices on the second moment of payoff with factors, and the second stage estimate is a GLS cross-sectional regression. What could be more sensible? The model (13.189) says that average prices should be a linear function of the second moment of payoff with factors, so the estimate runs a linear regression. These are cross-sectional regressions since they operate across assets on sample averages. The “data points” in the regression are sample average prices (y) and second moments of payoffs with factors (x) across test assets. We are picking the parameter b to make the model fit explain the cross-section of asset prices as well as possible.

We find the distribution theory from the standard GMM standard error formulas (11.144) and (11.150). In the first stage, a = d0 . First stage : Second stage :

1 cov(ˆb1 ) = (d0 d)−1 d0 Sd(d0 d)−1 T 1 cov(ˆb2 ) = (d0 S −1 d)−1 . T

(13.190)

Unsurprisingly, these are exactly the formulas for OLS and GLS regression errors with error covariance S . The pricing errors are correlated across assets, since the payoffs are correlated. Therefore the OLS cross-sectional regression standard errors need to be corrected for correlation, as they are in (13.190) and one can pursue an efficient estimate as in GLS. The analogy is GLS is close, since S is the covariance matrix of E(p) − E(xf 0 )b; S is the covariance matrix of the “errors” in the cross-sectional regression. 236

S ECTION 13.2

T HE CASE OF EXCESS RETURNS

The covariance matrix of the pricing errors is, from (11.147), (11.151) and (11.152) h i ¡ ¢ ¡ ¢ First stage : T cov gT (ˆb) = I − d(d0 d)−1 d0 S I − d(d0 d)−1 d0 (13.191) h i Second stage : T cov gT (ˆb) = S − d(d0 S −1 d)−1 d0 .

These are obvious analogues to the standard regression formulas for the covariance matrix of regression residuals. The model test gT (b)0 cov(gT )−1 gT (b) ∼ χ2 (#moments − #parameters)

which specializes for the second-stage estimate as T gT (ˆb)0 S −1 gT (ˆb) ∼ χ2 (#moments − #parameters).

There is not much point in writing these out, other than to point out that the test is a quadratic form in the vector of pricing errors. It turns out that the χ2 test has the same value for first and second stage for this model, even though the parameter estimates, pricing errors and covariance matrix are not the same.

13.2

The case of excess returns

When mt+1 = a − b0 ft+1 and the test assets are excess returns, the GMM estimate is a GLS cross-sectional regression of average returns on the second moments of returns with factors, First stage : Second stage :

ˆb1 = (d0 d)−1 d0 ET (Re ) ˆb2 = (d0 S −1 d)d0 S −1 ET (Re ).

where d is the covariance matrix between returns and factors. The other formulas are the same. The analysis of the last section requires that at least one asset has a nonzero price. If all assets are excess returns then ˆb1 = (d0 d)−1 d0 ET (p) = 0. Linear factor models are most often applied to excess returns, so this case is important. The trouble is that in this case the mean discount factor is not identified. If E(mRe ) = 0 then E((2 × m)Re ) = 0. Analogously in expected return-beta models, if all test assets are excess returns, then we have no information on the level of the zero-beta rate. Writing out the model as m = a−b0 f , we cannot separately identify a and b so we have to choose some normalization. The choice is entirely one of convenience; lack of identification 237

C HAPTER 13

GMM FOR LINEAR FACTOR MODELS IN DISCOUNT FACTOR FORM

means precisely that the pricing errors do not depend on the choice of normalization. The easiest choice is a = 1. Then gT (b) = ET (mRe ) = ET (Re ) − E(Re f 0 )b.

We have d0 =

∂gT (b) = E(fRe0 ), ∂b0

the second moment matrix of returns and factors. The first order condition to min gT0 W gT is d0 W [d b + ET (Re )] = 0.

Then, the GMM estimates of b are ˆb1 = (d0 d)−1 d0 ET (Re ) ˆb2 = (d0 S −1 d)d0 S −1 ET (Re ).

First stage : Second stage :

The GMM estimate is a cross-sectional regression of mean excess returns on the second moments of returns with factors. From here on in, the distribution theory is unchanged from the last section. Mean returns on covariances

We can obtain a cross-sectional regression of mean excess returns on covariances, which are just a heartbeat away from betas, by choosing the normalization a = 1 + b0 E(f) rather than a = 1. Then, the model is m = 1 − b0 (f − E(f )) with mean E(m) = 1. The pricing errors are gT (b) = ET (mRe ) = ET (Re ) − ET (Re f˜0 )b

where I denote f˜ ≡ f − E(f). We have d0 =

∂gT (b) ˜ e0 ), = ET (fR ∂b0

which now denotes the covariance matrix of returns and factors. The first order condition to min gT0 W gT is now d0 W [d b + ET (Re )] = 0.

Then, the GMM estimates of b are First stage : Second stage :

ˆb1 = (d0 d)−1 d0 ET (Re ) ˆb2 = (d0 S −1 d)d0 S −1 ET (Re ).

238

S ECTION 13.3

H ORSE R ACES

The GMM estimate is a cross-sectional regression of expected excess returns on the covariance between returns and factors. Naturally, the model says that expected excess returns should be proportional to the covariance between returns and factors, and the estimate estimates that relation by a linear regression. The standard errors and variance of the pricing errors are the same as in (13.190) and (13.191), with d now representing the covariance matrix. The formulas are almost exactly identical to those of the cross-sectional regressions in section 12.2. The p = E(mx) formulation of the model for excess returns is equivalent to E(Re ) = −Cov(Re , f 0 )b; thus covariances enter in place of betas β .

There is one fly in the ointment; the mean of the factor E(f ) is estimated, and the distribution theory should recognize sampling variation induced by this fact, as we did for the fact that betas are generated regressors in the cross-sectional regressions of section 2.3. The distribution theory is straightforward, and a problem at the end of the chapter guides you through it. However, I think it is better to avoid the complication and just use the second moment approach, or some other non-sample dependent normalization for a. The pricing errors are identical – the whole point is that the normalization of a does not matter to the pricing errors. Therefore, the χ2 statistics are also identical. As you change the normalization for a, you change the estimate of b. Therefore, the only effect is to add a term in the sampling variance of the estimated parameter b.

13.3

Horse Races

How to test whether one set of factors drives out another. Test b2 = 0 in m = b01 f1 + b02 f2 using the standard error of ˆb2 , or the χ2 difference test. It’s often interesting to test whether one set of factors drives out another. For example, Chen Roll and Ross (1986) test whether their five macroeconomic factors price assets so well that one can ignore even the market return. Given the large number of factors that have been proposed, a statistical procedure for testing which factors survive in the presence of the others is desirable. In this framework, such a test is very easy. Start by estimating a general model m = b01 f1 + b02 f2 .

(192)

We want to know, given factors f1 , do we need the f2 to price assets – i.e. is b2 = 0? There are two ways to do this. First and most obviously, we have an asymptotic covariance matrix for [b1 b2 ], so we can form a t test (if b2 is scalar) or χ2 test for b2 = 0 by forming the statistic ˆb0 var(ˆb2 )−1ˆb2 ∼ χ2 2 #b2

239

C HAPTER 13

GMM FOR LINEAR FACTOR MODELS IN DISCOUNT FACTOR FORM

where #b2 is the number of elements in the b2 vector. This is a Wald test.. Second, we can estimate a restricted system m = b01 f1 . Since there are fewer free parameters and the same number of moments than in (13.192), we expect the criterion JT to rise. If we use the same weighting matrix, (usually the one estimated from the unrestricted model (13.192)) then the JT cannot in fact decline. But if b2 really is zero, it shouldn’t rise “much.” How much? The χ2 difference test answers that question; T JT (restricted) − T JT (unrestricted) ∼ χ2 (#of restrictions)

This is very much like a likelihood ratio test.

13.4

Testing for characteristics

How to check whether an asset pricing model drives out a characteristic such as size, book/market or volatility. Run cross sectional regressions of pricing errors on characteristics; use the formulas for covariance matrix of the pricing errors to create standard errors. It’s often interesting to characterize a model by checking whether the model drives out a characteristic. For example, portfolios organized by size or market capitalization show a wide dispersion in average returns (at least up to 1979). Small stocks gave higher average returns than large stocks. The size of the portfolio is a characteristic. A good asset pricing model should account for average returns by betas. It’s ok if a characteristic is associated with average returns, but in the end betas should drive out the characteristic; the alphas or pricing errors should not be associated with the characteristic. The original tests of the CAPM similarly checked whether the variance of the individual portfolio had anything to do with average returns once betas were included. Denote the characteristic of portfolio i by yi . An obvious idea is to include both betas and the characteristic in a multiple, cross-sectional regression, E(Rei ) = (α0 ) + β 0i λ + γyi + εi ; i = 1, 2, ...N

Alternatively, subtract βλ from both sides and consider a cross-sectional regression of alphas on the characteristic, αi = (α0 ) + γyi + εi ; i = 1, 2, ...N.

(The difference is whether you allow the presence of the size characteristic to affect the λ estimate or not.) We can always run such a regression, but we don’t want to use the OLS formulas for the sampling error of the estimates, since the errors εi are correlated across assets. Under the null that γ = 0, ε = α, so we can simply use the covariance matrix of the alphas to generate 240

S ECTION 13.5

T ESTING FOR PRICED FACTORS :

LAMBDAS OR B ’ S ?

standard errors of the γ . Let X denote the vector of characteristics, then the estimate is γˆ = (X 0 X)−1 X 0 α ˆ

with standard error σ(ˆ γ ) = (X 0 X)−1 X 0 cov(ˆ α)X(X 0 X)−1

At this point, simply use the formula for cov(ˆ α) or cov(gT ) as appropriate for the model that you tested. Sometimes, the characteristic is also estimated rather than being a fixed number such as the size rank of a size portfolio, and you’d like to include the sampling uncertainty of its estimation in the standard errors of γˆ. Let yti denote the time series whose mean E(yti ) determines the characteristic. Now, write the moment condition for the ith asset as gT = ET (mt+1 (b)xt+1 − pt − γyti ).

The estimate of γ tells you how the characteristic E(yi ) is associated with model pricing errors E(mt+1 (b)xt+1 − pt ). The GMM estimate of γ is E(y)0 W (E(mx) − p − γy) −1

γˆ = (ET0 (y)W ET (y))

ET0 (y)W gT

a OLS or GLS regression of the pricing errors on the estimated characteristics. The standard GMM formulas for the standard deviation of γ or the χ2 difference test for γ = 0 tell you whether the γ estimate is statistically significant, including the fact that E(y) must be estimated.

13.5

Testing for priced factors: lambdas or b’s?

bj asks whether factor j helps to price assets given the other factors. bj gives the multiple regression coefficient of m on fj given the other factors. λj asks whether factor j is priced, or whether its factor-mimicking portfolio carries a positive risk premium. λj gives the single regression coefficient of m on fj .

Therefore, when factors are correlated, one should test bj = 0 to see whether to include factor j given the other factors rather than test λj = 0. Expected return-beta models defined with single regression betas give rise to λ with multiple regression interpretation that one can use to test factor pricing.

241

C HAPTER 13

GMM FOR LINEAR FACTOR MODELS IN DISCOUNT FACTOR FORM

In the context of expected return-beta models, it has been more traditional to evaluate the relative strengths of models by testing the factor risk premia λ of additional factors, rather than test whether their b is zero. (The b’s are not the same as the β ’s. b are the regression coefficient of m on f, β are the regression coefficients of Ri on f.) To keep the equations simple, I’ll use mean-zero factors, excess returns, and normalize to E(m) = 1, since the mean of m is not identified with excess returns. The parameters b and λ are related by λ = E(f f 0 )b.

See section 6.3. Briefly, 0 = E(mRe ) = E [Re (1 − f 0 b)] E(Re ) = cov(Re , f 0 )b = cov(Re , f 0 )E(f f 0 )−1 E(f f 0 )b = β 0 λ.

Thus, when the factors are orthogonal, E(f f 0 ) is diagonal, and each λj = 0 if and only if the corresponding bj = 0. The distinction between b and λ only matters when the factors are correlated. Factors are often correlated however. λj captures whether factor fj is priced. We can write λ = E [f (f 0 b)] = −E(mf) to see that λ is (the negative of) the price that the discount factor m assigns to f . b captures whether factor fj is marginally useful in pricing assets, given the presence of other factors. If bj = 0, we can price assets just as well without factor fj as with it. λj is proportional to the single regression coefficient of m on f . λj = cov(m, fj ). λj = 0 asks the corresponding single regression coefficient question—“is factor j correlated with the true discount factor?” bj is the multiple regression coefficient of m on fj given all the other factors. This just follows from m = b0 f . (Regressions don’t have to have error terms!) A multiple regression coefficient β j in y = xβ + ε is the way to answer “does xj help to explain variation in y given the presence of the other x’s?” When you want to ask the question, “should I include factor j given the other factors?” you want to ask the multiple regression question.

For example, suppose the CAPM is true, which is the single factor model m = a − bRem

where Rem is the market excess return. Consider any other excess return Rex , positively correlated with Rem (x for extra). If we try a factor model with the spurious factor Rex , the answer is m = a − bRem + 0 × Rex . bx is obviously zero, indicating that adding this factor does not help to price assets.

However, since the correlation of Rex with Rem is positive, the beta of Rex on Rem is positive, Rex earns a positive expected excess return, and λx = E(Rex ) > 0. In the expected 242

S ECTION 13.5

T ESTING FOR PRICED FACTORS :

LAMBDAS OR B ’ S ?

return - beta model E(Rei ) = β im λm + β ix λx λm = E(Rem ) is unchanged by the addition of the spurious factor. However, since the factors Rem , Rex are correlated, the multiple regression betas of Rei on the factors change when we add the extra factor Rex . If β ix is positive, β im will decline from its single-regression value, so the new model explains the same expected return E(Rei ). The expected return beta model will indicate a risk premium for β x exposure, and many assets will have β x exposure (Rx for example!) even though factor Rx is spurious. In particular, Rex will of course have multiple regression coefficients β x,m = 0 and β x,x = 1, and its expected return will be entirely explained by the new factor x.

So, as usual, the answer depends on the question. If you want to know whether factor i is priced, look at λ (or E(mf i )). If you want to know whether factor i helps to price other assets, look at bi . This is not an issue about sampling error or testing. All moments above are population values. Of course, testing b = 0 is particularly easy in the GMM, p = E(mx) setup. But you can always test the same ideas in any expression of the model. In an expected return-beta model, estimate b by E(ff 0 )−1 λ and test the elements of that vector rather than λ itself. You can write an asset pricing model as ERe = β 0 λ and use the λ to test whether each factor can be dropped in the presence of the others, if you use single regression betas rather than multiple regression betas. In this case each λ is proportional to the corresponding b. Problem 2 at the end of this chapter helps you to work out this case. 13.5.1

Mean-variance frontier and performance evaluation

A GMM, p = E(mx) approach to testing whether a return expands the mean-variance frontier. Just test whether m = a + bR prices all returns. If there is no risk free rate, use two values of a. We often summarize asset return data by mean-variance frontiers. For example, a large literature has examined the desirability of international diversification in a mean-variance context. Stock returns from many countries are not perfectly correlated, so it looks like one can reduce portfolio variance a great deal for the same mean return by holding an internationally diversified portfolio. But is this real or just sampling error? Even if the value-weighted portfolio were ex-ante mean-variance efficient, an ex-post mean-variance frontier constructed from historical returns on the roughly NYSE stocks would leave the value-weighted portfolio well inside the ex-post frontier. So is “I should have bought Japanese stocks in 1960” (and sold them in 1990!) a signal that broad-based international diversification a good idea now, or is it simply 20/20 hindsight regret like “I should have bought Microsoft in 1982?” Sim243

C HAPTER 13

GMM FOR LINEAR FACTOR MODELS IN DISCOUNT FACTOR FORM

Frontiers intersect E(R)

1/E(m)

σ(R) Figure 27. Mean variance frontiers might intersect rather than coincide. ilarly, when evaluating fund managers, we want to know whether the manager is truly able to form a portfolio that beats mean-variance efficient passive portfolios, or whether better performance in sample is just due to luck. Since a factor model is true if and only if a linear combination of the factors (or factormimicking portfolios if the factors are not returns) is mean-variance efficient, one can interpret a test of any factor pricing model as a test whether a given return is on the mean-variance frontier. Section 12.1 showed how the Gibbons Ross and Shanken pricing error statistic can be interpreted as a test whether a given portfolio is on the mean-variance frontier, when returns and factors are i.i.d., and the GMM distribution theory of that test statistic allows us to extend the test to non-i.i.d. errors. A GMM, p = E(mx), m = a − bRp test analogously tests whether Rp is on the mean-variance frontier of the test assets. We may want to go one step further, and not just test whether a combination of a set of assets Rd (say, domestic assets) is on the mean-variance frontier, but whether the Rd assets span the mean-variance frontier of Rd and Ri (say, foreign or international) assets. The trouble is, that if there is no riskfree rate, the frontier generated by Rd might just intersect the frontier generated by Rd and Ri together, rather than span or coincide with the latter frontier, as shown in Figure 27. Testing that m = a − b0 Rd prices both Rd and Ri only checks for intersection.

244

S ECTION 13.6

P ROBLEMS

DeSantis (1992) and Chen and Knez (1992,1993) show how to test for spanning as opposed to intersection. For intersection, m = a − b0d Rd will price both Rd and Rf only for one value of a, or equivalently E(m) or choice of the intercept, as shown. If the frontiers coincide or span, then m = a + b0d Rd prices both Rd and Rf for any value of a. Thus,we can test for coincident frontiers by testing whether m = a + b0d Rd prices both Rd and Rf for two prespecified values of a simultaneously. To see how this work, start by noting that there must be at least two assets in Rd . If not, there is no mean-variance frontier of Rd assets; it is simply a point. If there are two assets in Rd ,Rd1 and Rd2 , then the mean-variance frontier of domestic assets connects them; they are each on the frontier. If they are both on the frontier, then there must be discount factors m1 = a1 − ˜b1 Rd1

and m2 = a2 − ˜b2 Rd2

and, of course, any linear combination, i ¤ h £ m = λa1 + (1 − λ)a2 − λ˜b1 Rd1 + (1 − λ)˜b2 Rd2 .

Equivalently, for any value of a, there is a discount factor of the form ¢ ¡ m = a − b1 Rd1 + b2 Rd2 . Thus, you can test for spanning with a JT test on the moments ¤ £ E (a1 − b10 Rd )Rd = 0 £ ¤ E (a1 − b10 Rd )Ri = 0 ¤ £ E (a2 − b20 Rd )Rd = 0 ¤ £ E (a2 − b20 Rd )Ri = 0

for any two fixed values of a1 , a2 .

13.6 1.

Problems

Work out the GMM distribution theory for the model m = 1 − b0 (f − E(f )) and test assets are excess returns. The distribution should recognize the fact that E(f ) is estimated in sample. To do this, set up · ¸ ET (Re − Re (f 0 − Ef 0 ) b) gT = ET (f − Ef ) 245

C HAPTER 13

GMM FOR LINEAR FACTOR MODELS IN DISCOUNT FACTOR FORM

aT =

2.

"

³ ´ # ˜ e0 ET fR 0 . 0 IK

The estimated parameters are b, E(f). You should end up with a formula for the standard error of b that resembles the Shanken correction (12.184), and an unchanged JT test. Show that if you use single regression betas, then the corresponding λ can be used to test for the marginal importance of factors. However, the λ are no longer the expected return of factor mimicking portfolios.

246

Chapter 14.

Maximum likelihood

Maximum likelihood is, like GMM, a general organizing principle that is a useful place to start when thinking about how to choose parameters and evaluate a model. It comes with an asymptotic distribution theory, which, like GMM, is a good place to start when you are unsure about how to treat various problems such as the fact that betas must be estimated in a cross-sectional regression. As we will see, maximum likelihood is a special case of GMM. Given a statistical description of the data, it prescribes which moments are statistically most informative. Given those moments, ML and GMM are the same. Thus, ML can be used to defend why one picks a certain set of moments, or for advice on which moments to pick if one is unsure. In this sense, maximum likelihood (paired with carefully chosen statistical models) justifies the regression tests above, as it justifies standard regressions. On the other hand, ML does not easily allow you to use other non-“efficient” moments, if you suspect that ML’s choices are not robust to misspecifications of the economic or statistical model. For example, ML will tell you how to do GLS, but it will not tell you how to adjust OLS standard errors for non-standard error terms. Hamilton (1994) p.142-148 and the appendix in Campbell Lo MacKinlay (1997) give nice summaries of maximum likelihood theory. Campbell Lo and MacKinlay’s Chapter 5 and 6 treat many more variations of regression based tests and maximum likelihood.

14.1

Maximum likelihood

The maximum likelihood principle says to pick the parameters that make the observed data most likely. Maximum likelihood estimates are asymptotically efficient. The information matrix gives the asymptotic standard errors of ML estimates. The maximum likelihood principle says to pick that set of parameters that makes the observed data most likely. This is not “the set of parameters that are most likely given the data” – in classical (as opposed to Bayesian) statistics, parameters are numbers, not random variables. To implement this idea, you first have to figure out what the probability of seeing a data set {xt } is, given the free parameters θ of a model. This probability distribution is called the likelihood function f({xt } ; θ). Then, the maximum likelihood principle says to pick ˆθ = arg max f ({xt } ; θ). {θ}

For reasons that will soon be obvious, it’s much easier to work with the log of this probability 247

C HAPTER 14

MAXIMUM LIKELIHOOD

distribution L({xt } ; θ) = ln f ({xt } ; θ),

Maximizing the log likelihood is the same thing as maximizing the likelihood. Finding the likelihood function isn’t always easy. In a time-series context, the best way to do it is often to first find the log conditional likelihood function f (xt |xt−1 , xt−2 , ...x0 ; θ), the chance of seeing xt+1 given xt , xt−1 , ... and given values for the parameters, . Since joint probability is the product of conditional probabilities, the log likelihood function is just the sum of the conditional log likelihood functions, L({xt } ; θ) =

T X t=1

ln f(xt |xt−1 , xt−2 ...x0 ; θ).

(193)

More concretely, we usually assume normal errors, so the likelihood function is T

L=−

1 X 0 −1 T ln (2π |Σ|) − ε Σ εt 2 2 t=1 t

(194)

where εt denotes a vector of shocks; εt = xt − E(xt |xt−1 , xt−2 ...x0 ; θ).

This expression gives a simple recipe for constructing a likelihood function. You usually start with a model that generates xt from errors, e.g. xt = ρxt−1 + εt . Invert that model to express the errors εt in terms of the data {xt } and plug in to (14.194). There is a small issue about how to start off a model such as (14.193). Ideally, the first observation should be the unconditional density, i.e. L({xt } ; θ) = ln f (x1 ; θ) + ln f (x2 |x1 ; θ) + ln f (x3 |x2 , x1 ; θ)...

However, it is usually hard to evaluate the unconditional density or the first terms with only a few lagged xs. Therefore, if as usual the conditional density can be expressed in terms of a finite number k of lags of xt , one often maximizes the conditional likelihood function (conditional on the first k observations), treating the first k observations as fixed rather than random variables. L({xt } ; θ) = ln f (xk+1 |xk , xk−1 ...x1 ; θ) + ln f(xk+2 |xk , xk−1... x2 ; θ) + ...

Alternatively, one can treat k pre-sample values {x0 , x−1 , ...x−k+1 } as additional parameters over which to maximize the likelihood function. Maximum likelihood estimators come with a useful asymptotic (i.e. approximate) distri248

S ECTION 14.2

ML IS GMM ON THE SCORES

bution theory. First, the distribution of the estimates is ˆθ∼N

Ã · ¸−1 ! ∂ 2L θ, − ∂θ∂θ0

(195)

If the likelihood L has a sharp peak at ˆθ, then we know a lot about the parameters, while if the peak is flat, other parameters are just as plausible. The maximum likelihood estimator is asymptotically efficient meaning that no other estimator can produce a smaller covariance matrix. The second derivative in (14.195) is known as the information matrix, I=−

T 1 ∂ 2L 1 X ∂ 2 ln f (xt+1 |xt , xt−1 , ...x0 ; θ) = − . T ∂θ∂θ0 T t=1 ∂θ∂θ0

(196)

(More precisely, the information matrix is defined as the expected value of the second partial, which is estimated with the sample value.) The information matrix can also be estimated as a product of first derivatives. The expression ¶µ ¶0 T µ 1 X ∂ ln f (xt+1 |xt , xt−1 , ...x0 ; θ) ∂ ln f(xt+1 |xt , xt−1 , ...x0 ; θ) I =− . T t=1 ∂θ ∂θ

converges to the same value as (14.196). (Hamilton 1994 p.429 gives a proof.) If we estimate a model restricting the parameters, the maximum value of the likelihood function will necessarily be lower. However, if the restriction is true, it shouldn’t be that much lower. This intuition is captured in the likelihood ratio test 2(Lunrestricted − Lrestricted )∼χ2number of restrictions

(197)

The form and idea of this test is much like the χ2 difference test for GMM objectives that we met in section 11.1.

14.2

ML is GMM on the scores

ML is a special case of GMM. ML uses the information in the auxiliary statistical model to derive statistically most informative moment conditions. To see this fact, start with the first order conditions for maximizing a likelihood function T

∂L({xt } ; θ) X ∂ ln f (xt |xt−1 xt−2 ...; θ) = = 0. ∂θ ∂θ t=1

249

(198)

C HAPTER 14

MAXIMUM LIKELIHOOD

This is a GMM estimate. It is the sample counterpart to a population moment condition g(θ) = E

µ

∂ ln f (xt |xt−1 xt−2 ...; θ) ∂θ

¶

= 0.

(199)

The term ∂ ln f (xt |xt−1 xt−2 ...; θ)/∂θ is known as the “score.” It is a random variable, formed as a combination of current and past data (xt , xt−1 ...). Thus, maximum likelihood is a special case of GMM, a special choice of which moments to examine. For example, suppose that x follows an AR(1) with known variance, xt = ρxt−1 + εt ,

and suppose the error terms are i.i.d. normal random variables. Then, ln f (xt |xt−1 , xt−2 ...; ρ) = const. −

ε2t (xt − ρxt−1 )2 = const − 2σ 2 2σ 2

and the score is (xt − ρxt−1 ) xt−1 ∂ ln f (xt |xt−1 xt−2 ...; ρ) = . ∂ρ σ2

The first order condition for maximizing likelihood is T 1X (xt − ρxt−1 ) xt−1 = 0. T t=1

This expression is a moment condition, and you’ll recognize it as the OLS estimator of ρ, which we have already regarded as a case of GMM. The example shows another property of scores: The scores should be unforecastable. In the example, Et−1

·

¸ hε x i (xt − ρxt−1 ) xt−1 t t−1 = 0. = E t−1 σ2 σ2

(200)

Intuitively, if we used a combination of the x variables E(h(xt , xt−1 , ...)) = 0 that was predictable, we could form another moment – an instrument – that described the predictability of the h variable and use that moment to get more information about the parameters. To prove this property more generally, start with the fact that f (xt |xt−1 , xt−2 , ...; θ) is a conditional 250

S ECTION 14.3

WHEN FACTORS ARE RETURNS , ML PRESCRIBES A TIME - SERIES REGRESSION

density and therefore must integrate to one, Z 1 = f (xt |xt−1 , xt−2 , ...; θ)dxt Z ∂f (xt |xt−1 , xt−2 , ...; θ) dxt 0 = ∂θ Z ∂ ln f (xt |xt−1 , xt−2 , ...; θ) f (xt |xt−1 , xt−2 , ...; θ)dxt 0 = ∂θ · ¸ ∂ ln f (xt |xt−1 , xt−2 , ...; θ) 0 = Et−1 . ∂θ Furthermore, as you might expect, the GMM distribution theory formulas give the same result as the ML distribution, i.e., the information matrix is the asymptotic variance-covariance matrix. To show this fact, apply the GMM distribution theory (11.144) to (14.198). The derivative matrix is T ∂gT (θ) 1 X ∂ 2 ln f (xt |xt−1 xt−2 ...; θ) d= = =I 0 T t=1 ∂θ0 ∂θ∂θ

This is the second derivative expression of the information matrix. The S matrix is · ¸ ∂ ln f (xt |xt−1 xt−2 ...; θ) ∂ ln f (xt |xt−1 xt−2 ...; θ) 0 E =I ∂θ ∂θ The lead and lag terms in S are all zero since we showed above that scores should be unforecastable. This is the outer product definition of the information matrix. There is no a matrix, since the moments themselves are set to zero. The GMM asymptotic distribution of ˆ θ is therefore √ ¤ £ ¤ £ T (ˆθ − θ) → N 0, d−1 Sd−10 = N 0, I −1 .

We recover the inverse information matrix, as specified by the ML asymptotic distribution theory.

14.3

When factors are returns, ML prescribes a time-series regression

I add to the economic model E (Re ) = βE(f) a statistical assumption that the regression errors are independent over time and independent of the factors. ML then prescribes a timeseries regression with no constant. To prescribe a time series regression with a constant, we drop the model prediction α = 0. I show how the information matrix gives the same result as the OLS standard errors. 251

C HAPTER 14

MAXIMUM LIKELIHOOD

Given a linear factor model whose factors are also returns, as with the CAPM, ML prescribes a time-series regression test. To keep notation simple, I again treat a single factor f . The economic model is (201)

E (Re ) = βE(f )

Re is an N × 1 vector of test assets, and β is an N × 1 vector of regression coefficients of these assets on the factor (the market return Rem in the case of the CAPM).

To apply maximum likelihood, we need to add an explicit statistical model that fully describes the joint distribution of the data. I assume that the market return and regression errors are i.i.d. normal, i.e. Ret ft

(14.202)

= α + βft + εt = E(f ) + ut · ¸ µ· ¸ · ¸¶ εt 0 Σ 0 ∼N , ut 0 0 σ2u

(We can get by with non-normal factors, but it is easier not to present the general case.) Equation (14.202) has no content other than normality. The zero correlation between ut and εt identifies β as a regression coefficient. You can just write Re , Rem as a general bivariate normal, and you will get the same results. The economic model (14.201) implies restrictions on this statistical model. Taking expectations of (14.202), the CAPM implies that the intercepts α should all be zero. Again, this is also the only restriction that the CAPM places on the statistical model (14.202). The most principled way to apply maximum likelihood is to impose the null hypothesis throughout. Thus, we write the likelihood function imposing α = 0. To construct the likelihood function, we reduce the statistical model to independent error terms, and then add their log probability densities to get the likelihood function. T

L = (const.) −

T

1X e 1 X (ft − E(f ))2 (Rt − βft )0 Σ−1 (Ret − βft ) − 2 t=1 2 t=1 σ 2u

The estimates follow from the first order conditions, ∂L ∂β ∂L ∂E(f )

−1

= Σ

T X t=1

=

(Ret

ˆ= − βft ) ft = 0 ⇒ β

Ã T X t=1

ft2

!−1

T X

Ret ft

t=1

T T 1 X 1X [ ˆ (f − E(f)) = 0 ⇒ E(f) = λ = ft t σ 2u t=1 T t=1

(∂L/∂Σ and ∂L/∂σ2 also produce ML estimates of the covariance matrices, which turn out 252

S ECTION 14.3

WHEN FACTORS ARE RETURNS , ML PRESCRIBES A TIME - SERIES REGRESSION

to be the standard averages of squared residuals.) The ML estimate of β is the OLS regression without a constant. The null hypothesis says to leave out the constant, and the ML estimator uses that fact to avoid estimating a constant. Since the factor risk premium is equal to the market return, it’s not too surprising that the λ estimate is the same as that of the average market return. We know that the ML distribution theory must give the same result as the GMM distribution theory which we already derived in section 12.1, but it’s worth seeing it explicitly. The asymptotic standard errors follow from either estimate of the information matrix, for example T X ∂ 2L −1 = −Σ ft2 = 0 ∂β∂β 0 t=1

Thus, ˆ = cov(β)

1 1 1 1 Σ= Σ. T E(f 2 ) T E(f )2 + σ2 (f )

(203)

This is the standard OLS formula. We also want pricing error measurements, standard errors and tests. We can apply maximum likelihood to estimate an unconstrained model, containing intercepts, and then use Wald tests (estimate/standard error) to test the restriction that the intercepts are zero. We can also use the unconstrained model to run the likelihood ratio test. The unconstrained likelihood function is T

1X e L = (const.) − (R − α − βft )0 Σ−1 (Rte − α − βft ) + ... 2 t=1 t

(I ignore the term in the factor, since it will again just tell us to use the sample mean to estimate the factor risk premium.) The estimates are now ∂L ∂α

= Σ−1

∂L ∂β

= Σ−1

T X t=1

ˆ T (ft ) (Ret − α − βft ) = 0 ⇒ α ˆ = ET (Rte ) − βE

T X

e ˆ = covT (Rt , ft ) (Ret − α − βft ) ft = 0 ⇒ β σ2T (ft ) t=1

Unsurprisingly, the maximum likelihood estimates of α and β are the OLS estimates, with a constant. The inverse of the information matrix gives the asymptotic distribution of these estimates. Since they are just OLS estimates, we’re going to get the OLS standard errors, but it’s worth 253

C HAPTER 14 seeing it come out of ML.

MAXIMUM LIKELIHOOD

−1

∂ 2L · ¸ − £ ¤ α ∂ ∂ α β β

=

·

=

1 2 σ (f)

Σ−1 Σ−1 E(f ) −1 Σ E(f ) Σ−1 E(f 2 )

¸−1

¸

⊗Σ

·

E(f 2 ) E(f ) E(f ) 1

ˆ are thus The covariance matrices of α ˆ and β cov(ˆ α) = ˆ = cov(β)

" µ ¶2 # E(f ) 1 1+ Σ T σ(f ) 1 1 Σ. T σ 2 (f )

(14.204)

These are just the usual OLS standard errors, which we derived in section 12.1 as a special case of GMM standard errors for the OLS time-series regressions when errors are uncorrelated over time and independent of the factors, or by specializing σ2 (X 0 X)−1 . You cannot just invert ∂ 2 L/∂α∂α0 to find the covariance of α ˆ . That attempt would give ˆ , which would be wrong. You have to invert the entire just Σ as the covariance matrix of α information matrix to get the standard error of any parameter. Otherwise, you are ignoring ˆ . In fact, what I presented is really the effect that estimating β has on the distribution of α ˆ is independent of α ˆ and wrong, since we also must estimate Σ. However, it turns out that Σ ˆ β – the information matrix is block-diagonal – so the top left two elements of the true inverse information matrix are the same as I have written here. The variance of βˆ in (14.204) is larger than it is in (14.203) was when we impose the null of no constant. ML uses all the information it can to produce efficient estimates – estimates with the smallest possible covariance matrix. The ratio of the two formulas is equal to the familiar term 1 + E(f )2 /σ2 (f ). In annual data for the CAPM, σ(Rem ) = 16%, E(Rem ) = 8%, means that unrestricted estimate (14.204) has a variance 25% larger than the restricted estimate (14.203), so the gain in efficiency can be important. In monthly data, however the gain is smaller since variance and mean both scale with the horizon. We can also view this fact as a warning: ML will ruthlessly exploit the null hypothesis and do things like running regressions without a constant in order to get any small improvement in efficiency. We can use these covariance matrices to construct a Wald (estimate/standard error) test the restriction of the model that the alphas are all zero, Ã µ ¶2 !−1 E(f) T 1+ α ˆ 0 Σ−1 α ˆ ∼χ2N . (205) σ(f) 254

S ECTION 14.4

WHEN FACTORS ARE NOT EXCESS RETURNS , ML PRESCRIBES A CROSS - SECTIONAL REGRESSION

Again, we already derived this χ2 test in (12.168), and its finite sample F counterpart, the GRS F test (12.169). The other test of the restrictions is the likelihood ratio test (14.197). Quite generally, likelihood ratio tests are asymptotically equivalent to Wald tests, and so gives the same result. Showing it in this case is not worth the algebra.

14.4

When factors are not excess returns, ML prescribes a cross-sectional regression

If the factors are not returns, we don’t have a choice between time-series and cross-sectional regression, since the intercepts are not zero. As you might suspect, ML prescribes a crosssectional regression in this case. The factor model, expressed in expected return beta form, is E(Rei ) = αi + β 0i λ; i = 1, 2, ..N

(206)

The betas are defined from time-series regressions 0 i Rei t = ai + β i ft + εt

(207)

The intercepts ai in the time-series regressions need not be zero, since the model does not apply to the factors. They are not unrestricted, however. Taking expectations of the timeseries regression (14.207) and comparing it to (14.206) (as we did to derive the restriction α = 0 for the time-series regression), the restriction α = 0 implies ai = β 0i (λ − E(ft ))

(208)

Plugging into (14.207), the time series regressions must be of the restricted form Rtei = β 0i λ + β 0i [ft − E(ft )] + εit .

(209)

In this form, you can see that β 0i λ determines the mean return. Since there are fewer factors than returns, this is a restriction on the regression (14.209). Stack assets i = 1, 2, ...N to a vector; and introduce the auxiliary statistical model that the errors and factors are i.i.d. normal and uncorrelated with each other. Then, the restricted model is Rte = Bλ + B [ft − E(ft )] + εt ft = E(f) + ut · ¸ µ ¶ εt Σ 0 ∼ N 0, ut 0 V

where B denotes a N × K matrix of regression coefficients of the N assets on the K factors. 255

C HAPTER 14

MAXIMUM LIKELIHOOD

The likelihood function is T

L = (const.) − εt

T

1 X 0 −1 1 X 0 −1 εt Σ εt − u V ut 2 t=1 2 t=1 t

= Rte − B [λ + ft − E(f)] ; ut = ft − E(f).

Maximizing the likelihood function, ∂L ∂E(f ) ∂L ∂λ

: 0=

T X

0

BΣ

t=1

: 0 = B0

−1

T X t=1

(Ret −B [λ

+ ft − E(f )]) +

T X t=1

V −1 (ft − E(f))

Σ−1 (Ret − B [λ + ft − E(f )])

The solution to this pair of equations is [) = ET (ft ) E(f ¡ ¢ ˆ = B 0 Σ−1 B −1 B 0 Σ−1 ET (Re ) . λ t

(14.210) (14.211)

The maximum likelihood estimate of the factor risk premium is a GLS cross-sectional regression of average returns on betas.

The maximum likelihood estimates of the regression coefficients B are again not the same as the standard OLS formulas. Again, ML imposes the null to improve efficiency. ∂L ∂B ˆ B

:

T X t=1

=

T X

Σ−1 (Rte −B [λ + ft − E(f)]) [λ + ft − E(f )]0 = 0 Ret

t=1

[ft + λ − E(f)]

0

Ã T X t=1

(14.212) 0

[ft + λ − E(f)] [ft + λ − E(f )]

!−1

This is true, even though the B are defined in the theory as population regression coefficients. (The matrix notation hides a lot here! If you want to rederive thesePformulas, it’s helpful to start with scalar parameters, e.g. Bij , and to think of it as ∂L/∂θ = Tt=1 (∂L/∂εt )0 ∂εt /∂θ. ) Therefore, to really implement ML, you have to solve (14.211) and (14.212) simultaneously ˆ, B ˆ , along with Σ ˆ whose ML estimate is the usual second moment matrix of the residfor λ ˆ , run an OLS cross-sectional uals. This can usually be done iteratively: Start with OLS B ˆ , form Σ ˆ , and iterate. regression for λ

14.5

Problems

256

S ECTION 14.5 1.

2.

P ROBLEMS

Why do we use restricted ML when the factor is a return, but unrestricted ML when the factor is not a return? To see why, try to formulate a ML estimator based on an unrestricted regression when factors are not returns, equation (12.166). Add pricing errors αi to the regression as we did for the unrestricted regression in the case that factors are returns, and then find ML estimators for B, λ, α, E(f ). (Treat V and Σ as known to make the problem easier.) Instead of writing a regression, build up the ML for the CAPM a little more formally. Write the statistical model as just the assumption that individual returns and the market return are jointly normal, ¸ · ¸ µ· ¶ Re E(Re ) Σ cov(Rem , Re0 ) , ∼N Rem σ2m E(Rem ) cov(Rem , Re ) The model’s restriction is E(Re ) = γcov(Rem , Re ).

Estimate γ and show that this is the same time-series estimator as we derived by presupposing a regression.

257

Chapter 15. Time series, cross-section, and GMM/DF tests of linear factor models The GMM/DF, time-series and cross-sectional regression procedures and distribution theory are similar, but not identical. Cross-sectional regressions on betas are not the same thing as cross sectional regressions on second moments. Cross-sectional regressions weighted by the residual covariance matrix are not the same thing as cross-sectional regressions weighted by the spectral density matrix. GLS cross-sectional regressions and second stage GMM have a theoretical efficiency advantage over OLS cross sectional regressions and first stage GMM, but how important is this advantage, and is it outweighed by worse finite-sample performance? The time-series regression, as ML estimate, has a potential gain in efficiency when returns are factors and the residuals are i.i.d. normal. Why does ML prescribe a time-series regression when the return is a factor and a cross-sectional regression when the return is not a factor? The time-series regression seems to ignore pricing errors and estimate the model by entirely different moments. How does adding one test asset make such a seemingly dramatic difference to the procedure? Finally, and perhaps most importantly, the GMM/discount factor approach is still a “new” procedure. Many authors still do not trust it. It is important to verify that it produces similar results and well-behaved test statistics in the setups of the classic regression tests. To address these questions, I first apply the various methods to a classic empirical question. How do time-series regression, cross-sectional regression and GMM/stochastic discount factor compare when applied to a test of the CAPM on CRSP size portfolios? I find that three methods produce almost exactly the same results for this classic exercise. They produce almost exactly the same estimates, standard errors, t-statistics and χ2 statistics that the pricing errors are jointly zero. Then I conduct a Monte Carlo and Bootstrap evaluation. Again, I find little difference between the methods. The estimates, standard errors, and size and power of tests are almost identical across methods. The Bootstrap does reveal that the traditional i.i.d. assumption generates χ2 statistics with about 1/2 the correct size – they reject half as often as they should under the null. Simple GMM corrections to the distribution theory repair this size defect. Also, you can ruin any estimate and test with a bad spectral density matrix estimate. I try an estimate with 24 lags and no Newey-West weights. It is singular in the data sample and many Monte Carlo replications. Interestingly, this singularity has minor effects on standard errors, but causes disasters when you use the spectral density matrix to weight a second-stage GMM. I also find that second stage “efficient” GMM is only very slightly more efficient than first stage GMM, but is somewhat less robust; it is more sensitive to the poor spectral density matrix and its asymptotic standard errors can be slightly misleading. As OLS is often better 258

S ECTION 15.1

T HREE APPROACHES TO THE CAPM IN SIZE PORTFOLIOS

than GLS, despite the theoretical efficiency advantage of GLS, first-stage GMM may be better than second stage GMM in many applications. This section should give comfort that the apparently “new” GMM/discount factor formulation is almost exactly the same as traditional methods in the traditional setup. There is a widespread impression that GMM has difficulty in small samples. The literature on the small sample properties of GMM (for example, Ferson and Foerster, 1994, Fuhrer, Moore, and Schuh, 1995) naturally tries hard setups, with highly nonlinear models, highly persistent and heteroskedastic errors, conditioning information, potentially weak instruments and so forth. Nobody would write a paper trying GMM in a simple situation such as this one, correctly foreseeing that the answer would not be very interesting. Unfortunately, many readers take from this literature a mistaken impression that GMM always has difficulty in finite samples, even in very standard setups. This is not the case. The point of the GMM/discount factor method, of course, is not a new way to handle the simple i.i.d. normal CAPM problems, which are already handled efficiently by regression techniques. The point of the GMM/discount factor method is its ability to transparently handle situations that are very hard with expected return - beta models and ML techniques, including the incorporation of conditioning information and nonlinear models. With the reassurance of this section, we can proceed to those more exiting applications. Cochrane (2000) presents a more in-depth analysis, including estimation and Monte Carlo evaluation of individual pricing error estimates and tests. Jagannathan and Wang (2000) compare the GMM/discount factor approach to classic regression tests analytically. They show that the parameter estimates, standard errors and χ2 statistics are asymptotically identical to those of an expected return- beta cross-sectional regression when the factor is not a return.

15.1

Three approaches to the CAPM in size portfolios

The time-series approach sends the expected return - beta line through the market return, ignoring other assets. The OLS cross -sectional regression minimizes the sum of squared pricing errors, so allows some market pricing error to fit other assets better. The GLS crosssectional regression weights pricing errors by the residual covariance matrix, so reduces to the time-series regression when the factor is a return and is included in the test assets. The GMM/discount factor estimates, standard errors and χ2 statistics are very close to time-series and cross-sectional regression estimates in this classic setup. Time series and cross section

Figures 28 and 29 illustrate the difference between time-series and cross-sectional regressions, in an evaluation of the CAPM on monthly size portfolios.

259

C HAPTER 15

T IME SERIES , CROSS - SECTION , AND GMM/DF TESTS OF LINEAR FACTOR MODELS

Figure 28 presents the time-series regression. The time-series regression estimates the factor risk premium from the average of the factor, ignoring any information in the other ˆ = ET (Rem ). Thus, a time-series regression draws the expected return-beta line assets, λ across assets by making it fit precisely on two points, the market return and the riskfree rate– The market and riskfree rate have zero estimated pricing error in every sample. (The far right portfolios are the smallest firm portfolios, and their positive pricing errors are the small firm anomaly – this data set is the first serious failure of the CAPM. I come back to the substantive issue in Chapter 20.) The time-series regression is the ML estimator in this case, since the factor is a return. Why does ML ignore all the information in the test asset average returns, and estimate the factor premium from the average factor return only? The answer lies in the structure that we told ML to assume when looking at the data. When we write Re = a + βft + εt and ε independent of f , we tell ML that a sample of returns already includes the same sample of the factor, plus extra noise. Thus, the sample of test asset returns cannot possibly tell ML anything more than the sample of the factor alone about the mean of the factor. Second, we tell ML that the factor risk premium equals the mean of the factor, so it may not consider the possibility that the two are different in trying to match the data.

Figure 28. Average excess returns vs. betas on CRSP size portfolios, 1926-1998. The line gives the predicted average return from the time-series regression, E(Re ) = βE(Rem ). The OLS cross-sectional regression in Figure 29 draws the expected return-beta line by 260

S ECTION 15.1

T HREE APPROACHES TO THE CAPM IN SIZE PORTFOLIOS

Figure 29. Average excess returns vs betas of CRSP size portfolios 1926-1998, and the fit of cross-sectional regressions. minimizing the squared pricing error across all assets. Therefore, it allows some pricing error for the market return, if by doing so the pricing errors on other assets can be reduced. Thus, the OLS cross-sectional regression gives some pricing error to the market return in order to lower the pricing errors of the other portfolios. When the factor is not also a return, ML prescribes a cross-sectional regression. ML still [ = ET (ft ). ignores anything but the factor data in estimating the mean of the factor–E(f) However, ML is now allowed to us a different parameter for the factor risk premium that fits average returns to betas, which it does by cross-sectional regression. However, ML is a GLS cross sectional regression, not an OLS cross-sectional regression. The GLS cross-sectional regression in Figure 29 is almost exactly identical to the time-series regression result – it passes right through the origin and the market return ignoring all the other pricing errors. The GLS cross-sectional regression ¡ ¢ ˆ = β 0 Σ−1 β −1 β 0 Σ−1 ET (Re ). λ

weights the various portfolios by the inverse of the residual covariance matrix Σ. If we include the market return as a test asset, it obviously has no residual variance–Rem = 0+1×Rem t t +0– so the GLS estimate pays exclusive attention to it in fitting the market line. The same thing 261

C HAPTER 15

T IME SERIES , CROSS - SECTION , AND GMM/DF TESTS OF LINEAR FACTOR MODELS

happens if the test assets span the factors – if a linear combination of the test assets is equal to the factor and hence has no residual variance. The size portfolios nearly span the market return, so the GLS cross-sectional regression is visually indistinguishable from the timeseries regression in this case. This observation wraps up one mystery, why ML seemed so different when the factor is and is not a return. As the residual variance of a portfolio goes to zero, the GLS regression pays more and more attention to that portfolio, until you have achieved the same result as a time-series regression. If we allow a free constant in the OLS cross-sectional regression, thus allowing a pricing error for the risk free rate, you can see from Figure 29 that the OLS cross-sectional regression line will fit the size portfolios even better, though allowing a pricing error in the risk free rate as well as the market return. However, a free intercept in an OLS regression on excess returns puts no weight at all on the intercept pricing error. It is a better idea to include the riskfree rate as a test asset, either directly by doing the whole thing in levels of returns rather than excess returns or by adding E(Re ) = 0, β = 0 to the cross-sectional regression. The GLS cross-sectional regression will notice that the T-bill rate has no residual variance and so will send the line right through the origin, as it does for the market return. GMM/discount factor first and second stage

Figure 30 illustrates the GMM/discount factor estimate with the same data. The horizontal axis is the second moment of returns and factors rather than beta, but you would not know it from the placement of the dots. The first stage estimate is an OLS cross-sectional regression of average returns on second moments. It minimizes the sum of squared pricing errors, and so produces pricing errors almost exactly equal to those of the OLS cross-sectional regression of returns on betas. The second stage estimate minimizes pricing errors weighted by the spectral density matrix. The spectral density matrix is not the same as the residual covariance matrix, so the second stage GMM does not go through the market portfolio as does the GLS cross-sectional regression. In fact, the slope of the line is slightly higher for the second stage estimate. (The spectral density matrix of the discount factor formulation does not reduce to the residual covariance matrix even if we assume the regression model, the asset pricing model is true, and factors and residuals are i.i.d. normal. In particular, when the market is a test asset, the GLS cross-sectional regression focuses all attention on the market portfolio but the second stage GMM/DF does not do so. The parameter b is related to λ by b = λ/E(Rem2 ). The other assets still are useful in determining the parameter b, even though Given the market return and the regression model Rtei = β i Rtem + εit , seeing the other assets does not help to determine the mean of the market return, ) Overall, the figures do not suggest any strong reason to prefer first and second stage GMM/discount factor, time-series, OLS or GLS cross sectional regression in this standard model and data set. The results are affected by the choice of method. In particular, the size of the small firm anomaly is substantially affected by how one draws the market line. But 262

S ECTION 15.1

T HREE APPROACHES TO THE CAPM IN SIZE PORTFOLIOS

Figure 30. Average excess return vs. predicted value of 10 CRSP size portfolios, 1926-1998, based on GMM/SDF estimate. The model predicts E(Re ) = bE(Re Rem ). The second stage estimate of b uses a spectral density estimate with zero lags. the graphs and analysis do not strongly suggest that any method is better than any other for purposes other than fishing for the answer one wants. Parameter estimates, standard errors, and tests

Table c1 presents the parameter estimates and standard errors from time-series, crosssection, and GMM/discount factor approach in the CAPM size portfolio test illustrated by Figures 28 and 29. The main parameter to be estimated is the slope of the lines in the above figures, the market price of risk λ in the expected return-beta model and the relation between mean returns and second moments b in the stochastic discount factor model. The big point of Table c1 is that the GMM/discount factor estimate and standard errors behave very similarly to the traditional estimates and standard errors. The rows compare results with various methods of calculating the spectral density matrix. i.i.d. imposes no serial correlation and regression errors independent of right hand variables, and is identical to the Maximum Likelihood based formulas. The 0 lag estimate allows conditional heteroskedasticity, but no correlation of residuals. The 3 lag, Newey West estimate is a sensible correction for short order autocorrelation. I include the 24 lag spectral density matrix to show how things can go wrong if you use a ridiculous spectral density matrix. 263

C HAPTER 15

Estimate i.i.d. 0 lags 3 lags, NW 24 lags

T IME SERIES , CROSS - SECTION , AND GMM/DF TESTS OF LINEAR FACTOR MODELS

TimeSeries 0.66 0.18 (3.67) 0.18 (3.67) 0.20 (3.30) 0.16 (4.13)

Beta model λ Cross section OLS 0.71 0.20 (3.55) 0.19 (3.74) 0.21 (3. 38) 0.16 (4.44)

GLS 0.66 0.18 (3.67) 0.18 (3.67) 0.20 (3.30) 0.16 (4.13)

1st stage 2.35

GMM/DF b 2nd stage Est. Std. Err.

0.63 (3.73) 0.69 (3.41) 1.00 (2.35)

2.46 2.39 2.15

0.61 (4.03) 0.64 (3.73) 0.69 (3.12)

Table c1. Parameter estimates and standard errors. Estimates are shown in italic, standard errors in regular type, and t-statistics in parentheses. The time-series estimate is the mean market return in percent per month. The cross-sectional estimate is the slope coefficient λ in E(Re ) = βλ. The GMM estimate is the parameter b in E(Re ) = E(Re f )b. CRSP monthly data 1926-1998. “Lags” gives the number of lags in the spectral density matrix. “NW” uses Newey-West weighting in the spectral density matrix. The OLS cross-sectional estimate 0.71 is a little higher than the mean market return 0.66, in order to better fit all of the assets, as seen in Figure 29. The GLS cross-sectional estimate is almost exactly the same as the mean market return, and the GLS standard errors are almost exactly the same as the time-series standard errors. The Shanken correction for generated regressors is very important to standard errors of the cross-sectional regressions. Without the Σf term in the standard deviation of λ (12.184)– i.e. treating the β as fixed right hand variables – the standard errors come out to 0.07 for OLS and √ 0.00 for GLS – far less than the correct 0.20 and 0.18 shown in the table, and far less than σ/ T . The b estimates are not directly comparable to the risk premium estimates, but it is easy to translate their units. Applying the discount factor model with normalization a = 1 to the market return itself, b=

E(Rem ) . E(Rem2 )

With E(Rem ) = 0.66% and σ(Rem ) = 5.47%, we have 100 × b = 100 (0.66) /(0.662 + 5.472 ) = 2.17. The entries in Table c1 are close to this magnitude. Most are slightly larger, as is the OLS cross-sectional regression, in order to better fit the other portfolios. The t-statistics are quite close across methods, which is another way to correct the units. The second-stage GMM/DF estimates (as well as standard errors) depend on which spectral density weighting matrix is used as a weighting matrix. The results are quite similar for all the sensible spectral density estimates. The 24 lag spectral density matrix starts to produce unusual estimates. This spectral density estimate will cause lots of problems below. Table c2 presents the χ2 and F statistics that test whether the pricing errors are jointly significant. The OLS and GLS cross-sectional regression, and the first and second stage GMM/discount factor tests give exactly the same χ2 statistic, though the individual pricing 264

S ECTION 15.2

MONTE CARLO AND B OOTSTRAP

errors and covariance matrix are not the same so I do not present them separately. The big point of Table c2 is that the GMM/discount factor method gives almost exactly the same result as the cross-sectional regression.

i.i.d. GRS F 0 lags 3 lags NW 24 lags

Time series χ2(10) % p 8.5 58 0.8 59 10.5 40 11.0 36 -432 -100

Cross section χ2(9) %p 8.5 49

GMM/DF χ2(9) % p

10.6 11.1 7.6

10.5 11.1 7.7

31 27 57

31 27 57

Table c2. χ2 tests that all pricing errors are jointly equal to zero. For the time-series regression, the GRS F test gives almost exactly the same rejection probability as does the asymptotic χ2 test. Apparently, the advantages of a statistic that is valid in finite samples is not that important in this data set. The χ2 tests for the time-series case without the i.i.d. assumption are a bit more conservative, with 30-40% p value rather than almost 60%. However, this difference is not large. The one exception is the χ2 test using 24 lags and no weights in the spectral density matrix. That matrix turns out not to be positive definite in this sample, with disastrous results for the χ2 statistic. (Somewhat surprisingly, the CAPM is not rejected. This is because the small firm effect vanishes in the latter part of the sample. I discuss this fact further in Chapter 20. See in particular Figure 28.) Looking across the rows, the χ2 statistic is almost exactly the same for each method. The cross-sectional regression and GMM/DF estimate have one lower degree of freedom (the market premium is estimated from the cross-section rather than from the market return), and so show slightly greater rejection probabilities. For a given spectral density estimation technique, the cross-sectional regression and the GMM/DF approach give almost exactly the same χ2 values and rejection probabilities. The 24 lag spectral density matrix is a disaster as usual. In this case, it is a greater disaster for the time-series test than for the cross-section or GMM/discount factor test. It turns out not to be positive definite, so the sample pricing errors ˆ 0 cov(ˆ α)−1 α ˆ produce a nonsensical negative value of α

15.2

Monte Carlo and Bootstrap

The parameter distribution for the time-series regression estimate is quite similar to that from the GMM/discount factor estimate. The size and power of χ2 test statistics is nearly identical for time-series regression test 265

C HAPTER 15

T IME SERIES , CROSS - SECTION , AND GMM/DF TESTS OF LINEAR FACTOR MODELS

and the GMM/discount factor test. A bad spectral density matrix can ruin either time-series or GMM/discount factor estimates and tests. There is enough serial correlation and heteroskedasticity in the data that conventional i.i.d. formulas produce test statistics with about 1/2 the correct size. If you want to do classic regression tests, you should correct the distribution theory rather than use the ML i.i.d. distributions. Econometrics is not just about sensible point estimates, it is about sampling variability of those estimates, and whether standard error formulas correctly capture that sampling variability. How well do the various standard error and test statistic formulas capture the true sampling distribution of the estimates? To answer this question I conduct two Monte Carlos and two bootstraps. I conduct one each under the null that the CAPM is correct, to study size, and one each under the alternative that the CAPM is false, to study power. The Monte Carlo experiments follow the standard ML assumption that returns and factors are i.i.d. normally distributed, and the factors and residuals are independent as well as uncorrelated. I generate artificial samples of the market return from an i.i.d. normal, using the sample mean and variance of the value-weighted return. I then generate artificial size em + εit , using the sample residual covaridecile returns under the null by Rei t = 0 + β i Rt ance matrix Σ to draw i.i.d. normal residuals εit and the sample regression coefficients β i . To generate data under the alternative, I add the sample αi . draw 5000 artificial samples. I try a long sample of 876 months, matching the CRSP sample analyzed above. I also draw a short sample of 240 months or 20 years, which is about as short as one should dare try to test a factor model. The bootstraps check whether non-normalities, autocorrelation, heteroskedasticity, and non-independence of factors and residuals matters to the sampling distribution in this data set. I do a block-bootstrap, resampling the data in groups of three months with replacement, to preserve the short-order autocorrelation and persistent heteroskedasticity in the data. To impose the CAPM, I draw the market return and residuals in the time-series regression, and then compute artificial data on decile portfolio returns by Rtei = 0 + β i Rtem + εit . To study the alternative, I simply redraw all the data in groups of three. Of course, the actual data may display conditioning information not displayed by this bootstrap, such as predictability and conditional heteroskedasticity based on additional variables such as the dividend/price ratio, lagged squared returns, or implied volatilities. The first-stage GMM/discount factor and OLS cross-sectional regression are nearly identical in every artificial sample, as the GLS cross-sectional regression is nearly identical to the time-series regression in every sample. Therefore, the important question is to compare the time series regression – which is ML with i.i.d. normal returns and factors – to the first and second stage GMM/DF procedure. For this reason and to save space, I do not include the cross-sectional regressions in the Monte Carlo and bootstrap. 266

S ECTION 15.2

MONTE CARLO AND B OOTSTRAP

χ2 tests

Table 6c presents the χ2 tests of the hypothesis that all pricing errors are zero under the null that the CAPM is true, and Table 7c presents the χ2 tests under the null that the CAPM is false. Each table presents the percentage of the 5000 artificial data sets in which the χ2 tests rejected the null at the indicated level. The central point of these tables is that the GMM/discount factor test performs almost exactly the same way as the time-series test. Compare the GMM/DF entry to its corresponding Time series entry; they are all nearly identical. Neither the small efficiency advantage of time-series vs. cross section, nor the difference between betas and second moments seems to make any difference to the sampling distribution.

Sample size: level (%): i.i.d. 0 lags 3 lags, NW 24 lags

Monte Carlo Time series GMM/DF 240 876 240 876 5 5 1 5 5 1 7.5 6.0 1.1 7.7 6.1 1.1 7.5 6.3 1.0 10.7 6.5 1.4 9.7 6.6 1.3 25 39 32 25 41 31

Block-Bootstrap Time series GMM/DF 240 876 240 876 5 5 1 5 5 1 6.0 2.8 0.6 7.7 4.3 1.0 6.6 3.7 0.9 10.5 5.4 1.3 9.5 5.3 1.3 23 38 31 24 41 32

Table 6c. Size. Probability of rejection for χ2 statistics under the null that all pricing errors are zero

Sample size: level (%): i.i.d. 0 lags 3 lags, NW 24 lags

Monte Carlo Time-Series GMM/DF 240 876 240 876 5 1 5 1 17 48 26 17 48 26 17 50 27 22 49 27 21 51 29 29 60 53 29 66 57

Block-Bootstrap Time-Series GMM/DF 240 876 240 876 5 1 5 1 11 40 18 15 54 28 14 55 29 18 57 31 17 59 33 27 63 56 29 68 60

Table 7c. Power. Probability of rejection for χ2 statistics under the null that the CAPM is false, and the true means of the decile portfolio returns are equal to their sample means. Start with the Monte Carlo evaluation of the time-series test in Table 6c. The i.i.d. and 0 lag distributions produce nearly exact rejection probabilities in the long sample and slightly too many (7.5%) rejections in the short sample. Moving down, GMM distributions here correct for things that aren’t there. This has a small but noticeable effect on the sensible 3 lag test, which rejects slightly too often under this null. Naturally, this is worse for the short sample, but looking across the rows, the time-series and discount factor tests are nearly identical in every case. The variation across technique is almost zero, given the spectral 267

C HAPTER 15

T IME SERIES , CROSS - SECTION , AND GMM/DF TESTS OF LINEAR FACTOR MODELS

density estimate. The 24 lag unweighted spectral density is the usual disaster, rejecting far too often. It is singular in many samples. In the long sample, the 1% tail of this distribution occurs at a χ2 value of 440 rather than the 23.2 of the χ2(10) distribution! The long sample block-bootstrap in the right half of the tables shows even in this simple setup how i.i.d. normal assumptions can be misleading. The traditional i.i.d. χ2 test has almost half the correct size – it rejects a 10% test 6% of the time, a 5% test 2.8% of the time and a 1% test 0.6% of the time. Removing the assumption that returns and factors are independent, going from i.i.d. to 0 lags, brings about half of the size distortion back, while adding one of the sensible autocorrelation corrections does the rest. In each row, the time-series and GMM/DF methods produce almost exactly the same results again. The 24 lag spectral density matrices are a disaster as usual. Table 7c shows the rejection probabilities under the alternative. The most striking feature of the table is that the GMM/discount factor test gives almost exactly the same rejection probability as the time-series test, for each choice of spectral density estimation technique. When there is a difference, the GMM/discount factor test rejects slightly more often. The 24 lag tests reject most often, but this is not surprising given that they reject almost as often under the null. Parameter estimates and standard errors

Table c5 presents the sampling variation of the λ and b estimates. The rows and columns ˆ , σ(ˆb), and in italic font, give the variation of the estimated λ or b across the market σ(λ) 5000 artificial samples. The remaining rows and columns give the average across samples of the standard errors. The presence of pricing errors has little effect on the estimated b or λ and their standard errors, so I only present results under the null that the CAPM is true. The parameters are not directly comparable – the b parameter includes the variance as well as the mean of the factor, and ET (Rem ) is the natural GMM estimate of the mean market return as it is the Time-series estimate of the factor risk premium. Still, it is interesting to know and to compare how well the two methods do at estimating their central parameter.

268

S ECTION 15.2

Time series T=876: ˆ σ(λ), σ(ˆb) i.i.d. 0 lags 3 lags NW 24 lags T=240: ˆ , σ(ˆb) σ(λ) i.i.d. 0 lags 3 lags NW 24 lags

MONTE CARLO AND B OOTSTRAP

Monte Carlo GMM/DF 1st 2nd stage stage σ(ˆb) E (s.e.)

0.19 0.18 0.18 0.18 0.18

0.64

0.35 0.35 0.35 0.35 0.29

1.25

0.65 0.65 0.62

1.23 1.22 1.04

0.61 0.62 130

1.24 1.26 191

Time series

Block-Bootstrap GMM/DF 1st 2nd stage stage σ(ˆb) E (s.e.) 0.69

0.60 0.59 0.27

0.20 0.18 0.18 0.19 0.19

1.40

1.14 1.11 0.69

0.37 0.35 0.35 0.36 0.31

0.63 0.67 0.66

1.24 1.31 1.15

0.67 0.67 1724

0.60 0.62 0.24

1.45 1.48 893

1.15 1.14 0.75

Table 5. Monte Carlo and block-bootstrap evaluation of the sampling variability of parameter estimates b and λ. The Monte Carlo redraws 5000 artificial data sets of length T=876 from a random normal assuming that the CAPM is true. The blockbootstrap redraws the data in groups of 3 with replacement. The row and columns ˆ and σ(ˆb) and using italic font give the variation across samples of the marked σ(λ) estimated λ and b. The remaining entries of “Time series” “1st stage” and “E (s.e.)” columns in roman font give the average value of the computed standard error of the parameter estimate, where the average is taken over the 5000 samples. The central message of this table is that the GMM/DF estimates behave almost exactly as the time-series estimate, and the asymptotic standard error formulas almost exactly capture the sampling variation of the estimates. The second stage GMM/DF estimate is a little bit more efficient at the cost of slightly misleading standard errors. Start with the long sample and the first column. All of the standard error formulas give essentially identical and correct results for the time-series estimate. Estimating the sample mean is not rocket science. The first stage GMM/DF estimator in the second column behaves the same way, except the usually troublesome 24 lag unweighted estimate. The second stage GMM/DF estimate in the third and fourth columns uses the inverse spectral density matrix to weight, and so the estimator depends on the choice of spectral density estimate. The sensible spectral density estimates (not 24 lags) produce second-stage estimates that vary less than the first-stage estimates, 0.61 − 0.62 rather than 0.64. Second stage GMM is more efficient, meaning that it produces estimates with smaller sampling variation. However, the table shows that the efficiency gain is quite small, so not much is lost if one prefers first stage OLS estimates. The sensible spectral density estimates produce second-stage standard errors that again almost exactly capture the sampling variation of the 269

C HAPTER 15

T IME SERIES , CROSS - SECTION , AND GMM/DF TESTS OF LINEAR FACTOR MODELS

estimated parameters. The 24 lag unweighted estimate produces hugely variable estimates and artificially small standard errors. Using bad or even singular spectral density estimates seems to have a secondary effect on standard error calculations, but using its inverse as a weighting matrix can have a dramatic effect on estimation. With the block-bootstrap in the right hand side of Table 5c, the time-series estimate is slightly more volatile as a result of the slight autocorrelation in the market return. The i.i.d. and zero lag formulas do not capture this effect, but the GMM standard errors that allow autocorrelation do pick it up. However, this is a very minor effect as there is very little autocorrelation in the market return. The effect is more pronounced in the first stage GMM/DF estimate, since the smaller firm portfolios depart more from the normal i.i.d. assumption. The true variation is 0.69, but standard errors that ignore autocorrelation only produce 0.63. The standard errors that correct for autocorrelation are nearly exact. In the second-stage GMM/DF, the sensible spectral density estimates again produce slightly more efficient estimates than the first stage, with variation of 0.67 rather than 0.69. This comes at a cost, though, that the asymptotic standard errors are a bit less reliable. In the shorter sample, we see that standard errors for the mean market return in the Time series column are all quite accurate, except the usual 24 lag case. In the GMM/DF case, we see that the actual sampling variability of the b estimate is no longer smaller for the second stage. The second stage estimate is not more efficient in this “small” sample. Furthermore, while the first stage standard errors are still decently accurate, the second stage standard errors substantially understate the true sampling variability of the parameter estimate. They represent a hoped-for efficiency that is not present in the small sample. Even in this simple setup, first-stage GMM is clearly a better choice for estimating the central parameter, and hence for examining individual pricing errors and their pattern across assets.

270

Chapter 16.

Which method?

Of course, the point of GMM/discount factor methods is not a gain in efficiency or simplicity in a traditional setup – linear factor model, i.i.d. normally distributed returns and factors, etc. It’s hard to beat the efficiency or simplicity of regression methods in those setups. The point of the GMM/discount factor approach is that it allows a simple technique for evaluating nonlinear or otherwise complex models, for including conditioning information while not requiring the econometrician to see everything that the agent sees, and for allowing the researcher to circumvent inevitable model misspecifications or simplifications and data problems by keeping the econometrics focused on interesting issues. The alternative is usually some form of maximum likelihood. This is much harder in most circumstances, since you have to write down a complete statistical model for the conditional distribution of your data. Just evaluating, let alone maximizing, the likelihood function is often challenging. Whole series of papers are written just on the econometric issues of particular cases, for example how to maximize the likelihood functions of specific classes of univariate continuous time models for the short interest rate. Of course, there is no necessary pairing of GMM with the discount factor expression of a model, and ML with the expected return-beta formulation. Many studies pair discount factor expressions of the model with ML, and many others evaluate expected return-beta model by GMM, as we have done in adjusting regression standard errors for non-i.i.d. residuals. Advanced empirical asset pricing faces an enduring tension between these two philosophies. The choice essentially involves tradeoffs between statistical efficiency, the effects of misspecification of both the economic and statistical models, and the clarity and economic interpretability of the results. There are situations in which it’s better to trade some small efficiency gains for the robustness of simpler procedures or more easily interpretable moments; OLS can be better than GLS. The central reason is specification errors; the fact that our statistical and economic models are at best quantitative parables. There are other situations in which one may really need to squeeze every last drop out of the data, intuitive moments are statistically very inefficient, and more intensive maximum-likelihood approaches are more appropriate. Unfortunately, the environments are complex, and differ from case to case. We don’t have universal theorems from statistical theory or generally applicable Monte Carlo evidence. Specification errors by their nature resist quantitative modeling – if you knew how to model them, they wouldn’t be there. We can only think about the lessons of past experiences. In my experience, in the limited range of applications I have worked with, a GMM approach based on simple easily interpretable moments has proved far more fruitful than formal maximum likelihood. In addition, I have found first stage GMM – OLS cross sectional regressions – to be more trustworthy than second-stage GMM, in any case where there was a substantial difference between the two approaches. The rest of this chapter collects some thoughts on the choice between formal ML and less formal GMM, focusing on economically interesting rather than statistically informative moments. 271

CHAPTER 16

WHICH METHOD ?

“ML” vs. “GMM”

The debate is often stated as a choice between “maximum likelihood” and “GMM.” This is a bad way to put the issue. ML is a special case of GMM: it suggests a particular choice of moments that are statistically optimal in a well-defined sense. Given the set of moments, the distribution theories are identical. Also, there is no such thing as “the” GMM estimate. GMM is a flexible tool; you can use any aT matrix and gT moments that you want to use. For example, we saw how to use GMM to derive the asymptotic distribution of the standard time-series regression estimator with autocorrelated returns. The moments in this case were not the pricing errors. It’s all GMM; the issue is the choice of moments. Both ML and GMM are tools that a thoughtful researcher can use in learning what the data says about a given asset pricing model, rather than as stone tablets giving precise directions that lead to truth if followed literally. If followed literally and thoughtlessly, both ML and GMM can lead to horrendous results. The choice is between moments selected by an auxiliary statistical model, even if completely economically uninterpretable, and moments selected for their economic or data summary interpretation, even if not statistically efficient. ML is often ignored

As we have seen, ML plus the assumption of normal i.i.d. disturbances leads to easily interpretable time-series or cross-sectional regressions, empirical procedures that are close to the economic content of the model. However, asset returns are not normally distributed or i.i.d.. They have fatter tails than a normal, they are heteroskedastic (times of high and times of low volatility), they are autocorrelated, and predictable from a variety of variables. If one were to take seriously the ML philosophy and its quest for efficiency, one should model these features of returns. The result would be a different likelihood function, and its scores would prescribe different moment conditions than the familiar and intuitive time-series or cross-sectional regressions. Interestingly, few empirical workers do this. (The exceptions tend to be papers whose primary point is illustration of econometric technique rather than empirical findings.) ML seems to be fine when it suggests easily interpretable regressions; when it suggests something else, people use the regressions anyway. For example, ML prescribes that one estimate β s without a constant. β s are almost universally estimated with a constant. Researchers often run cross-sectional regressions rather than time-series regressions, even when the factors are returns. ML specifies a GLS cross-sectional regression, but many empirical workers use OLS cross-sectional regressions instead, distrusting the GLS weighting matrix. Time-series regressions are almost universally run with a constant, though ML prescribes a regression with no constant. The true ML formulas for GLS regressions require one to iterate between non-OLS formulas for betas, covariance matrix estimate and the cross-sectional regression estimate. Empirical applications usually use the unconstrained estimates of all these quantities. And of course, any of the regression tests continue to be run at all, with ML justifications, despite the fact that returns are not i.i.d. The 272

regressions came first, and the maximum likelihood formalization came later. If we had to assume that returns had a gamma distribution to justify the regressions, it’s a sure bet that we would make that “assumption” behind ML instead of the normal i.i.d. assumption! The reason must be that researchers feel that omitting some of the information in the null hypothesis, the estimation and test is more robust, though some efficiency is lost if the null economic and statistical models are exactly correct. Researchers must not really believe that their null hypotheses, statistical and economic, are exactly correct. They want to produce estimates and tests that are robust to reasonable model mis-specifications. They also want to produce estimates and tests that are easily interpretable, that capture intuitively clear stylized facts in the data, and that relate directly to the economic concepts of the model. Such estimates are persuasive in large part because the reader can see that they are robust. (And following this train of thought, one might want to pursue estimation strategies that are even more robust than OLS, since OLS places a lot of weight on outliers. For example, Chen and Ready 1997 claim that Fama and French’s 1993 size and value effects depend crucially on a few outliers.) ML does not necessarily produce robust or easily interpretable estimates. It wasn’t designed to do so. The point and advertisement of ML is that it provides efficient estimates; it uses every scrap of information in the statistical and economic model in the quest for efficiency. It does the “right” efficient thing if model is true. It does not necessarily do the “reasonable” thing for “approximate” models. OLS vs. GLS cross-sectional regressions

One place in which this argument crystallizes is in the choice between OLS and GLS cross-sectional regressions, or equivalently between first and second stage GMM. The last chapter can lead to a mistaken impression that the doesn’t matter that much. This is true to some extent in that simple environment, but not in more complex environments. For example, Fama and French (1997) report important correlations between betas and pricing errors in a time-series test of a three-factor model on industry portfolios. This correlation cannot happen with an OLS cross-sectional estimate, as the cross-sectional estimate sets the cross-sectional correlation between right hand variables (betas) and error terms (pricing errors) to zero by construction. First stage estimates seem to work better in factor pricing models based on macroeconomic data. For example, Figure 5 presents the first stage estimate of the consumption-based model. The second-stage estimate produced much larger individual pricing errors, because by so doing it could lower pricing errors of portfolios with strong long-short positions required by the spectral density matrix. The same thing happened in the investment based factor pricing model Cochrane (1996), and the scaled consumption-based model of Lettau and Ludvigson (2000). Authors as far back as Fama and MacBeth (1973) have preferred OLS cross-sectional regressions, distrusting the GLS weights. GLS and second-stage GMM gain their asymptotic efficiency when the covariance and spectral density matrices have converged to their population values. GLS and second stage GMM use these matrices to find well-measured portfolios; portfolios with small residual 273

CHAPTER 16

WHICH METHOD ?

variance for GLS, and small variance of discounted return for GMM. The danger is that these quantities are poorly estimated in a finite sample, that sample minimum-variance portfolios bear little relation to population minimum-variance portfolios. This by itself should not create too much of a problem for a perfect model, one that prices all portfolios. But an imperfect model that does a very good job of pricing a basic set of portfolios may do a poor job of pricing strange linear combinations of those portfolios, especially combinations that involve strong long and short positions, positions that really are outside the payoff space given transactions, margin, and short sales constraints. Thus, the danger is the interaction between spurious sample minimum-variance portfolios and the specification errors of the model. Interestingly, Kandel and Stambaugh (1995) and Roll and Ross (1995) argue for GLS cross-sectional regressions also as a result of model misspecification. They start by observing that show that so long as there is any misspecification at all – so long as the pricing errors are not exactly zero; so long as the market proxy is not exactly on the mean-variance frontier – then there are portfolios that produce arbitrarily good and arbitrarily bad fits in plots of expected returns vs. betas. Since even a perfect model leaves pricing errors in sample, this is always true in samples. It’s easy to see the basic argument. Take a portfolio long the positive alpha securities and short the negative alpha securities; it will have a really big alpha! More precisely, if the original securities follow E(Re ) = α + λβ,

then consider portfolios of the original securities formed from a non-singular matrix A. They follow E(ARe ) = Aα + λAβ.

You can make all these portfolios have the same β with Aβ =constant, and then they will have a spread in alphas. You will see a plot in which all the portfolios have the same beta but the average returns are spread up and down. Conversely, you can pick A to make the expected return-beta plot look as good as you want. GLS has an important feature in this situation: the GLS cross-sectional regression is independent of such repackaging of portfolios. If you transform a set of returns Re to ARe , then the OLS cross-sectional regression is transformed from

to

¡ ¢ ˆ = β 0 β −1 β 0 E (Re ) λ ¡ ¢ ˆ = β 0 A0 Aβ −1 β 0 A0 AE (Re ) . λ

This does depend on the repackaging A. However, the residual covariance matrix of ARe is 274

AΣA0 , so the GLS regression ¡ ¢ ˆ = β 0 Σ−1 β −1 β 0 Σ−1 E (Re ) λ

is not affected so long as A is full rank and therefore does not throw away information ¡ ¢ ¡ ¢ ˆ = β 0 A0 (AΣA0 )−1 Aβ −1 β 0 A0 (A0 ΣA)−1 AE (Re ) = β 0 Σ−1 β −1 β 0 Σ−1 E (Re ) . λ

(The spectral density matrix and second stage estimate shares this property in GMM estimates. These are not the only weighting matrix choices that are invariant to portfolios. For example, Hansen and Jagannathan’s 1997 suggestion of the return second moment matrix has the same property.) This is a fact, but it does not show that OLS chooses a particularly good or bad set of portfolios. Perhaps you don’t think that GLS’ choice of portfolios is particularly informative. In this case, you use OLS precisely to focus attention on a particular set of economically interesting portfolios. The choice depends subtly on what you want your test to accomplish. If you want to prove the model wrong, then GLS helps you to focus on the most informative portfolios for proving the model wrong. That is exactly what an efficient test is supposed to do. But many models are wrong, but still pretty darn good. It is a shame to throw out the information that the model does a good job of pricing an interesting set of portfolios. The sensible compromise would seem to be to report the OLS estimate on “interesting” portfolios, and also to report the GLS test statistic that shows the model to be rejected. That is, in fact, the typical collection of facts. Additional examples of trading off efficiency for robustness

Here are some additional examples of situations in which it has turned out to be wise to trade off some apparent efficiency for robustness to model misspecifications. Low frequency time-series models. In estimating time-series modelsPsuch as the AR(1), maximum likelihood minimizes one-step ahead forecast error variance, ε2t . But any timeseries model is only an approximation, and the researcher’s objective may not be one-step ahead forecasting. For example, in making sense of the yield on long term bonds, one is interested in the long-run behavior of the short rate of interest. In estimating the magnitude of long-horizon univariate mean reversion in stock returns, we want to know only the sum of autocorrelations or moving average coefficients. (We will study this application in section 20.335.) The approximate model that generates the smallest one-step ahead forecast error variance may be quite different from the model that best matches long-run autocorrelations. ML can pick the wrong model and make very bad predictions for long-run responses. (Cochrane 1986 contains a more detailed analysis of this point in the context of long-horizon GDP forecasting.) Lucas’ money demand estimate. Lucas (1988) is a gem of an example. Lucas was in-

275

CHAPTER 16

WHICH METHOD ?

terested in estimating the income elasticity of money demand. Money and income trend upwards over time and over business cycles, but also have some high-frequency movement that looks like noise. If you run a regression in log-levels, mt = a + byt + εt

you get a sensible coefficient of about b = 1, but you find that the error term is strongly serially correlated. Following standard advice, most researchers run GLS, which amounts pretty much to first-differencing the data, mt − mt−1 = b(yt − yt−1 ) + η t .

This error term passes its Durbin-Watson statistic, but the b estimate is much lower, which doesn’t make much economic sense, and, worse, is unstable, depending a lot on time period and data definitions. Lucas realized that the regression in differences threw out all of the information in the data, which was in the trend, and focused on the high-frequency noise. Therefore, the regression in levels, with standard errors corrected for correlation of the error term, is the right one to look at. Of course, GLS and ML didn’t know there was any “noise” in the data, which is why they threw out the baby and kept the bathwater. Again, ML ruthlessly exploits the null for efficiency, and has no way of knowing what is “reasonable” or “intuitive.” Stochastic singularities and calibration. Models of the term structure of interest rates (we will study these models in section 19) and real business cycle models in macroeconomics give even more stark examples. These models are stochastically singular. They generate predictions for many time series from a few shocks, so the models predict that there are combinations of the time series that leave no error term. Even though the models have rich and interesting implications, ML will seize on this economically uninteresting singularity, refuse to estimate parameters, and reject any model of this form.

The simplest example of the situation is the linear-quadratic permanent income model paired with an AR(1) specification for income. The model is yt ct − ct−1

= ρyt−1 + εt = (Et − Et−1 )

∞ 1 X j 1 εt β yt+j = 1 − β j=0 (1 − βρ) (1 − β)

This model generates all sorts of important and economically interesting predictions for the joint process of consumption and income (and asset prices). Consumption should be roughly a random walk, and should respond only to permanent income changes; investment should be more volatile than income and income more volatile than consumption. Since there is only one shock and two series, however, the model taken literally predicts a deterministic relation between consumption and income; it predicts ct − ct−1 =

rβ (yt − ρyt−1 ) . 1 − βρ

276

ML will notice that this is the statistically most informative prediction of the model. There is no error term! In any real data set there is no configuration of the parameters r, β, ρ that make this restriction hold, data point for data point. The probability of observing a data set {ct , yt } is exactly zero, and the log likelihood function is −∞ for any set of parameters. ML says to throw the model out. The popular affine models of the term structure of interest rates act the same way. They specify that all yields at any moment in time are deterministic functions of a few state variables. Such models can capture much of the important qualitative behavior of the term structure, including rising, falling and humped shapes, and the information in the term structure for future movements in yields and the volatility of yields. They are very useful for derivative pricing. But it is never the case in actual yield data that yields of all maturities are exact functions of K yields. Actual data on N yields always require N shocks. Again, a ML approach reports a −∞ log likelihood function for any set of parameters.

Addressing model mis-specification

The ML philosophy offers an answer to model mis-specification: specify the right model, and then do ML. If regression errors are correlated, model and estimate the covariance matrix and do GLS. If you are worried about proxy errors in the pricing factor, short sales costs or other transactions costs so that model predictions for extreme long-short positions should not be relied on, time-aggregation or mismeasurement of consumption data, non-normal or non-i.i.d. returns, time-varying betas and factor risk premia, additional pricing factors and so on–don’t chat about them, write them down, and then do ML. Following this lead, researchers have added “measurement errors” to real business cycle models (Sargent 1989 is a classic example) and affine yield models in order to break the stochastic singularity (I discuss this case a bit more in section 19.6). The trouble is, of course, that the assumed structure of the measurement errors now drives what moments ML pays attention to. And seriously modeling and estimating the measurement errors takes us further away from the economically interesting parts of the model. (Measurement error augmented models will often wind up specifying sensible moments, but by assuming ad-hoc processes for measurement error, such as i.i.d. errors. Why not just specify the sensible moments in the first place?) More generally, authors tend not to follow this advice, in part because it is ultimately infeasible. Economics necessarily studies quantitative parables rather than completely specified models. It would be nice if we could write down completely specified models, if we could quantitatively describe all the possible economic and statistical model and specification errors, but we can’t. The GMM framework, used judiciously, allows us to evaluate misspecified models. It allows us to direct that the statistical effort focus on the “interesting” predictions while ignoring the fact that the world does not match the “uninteresting” simplifications. For example, ML only gives us a choice of OLS, whose standard errors are wrong, or GLS, which we may not trust in small samples or which may focus on uninteresting parts of the data. GMM allows 277

CHAPTER 16

WHICH METHOD ?

us to keep an OLS estimate, but correct the standard errors for non-i.i.d. distributions. More generally, GMM allows one to specify an economically interesting set of moments, or a set of moments that one feels will be robust to misspecifications of the economic or statistical model, without having to spell out exactly what is the source of model mis-specification that makes those moments “optimal” or even “interesting” and “robust.” It allows one to accept the lower “efficiency” of the estimates under some sets of statistical assumptions, in return for such robustness. At the same time, the GMM framework allows us to flexibly incorporate statistical model misspecifications in the distribution theory. For example, knowing that returns are not i.i.d. normal, one may want to use the time series regression technique to estimate betas anyway. This estimate is not inconsistent, but the standard errors that ML formulas pump out under this assumption are inconsistent. GMM gives a flexible way to derive at least and asymptotic set of corrections for statistical model misspecifications of the time-series regression coefficient. Similarly, a pooled time-series cross-sectional OLS regression is not inconsistent, but standard errors that ignore cross-correlation of error terms are far too small. The “calibration” of real business cycle models is often really nothing more than a GMM parameter estimate, using economically sensible moments such as average output growth, consumption/output ratios etc. to avoid the stochastic singularity that would doom a ML approach. (Kydland and Prescott’s 1982 idea that empirical microeconomics would provide accurate parameter estimates for macroeconomic and financial models has pretty much vanished.) Calibration exercises usually do not compute standard errors, nor do they report any distribution theory associated with the “evaluation” stage when one compares the model’s predicted second moments with those in the data. Following Burnside, Eichenbaum and Rebelo (1993) however, it’s easy enough to calculate such a distribution theory – to evaluate whether the difference between predicted “second moments” and actual moments is large compared to sampling variation, including the variation induced by parameter estimation in the same sample – by listing the first and second moments together in the gT vector. “Used judiciously” is an important qualification. Many GMM estimations and tests suffer from lack of thought in the choice of moments, test assets and instruments. For example, early GMM papers tended to pick assets and especially instruments pretty much at random. Industry portfolios have almost no variation in average returns to explain. Authors often included many lags of returns and consumption growth as instruments to test a consumptionbased model. However, the 7th lag of returns really doesn’t predict much about future returns given lags 1-12, and the first-order serial correlation in seasonally adjusted, ex-post revised consumption growth may be economically uninteresting. More recent work tends to emphasize a few well-chosen assets and instruments that capture important and economically interesting features of the data.. Auxiliary model

ML requires an auxiliary statistical model. For example, in the classic ML formalization of regression tests, we had to stopped to assume that returns and factors are jointly i.i.d. normal. As the auxiliary statistical model becomes more and more complex and hence realistic, 278

more and more effort is devoted to estimating the auxiliary statistical model. ML has no way of knowing that some parameters – a, b; β, λ, risk aversion γ – are more “important” than others – Σ, and parameters describing time-varying conditional moments of returns. A very convenient feature of GMM is that it does not require such an auxiliary statistical model. For example, in studying GMM we went straight from p = E(mx) to moment conditions, estimates, and distribution theory. This is an important saving of the researcher’s and the reader’s time, effort and attention. Finite sample distributions

Many authors say they prefer regression tests and the GRS statistic in particular because it has a finite sample distribution theory, and they distrust the finite-sample performance of the GMM asymptotic distribution theory. This argument does not have much force. The finite sample distribution only holds if returns really are normal and i.i.d., and if the factor is perfectly measured. Since these assumptions do not hold, it is not obvious that a finite-sample distribution that ignores non-i.i.d. returns will be a better approximation than an asymptotic distribution that corrects for them. All approaches give essentially the same answers in the classic setup of i.i.d. returns. The issue is how the various techniques perform in more complex setups, especially with conditioning information, and here there are no analytic finite-sample distributions. In addition, once you have picked the estimation method – how you will generate a number from the data; or which moments you will use – finding its finite sample distribution, given an auxiliary statistical model, is simple. Just run a Monte Carlo or bootstrap. Thus, picking an estimation method because it delivers analytic formulas for a finite sample distribution (under false assumptions) should be a thing of the past. Analytic formulas for finite sample distributions are useful for comparing estimation methods and arguing about statistical properties of estimators, but they are not necessary for the empiricists’ main task. Finite sample quality of asymptotic distributions, and “nonparametric” estimates

Several investigations (Ferson and Foerster 1994, Hansen, Heaton, and Yaron 1996) have found cases in which the GMM asymptotic distribution theory is a poor approximation to a finite-sample distribution theory. This is especially true when one asks “non-parametric” corrections for autocorrelation or heteroskedasticity to provide large corrections and when the number of moments is large compared to the sample size, or if the moments one uses for GMM turn out to be very inefficient (Fuhrer, Moore, and Schuh 1995) which can happen if you put in a lot of instruments with low forecast power The ML distribution is the same as GMM, conditional on the choice of moments, but typical implementations of ML also use the parametric time-series model to simplify estimates of the terms in the distribution theory as well as to derive the likelihood function. If this is the case – if the “nonparametric” estimates of the GMM distribution theory perform poorly in a finite sample, while the “parametric” ML distribution works well – there is no reason not to use a parametric time series model to estimate the terms in the GMM 279

CHAPTER 16

WHICH METHOD ?

P distribution as well. For example, rather than calculate ∞ j=−∞ E(ut ut−j ) from a large sum ofPautocorrelations, you can model ut = ρut−1 + εt , estimate ρ, and then calculate 1+ρ j 2 σ2 (u) ∞ j=−∞ ρ = σ (u) 1−ρ . Section 11.7 discussed this idea in more detail.

The case for ML

In the classic setup, the efficiency gain of ML over GMM on the pricing errors is tiny. However, several studies have found cases in which the statistically motivated choice of moments suggested by ML has important efficiency advantages. For example, Jacquier, Polson and Rossi (1994) study the estimation of a time-series model with stochastic volatility. This is a model of the form dSt /St dVt

= µdt + Vt dZ1t = µV (Vt )dt + σ(Vt )dZ2t ,

(16.213)

and S is observed but V is not. The obvious and easily interpretable moments include the autocorrelation of squared returns, or the autocorrelation of the absolute value of returns. However, Jacquier, Polson and Rossi find that the resulting estimates are far less efficient than those resulting from the ML scores. Of course, this study presumes that the model (16.213) really is exactly true. Whether the uninterpretable scores or the interpretable moments really perform better to give an approximate model of the form (16.213), given some other data-generating mechanism is open to discussion. Even in the canonical OLS vs. GLS case, a wildly heteroskedastic error covariance matrix can mean that OLS spends all its effort fitting unimportant data points. A “judicious” application of GMM (OLS) in this case would require at least some transformation of units so that OLS is not wildly inefficient. Statistical philosophy

The history of empirical work that has been persuasive – that has changed people’s understanding of the facts in the data and which economic models understand those facts – looks a lot different than the statistical theory preached in econometrics textbooks. The CAPM was taught and believed in and used for years despite formal statistical rejections. It only fell by the wayside when other, coherent views of the world were offered in the multifactor models. And the multifactor models are also rejected! It seems that “it takes a model to beat a model,” not a rejection. Even when evaluating a specific model, most of the interesting calculations come from examining specific alternatives rather than overall pricing error tests. The original CAPM tests focused on whether the intercept in a cross-sectional regression was higher or lower than the risk free rate, and whether individual variance entered into cross-sectional regressions. The CAPM fell when it was found that characteristics such as size and book/market do enter cross-sectional regressions, not when generic pricing error tests rejected. 280

Influential empirical work tells a story. The most efficient procedure does not seem to convince people if they cannot transparently see what stylized facts in the data drive the result. A test of a model that focuses on its ability to account for the cross section of average returns of interesting portfolios will in the end be much more persuasive than one that (say) focuses on the model’s ability to explain the fifth moment of the second portfolio, even if ML finds the latter moment much more statistically informative. Most recently, Fama and French (1988b) and (1993) are good examples of empirical work that changed many people’s minds, in this case that long-horizon returns really are predictable, and that we need a multifactor model rather than the CAPM to understand the cross-section of average returns. These papers are not stunning statistically: long horizon predictability is on the edge of statistical significance, and the multifactor model is rejected by the GRS test. But these papers made clear what stylized and robust facts in the data drive the results, and why those facts are economically sensible. For example, the 1993 paper focused on tables of average returns and betas. Those tables showed strong variation in average returns that was not matched by variation in market betas, yet was matched by variation in betas on new factors. There is no place in statistical theory for such a table, but it is much more persuasive than a table of χ2 values for pricing error tests. On the other hand, I can think of no case in which the application of a clever statistical models to wring the last ounce of efficiency out of a dataset, changing t statistics from 1.5 to 2.5, substantially changed the way people think about an issue. Statistical testing is one of many questions we ask in evaluating theories, and usually not the most important one. This is not a philosophical or normative statement; it is a positive or empirical description of the process by which the profession has moved from theory to theory. Think of the kind of questions people ask when presented with a theory and accompanying empirical work. They usually start by thinking hard about the theory itself. What is the central part of the economic model or explanation? Is it internally consistent? Do the assumptions make sense? Then, when we get to the empirical work, how were the numbers produced? Are the data definitions sensible? Are the concepts in the data decent proxies for the concepts in the model? (There’s not much room in statistical theory for that question!) Are the model predictions robust to the inevitable simplifications? Does the result hinge on power utility vs. another functional form? What happens if you add a little measurement error, or if agents have an information advantage, etc.? What are the identification assumptions, and do they make any sense – why is y on the left and x on the right rather than the other way around? Finally, someone in the back of the room might raise his hand and ask, “if the data were generated by a draw of i.i.d. normal random variables over and over again, how often would you come up with a number this big or bigger?” That’s an interesting and important check on the overall believability of the results. But it is not necessarily the first check, and certainly not the last and decisive check. Many models are kept that have economically interesting but statistically rejectable results, and many more models are quickly forgotten that have strong statistics but just do not tell as clean a story. The classical theory of hypothesis testing, its Bayesian alternative, or the underlying hypothesis-testing view of the philosophy of science are miserable descriptions of the way 281

CHAPTER 16

WHICH METHOD ?

science in general and economics in particular proceed from theory to theory. And this is probably a good thing too. Given the non-experimental nature of our data, the inevitable fishing biases of many researchers examining the same data, and the unavoidable fact that our theories are really quantitative parables more than literal descriptions of the way the data are generated, the way the profession settles on new theories makes a good deal of sense. Classical statistics requires that nobody ever looked at the data before specifying the model. Yet more regressions have been run than there are data points in the CRSP database. Bayesian econometrics can in principle incorporate the information of previous researchers, yet it never applied in this way – each study starts anew with a “uninformative” prior. Statistical theory draws a sharp distinction between the model – which we know is right; utility is exactly power; and the parameters which we estimate. But this distinction isn’t true; we are just as uncertain about functional forms as we are about parameters. A distribution theory at bottom tries to ask an unknowable question: If we turned the clock back to 1947 and reran the postwar period 1000 times, in how many of those alternative histories would (say) the average S&P500 return be greater than 9%? It’s pretty amazing in fact that a statistician can purport to give any answer at all to such a question, having observed only one history. These paragraphs do not contain original ideas, and they mirror changes in the philosophy of science more broadly. 50 years ago, the reigning philosophy of science focused on the idea that scientists provide rejectable hypotheses. This idea runs through philosophical writings exemplified by Popper (1959), classical statistical decision theory, and mirrored in economics by Friedman (1953) However, this methodology contains an important inconsistency. Though researchers are supposed to let the data decide, writers on methodology do not look at how actual theories evolved. It was, as in Friedman’s title, a “Methodology of positive economics,” not a “positive methodology of economics.” Why should methodology be normative, a result of philosophical speculation, and not an empirical discipline like everything else. In a very famous book, Kuhn (1970) looked at the actual history of scientific revolutions, and found that the actual process had very little to do with the formal methodology. McCloskey (1983, 1998) has gone even further, examining the “rhetoric” of economics; the kinds of arguments that persuaded people to change their minds about economic theories. Needless to say, the largest t-statistic did not win! Kuhn’s and especially McCloskey’s ideas are not popular in the finance and economics professions. Precisely, they are not popular in how people talk about their work, though they describe well how people actually do their work. Most people in the fields cling to the normative, rejectable-hypothesis view of methodology. But we need not suppose that they would be popular. The ideas of economics and finance are not popular among the agents in the models. How many stock market investors even know what a random walk or the CAPM is, let alone believing those models have even a grain of truth? Why should the agents in the models of how scientific ideas evolve have an intuitive understanding of the models? “As if” rationality can apply to us as well! Philosophical debates aside, a researcher who wants his ideas to be convincing, as well as right, would do well to study how ideas have in the past convinced people, rather than just study a statistical decision theorist’s ideas about how ideas should convince people. Kuhn, 282

and, in economics, McCloskey have done that, and their histories are worth reading. In the end, statistical properties may be a poor way to choose statistical methods. Summary

The bottom line is simple: It’s ok to do a first stage or simple GMM estimate rather than an explicit maximum likelihood estimate and test. Many people (and, unfortunately, many journal referees) seem to think that nothing less than a full maximum likelihood estimate and test is acceptable. This section is long in order to counter that impression; to argue that at least in many cases of practical importance, a simple first stage GMM approach, focusing on economically interpretable moments, can be adequately efficient, robust to model misspecifications, and ultimately more persuasive.

283

PART III Bonds and options

284

The term structure of interest rates and derivative pricing use closely related techniques. As you might have expected, I present both issues in a discount factor context. All models come down to a specification of the discount factor. The discount factor specifications in term structure and option pricing models are quite simple. So far, we have focused on returns, which reduces the pricing problem to a one-period or instantaneous problem. Pricing bonds and options forces us to start thinking about chaining together the one-period or instantaneous representations to get a prediction for prices of longlived securities. Taking this step is very important, and I forecast that we will see much more multiperiod analysis in stocks as well, studying price and stream of payoffs rather than returns. This step rather than the discount factor accounts for the mathematical complexity of some term structure and option pricing models. There are two standard ways to go from instantaneous or return representations to prices. First, we can chain the discount factors together, finding from a one period discount factor mt,t+1 a long-term discount factor mt,t+j = mt,t+1 mt+1,t+2 ...mt+j−1 mt+j that can price a j period payoff. In continuous time, we will find the discount factor increments dΛ satisfy the instantaneous pricing equation 0 = Et [d (ΛP )], and then solve its stochastic differential equation to find its level Λt+j in order to price a j− period payoff as Pt = Et [Λt+j /Λt xt+j ]. Second, we can chain the prices together. Conceptually, this is the same as chaining returns Rt,t+j = Rt,t+1 Rt+1,t+2 ..Rt+j−1,t+j instead of chaining together the discount factors. From 0 = Et [d (ΛP )], we find a differential equation for the prices, and solve that back. We’ll use both methods to solve interest rate and option pricing models.

285

Chapter 17.

Option pricing

Options are a very interesting and useful set of instruments, as you will see in the background section. In thinking about their value, we will adopt an extremely relative pricing approach. Our objective will be to find out a value for the option, taking as given the values of other securities, and in particular the price of the stock on which the option is written and an interest rate.

17.1 17.1.1

Background Definitions and payoffs

A call option gives you the right to buy a stock for a specified strike price on a specified expiration date. The call option payoff is CT = max(ST − X, 0).

Portfolios of options are called strategies. A straddle – a put and a call at the same strike price – is a bet on volatility Options allow you to buy and sell pieces of the return distribution. Before studying option prices, we need to start by understanding option payoffs. A call option gives you the right, but not the obligation, to buy a stock (or other “underlying” asset) for a specified strike price (X) on (or before) the expiration date (T). European options can only be exercised on the expiration date. American options can be exercised anytime before as well as on the expiration date. I will only treat European options. A put option gives the right to sell a stock at a specified strike price on (or before) the expiration date. I’ll use the standard notation, C = Ct = CT = S = St = ST = X=

call price today call payoff = value at expiration (T). stock price today stock price at expiration strike price

Our objective is to find the price C . The general framework is (of course) C = E(mx) where x denotes the option’s payoff. The option’s payoff is the same things as its value at expiration. If the stock has risen above the strike price, then the option is worth the difference between stock and strike. If the stock has fallen below the strike price, it expires worthless. 286

S ECTION 17.1

BACKGROUND

Thus, the option payoff is ½

ST − X if ST ≥ X if ST ≤ X 0 = max(ST − X, 0).

Call payoff = CT

A put works the opposite way: It gains value as the stock falls below the strike price, since the right to sell it at a high price is more and more valuable. Put payoff = PT = max(X − ST , 0). It’s easiest to keep track of options by a graph of their value as a function of stock price. Figure 31 graphs the payoffs from buying calls and puts, and the corresponding short positions, which are called writing call and put options. One of the easiest mistakes to make is to confuse the payoff, with the profit, which is the value at expiration less the cost of buying the option. I drew in profit lines, payoff - cost, to emphasize this difference. Call

Put Payoff

Profit

ST

ST

Write Call

Write Put ST

ST

Figure 31. Payoff diagrams for simple option strategies. Right away, you can see some of the interesting features of options. A call option allows you a huge positive beta. Typical at-the-money options (strike price = current stock price) give a beta of about 10; meaning that the option is equivalent to borrowing $10 to invest $11 in the stock. However, your losses are limited to the cost of the option, which is paid upfront. Options are obviously very useful for trading. Imagine how difficult it would be to buy stock on such huge margin, and how difficult it would be to make sure people paid if the bet went 287

CHAPTER 17

O PTION PRICING

bad. Options solve this problem. No wonder that active options trading started only a year or two after the first stocks started trading. The huge beta also means that options are very useful for hedging. If you have a large illiquid portfolio, you can offset the risks very cheaply with options. Finally, options allow you to shape the distribution of returns in interesting and sometimes dangerous ways. For example, if you buy a 20% out of the money put option as well as a stock, you have bought “catastrophe insurance” for your stock portfolio, at what turns out to be a remarkably small price. You cut off the left tail of the return distribution, at a small cost to the mean of the overall distribution. On the other side, by writing out of the money put options, you can earn a small fee year in and year out, only once in a while experiencing a huge loss. You have a large probability of a small gain and a small probability of a large loss. You are providing catastrophe insurance to the market, and it works much like, say, writing earthquake insurance. The distribution of returns from this strategy is extremely non-normal, and thus statistical evaluation of its properties will be difficult. This strategy is tempting to a portfolio manager who is being evaluated only by the statistics of his achieved return. If he writes far out of the money options in addition to investing in an index, the chance of beating the index for one or even five years is extremely high. If the catastrophe does happen and he loses a billion dollars or so, the worst you can do is fire him. (Your contract with the manager is always a call option.) This is why portfolio management contracts are not purely statistical, but also write down what kind of investments can and cannot be made. Straddle

Call payoff Put payoff P+C Straddle profit

Stock Price

Make money if stock ends up here

Figure 32. Payoff diagram for a straddle.

288

S ECTION 17.1

BACKGROUND

Portfolios of put and call options are called strategies, and have additional interesting properties. Figure 32 graphs the payoff of a straddle, which combines a put and call at the same strike price. This strategy pays off if the stock goes up or goes down. It loses money if the stock does not move. Thus the straddle is a bet on volatility. Of course, everyone else understands this, and will bid the put and call prices up until the straddle earns only an equilibrium rate of return. Thus, you invest in a straddle if you think that stock volatility is higher than everyone else thinks it will be. Options allow efficient markets and random walks to operate on the second and higher moments of stocks as well as their overall direction! You can also see quickly that volatility will be a central parameter in option prices. The higher the volatility, the higher both put and call prices. More generally, by combining options of various strikes, you can buy and sell any piece of the return distribution. A complete set of options – call options on every strike price – is equivalent to complete markets, i.e., it allows you to form payoffs that depend on the terminal stock price in any way; you can form any payoff of the form f (ST ).

289

CHAPTER 17 17.1.2

O PTION PRICING

Prices: one period analysis

I use the law of one price – existence of a discount factor – and no-arbitrage – existence of a positive discount factor – to characterize option prices. The results are: 1) Put-call parity: P = C − S + X/Rf . 2) Arbitrage bounds, best summarized by Figure 34 3) The proposition that you should never exercise an American call option early on a stock that pays no dividends. The arbitrage bounds are a linear program, and this procedure can be used to find them in more complex situations where clever identification of arbitrage portfolios may fail. We have a set of interesting payoffs. Now what can we say about their prices – their values at dates before expiration? Obviously, p = E(mx) as always. We have learned about x, now we have to think about m. We can start by imposing little structure – the law of one price and the absence of arbitrage, or, equivalently, the existence of some discount factor or a positive discount factor. . In the case of options, these two principles actually do tell you a good deal about the option price. Put-Call parity

The law of one price, or the existence of some discount factor that prices stock, bond, and a call option, allows us to deduce the value of a put in terms of the price of the stock, bond, and call. Consider the following two strategies: 1) Hold a call, write a put, same strike price. 2) Hold stock, promise to pay X. The payoffs of these two strategies are the same, as shown in Figure 33. Equivalently, the payoffs are related by PT = CT − ST + X.

Thus, so long as the law of one price holds, the prices of left and right hand sides must be equal. Applying E(m·) to both sides for any m, P = C − S + X/Rf .

(The price of ST is S . The price of the payoff X is X/Rf .) Arbitrage bounds

If we add the absence of arbitrage, or equivalently the restriction that the discount factor must be positive, we can deduce bounds on the call option price without needing to know the put price. In this case, it is easiest to cleverly notice arbitrage portfolios – situations in which 290

S ECTION 17.1

BACKGROUND

Buy call X

ST

=

+ Write put

Buy stock =

+ Borrow strike

-X

Figure 33. Put call parity. portfolio A dominates portfolio B. Then, either directly from the definition of no-arbitrage or from A > B, m > 0 ⇒ E(mA) > E(mB), you can deduce that the price of A must be greater than the price of B. The arbitrage portfolios are 1. 2.

3.

CT > 0 ⇒ C > 0. The call payoff is positive so the call price must be positive. CT ≥ ST − X ⇒ C ≥ S − X/Rf . The call payoff is better than hold stock - pay strike for sure, so the call price is greater than holding the stock and borrowing the strike, i.e. promising to pay it for sure. CT ≤ ST ⇒ C ≤ S . The call payoff is worse than stock payoff (because you have to pay the strike price. Thus, the call price is less than stock price.

Figure 34 summarizes these arbitrage bounds on the call option value. We have gotten somewhere – we have restricted the range of the option prices. However, the arbitrage bounds are too large to be of much use. Obviously, we need to learn more about the discount factor than pure arbitrage or m > 0 will allow. We could retreat to economic models, e.g. use the CAPM or other explicit discount factor model. Option pricing is famous because we don’t have to do that. Instead, if we open up dynamic trading–the requirement that the discount factor price the stock and bond at every date to expiration–it turns out that we can sometimes determine the discount factor and hence the option value precisely. This presentation is unsettling for two reasons. First, you may worry that you will not be clever enough to dream up dominating portfolios in more complex circumstances. Second, you may worry that we have not dreamed up all of the arbitrage portfolios in this circumstance. Perhaps there is another one lurking out there, which would reduce the unsettling large size of the bounds. It leaves us hungry for a constructive technique for finding arbi291

CHAPTER 17

O PTION PRICING

C Call value Today Call value C In here

X/Rf

S Stock value today

Figure 34. Arbitrage bounds for a call option trage bounds that would be guaranteed to work in general situations, and to find the tightest arbitrage bound. We want to know Ct = E(mt,T xcT ) wherexcT = max(ST −X, 0) denotes the call payoff, we want to use information in the observed stock and bond prices to learn about the option price, and we want to impose the absence of arbitrage. We can capture this search with the following problem: (214)

f max Ct = Et (mxC T ) s.t. m > 0, St = Et (mST ), 1 = Et (mR ) m

and the corresponding minimization. The first constraint implements absence of arbitrage. The second and third use the information in the stock and bond price to learn what we can about the option price. Write 17.214 out in state notation, max Ct =

{m(s)}

X

π(s)m(s)xC T (s) s.t. m(s) > 0, St =

s

X s

π(s)m(s)ST (s), 1 =

X

π(s)m(s)Rf

s

This is a linear program – a linear objective and linear constraints. In situations where you do not know the answer, you can calculate arbitrage bounds – and know you have them all – by solving this linear program. I don’t know how you would begin to check that for every portfolio A whose payoff dominates B , the price of A is greater than the price of B . The discount factor method lets you construct the arbitrage bounds. 292

S ECTION 17.2

BLACK -S CHOLES FORMULA

Early exercise

By applying the absence of arbitrage, we can show quickly that you should never exercise an American call option on a stock that pays no dividends before the expiration date. This is a lovely illustration because such a simple principle leads to a result that isn’t initially obvious. Follow the table: Payoffs Price Rf > 1

CT = max(ST − X, 0) C C

≥ ≥ ≥

ST − X S − X/Rf S−X

S − X is what you get if you exercise now. The value of the call is greater than this value, because you can delay paying the strike, and because exercising early loses the option value. Put-call parity let us concentrate on call options; this fact lets us concentrate on European options.

17.2

Black-Scholes formula

Write a process for stock and bond, then use Λ∗ to price the option. The Black-Scholes formula 17.220 results. You can either solve for the finite-horizon discount factor ΛT /Λ0 and find the call option price by taking the expectation C0 = E0 (ΛT /Λ0 xC T ), or you can find a differential equation for the call option price and solve it backward. Our objective, again, is to learn as much as we can about the value of an option, given the value of the underlying stock and bond. The one-period analysis led only to arbitrage bounds, at which point we had to start thinking about discount factor models. Now, we allow intermediate trading, which means we really are thinking about dynamic multiperiod asset pricing. The standard approach to the Black-Scholes formula rests on explicitly constructing portfolios: at each date we cleverly construct a portfolio of stock and bond that replicates the instantaneous payoff of the option; we reason that the price of the option must equal the price of the replicating portfolio. Instead, I follow the discount factor approach. The law of one price is the same thing as the existence of a discount factor. Thus, rather than construct law-of-one-price replicating portfolios, construct at each date a discount factor that prices the stock and bond, and use that discount factor to price the option. the discount factor approach shows how thinking of the world in terms of a discount factor is equivalent in the result and as easy in the calculation as other approaches. This case shows some of the interest and engineering complexity of continuous time models. Though at each instant the analysis is trivial law of one price, chaining it together 293

CHAPTER 17

O PTION PRICING

over time is not trivial. The call option payoff is CT = max(ST − X, 0)

where X denotes the strike price X and ST denotes the stock price on the expiration date T . The underlying stock follows dS = µdt + σdz. S

There is also a money market security that pays the real interest rate rdt. We want a discount factor that prices the stock and bond. All such discount factors are of the form m = x∗ + w, E(xw) = 0. In continuous time, all such discount factors are of the form (µ − r) dΛ = −rdt − dz − σ w dw; E(dwdz) = 0. Λ σ

(You can check that this set of discount factors does in fact price the stock and interest rate, or take a quick look back at section 4.3.) Now we price the call option with this discount factor, and show that the Black-Scholes equation results. Importantly, the choice of discount factor via choice of σ w dw turns out to have no effect on the resulting option price. Every discount factor that prices the stock and interest rate gives the same value for the option price. The option is therefore priced using the law of one price alone. There are two paths to follow. Either we solve the discount factor forward, and then find the call value by C = E(mxC ), or we characterize the price path and solve it backwards from expiration. 17.2.1

Method 1: Price using discount factor

Let us use the discount factor to price the option directly: ¾ Z ½ ΛT ΛT C0 = Et max (ST − X, 0) = max (ST − X, 0) df (ΛT , ST ) Λt Λt where ΛT and ST are solutions to dS S dΛ Λ

= µdt + σdz = −rdt −

(17.215)

µ−r dz − σ w dw. σ

I simplify the algebra by setting σ w dw to zero, anticipating that it does not matter. You can 294

S ECTION 17.2

BLACK -S CHOLES FORMULA

reason that since S does not depend on dw, CT depends only on ST , so C will only depend on S , and dw will have no effect on the answer. If this isn’t good enough, a problem asks you to include the dw, trace through the remaining steps and verify that the answer does not in fact depend on dw. “Solving” a stochastic differential equation such as (17.215) means finding the distribution of the random variables ST and ΛT , using information as of date 0. This is just what we do with difference equations. For example, if we solve xt+1 = ρxt + εt+1 with ε normal P forward to xT = ρT x0 + Tj=1 ρT −j εj , we have expressed xT as a normally distributed ranP dom variable with mean ρT x0 and standard deviation Tj=1 ρ2(T −j) . In the continuous time case, it turns out that we can solve some nonlinear specifications as well. Integrals of dz give us shocks, as integrals of dt give us deterministic functions of time. We can find analytical expressions for the solutions equations of the form (17.215). Start with the stochastic differential equation dY = µY dt + σ Y dz. Y

(216)

Write d ln Y =

µ ¶ 1 1 dY 1 2 2 − σ dY = µ − dt + σ Y dZ Y Y 2Y2 2 Y

Integrating from 0 to T , (17.216) has solution µ ¶ σ2Y ln YT = ln Y0 + µY − T + σY (zT − z0 ) 2

(217)

zT − z0 is a normally distributed random variable with´mean zero and variance T . Thus, ln Y ³ 2 is conditionally normal with mean ln Y0 + µY − σ2Y T and variance σ2Y T. You can check this solution by differentiating it – don’t forget the second derivative terms.

Applying the solution (17.217) to (17.215), we have µ ¶ √ σ2 ln ST = ln S0 + µ − T + σ Tε 2 Ã µ ¶2 ! 1 µ−r µ − r√ ln ΛT = ln Λ0 − r + Tε T− 2 σ σ

(17.218)

where the random variable ε is ε=

zT − z0 √ ∼ N (0, 1) . T

Having found the joint distribution of stock and discount factor, we evaluate the call 295

CHAPTER 17

O PTION PRICING

option value by doing the integral corresponding to the expectation, C0

Z

=

∞

ST =X Z ∞

=

ST =X

ΛT (ST − X) df (ΛT , ST ) Λt ΛT (ε) (ST (ε) − X) df (ε) Λt

(17.219)

We know the joint distribution of the terminal stock price ST and discount factor ΛT on the right hand side, so we have all the information we need to calculate this integral. This example has enough structure that we can find an analytical formula. In more general circumstances, you may have to resort to numerical methods. At the most basic level, you can simulate the Λ, S process forward and then take the integral by summing over many such simulations. Doing the integral

Start by breaking up the integral (17.219) into two terms, Z

C0 =

∞

ΛT (ε) ST (ε) df (ε) − Λt

ST =X

Z

∞

ST =X

ΛT (ε) X df (ε) . Λt

ST and ΛT are both exponential functions of ε. The normal distribution is also an exponential function of ε. Thus, we can approach this integral exactly as we approach the expectation of a lognormal; we can merge the two exponentials in ε into one term, and express the result as integrals against a normal distribution. Here we go. Plug in (17.218) for ST , ΛT , and simplify the exponentials in terms of ε, C0

=

Z

∞

´ ³ √ 2 T − µ−r − r+ 12 ( µ−r Tε σ ) σ

e

ST =X Z ∞

−X Z = S0 −X

´ ³ √ 2 T − µ−r − r+ 12 ( µ−r Tε σ ) σ

e

ST =X h ∞

e

ST =X Z ∞

1 2 S0 e(µ− 2 σ )T +σ

ST =X

´ ³ √ 2 T − µ−r − r+ 12 ( µ−r Tε σ ) σ

Now add the normal distribution formula for f(ε), f (ε) =

1 − 1 ε2 e 2 . 2π

296

f (ε)dε

f (ε) dε

³ ´i √ 2 µ−r− 12 σ2 +( µ−r T +(σ− µ−r σ ) σ ) Tε

e

√ Tε

f (ε)

f(ε)dε

S ECTION 17.2

BLACK -S CHOLES FORMULA

The result is C0

Z ∞ h ³ ´i √ 2 1 2 1 µ−r− 12 σ 2 +( µ−r T +(σ− µ−r σ ) σ ) T ε− 2 ε √ = e dε S0 2π ST =X Z ∞ i h √ 2 1 T − µ−r − r+ 12 ( µ−r T ε− 12 ε2 σ ) σ −√ X e dε 2π S =X Z ∞T √ 2 µ−r 1 1 = √ S0 e− 2 [ε−(σ− σ ) T ] dε 2π ST =X Z ∞ √ 2 µ−r 1 1 −rT − √ Xe e− 2 (ε+ σ T ) dε. 2π ST =X

Notice that the integrals have the form of a normal distribution with nonzero mean. The lower bound ST = X is, in terms of ε, µ ¶ √ σ2 ln X = ln ST = ln S0 + µ − T + σ Tε 2 ³ ln X − ln S0 − µ − √ ε= σ T

σ2 2

´

T .

Finally, we can express definite integrals against a normal distribution by the cumulative normal, Z ∞ 2 1 1 √ e− 2 (ε−µ) dε = Φ (µ − a) 2π a i.e., Φ() is the area under the left tail of the normal distribution.

C0

´ ³ 2 ¶ ln X − ln S0 − µ − σ2 T µ √ µ − r √ + σ− = S0 Φ − T σ σ T −t ´ ³ 2 ln X − ln S0 − µ − σ2 T √ µ−r √ −Xe−r(T −t) Φ − T − σ σ T

Simplifying, we get the Black-Scholes formula Ã Ã ¤ ! ¤ ! £ £ ln S0 /X + r + 12 σ 2 T ln S0 /X + r − 12 σ 2 T −rT √ √ C0 = S0 Φ Φ − Xe . (220) σ T σ T 297

CHAPTER 17 17.2.2

O PTION PRICING

Method 2: Derive Black-Scholes differential equation

Rather than solve the discount factor forward and then integrate, we can solve the price backwards from expiration. The instantaneous or expected return formulation of a pricing model amounts to a differential equation for prices. Guess that the solution for the call price is a function of stock price and time to expiration, Ct = C(S, t). Use Ito’s lemma to find derivatives of C(S, t), 1 dC = Ct dt + CS dS + CSS dS 2 2 · ¸ 1 dC = Ct + CS Sµ + CSS S 2 σ 2 dt + CS Sσdz 2

Plugging into the basic asset pricing equation 0 = Et (dΛC) = CEt dΛ + ΛEt dC + Et dΛdC,

using Et (dΛ/Λ) = −rdt and canceling Λdt, we get 1 0 = −rC + Ct + CS Sµ + CSS S 2 σ2 − S (µ − r) CS 2

or, 1 0 = −rC + Ct + SrCS + CSS S 2 σ 2 . 2

(221)

This is the Black-Scholes differential equation for the option price. We now know a differential equation for the price function C(S, t). We know the value of this function at expiration, C(ST , T ) = max(ST − X, 0). The remaining task is to solve this differential equation backwards through time. Conceptually, and numerically, this is easy. Express the differential equation as −

∂C(S, t) 1 ∂ 2 C(S, t) 2 2 ∂C(S, t) = −rC(S, t) + Sr + S σ . ∂t ∂S 2 ∂S 2

At any point in time, you know the values of C(S, t) for all S – for example, you can store them on a grid for S . Then, you can take the first and second derivatives with respect to S and form the quantity on the right hand side at each value of S . Now, you can find the option price at any value of S , one instant earlier in time. This differential equation, solved with boundary condition C = max {ST − X, 0}

has an analytic solution – the familiar formula (17.220). One standard way to solve differential equations is to guess and check; and by taking derivatives you can check that (17.220) 298

S ECTION 17.3

P ROBLEMS

does satisfy (17.221). Black and Scholes solved the differential equation with a fairly complicated Fourier transform method. The more elegant Feynman-Kac solution amounts to showing that solutions of the partial differential equation (17.221) can be represented as integrals of the form that we already derived independently as in (17.219). (See Duffie 1992 p.87)

17.3 1.

2.

Problems

We showed that you should never exercise an American call early if there are no dividends. Is the same true for American puts, or are there circumstances in which it is optimal to exercise American puts early? Retrace the steps in the integral derivation of the Black-Scholes formula and show that the dw does not affect the final result.

299

Chapter 18. Option pricing without perfect replication 18.1

On the edges of arbitrage

The Black-Scholes formula is justly famous and launched a thousand techniques for option pricing. The principle of no-arbitrage pricing is obvious, but its application leads to many subtle and unanticipated pricing relationships. However, in many practical situations, the law of one price arguments that we used in the Black-Scholes formula break down. If options really were redundant, it is unlikely that they would be traded as separate assets. It really is easy to synthesize forward rates from zero-coupon bonds, and forward rates are not separately traded or quoted. We really cannot trade continuously, and trying to do so would drown a strategy in transactions costs. As a practical example, at the time of the 1987 stock market crash, several prominent funds were trying to follow “portfolio insurance” strategies, essentially synthesizing put options by systematically selling stocks as prices declined. During the time of the crash, however, they found that the markets just dried up – they were unable to sell as prices plummeted. We model this situation mathematically as a Poisson jump, a discontinuous movement in prices. In the face of such jumps the call option payoff is not perfectly hedged by a portfolio of stock and bond, and cannot be priced as such. Generalizations of the stochastic setup lead to the same result. If the interest rate or stock volatility are stochastic, we do not have securities that allow us to perfectly hedge the corresponding shocks, so the law of one price again breaks down. In addition, many options are written on underlying securities that are not traded, or not traded continually and with sufficient liquidity. Real options in particular – the option to build a factory in a particular location – are not based on a tradeable underlying security, so the logic behind Black-Scholes pricing does not apply. Executives are specifically forbidden to short stock in order to hedge executive options. Furthermore, applications of option pricing formulas to trading activities seem to suffer a strange inconsistency. We imagine that the stock and bond are perfectly priced and perfectly liquid – available for perfect hedging. Then, we search for options that are priced incorrectly as trading opportunities. If the options can be priced wrong, why can’t the stock and bond be priced wrong? We should treat all assets symmetrically in evaluating trading opportunities. Trading opportunities also involve risk, and a theory that pretends they are arbitrage opportunities does not help much to quantify that risk. In all of these situations, an unavoidable “basis risk” creeps in between the option payoff and the best possible hedge portfolio. Holding the option entails some risk, and the value of the option depends on the “market price” of that risk – the covariance of the risk with an 300

S ECTION 18.2

O NE - PERIOD GOOD DEAL BOUNDS

appropriate discount factor. Nonetheless, we would like not to give up and go back to the consumption-based model, factor models, or other “absolute” methods that try to price all assets. We are still willing to take as given the prices of lots of assets in determining the price of an option, and in particular assets that will be used to hedge the option. We can form an “approximate hedge” or portfolio of basis assets “closest to” the focus payoff, and we can hedge most of the option’s risk with that approximate hedge. Then, the uncertainty about the option payoff is reduced only to figuring out the price of the residual. In addition, since the residuals are small, we might be able to say a lot about option prices with much weaker restrictions on the discount factor than those suggested by absolute models. In this chapter, I survey “good deal” option price bounds, a technique that Jesus SaáRequejo and I (1999) advocated for this situation. The good deal bounds amount to systematically searching over all possible assignments of the “market price of risk” of the residual, constraining the total market price of risk to a reasonable value, and imposing no arbitrage opportunities, to find upper and lower bounds on the option price. It is not equivalent to pricing options with pure Sharpe ratio arguments. The concluding section of this chapter surveys some alternative and additional techniques.

18.2

One-period good deal bounds

We want to price the payoff xC , for example, xC = max(ST − K, 0) for a call option. We have in hand a N −dimensional vector of basis payoffs x, whose prices p we can observe, for example the stock and bond. The good deal bound finds the minimum and maximum value of xC by searching over all positive discount factors that price the basis assets and have limited volatility: C = max E(mxC ) s.t. p = E(mx), m ≥ 0, σ2 (m) ≤ h/Rf {m}

(222)

The corresponding minimization yields the lower bound C . This is a one-period discretetime problem. The Black-Scholes formula does not apply because you can’t trade between the price and payoff periods. The first constraint on the discount factor imposes the price of the basis assets. We want to do as much relative pricing as possible; we want to extend what we know about the prices of x to price xC , without worrying about where the prices of x come from. The second constraint imposes the absence of arbitrage. This problem without the last constraint yields the arbitrage bounds that we studied in section 2. In most situations, the arbitrage bounds are too wide to be of much use. The last is an additional constraint on discount factors, and the extra content of gooddeal vs. arbitrage bounds. It is a relatively weak restriction. We could obtain closer bounds 301

C HAPTER 18

O PTION PRICING WITHOUT PERFECT REPLICATION

on prices with more information about the discount factor. In particular, if we know the correlation of the discount factor with the payoff xC we could price the option a lot better! As m > 0 means that no portfolios priced by m may display an arbitrage opportunity, σ2 (m) ≤ h/Rf means that no portfolio priced by m may have a Sharpe ratio greater than h. Recall E(mRe ) = 0 implies E(m)E(Re ) = −ρσ(m)σ(Re ) and |ρ| ≤ 1.

It is a central advantage of a discount factor approach that we can easily impose both the discount factor volatility constraint and positivity, merging the lessons of factor models and option pricing models. The prices and payoffs generated by discount factors that satisfy both m ≥ 0 and σ(m) ≤ h/Rf do more than rule out arbitrage opportunities and high Sharpe ratios. I’ll treat the case that there is a riskfree rate, so we can write E(m) = 1/Rf . In this case, it is more convenient to express the volatility constraint as a second moment, so the bound (18.222) becomes ¡ ¢ C = min E (m xc ) s.t. p = E (mx) , E m2 ≤ A2 , m ≥ 0, {m}

(223)

where A2 ≡ (1 +h2 )/Rf 2 . The problem is a standard minimization with two inequality constraints. Hence we find a solution by trying all the combinations of binding and nonbinding constraints, in order of their ease of calculation. 1) Assume the volatility constraint binds and the positivity constraint is slack. This one is very easy to calculate, since we will find analytic formulas for the solution. If the resulting discount factor m is nonnegative, this is the solution. If not, 2) assume that the volatility constraint is slack and the positivity constraint binds. This is the classic arbitrage bound. Find the minimum variance discount factor that generates the arbitrage bound. If this discount factor satisfies the volatility constraint, this is the solution. If not, 3) solve the problem with both constraints binding. 18.2.1

Volatility constraint binds, positivity constraint is slack

If the positivity constraint is slack, the problem reduces to ¡ ¢ C = min E(m xc ) s.t. p = E (mx) , E m2 ≤ A2 . {m}

(224)

We could solve this problem directly, choosing m in each state with Lagrange multipliers on the constraints. But as with the mean-variance frontier it is much more elegant to set up orthogonal decompositions and then let the solution pop out. Figure 35 describes the idea. X denotes the space of payoffs of portfolios of the basis assets x, a stock and a bond in the classic Black-Scholes setup. Though graphed as a line, X is typically a larger space. We know all prices in X , but the payoff xc that we wish to value does not lie in X . 302

S ECTION 18.2

O NE - PERIOD GOOD DEAL BOUNDS

Start by decomposing the focus payoff xc into an approximate hedge x ˆc and a residual w, xc x ˆc w

= x ˆc + w, ≡ proj(xc |X) = E(xc x0 )E(xx0 )−1 x, ˆc . ≡ xc − x

(18.225)

We know the price of x ˆc . We want to bound the price of the residual w to learn as much as we can about the price of xc . All discount factors that price x – that satisfy p = E(mx)– lie in the plane through x∗ . As we sweep through these discount factors, we generate any price from −∞ to ∞ for the residual w and hence payoff xc . All positive discount factors m > 0 lie in the intersection of the m plane and the positive orthant – the triangular region. Discount factors m in this range generate a limited range of prices for the focus payoff – the arbitrage bounds. Since second moment defines distance in Figure 35, the set of discount factors that satisfies the volatility constraint E(m2 ) ≤ A2 lies inside a sphere around the origin. The circle in Figure 35 shows the intersection of this sphere with the set of discount factors. This restricted range of discount factors will produce a restricted range of values for the residual w and hence a restricted range of values for the focus payoff xc . In the situation I have drawn, the positivity constraint is slack, since the E(m2 ) ≤ A2 circle lies entirely in the positive orthant.

We want to find the discount factors in the circle that minimize or maximize the price of the residual w. The more a discount factor points in the w direction, the larger a price E(mw) it assigns to the residual. Obviously, the discount factors that maximize or minimize the price of w point as much as possible towards and away from w. If you add any movement ε orthogonal to w, this increases discount factor volatility without changing the price of w. Hence, the discount factor that generates the lower bound is m = x∗ − vw

(226)

where v=

s

A2 − E(x∗2 ) E(w2 )

(227)

is picked to just satisfy the volatility constraint. The bound is C = E(mxc ) = E(x∗ xc ) − vE(w2 )

(228)

The upper bound is given by v = −v

The first term in equation (18.228) is the value of the approximate hedge portfolio, and can be written several ways, including E(x∗ xc ) = E(x∗ x ˆc ) = E(mˆ xc )

303

(229)

C HAPTER 18

O PTION PRICING WITHOUT PERFECT REPLICATION

xc

X

w

x xˆ c

x * + vw

E (m 2 ) < A2

x*

x * −vw

m>0

Figure 35. Construction of a discount factor to solve the one-period good deal bound when the positivity constraint is slack. for any discount factor m that prices basis assets. (Don’t forget, E(xy) = E[x proj(y|X)].) The second term in equation (18.228) is the lowest possible price of the residual w consistent with the discount factor volatility bound: vE(w2 ) = E(vw w) = E[(x∗ + vw)w] = E(mw).

For calculations you can substitute the definitions of x∗ and w in equation (18.228) to obtain an explicit, if not very pretty, formula: C = p0 E(xx0 )−1 E(xxc ) −

p p A2 − p0 E(xx0 )−1 p E(xc2 ) − E(xc x0 )E(xx0 )−1 E(xxc ). (230)

304

S ECTION 18.2

O NE - PERIOD GOOD DEAL BOUNDS

The upper bound C is the same formula with a + sign in front of the square root. Using (18.226), check whether the discount factor is positive in every state of nature. If so, this is the good-deal bound, and the positivity constraint is slack. If not, proceed to the next step. If you prefer an algebraic and slightly more formal argument, start by noticing that any discount factor that satisfies p = E(mx) can be decomposed as m = x∗ + vw + ε

where E(x∗ w) = E(x∗ ε) = E(wε). Check these properties from the definition of w and ε; this is just like R = R∗ + wRe∗ + n. Our minimization problem is then min E(mxc ) s.t. E(m2 ) ≤ A2 ¡ ¢ min E [((x∗ + vw + ε) (ˆ xc + w)] s.t. E x∗2 + v2 E(w2 ) + E(ε2 ) ≤ A2 {v,ε} ¡ ¢ min E(x∗ x ˆc ) + vE(w2 ) s.t. E x∗2 + v2 E(w2 ) + E(ε2 ) ≤ A2 {v,ε}

{v,ε}

q 2 −E(x∗2 ) The solution is ε = 0 and v = ± A E(w . 2) 18.2.2

Both constraints bind

Next, I find the bounds when both constraints bind. Though this is the third step in the procedure, it is easiest to describe this case first. Introducing Lagrange multipliers, the problem is C = min

max E (m xc ) + λ0 [E (mx) − p] +

{m>0} {λ,δ>0}

¤ δ £ ¡ 2¢ E m − A2 2

The first order conditions yield a discount factor that is a truncated linear combination of the payoffs, ¶ · c µ c ¸+ x + λ0 x x + λ0 x ,0 = − m = max − . δ δ

(231)

The last equality defines the []+ notation for truncation. In finance terms, this is a call option with zero strike price. You can derive (18.231) by introducing a Kuhn-Tucker multiplier π(s)ν(s) on m > 0 305

C HAPTER 18

O PTION PRICING WITHOUT PERFECT REPLICATION

and taking partial derivatives with respect to m in each state, # " X X 0 c C = min π(s)m(s)x (s) + λ π(s)m(s)x(s) − p {m}

s

s

# X δ X 2 2 + π(s)m(s) − A + π(s)ν(s)m(s) 2 s s "

1 ∂ : xc (s) + λ0 x(s) + δm(s) + ν(s) = 0 π(s) ∂s

(232)

If the positivity constraint is slack, the Kuhn-Tucker multiplier ν(s) is zero, m(s) = −

xc (s) + λ0 x(s) . δ

If the positivity constraint binds, then m(s) = 0, and ν(s) is just enough to make (18.232) hold. In sum, we have (18.231). We could plug expression (18.231) into the constraints, and solve numerically for Lagrange multipliers λ and δ that enforce the constraints. Alas, this procedure requires the solution of a system of nonlinear equations in (λ, δ ), which is often a numerically difficult or unstable problem. Hansen, Heaton and Luttmer (1995) show how to recast the problem as a maximization, which is numerically much easier. Interchanging min and max, C = max

min E(m xc ) + λ0 [E (mx) − p] +

{λ,δ>0} {m>0}

¤ δ £ ¡ 2¢ E m − A2 . 2

(233)

The inner minimization yields the same first order conditions (18.231). Plugging those firstorder conditions into the outer maximization of (18.233) and simplifying, we obtain ( · c ¸+2 ) δ δ x + λ0 x C = max E − − (234) − λ0 p− A2 . 2 δ 2 {λ,δ>0} You can search numerically over (λ, δ) to find the solution to this problem. The upper bound is found by replacing max with min and replacing δ > 0 with δ < 0. 18.2.3

Positivity binds, volatility is slack

If the volatility constraint is slack and the positivity constraint binds, the problem reduces to C = min E (m xc ) s.t. p = E (mx) , m > 0. {m}

306

(235)

S ECTION 18.2

O NE - PERIOD GOOD DEAL BOUNDS

These are the arbitrage bounds. We found these bounds in section 2 for a call option by just being clever. If you can’t be clever, (18.235) is a linear program. We still have to check that the discount factor volatility constraint can be satisfied at the arbitrage bound. Denote the lower arbitrage bound by Cl . The minimum variance (second moment) discount factor that generates the arbitrage bound Cl solves · ¸ µ · ¸¶ p x E(m2 )min = min E(m2 ) s.t =E m , m > 0. Cl xc {m} Using the same conjugate method, this problem is equivalent to o n +2 − 2v0 p−2µCl . E(m2 )min = max −E [− (µxc + v0 x)] {v,µ}

Again, search numerically for (v,µ) to solve this problem. If E(m2 )min ≤ A, Cl is the solution to the good-deal bound; if not we proceed with the case that both constraints are binding described above. 18.2.4

Application to Black-Scholes

The natural first exercise with this technique is to see how it applies in the Black-Scholes world. Keep in mind, this is the Black-Scholes world with no intermediate trading; compare the results to the arbitrage bounds, not to the Black-Scholes formula. Figure 36, taken from Cochrane and Saá-Requejo (1999) presents the upper and lower good-deal bounds for a call option on the S&P500 index with strike price K = $100, and three months to expiration. We used parameter values E(R) = 13%, σ(R) = 16% for the stock index return and an riskfree rate Rf = 5%. The discount factor volatility constraint is twice the historical market Sharpe ratio, h = 2 × E(R − Rf )/σ(R) = 1.0. To take the expectations required in the formula, we evaluated integrals against the lognormal stock distribution. The figure includes the lower arbitrage bounds C ≥ 0, C ≥ K/Rf . The upper arbitrage bound states that C ≤ S , but this 45◦ line is too far up to fit on the vertical scale and still see anything else. As in many practical situations, the arbitrage bounds are so wide that they are of little use. The upper good-deal bound is much tighter than the upper arbitrage bound. For example, if the stock price is $95, the entire range of option prices between the upper bound of $2 and the upper arbitrage bound of $95 is ruled out. The lower good-deal bound is the same as the lower arbitrage bound for stock prices less than about $90 and greater than about $110. In this range, the positivity constraint binds and the volatility constraint is slack. This range shows that it is important to impose both volatility and positivity constraints. Good deal bounds are not just the imposition of low Sharpe ratios on options. (I emphasize it because this point causes a lot of confusion.) The volatility bound alone admits negative prices. A free out of the money call option is like a lottery ticket: it is an arbitrage opportunity, but its expected return/standard deviation ratio is terrible, because the standard deviation is so high. A Sharpe ratio criterion alone will not rule it out. 307

C HAPTER 18

O PTION PRICING WITHOUT PERFECT REPLICATION

Figure 36. Good deal option price bounds as a function of stock price. Options have three months to expiration and strike price K = $100. The bounds assume no trading until expiration, and a discount factor volatility bound h = 1.0 corresponding to twice the market Sharpe ratio. The stock is lognormally distributed with parameters calibrated to an index option. In between $90 and $110, the good-deal bound improves on the lower arbitrage bound. It also improves on a bound that only imposes only the volatility constraint. In this region, both positivity and volatility constraints bind. This fact has an interesting implication: Not all values outside the good-deal bounds imply high Sharpe ratios or arbitrage opportunities. Such values might be generated by a positive but highly volatile discount factor, and generated by another less volatile but sometimes negative discount factor, but no discount factor generates these values that is simultaneously nonnegative and respects the volatility constraint. It makes sense rule out these values. If we know that an investor will invest in any arbitrage opportunity or take any Sharpe ratio greater than h, then we know that his unique marginal utility satisfies both restrictions. He would find a utility-improving trade for values outside the good-deal bounds, even though those values may not imply a high Sharpe ratio, an arbitrage opportunity, or any other simple portfolio interpretation. The right thing to do is to intersect restrictions on the discount factor. Simple portfolio 308

S ECTION 18.3

MULTIPLE PERIODS AND CONTINUOUS TIME

interpretations, while historically important, are likely to fall by the wayside as we add more discount factor restrictions or intersect simple ones.

18.3

Multiple periods and continuous time

Now, on to the interesting case. Option pricing is all about dynamic hedging, even if imperfect dynamic hedging. Good deal bounds would be of little use if we could only apply them to one-period environments. 18.3.1

The bounds are recursive

The central fact that makes good deal bounds tractable in dynamic environments is that the bounds are recursive. Today’s bound can be calculated as the minimum price of tomorrow’s bound, just as today’s option price can be calculated as the value of tomorrow’s option price. To see that the bounds are recursive, consider a two-period version of the problem, C0 =

min

{m1 , m2 }

E0 (m1 m2 xc2 ) s.t.

pt = Et (mt+1 pt+1 ); Et (m2t+1 ) ≤ A2t , mt+1 > 0, t = 0, 1.

(236)

This two period problem is equivalent to a series of one period problems, in which the C0 problem finds the lowest price of the C1 lower bound, C 1 = min E1 (m2 xc2 ) ; C 0 = min E0 (m1 C 1 ) {m2 }

{m1 }

subject to (18.236). Why? The solution to the two-period problem min E0 (m1 E1 (m2 xc )) must minimize E1 (m2 xc ) in each state of nature at time 1. If not, you could lower E1 (m2 xc ) without affecting the constraints, and lower the objective. Note that this recursive property only holds if we impose m > 0. If m1 < 0 were possible we would want to maximize E1 (m2 xc ) in some states of nature. 18.3.2

Basis risk and real options

The general case leads to some dense formulas, so a simple example will let us understand the idea most simply. Let’s value a European call option on an event V that is not a traded asset, but is correlated with a traded asset that can be used as an approximate hedge. This situation is common with real options and nonfinancial options and describes some financial options on illiquid assets. 309

C HAPTER 18

O PTION PRICING WITHOUT PERFECT REPLICATION

The terminal payoff is xcT = max(VT − K, 0).

Model the joint evolution of the traded asset S and the event V on which the option is written as dS S dV V

= µS dt + σS dz, = µV dt + σ V z dz + σ V w dw.

The dw risk cannot be hedged by the S asset, so the market price of dw risk – its correlation with the discount factor – will matter to the option price. We are looking for a discount factor that prices S and rf , has instantaneous volatility A, and generates the largest or smallest price for the option. Hence, it will have the largest loading on dw possible. By analogy with the one period case (18.226), you can quickly see that the discount factor will have the form q dΛ∗ dΛ = ± A2 − h2S dw Λ Λ∗ dΛ∗ = −rdt − hS dz Λ∗ µS − r hS = . σS dΛ∗ /Λ∗ is the familiar analogue to x∗ that prices stock and bond. We add a loading on the orthogonal shock dw just sufficient to satisfy the constraint Et (dΛ2 /Λ2 ) = A2 . One of ± will generate the upper bound, and one will generate the lower bound.

Now that we have the discount factor, the good deal bound is given by ¸ · ΛT C t = Et max(VT − K) . Λt St , Vt , and Λt are all diffusions with constant coefficients. Therefore, ST , VT and ΛT are jointly lognormally distributed, so the double integral defining the expectation is straightforward to perform, and works very similarly to the integral we evaluated to solve the BlackScholes formula in section 2.1. (If you get stuck, see Cochrane and Saá-Requejo 1999 for the algebra.)

The result is C or C = V0 e

ηT

µ ¶ µ ¶ 1 √ 1 √ −rT φ d + σ V T − Ke φ d − σV T 2 2

310

(237)

S ECTION 18.3

MULTIPLE PERIODS AND CONTINUOUS TIME

where φ(·) denotes the left tail of the normal distribution and dV 2 = σ 2V z + σ2V w V2 ln(V0 /K) + (η + r) T √ d ≡ σV T s Ã !# " p A2 − 1 1 − ρ2 σV η ≡ hV − hS ρ − a h2S

σ 2V

≡ Et

µS − r µ −r ; hV ≡ V σS σ µ ¶ V dV dS σV z , ρ ≡ corr = V S σV ½ +1 upper bound . a = −1 lower bound

hS

≡

This expression is exactly the Black-Scholes formula with the addition of the η term. µV enters the formula because the event V may not grow at the same rate as the asset S . Obviously, the correlation ρ between V shocks and asset shocks enters the formula, and as this correlation declines, the bounds widen. The bounds also widen as the volatility constraint A becomes larger relative to the asset Sharpe ratios hS . Market prices of risk

Continuous-time pricing problems are often specified in terms of “market prices of risk” rather than discount factors. This is the instantaneous Sharpe ratio that an asset must earn if it loads on a specific shock. If an asset has a price process P that loads on a shock σdw, then its expected return must be ¶ µ dP dΛ f − r dt = −σEt dw Et P Λ with Sharpe ratio f Et dP P − r dt = −Et λ= σ

µ

¶ dΛ dw . Λ

I have introduced the common notation λ for the market price of risk. Thus, problems are often attacked by making assumptions about λ directly and then proceeding from Et

dP − rf dt = λσ. P

In this language, the market price of stock risk is hS and can be measured by observing the stock, and does not matter when you can price by arbitrage (notice it is missing from the Black-Scholes formula). Our problem comes down to choosing the market price of dw 311

C HAPTER 18

O PTION PRICING WITHOUT PERFECT REPLICATION

risk, which cannot be measured by observing a traded asset, in such a way as qto minimize or maximize the option price, subject to a constraint that the total price of risk 18.3.3

h2S + λ2 ≤ A.

Continuous time

Now, a more systematic expression of the same ideas in continuous time. As in the option pricing case in the last chapter and the term structure case in the next chapter, we will obtain a differential characterization. To actually get prices, we have either to solve the discount factor forward, or to find a differential equation for prices which we solve backward. Basis assets

In place of E(x), E(xx0) etc., model the price processes of an nS -dimensional vector of basis assets by a diffusion, dS = µS (S, V, t)dt + σS (S, V, t)dz; E(dz dz 0 ) = I S

(238)

Rather than complicate£the notation, understand division to operate element by element on ¤ vectors, e.g., dS/S = dS1 /S1 dS2 /S2 · · · . The basis assets may pay dividends at rate D(S, V, t)dt. V represents an nV -dimensional vector of additional state variables that follow

dV = µV (S, V, t)dt + σ V z (S, V, t)dz + σV w (S, V, t)dw; E(dw dw0 ) = I; E(dw dz 0 ) = 0. (239)

This could include a stochastic stock volatility or stochastic interest rate – classic cases in which the Black-Scholes replication breaks down. Again, I keep it simple by assuming there is a risk free rate r(S, V, t)dt. The problem

We want to value an asset that pays continuous dividends at rate xc (S, V, t)dt and with a terminal payment xcT (S, V, T ). Now we must choose a discount factor process to minimize the asset’s value Ct =

min

{Λs , t<s≤T }

Et

Z

T

s=t

Λs c x ds + Et Λt s

µ

ΛT c x Λt T

¶

(240)

subject to the constraints that 1) the discount factor prices the basis assets S, r at each moment in time, 2) the instantaneous volatility of the discount factor process is less than a prespecified value A2 and 3) the discount factor is positive Λs > 0, t ≤ s ≤ T .

One period at a time; differential statement

312

S ECTION 18.3

MULTIPLE PERIODS AND CONTINUOUS TIME

Since the problem is recursive, we can study how to move one step back in time, C t Λt = min Et {Λs }

Z

t+∆t

s=t

or, for small time intervals,

¡ ¢ Λs xcs ds + Et Λt+∆t C t+∆t

C t Λt = min Et {xc ∆t + (C t + ∆C) (Λt + ∆Λ)} . {∆Λ}

Letting ∆t → 0, we can write the objective in differential form, 0=

xct Et [d (ΛC)] dt + min , C {dΛ} ΛC

(241)

subject to the constraints. We can also write (18.241) as dC xct Et + dt − rf dt = − min Et C C {dΛ}

µ

dΛ dC Λ C

¶

.

(242)

Since the second and third terms on the left hand side are fixed, the condition sensibly tells us to find the lowest value C by maximizing the drift of the bound at each date. You should recognize the form of (18.241) and (18.242) as the basic pricing equations in continuous time, relating expected returns to covariance with discount factors. Constraints

Now we express the constraints. As in the discrete time case, we orthogonalize the discount factor in m = x∗ + ε form, and then the solution pops out. Any discount factor that prices the basis assets is of the form dΛ∗ dΛ = ∗ − vdw Λ Λ

(243)

where dΛ∗ Λ∗ µ ˜S

≡ −rdt − µ ˜ 0S Σ−1 S σ S dz ≡ µS +

D − r; ΣS = σS σ0S . S

and v is a 1 × nV matrix. We can add shocks orthogonal to dw if we like, but they will have no effect on the answer; the minimization will say to set such loadings to zero. The volatility constraint is 1 dΛ2 Et 2 ≤ A2 dt Λ

313

C HAPTER 18

O PTION PRICING WITHOUT PERFECT REPLICATION

and hence, using (18.243), vv0 ≤ A2 −

1 dΛ∗2 Et ∗2 = A2 − µ ˜ 0S Σ−1 ˜S . S µ dt Λ

(244)

By expressing the constraints via (18.243) and (18.244), we have again reduced the problem of choosing the stochastic process for Λ to the choice of loadings v on the noises dw with unknown values, subject to a quadratic constraint on vv0 . Since we are picking differentials and have ruled out jumps, the positivity constraint is slack so long as Λ > 0. Market prices of risk

Using equation (18.243), v is the vector of market prices of risks of the dw shocks – the expected return that any asset must offer if its shocks are dw: −

1 E dt

µ

dΛ dw Λ

¶

= v.

Thus, the problem is equivalent to: find at each date the assignment of market prices of risk to the dw shocks that minimizes (maximizes) the focus payoff value, subject to the constraint that the total (sum of squared) market price of risk is bounded by A2 . Now, we’re ready to follow the usual steps. We can characterize a differential equation for the option price that must be solved back from expiration, or we can try to solve the discount factor forward and take an expectation. Solutions: the discount factor and bound drift at each instant

We can start by characterizing the bound’s process, just as the basis assets follow (18.238). This step is exactly the instantaneous analogue of the one-period bound without a positivity constraint, so remember that logic if the equations start to get a bit forbidding. Guess that lower bound C follows a diffusion process, and figure out what the coefficients must look like. Write dC = µC (S, V, t)dt + σCz (S, V, t)dz + σCw (S, V, t)dw. C

(245)

σCz and σ Cw capture the stochastic evolution of the bound over the next instant – the analogues to E(xxc ), etc. that were inputs to the one period problem. Therefore, a differential or moment-to-moment characterization of the bound will tell us µC and dΛin terms of σCz and σ Cw .

Theorem:The lower bound discount factor Λt follows dΛ∗ dΛ = ∗ − v dw Λ Λ

314

(246)

S ECTION 18.3

MULTIPLE PERIODS AND CONTINUOUS TIME

and µC , σ Cz and σCw satisfy the restriction ¶ µ ∗ xc 1 dΛ µC + − r = − Et σCz dz + vσ0Cw C dt Λ∗

(247)

where v=

r

A2 −

σCw 1 dΛ∗2 Et ∗2 q dt Λ σ Cw σ 0

(248)

Cw

The upper bound process C t and discount factor Λt have the same representation with v = −v. This theorem has the same geometric interpretation as shown in Figure 35. dΛ∗ /Λ∗ is the combination of basis asset shocks that prices the basis assets by construction, in analogy to x∗ . The term σCw dw corresponds to the error w, and σCw σ0Cw corresponds to E(w2 ). The proposition looks a little different because now we choose a vector v rather than a number. We could define a residual σCw dw and then the problem would reduce to choosing a number, the loading of dΛ on this residual. It is not convenient to do so in this case since σCw potentially changes over time. In the geometry of Figure 35, the w direction may change over time. The algebraic proof just follows this logic. Proof: Substituting equation (18.243) into the problem (18.241) in order to impose the pricing constraint, the problem is · ¸ µ ¶ µ ∗2 ¶ xc 1 d(Λ∗ C) dC dΛ 0 2 0 = dt + Et s.t. vv ≤ A − Et − min vEt dw . C Λ∗ C C dt Λ∗2 {v}

Using equation (18.245) for dC/C in the last term, the problem is · ¸ µ ∗2 ¶ xc 1 1 d(Λ∗ C) dΛ 0 0 2 E 0= + Et vσ s.t. vv ≤ A − − min . t Cw ∗ C dt Λ C dt Λ∗2 {v}

(249)

This is a linear objective in v with a quadratic constraint. Therefore, as long as σCw 6= 0, the constraint binds and the optimal v is given by (18.248). v = −v gives the maximum since σ Cw σ0Cw > 0. Plugging the optimal value for v in (18.249) gives · ¸ xc 1 d (Λ∗ C) 0= + Et − vσ 0Cw . C dt Λ∗ C For clarity, and exploiting the fact that dΛ∗ does not load on dw, write the middle term as ¶ · ¸ µ ∗ 1 d (Λ∗ C) dΛ 1 Et σCz dz = µC − r + Et dt Λ∗ C dt Λ∗ 315

C HAPTER 18

O PTION PRICING WITHOUT PERFECT REPLICATION

If σ Cw = 0, any v leads to the same price bound. In this case we can most simply ¥ take v = 0. As in the discrete-time case, we can plug in the definition of Λ∗ to obtain explicit, if less intuitive, expressions for the optimal discount factor and the resulting lower bound, q σCw dΛ = −rdt + µ ˜ 0S Σ−1 σ dz − A2 − µ ˜ 0S Σ−1 ˜S q dw (250) S S S µ Λ σ Cw σ 0Cw µC +

xc −r =µ ˜ 0S Σ−1 S σ S σ Cz + C

q q A2 − µ ˜ 0S Σ−1 µ ˜ σCw σ0Cw . S S

(251)

A partial differential equation

Now we are ready to apply the standard method; find a partial differential equation and solve it backwards to find the price at any date. The method proceeds exactly as for the BlackScholes formula: Guess a solution C(S, V, t). Use Ito’s lemma to derive expressions for µC and σ Cz , σCw in terms of the partial derivatives of C(S, V, t). Substitute these expressions into restriction (18.251). The result is ugly, but straightforward to evaluate numerically. Just like the Black-Scholes partial differential equation, it expresses the time derivative ∂C/∂t in terms of derivatives with respect to state variables, and thus can be used to work back from a terminal period. Theorem. The lower bound C(S, V, t) is the solution to the partial differential equation xc − rC +

∂C + ∂t

+

X ∂ 2C 1 X ∂2C 1 X ∂ 2C Si Sj σSi σ 0Sj + (σ V zi σ 0V zj + σ V wj σ0V wj ) + Si σSi σ0V zj = 2 i,j ∂Si ∂Sj 2 i,j ∂Vi ∂Vj ∂S ∂V i j i,j

=

µ

D −r S

¶0

q q ¢ ¡ 0 −1 (SC S ) + µ ˜ S ΣS σ S σ0V z − µ0V C V + A2 − µ ˜ 0S Σ−1 ˜ S C 0V σV w σ 0V w C V S µ

subject to the boundary conditions provided by the focus asset payoff xcT . C V denotes the vector with typical element ∂C/∂Vj and (SC S ) denotes the vector with typical element Si ∂C/∂Si . Replacing + with − before the square root gives the partial differential equation satisfied by the upper bound. The discount factor

316

S ECTION 18.4

EXTENSIONS , OTHER APPROACHES , AND BIBLIOGRAPHY

In general, the Λ process (18.246) or (18.250) depends on the parameters σCw . Hence, without solving the above partial differential equation we do not know how to spread the loading of dΛ across the multiple sources of risk dw whose risk prices we do not observe. Equivalently, we do not know how to optimally spread the total market price of risk across the elements of dw. Thus, in general we cannot use the integration approach– solve the discount factor forward – to find the bound by µ ¶ Z T Λs c ΛT c C t = Et xs ds + Et x . Λt T s=t Λt However, if there is only one shock dw, then we don’t have to worry about how the loading of dΛ spreads across multiple sources of risk. v can be determined simply by the volatility constraint. In this special case, dw and σCw are scalars. Hence equation (18.246) simplifies as follows: Theorem: In the special case that there is only one extra noise dw driving the V process, we can find the lower bound discount factor Λ from directly from q dΛ = −rdt − µ ˜ 0S Σ−1 σ dz − A2 − µ ˜ 0S Σ−1 ˜ S dw. (252) S S S µ Λ

I used this characterization to solve for the case of a non-traded underlying in the last section. In some applications, the loading of dΛ on multiple shocks dw may be constant over time. In such cases, one can again construct the discount factor and solve for bounds by (possibly numerical) integration, avoiding the solution of a partial differential equation.

18.4

Extensions, other approaches, and bibliography

The roots of the good deal good deal idea go a long way back. Ross (1976) bounded APT residuals by assuming that no portfolio can have more than twice the market Sharpe ratio, and I used the corresponding idea that discount factor volatility should be bounded to generate a robust approximate APT in Chapter 9.4. Good deal bounds apply the same idea to option payoffs. However, the good deal bounds also impose positive discount factors, and this constraint is important in an option pricing context. We also study dynamic models that chain discount factors together as in the option pricing literature. The one-period good-deal bound is the dual to the Hansen-Jagannathan (1991) bound with positivity – Hansen and Jagannathan study the minimum variance of positive discount factors that correctly price a given set of assets. The good deal bound interchanges the position of the option pricing equation and the variance of the discount factor. The techniques for solving the bound, therefore, are exactly those of the Hansen-Jagannathan bound in this one-period setup. There is nothing magic about discount factor volatility. This kind of problem needs weak 317

C HAPTER 18

O PTION PRICING WITHOUT PERFECT REPLICATION

but credible discount factor restrictions that lead to tractable and usefully tight bounds. Several other similar restrictions have been proposed in the literature. 1) Levy (1985) and Constantinides (1998) assume that the discount factor declines monotonically with a state variable; marginal utility should decline with wealth. 2) The good deal bounds allow the worst case that marginal utility growth is perfectly correlated with a portfolio of basis and focus assets. In many cases one could credibly impose a sharper limit than −1 ≤ ρ ≤ 1 on this correlation to obtain tighter bounds. 3) Bernardo and Ledoit (1999) use the restriction a ≥ m ≥ b to sharpen the no-arbitrage restriction ∞ ≥ m > 0. They show that this restriction has a beautiful portfolio interpretation – a < m < b corresponds to limited “gain - loss ratios” just as σ(m)/E(m) corresponds to limited Sharpe ratios. Define [Re ]+ = max(Re , 0) and [Re ]− = − min(−Re , 0) as the “gains and losses” of an excess return Re . Then, max e e

{R ∈R

[Re ]+ sup(m) min e − = e } [R ] {m:0=E(mR )} inf(m)

(253)

(The sup and inf ignore measure zero states.) This is exactly analogous to max e e

{R ∈R

|E(Re )| σ(m) = min } σ(Re ) {m:0=E(mRe )} E(m)

and hints at an interesting restatement of asset pricing theory in L1 with sup norm rather than L2 with second moment norm. Since m ≥ a, the call option price generated by this restriction in a one-period model is strictly greater than the lower arbitrage bound generated by m = 0; as in this case, the gain-loss bound can improve on the good-deal bound. 4) Bernardo and Ledoit also suggest a ≥ m/y ≥ b where y is an explicit discount factor model such as the consumption-based model or CAPM, as a way of imposing a “weak implication” of that particular model. These alternatives are really not competitors. Add all the discount factor restrictions that are appropriate and useful for a given problem. This exercise seems to me a strong case for discount factor methods as opposed to portfolio methods. The combination of positivity and volatility constraints on the discount factor leads to a sharper bound than the intersection of no-arbitrage and limited Sharpe ratios. I don’t know of a simple portfolio characterization of the set of prices that are ruled out by the good deal bound when both constraints bind. The same will be true as we add, say, gain-loss restrictions, monotonicity restrictions, etc. In continuous time, option pricing and terms structure problems increasingly feature assumptions about the “market price of risk” of the non-traded shocks. The good-deal bounds treat these rather formally; they choose the market prices of risks at each instant to minimize or maximize the option price subject to a constraint that the total market price of risk is less 318

S ECTION 18.5

P ROBLEMS

than a reasonable value, compared to the Sharpe ratios of other trading opportunities. One needn’t be this formal. Many empirical implementations of option pricing and term structure models feature unbelievable sizes and time-variation in market prices of risk. Just imposing sensible values for the market prices of risk, and trying on a range of sensible values may be good enough for many practical situations. The continuous-time treatment has not yet been extended to the important case of jumps rather than diffusion processes. With jumps, both the positivity and volatility constraints will bind.

18.5 1.

Problems

Prove (18.253), max e e

{R ∈R

2.

[Re ]+ sup(m) min e − = e } [R ] {m:0=E(mR )} inf(m)

Start with a finite state space. Binomial models are very popular in option pricing. This simple problem illustrates the technique. A stock currently selling at S will either rise to ST = uS with probability πu or decline to ST = dS with probability π d , paying no dividends in the interim. There is a constant gross interest rate Rf . (a) (b) (c) (d)

Find a discount factor that prices stock and bond. This means, find its value in each state of nature. Use the discount factor to price a call option one step before expiration. Express your results as an expected value using risk-neutral probabilities Do the same thing two steps before expiration. Cox, Ross and Rubinstein (1979) derive these formulas by setting up a hedge portfolio of stocks and bonds, and finding portfolio weights to exactly synthesize the option. Rederive your result with this method.

319

Chapter 19. rates

Term structure of interest

Term structure models are particularly simple, since bond prices are just the expected value of the discount factor. In equations, the price at time t of a zero coupon bond that comes (j) due at time t + j is Pt = Et (mt,t+j ). Thus, once you specify a time-series process for a one-period discount factor mt,t+1 , you can in principle find the price of any bond by chaining (j) together the discount factors and finding Pt = Et (mt,t+1 mt+1,t+2 ...mt+j−1,t+j ). As with option pricing models, this chaining together can be hard to do, and much of the analytical machinery in term structure models centers on this technical question. As with option pricing models, there are two equivalent ways to do the chaining together: Solve the discount factor forward and take an integral, or find a partial differential equation for prices and solve it backwards from the maturity date.

19.1

Definitions and notation

A quick introduction to bonds, yields, holding period returns, forward rates, and swaps. (N)

= log price of N period zero-coupon bond at time t.

(N)

= − N1 p(N) = log yield.

pt y

(N)

(N−1)

hprt+1 = pt+1 hpr =

dP (N,t) P

(N→N+1) ft

f(N, t) =

19.1.1

=

−

(N)

− pt

= log holding period return.

1 ∂P (N,t) P ∂N dt

(N) pt

−

(N+1) pt

(N,t) − P1 ∂P∂N

= instantaneous return.

= forward rate.

= instantaneous forward rate.

Bonds

The simplest fixed-income instrument is a zero-coupon bond. A zero-coupon bond is a promise to pay one dollar (a nominal bond) or one unit of the consumption good (a real (3) bond) on a specified date. I use a superscript in parentheses to denote maturity: Pt is the price of a three year zero-coupon bond. I will suppress the t subscript when it isn’t necessary. (N)

(N)

I denote logs by lowercase symbols, pt = ln Pt . The log price has a nice interpretation. If the price of a one-year zero coupon bond is 0.95, i.e. 95/c per dollar face value, the log price is ln(0.95) = −0.051. This means that the bond sells at a 5% discount. Logs also 320

S ECTION 19.1

D EFINITIONS AND NOTATION

give the continuously compounded rate. If we write erN = 1/P (N) then the continuously compounded rate is rN = − ln P (N) .

Coupon bonds are common in practice. For example, a $100 face value 10 year coupon bond may pay $5 every year for 10 years and $100 at 10 years. (Coupon bonds are often issued with semiannual or more frequent payments, $2.50 every six months for example.) We price coupon bonds by considering them as a portfolio of zeros.

Yield. The yield of a bond is the fictional, constant, known, annual, interest rate that justifies the quoted price of a bond, assuming that the bond does not default. It is not the rate of return of the bond. From this definition, the yield of a zero coupon bond is the number Y (N) that satisfies

Hence

1 P (N) = £ ¤N . (N) Y 1 1 (N) Y (N) = £ = − p(N) . ¤ N1 ; y N P (N)

The latter expression nicely connects yields and prices. If the price of a 4 year bond is -0.20 or a 20% discount, that is 5% discount per year, or a yield of 5%. The yield of any stream of cash flows is the number Y that satisfies P =

N X CFj j=1

Yj

.

In general, you have to search for the value Y that solves this equation, given the cash flows and the price. So long as all cash flows are positive, this is fairly easy to do. As you can see, the yield is just a convenient way to quote the price. In using yields we make no assumptions. We do not assume that actual interest rates are known or constant; we do not assume the actual bond is default-free. Bonds that may default trade at lower prices or higher yields than bonds that are less likely to default. This only means a higher return if the bond happens not to default. 19.1.2

Holding Period Returns

If you buy an N period bond and then sell it—it has now become an N −1 period bond—you achieve a return of (N−1)

(N)

HP Rt+1 =

P $back = t+1 (N) $paid Pt

321

(254)

CHAPTER 19

T ERM STRUCTURE OF INTEREST RATES

or, of course, (N)

(N−1)

hprt+1 = pt+1

(N)

− pt

.

We date this return (from t to t + 1) as t + 1 because that is when you find out its value. If this is confusing, take the time to write returns as HP Rt→t+1 and then you’ll never get lost. In continuous time, we can easily find the instantaneous holding period return of bonds with fixed maturity date P (T, t) hpr =

P (T, t + ∆) − P (T, t) P (T, t)

and, taking the limit, hpr =

dP (T, t) . P

However, it’s nicer to look for a bond pricing function P (N, t) that fixes the maturity rather than the date. As in (19.254), we have to account for the fact that you sell bonds that have shorter maturity than you buy. hpr = =

P (N − ∆, t + ∆) − P (N, t) P (N, t) P (N − ∆, t + ∆) − P (N, t + ∆) + P (N, t + ∆) − P (N, t) P (N, t)

and, taking the limit hpr =

19.1.3

1 ∂P (N, t) dP (N, t) − dt P P ∂N

(255)

Forward rate

The forward rate is defined as the rate at which you can contract today to borrow or lend money starting at period N, to be paid back at period N + 1. You can synthesize a forward contract from a spectrum of zero coupon bonds, so the forward rate can be derived from the prices of zero-coupon bonds. Here’s how. Suppose you buy one N period zero and simultaneously sell x N + 1 period zero coupon bonds. Let’s track your cash flow at every date. 322

S ECTION 19.1

D EFINITIONS AND NOTATION

Buy N-Period zero −P (N) 1

Today 0: Time N: Time N+1:

Sell x N+1 Period zeros +xP (N+1) -x

Net cash flow xP (N+1) − P (N) 1 -x

Now, choose x so that today’s cash flow is zero: x=

P (N) P (N+1)

You pay or get nothing today, you get $1.00 at N , and you pay P (N) /P (N+1) at N + 1. You have synthesized a contract signed today for a loan from N to N + 1—a forward rate! Thus, (N)

(N→N+1)

Ft

= Forward rate at t for N → N + 1 =

Pt

(N+1)

Pt

and of course (N→N+1)

ft

(N)

= pt

(N+1)

− pt

(256)

. (N)

People sometimes identify forward rates by the initial date, ft , and sometimes by the (N+1) ending date, ft . I use the arrow notation when I want to be really clear about dating a return. Forward rates have the lovely property that you can always express a bond price as its discounted present value using forward rates, (N)

(N)

pt

= pt

(N−1)

− pt

(N−1→N)

= −ft (1)

(yt

(0→1)

= ft

(N−1)

+ pt

(N−2)

− pt

(N−2→N−1)

− ft

(2)

(1)

(1)

− ...pt − pt + pt (1→2)

− ... − ft

(1)

− yt

of course), so (N)

Pt

(N) pt

=e

= e−

P N−1

(j→j+1) j=0 ft

N−1 Y

=

j=0

−1

(j→j+1)

Ft

.

Intuitively, the price today must be equal to the present value of the payoff at rates you can lock in today. In continuous time, we can define the instantaneous forward rate f(N, t) = −

∂p(Nt ) 1 ∂P (N, t) =− P ∂N ∂N

(257)

Then, forward rates have the same property that you can express today’s price as a discounted 323

CHAPTER 19

T ERM STRUCTURE OF INTEREST RATES

value using the forward rate, p(N, t) = −

Z

P (N, t) = e−

N

f(x, t)dx

x=0 RN x=0

f (x,t)dx

.

Equations 19.256 and 19.257 express forward rates as derivatives of the price vs. maturity curve. Since yield is related to price, we can relate forward rates to the yield curve directly. Differentiating the definition of yield y(N, t) = −p(N, t)/N, 1 1 1 1 ∂p(N, t) ∂y(N, t) = 2 p(N, t) − = − y(N, t) + f(N, t) ∂N N N ∂N N N

Thus, f(N, t) = y(N, t) + N

∂y(N, t) . ∂N

In the discrete case, (19.256) implies (N→N+1)

ft

19.1.4

(N)

= −N yt

(N+1)

+ (N + 1)yt

(N+1)

= yt

´ ³ (N+1) (N) . + N yt − yt

Swaps and options

Swaps are an increasingly popular fixed income instrument. The simplest example is a fixedfor-floating swap. Party A may have issued a 10 year fixed coupon bond. Party B may have issued a 10 year variable rate bond – a bond that promises to pay the current one year rate. (For example, if the current rate is 5%, the variable rate issuer would pay $5 for every $100 of face value. A long-term variable rate bond is the same thing as rolling over one-period debt.) They may be unhappy with these choices. For example, the fixed-rate payer may not want to be exposed to interest rate risk that the present value of his promised payments rises if interest rates decline. The variable-rate issuer may want to take on this interest rate risk, betting that rates will rise or to hedge other commitments. If they are unhappy with these choices, they can swap the payments. The fixed rate issuer pays off the variable rate coupons, and the variable rate issuer pays off the fixed rate coupons. Obviously, only the difference between fixed and variable rate actually changes hands. Swapping the payments is much safer than swapping the bonds. If one party defaults, the other can drop out of the contract, losing the difference in price resulting from intermediate interest rate changes, but not losing the principal. For this reason, and because they match the patterns of cashflows that companies usually want to hedge, swaps have become very popular tools for managing interest rate risk. Foreign exchange swaps are also popular: Party A may swap dollar payments for party B’s yen payments. Obviously, you don’t need to have issued 324

S ECTION 19.2

Y IELD CURVE AND EXPECTATIONS HYPOTHESIS

the underlying bonds to enter into a swap contract – you simply pay or receive the difference between the variable rate and the fixed rate each period. The value of a pure floating rate bond is always exactly one. The value of a fixed rate bond varies. Swaps are set up so no money changes hands initially, and the fixed rate is calibrated so that the present value of the fixed payments is exactly one. Thus, the “swap rate” is the same thing as a the yield on a comparable coupon bond. Many fixed income securities contain options, and explicit options on fixed income securities are also popular. The simplest example is a call option. The issuer may have the right to buy the bonds back at a specified price. Typically, he will do this if interest rates fall a great deal, making a bond without this option more valuable. Home mortgages contain an interesting prepayment option: if interest rates decline, the homeowner can pay off the loan at face value, and refinance. Options on swaps also exist; you can buy the right to enter into a swap contract at a future date. Pricing all of these securities is one of the tasks of term structure modeling.

19.2

Yield curve and expectations hypothesis

The expectations hypothesis is three equivalent statements about the pattern of yields across maturity, 1. 2. 3.

The N period yield is the average of expected future one-period yields The forward rate equals the expected future spot rate. The expected holding period returns are equal on bonds of all maturities. The expectations hypothesis is not quite the same thing as risk neutrality, since it ignores 1/2σ 2 terms that arise when you move from logs to levels. The yield curve is a plot of yields of zero coupon bonds as a function of their maturity. Usually, long-term bond yields are higher than short therm bond yields – a rising yield curve – but sometimes short yields are higher than long yields – an inverted yield curve. The yield curve sometimes has humps or other shapes as well. The expectations hypothesis is the classic theory for understanding the shape of the yield curve. More generally, we want to think about the evolution of yields – the expected value and conditional variance of next period’s yields. This is obviously the central ingredients for portfolio theory, hedging, derivative pricing, and economic explanation. The expectations hypothesis is the traditional benchmark for thinking about the expected value of future yields. We can state the expectations hypothesis in three mathematically equivalent forms:

325

CHAPTER 19 1.

2.

3.

T ERM STRUCTURE OF INTEREST RATES

The N period yield is the average of expected future one-period yields ³ ´ 1 (N) (1) (1) (1) (1) yt = Et yt + yt+1 + yt+2 + ... + yt+N−1 (+ risk premium) N The forward rate equals the expected future spot rate ³ ´ (1) ftN→N+1 = Et yt+N (+ risk premium)

(258)

(259)

The expected holding period returns are equal on bonds of all maturities (N)

(1)

Et (hprt+1 ) = yt

(+ risk premium).

(260)

You can see how the expectations hypothesis explains the shape of the yield curve. If the yield curve is upward sloping – long term bond yields are higher than short term bond yields – the expectations hypothesis says this is because short term rates are expected to rise in the future. You can view the expectations hypothesis as a response to a classic misconception. If long term yields are 10% but short term yields are 5%, an unsophisticated investor might think that long-term bonds are a better investment. The expectations hypothesis shows how this may not be true. Future short rates are expected to rise: this means that you will roll over the short term bonds at a really high rate, say 20%, giving the same long-term return. When the short term interest rates rise in the future, long-term bond prices decline. Thus, the long-term bonds will only give a 5% rate of return for the first year. You can see from the third statement that the expectations hypothesis is roughly the same as risk-neutrality. If we had said that the expected level of returns was equal across maturities, that would be the same as risk-neutrality. The expectations hypothesis specifies that the expected log return is equal across maturities. This is typically a close approximation to risk 1 2 neutrality, but not the same thing. If returns are log-normal, then E(R) = eE(r)+ 2 σ (r) . If mean returns are about 10% or 0.1 and the standard deviation of returns is about 0.1, then 1/2σ 2 is about 0.005, which is very small but not zero. We could easily specify risk-neutrality in the third expression of the expectations hypothesis, but then it would not imply the other two – 12 σ 2 terms would crop up. The intuition of the third form is clear: risk-neutral investors will adjust positions until the expected one-period returns are equal on all securities. Any two ways of getting money from t to t + 1 must offer the same expected return. The second form adapts the same idea to the choice of locking in a forward contract vs. waiting and borrowing and lending at the spot rate. Risk-neutral investors will load up on one or the other contract until the expected returns are the same. Any two ways of getting money from t + N to t + N + 1 must give the same expected return. The first form reflects a choice between two ways of getting money from t to N. You can buy a N period bond, or roll-over N one-period bonds. Risk neutral investors will choose one over the other strategy until the expected N − period return is the same. 326

S ECTION 19.3

T ERM STRUCTURE MODELS – A DISCRETE - TIME INTRODUCTION

The three forms are mathematically equivalent. If every way of getting money from t to t + 1 gives the same expected return, then so must every way of getting money from t + 1 to t + 2, and, chaining these together, every way of getting money from t to t + 2. For example, let’s show that forward rate = expected future spot rate implies the yield curve. Start by writing (1)

ftN−1→N = Et (yt+N−1 ).

Add these up over N, ³ ´ (1) (1) (1) (1) ft0→1 + ft1→2 + ... + ftN−2→N−1 + ftN−1→N = Et yt + yt+1 + yt+2 + ... + yt+N−1 .

The right hand side is already what we’re looking for. Write the left hand side in terms of the definition of forward rates, remembering P (0) = 1 so p(0) = 0, ´ ³ ´ ³ ´ ³ (0) (1) (1) (2) (N−1) (N) + pt − pt + ... + pt ft0→1 + ft1→2 + ... + ftN−2→N−1 + ftN−1→N = pt − pt − pt (N)

= −pt

(N)

= Nyt

.

You can show all three forms are equivalent by following similar arguments. (This is a great problem.) It is common to add a constant risk premium and still refer to the resulting model as the expectations hypothesis, and I include a risk premium in parentheses to remind you of this idea. One end of each of the three statements does imply more risk than the other. A forward rate is known while the future spot rate is not. Long-term bond returns are more volatile than short term bond returns. Rolling over short term real bonds is a riskier long-term investment than buying a long term real bond. If real rates are constant, and the bonds are nominal, then the converse can hold: short term real rates can adapt to inflation, so rolling over short nominal bonds can be a safer long-term real investment than long-term nominal bonds. These risks will generate expected return premia if they covary with the discount factor, and our theory should reflect this fact. If you allow an arbitrary, time-varying risk premium, the model is a tautology, of course. Thus, the entire content of the “expectations hypothesis” augmented with risk premia is in the restrictions that the risk premium is constant over time. We will see that the constant risk premium model does not do that well empirically. One of the main points of term structure models is – or at least ought to be – to quantify the size and movement over time in the risk premium.

19.3

Term structure models – a discrete-time introduction

327

CHAPTER 19

T ERM STRUCTURE OF INTEREST RATES

Term structure models specify the evolution of the short rate and potentially other state variables, and the prices of bonds of various maturities at any given time as a function of the short rate and other state variables. I examine a very simple example based on an AR(1) for the short rate and the expectations hypothesis, which gives a geometric pattern for the yield curve. A good way to generate term structure models is to write down a process for the discount factor, and then price bonds as the conditional mean of the discount factor. This procedure guarantees the absence of arbitrage. I give a very simple example of an AR(1) model for the log discount factor, which also results in geometric yield curves. . A natural place to start in modeling the term structure is to model yields statistically. You might run regressions of changes in yields on the levels of lagged yields, and derive a model of the mean and volatility of yield changes. You would likely start with a factor analysis of yield changes and express the covariance matrix of yields in terms of a few large factors that describe their common movement. The trouble with this approach is that you can quite easily reach a statistical representation of yields that admits an arbitrage opportunity, and you would not want to use such a statistical characterization for economic understanding of yields, for portfolio formation, or for derivative pricing. For example, a statistical analysis would strongly suggest that a first factor should be a “level” factor, in which all yields move up and down together. It turns out that this assumption violates arbitrage: the long-maturity yield must converge to a constant. (More precisely, the long-term forward rate, if it exists, must never fall.) How do you model yields without arbitrage? An obvious solution is to use the discount factor existence theorem: Write a statistical model for a positive discount factor, and find bond prices as the expectation of this discount factor. Such a model will be, by construction, arbitrage free. Conversely, any arbitrage-free distribution of yields can be captured by some positive discount factor, so you don’t lose any generality with this approach. 19.3.1

A term structure model based on the expectations hypothesis

We can use the expectations hypothesis to give the easiest example of a term structure model. (This one does not start from a discount factor and so may not be arbitrage-free.) Suppose the one-period yield follows an AR(1),

(1)

(1)

yt+1 − δ = ρ(yt

− δ) + εt+1 .

Now, we can use the expectations hypothesis (19.258) to calculate yields on bonds of all 328

S ECTION 19.3

T ERM STRUCTURE MODELS – A DISCRETE - TIME INTRODUCTION

maturities as a function of today’s one period yield, (2)

yt

i 1 h (1) (1) Et yt + yt+1 2 i 1 h (1) (1) yt + δ + ρ(yt − δ) = 2 1 + ρ (1) (yt − δ) = δ+ 2

=

Continuing in this way, ³ ´ 1 1 − ρN+1 (1) (N) yt − δ = (yt − δ). N 1−ρ

(261)

You can see some features that will recur throughout the term structure models. First, the model (19.261) can describe some movements in the yield curve over time. If the short rate is below its mean, then there is a smoothly upward sloping yield curve. Long term bond yields are higher as short rates are expected to increase in the future. If the short rate is above its mean, we get a smoothly inverted yield curve. This particular model cannot produce humps or other interesting shapes that we sometimes see in the term structure. Second, this model (N) (1) predicts no average slope of the term structure – E(yt ) = E(yt ) = δ . In fact, the average term structure seems to slope up slightly. Third, all bond yields move together in the model. If we were to stack the yields up in a VAR representation, it would be ³ ´ (1) (1) yt+1 − δ = ρ yt − δ + εt+1 ³ ´ (2) (2) 1+ρ yt+1 − δ = ρ yt − δ + 2 εt+1 . ... ³ ´ N +1 (N) (N) + N1 1−ρ yt+1 − δ = ρ yt − δ 1−ρ εt+1 (1)

(You can write the right hand variable in terms of yt if you want – any one yield carries the same information as any other.) The error terms are all the same. We can add more factors to the short rate process, to improve on this prediction, but most tractable term structure models maintain less factors than there are bonds, so some perfect factor structure is a common prediction of term structure models. Finally, this model has a problem in that the short rate, following an AR(1), can be negative. Since people can always hold cash, nominal short rates are never negative, so we want to start with a short rate process that does not have this feature. With this simple model in hand, you can see some obvious directions for generalization. First, we will want more complex driving processes than an AR(1). For example, a humpshape in the conditionally expected short rate will result in a hump-shaped yield curve. If there are multiple state variables driving the short rate, then we will have multiple factors driving the yield curve which will also result in more interesting shapes. We also want processes that keep the short rate positive in all states of nature. Second, we will want to add 329

CHAPTER 19

T ERM STRUCTURE OF INTEREST RATES

some “market prices of risk” – some risk premia. This will allow us to get average yield curves to not be flat, and time-varying risk premia seem to be part of the yield data. The yield curve literature proceeds in exactly this way: specify a short rate process and the risk premia, and find the prices of long term bonds. The trick is to specify sufficiently complex assumptions to be interesting, but preserve our ability to solve the models. 19.3.2

The simplest discrete time model

The simplest nontrivial model I can think of is to let the log of the discount factor follow an AR(1) with normally distributed shocks. I write the AR(1) for the log rather than the level in order to make sure the discount factor is positive, precluding arbitrage. Log discount factors are typically slightly negative, so I denote the unconditional mean E (ln m) = −δ (ln mt+1 + δ) = ρ (ln mt + δ) + εt+1 .

In turn, you can think of this discount factor model as arising from a consumption based power utility model with normal errors. mt+1 ct+1 − ct

¶γ Ct+1 = e Ct = ρ(ct − ct−1 ) + δ t+1 . −δ

µ

The term structure literature has only started to think whether the empirically successful discount factor processes can be connected empirically back to macroeconomic events in this way. From this discount factor, we can find bond prices and yields. This is easy because the conditional mean and variance of an AR(1) are easy to find. (I am following the strategy of solving the discount factor forward rather than solving the price backward.) We need (1)

yt

(2)

yt

(1)

= −pt = − ln Et (eln mt+1 ) 1 (2) 1 = − pt = − ln Et (eln mt+1 +ln mt+2 ) 2 2

and so on. Iterating the AR(1) forward, (ln mt+2 + δ) = ρ2 (ln mt + δ) + ρεt+1 + εt+2 (ln mt+3 + δ) = ρ3 (ln mt + δ) + ρ2 εt+1 + ρεt+2 + εt+3

so (ln mt+1 + δ) + (ln mt+2 + δ) = (ρ + ρ2 ) (ln mt + δ) + (1 + ρ)εt+1 + εt+2

330

S ECTION 19.3

T ERM STRUCTURE MODELS – A DISCRETE - TIME INTRODUCTION

(ln mt+1 + δ) + (ln mt+2 + δ) + (ln mt+3 + δ) = (ρ + ρ2 + ρ3 ) (ln mt + δ) + (1 + ρ + ρ2 )εt+1 + (1 + ρ)εt+2 + εt+3 . 1

2

Using the rule for a lognormal E(ex ) = eE(x)+ 2 σx , we have finally (1)

yt

(2)

yt

(3)

yt

1 = δ − ρ(ln mt + δ) − σ2ε 2 (ρ + ρ2 ) 1 + (1 + ρ)2 2 (ln mt + δ) − σε = δ− 2 4 (ρ + ρ2 + ρ3 ) 1 + (1 + ρ)2 + (1 + ρ + ρ2 )2 2 (ln mt + δ) − σε . = δ− 3 6

Notice all yields move as linear functions of a single state variable, ln mt + δ . Therefore, we can substitute out the discount factor and express the yields on bonds of any maturity as functions of the yields on bonds of one maturity. Which one we choose is arbitrary, but it’s conventional to use the shortest interest rate as the state variable. With E(y (1) ) = δ − 12 σ2ε , we can write our term structure model as h i (1) (1) yt − E(y (1) ) = ρ yt−1 − E(y (1) ) + εt (19.262) ´ 1 + (1 + ρ)2 (ρ + ρ2 ) ³ (1) (2) yt − E(y(1) ) − σ2ε yt = δ− 2 4 ´ 1 + (1 + ρ)2 + (1 + ρ + ρ2 )2 (ρ + ρ2 + ρ3 ) ³ (1) (3) yt − E(y (1) ) − σ2ε . yt = δ− 3 6 j ∞ X ´ X ρ 1 − ρN ³ (1) (1 + ρk )2 2 (N) yt − E(y (1) ) + σε = δ− yt N 1−ρ 2j j=1 k=1

This is the form in which term structure models are usually written – an evolution equation for the short rate process (together, in general, with other factors or other yields used to identify those factors), and then longer rates written as functions of the short rate, or the other factors. This is not a very realistic term structure model. In the data, the average yield curve – (N) the plot of {E[yt ]} versus N – is slightly upward sloping. The average yield curve from this model is slightly downward sloping as the σ2ε terms pile up. The effect is not large; with (2) (1) (3) (1) ρ = 0.9 and σ ε = 0.02, I find E(yt ) = E(yt ) − 0.02% and E(yt ) = E(yt ) − 0.06%. Still, it does not slope up. More importantly, this model only produces smoothly upward sloping or downward sloping term structures. For example, with ρ = 0.9, the first three terms multiplying the one period rate in (19.262) are 0. 86, 0. 81, 0.78. Two, three and four period bonds move exactly with one-period bonds using these coefficients. This model shows no conditional heteroskedasticity – the conditional variance of yield changes is always the same. The term structure data show times of high and low volatility, and times of high yields and high yield spreads seem to track these changes in volatility Finally, this model shows a 331

CHAPTER 19

T ERM STRUCTURE OF INTEREST RATES

weakness of almost all term structure models – all the yields move together; they follow an exact one-factor structure, they are perfectly conditionally correlated. The solution, of course, is to specify more complex discount rate processes that give rise to more interesting term structures.

19.4

Continuous time term structure models

The basic steps. 1.

Write a time-series model for the discount factor, typically in the form dΛ = −rdt − σ Λ (·)dz Λ dr = µr (·)dt + σ r (·)dz

2.

Solve the discount factor model forward and take expectations, to find bond prices (N)

Pt

3.

= Et

µ

Λt+N Λt

¶

.

Alternatively, from the basic pricing equation 0 = E [d(ΛP )] we can find a differential equation that the price must follow, ∂P 1 ∂P 2 ∂P ∂P µr + − rP = σr σΛ. σr − 2 ∂r 2 ∂r ∂N ∂r (0)

You can solve this back from PN = 1. I contrast the discount factor approach to the market price of risk and arbitrage pricing approaches. Term structure models are usually more convenient in continuous time. As with the last model, I specify a discount factor process and then find bond prices. A wide and popular class of term structure models are based on discount factor process of the form dΛ Λ dr

= −rdt − σΛ (·)dz = µr (·)dt + σ r (·)dz

332

(19.263)

S ECTION 19.4

CONTINUOUS TIME TERM STRUCTURE MODELS

This specification is analogous to a discrete time model of the form mt+1 xt+1

= xt + σεt+1 = ρxt + εt+1 .

This is a convenient representation since the state variable x carries the mean discount factor information. The r variable starts out as a state variable for the drift of the discount factor. However, you can see quickly that it will become the short rate process since Et (dΛ/Λ) = −rtf dt. The dots (·) remind you that these terms can be functions of state variables. Of course, we can add orthogonal components to the discount factor with no effect on the bond prices. Thus, the perfect correlation between interest rate and discount factor shocks is not essential. Term structure models differ in the specification of the functional forms for µr , σ r , σ Λ . We will study three famous examples, the Vasicek model, the Cox-Ingersoll-Ross model and the general affine specification. The first two are Vasicek

CIR

dΛ Λ dr

dΛ Λ dr

= −rdt − σΛ dz

(19.264)

= φ(¯ r − r)dt + σ r dz

√ = −rdt − σΛ Λdz √ = φ(¯ r − r)dt + σr rdz

(19.265)

The Vasicek model is quite similar to the AR(1) we studied in the last section. The CIR model adds the square root terms in the volatility. This specification captures the fact in US data that higher interest rates seem to be more volatile. In the other direction, it keeps the level of r guarantee that the square root the interest rate from falling below zero. (We need σr ≤ 2φ¯ process does not get stuck at zero.) Having specified a discount factor process, it’s a simple matter to find bond prices once again, µ ¶ Λt+N (N) Pt = Et . Λt We can simply solve the discount factor forward and take the expectation. We can also use the instantaneous pricing condition 0 = E(d(ΛP )) to find a partial differential equation for prices, and solve that backward. Both methods naturally adapt to pricing term structure derivatives – call options on bonds, interest rate floors or caps, “swaptions” that give you the right to enter a swap, and so forth. We simply put any payoff xC that depends on interest rates or interest rate state variables 333

CHAPTER 19

T ERM STRUCTURE OF INTEREST RATES

inside the expectation (N) Pt

= Et

Z

∞

s=t

Λs C x (s)ds. Λt

Alternatively, the price of such options will also be a function of the state variables that drive the term structure, so we can solve the bond pricing differential equation backwards using the option payoff rather than one as the boundary condition. 19.4.1

Expectation approach

As with the Black-Scholes option pricing model, we can think of solving the discount factor forward, and then taking the expectation. We can write the solution8 to (19.263) RT RT 1 2 ΛT = e− s=0 (rs + 2 σΛs )ds − s=0 σΛs dz Λ0

and thus, (N)

P0

³ RT ´ RT 1 2 = E0 e− s=0 (rs + 2 σΛs )ds − s=0 σΛs dz .

(266)

For example, in a riskless economy σΛ = 0, we obtain the continuous-time present value formula, (N)

P0

= e−

RT

s=0

rs ds.

With a constant interest rate r, P0 = e−rT .

In more interesting situations, solving the Λ equation forward and taking the expectation analytically is not so easy. Conceptually and numerically, it is easy, of course. Just simulate the system (19.263) forward a few thousand times, and take the average. 8

If this is mysterious, write first

d ln Λ =

µ ¶ 1 dΛ2 dΛ 1 2 − σ = − r + dt − σ Λ dz Λ 2 Λ2 2 Λ

and then integrate both sides from zero to T .

334

S ECTION 19.4 19.4.2

CONTINUOUS TIME TERM STRUCTURE MODELS

Differential equation approach

Recall the basic pricing equation for a security with price S and no dividends is µ

Et

dS S

¶

− rdt = −Et

µ

dS dΛ S Λ

¶

.

(267)

The left hand side is the expected excess return. As we guessed an option price C(S, t) and used (19.267) to derive a differential equation for the call option price, so we will guess a bond price P (N, t) and use this equation to derive a differential equation for the bond price. If we specified bonds by their maturity date T , P (t, T ), we could apply (19.267) directly. However, it’s nicer to look for a bond pricing function P (N, t) that fixes the maturity rather than the date. Equation (19.255) gives the holding period return for this case, which adds an extra term to correct for the fact that you sell younger bonds than you buy, return =

1 ∂P (N, t) dP (N, t) − dt P P ∂N

Thus, the fundamental pricing equation, applied to the price of bonds of given maturity P (N, t) is Et

µ

dP P

¶

−

µ

¶ µ ¶ dP dΛ 1 ∂P (N, t) + r dt = −Et . P ∂N P Λ

(268)

Now, we’re ready to find a differential equation for the bond price, just as we did for the option price to derive the Black-Scholes formula. Guess that all the time dependence comes through the state variable r, so P (N, r). Using Ito’s lemma dP =

µ

¶ ∂P 1 ∂ 2P 2 ∂P µr + σ r dz. σr dt + 2 ∂r 2 ∂r ∂r

(There is no ∂P/∂N because we defined the P function as the price of bonds with constant maturity N . If we had defined P (T, r) as the price of bonds that come due on date T , then there would be a −∂P/∂T term. This would take the place of the ∂P/∂N term in (19.268).) Plugging in to (19.268) and canceling dt, we obtain the fundamental differential equation for bonds, ∂P 1 ∂P 2 ∂P ∂P µr + − rP = σr σΛ. σr − 2 ∂r 2 ∂r ∂N ∂r

(269)

All you have to do is specify the functions µr (·), σr (·), σ Λ (·) and solve the differential equation. 335

CHAPTER 19 19.4.3

T ERM STRUCTURE OF INTEREST RATES

Market price of risk and risk-neutral dynamic approaches

The bond pricing differential equation (19.269) is conventionally derived without discount factors. One conventional approach is to write the short rate process dr = µr (·)dt + σ r (·)dz , and then specify that any asset whose payoffs have shocks σr dz must offer a Sharpe ratio of λ(·). We would then write ∂P 1 ∂P 2 ∂P ∂P µ + − rP = σr λ. σ − ∂r r 2 ∂r2 r ∂N ∂r

With λ = σΛ , this is just (19.269) of course. (If the discount factor and shock are imperfectly correlated, then λ = σ Λ ρ.) Different authors use “market price of risk” in different ways. Cox Ingersoll and Ross (1985, p.398) warn against modeling the right hand side as ∂P/∂r ψ (·) directly; this specification could lead to a positive expected return when σr = 0 and hence an infinite Sharpe ratio or arbitrage opportunity. By generating expected returns as the covariance of payoff shocks and discount factor shocks, we naturally avoid this mistake and other subtle ways to introduce arbitrage opportunities without realizing that you have done so. A second conventional approach is to use an alternative interest rate and discount factor process, dΛ Λ dr

= −rdt

(19.270)

= (µr − σ r λ)dt + σr dzr .

If we use this alternative process, we obtain 1 ∂P 2 ∂P ∂P (µr − σr λ) + − rP = 0 σ − ∂r 2 ∂r2 r ∂N

which is of course the same thing. This is the “risk neutral probability” approach, since the drift term in (19.270) is not the true drift. Since (19.270) gives the same prices, we can find and represent the bond price via the integral h RT i (N) Pt = Et∗ e− s=0 rs ds where E ∗ represents expectation with respect to the risk-neutral process defined in (19.270) rather than the true probabilities defined by the process (19.263).

When we derive the model from a discount factor, the single discount factor carries two pieces of information. The drift or conditional mean of the discount factor gives the short rate of interest, while the covariance of the discount factor shocks with asset payoff shocks generates expected returns or “market prices of risk.” I find it useful to write the discount factor model to keep the term structure connected with the rest of asset pricing, and to remind 336

S ECTION 19.5

T HREE LINEAR TERM STRUCTURE MODELS

myself where “market prices of risk” come from, and reasonable values for their magnitude. Of course, the result is the same no matter which method you follow. The fact that there are fewer factors than bonds means that once you have one bond price you can derive all the others by “no arbitrage” arguments and make this look like option pricing. Some derivations of term structure models follow this approach, setting up arbitrage portfolios. 19.4.4

Solving the bond price differential equation

Now we have to solve the partial differential equation (19.269) subject to the boundary condition P (N = 0, r) = 1. Solving this equation is straightforward conceptually and numerically. Express (19.269) as ∂P 1 ∂ 2P 2 ∂P = (µr − σ r σ Λ ) + σ − rP. ∂N ∂r 2 ∂r2 r

We can start at N = 0 on a grid of r, and P (0, r) = 1. For fixed N , we can work to one step larger N by evaluating the derivatives on the right hand side. The first step is P (∆N, r) =

∂P ∆N = −r∆N. ∂N

At the second step, ∂P/∂r = ∆N , ∂ 2 P/∂r2 = 0, so P (2∆N, r) = ∆N 2 (µr − σr σ Λ ) − r2 ∆N 2 .

Now the derivatives of µr and σr with respect to r will start to enter, and we let the computer take it from here. Analytic solutions only exist in special cases, which we study next.

19.5

Three linear term structure models

I solve the Vasicek, Cox Ingersoll Ross, and Affine model. Each model gives a linear function for log bond prices and yields, for example,

ln P (N, r) = A(N ) − B(N )r

As we have seen, term structure models are easy in principle and numerically: specify a discount factor process and find its conditional expectation or solve the bond pricing partial differential equation back from maturity. In practice, the computations are hard. I present 337

CHAPTER 19

T ERM STRUCTURE OF INTEREST RATES

next three famous special cases of term structure models – specifications for the discount factor process – that allow analytical or quickly calculable solutions. Analytical or close to analytical solutions are still important, because we have not yet found good techniques for reverse-engineering the term structure. We know how to start with a discount factor process and find bond prices. We don’t know how to start with the characteristics of bond prices that we want to model and construct an appropriate discount factor process. (At least one whose parameters do not change every day, as in the “arbitrage free” literature.) Thus, in evaluating term structure models, we will have to do lots of the “forward” calculations – from assumed discount factor model to bond prices – and it is important that we should be able to do them quickly. Obviously, this field would be dramatically changed if we could find a way to reverseengineer the term structure to directly estimate the discount factor process. Also, the ad-hoc time-series models of the discount factor obviously need to be connected to macroeconomics; if not to consumption, then at least to inflation, marginal products of capital, and macroeconomic variables used by the Federal Reserve in its interventions in the interest rate markets. 19.5.1

Vasicek model via by pde

The Vasicek (1977) model is a special case that allows a fairly easy analytic solution. The method is the same as the more complex analytic solution in the CIR and affine classes, but the algebra is easier, so this is a good place to start. The Vasicek discount factor process is dΛ Λ dr

= −rdt − σΛ dz = φ(¯ r − r)dt + σr dz

Using this process in the basic bond differential equation (19.269), we obtain 1 ∂ 2 P 2 ∂P ∂P ∂P φ(¯ r − r) + − rP = σr σΛ . σ − ∂r 2 dr2 r ∂N ∂r

(271)

I’ll solve this equation with the usual unsatisfying non-constructive technique – guess the functional form of the answer and show it’s right. I guess that log yields and hence log prices are a linear function of the short rate, P (N, r) = eA(N)−B(N)r .

(272)

I take the partial derivatives required in (19.271) and see if I can find A(N ) and B(N ) to make (19.271) work. The result is a set of ordinary differential equations for A(N) and B(N ), and these are of a particularly simple form that can be solved by integration. I solve 338

S ECTION 19.5

T HREE LINEAR TERM STRUCTURE MODELS

them, subject to the boundary condition imposed by P (0, r) = 1. The result is ¢ 1¡ 1 − e−φN φ ¶ µ 2 σ2 1 σr σ r σΛ − r ¯ (N − B(N)) − r B(N )2 . A(N ) = + 2 2φ φ 4φ

B(N ) =

(19.273) (19.274)

The exponential form of (19.272) means that log prices and log yields are linear functions of the interest rate, p(N, r) = A(N ) − B(N )r A(N ) B(N ) + r. y(N, r) = − N N

Solving the pde: details.

The boundary condition P (0, r) = 1 will be satisfied if A(0) − B(0)r = 0.

Since this must hold for every r, we will need A(0) = 0; B(0) = 0.

Given the guess (19.272), the derivatives that appear in (19.271) are 1 ∂Pr P ∂r 1 ∂2P P ∂r2 1 ∂P P ∂N

= −B(N ) = B(N )2 = A0 (N) − B 0 (N )r.

Substituting these derivatives in (19.271), 1 −B(N )φ(¯ r − r) + B(N)2 σ2r − A0 (N ) + B 0 (N )r − r = −B(N )σr σΛ . 2

This equation has to hold for every r, so the terms multiplying r and the constant terms must separately be zero. 1 B(N )2 σ 2r − (φ¯ r − σ r σΛ ) B(N ) 2 B 0 (N) = 1 − B(N)φ. A0 (N) =

(19.275)

We can solve this pair of ordinary differential equations by simple integration. The second 339

CHAPTER 19

T ERM STRUCTURE OF INTEREST RATES

one is dB dN Z dB 1 − φB

= 1 − φB = dN

1 − ln (1 − φB) = N φ

and hence B(N ) =

¢ 1¡ 1 − e−φN . φ

(276)

Note B(0) = 0 so we did not need a constant in the integration. We solve the first equation in (19.275) by simply integrating it, and choosing the constant to set A(0) = 0. Here we go. A0 (N ) = A(N ) = A(N ) = A(N ) =

1 B(N )2 σ 2r − (φ¯ r − σ r σ Λ ) B(N ) 2 Z 2 Z σr B(N)2 dN − (φ¯ r − σ r σΛ ) B(N )dN + C 2 µ ¶Z Z ¡ ¡ ¢ ¢ σ r σΛ σ 2r −φN −2φN 1 − 2e 1 − e−φN dN + C dN − r¯ − +e 2 φ 2φ µ ¶ µ ¶µ ¶ σ 2r e−2φN 2e−φN σr σ Λ e−φN − N + − r ¯ − N + +C φ 2φ φ φ 2φ2

We pick the constant of integration to give A(0) = 0. You can do this explicitly, or figure out directly that the result is achieved by subtracting one from the e−φN terms, Ã ¢ ¡ −2φN ¢! µ ¢! ¡ ¡ −φN ¶Ã e −1 −1 2 e−φN − 1 e σ 2r σr σΛρ − A(N ) = 2 N + − r¯ − N+ φ 2φ φ φ 2φ Now, we just have to make it pretty. I’m aiming for the form given in (19.274). Note ¢ 1 ¡ −φN + e−2φN 2 1 − 2e φ e−2φN − 1 1 − e−φN + φB(N)2 = 2 φ φ −2φN e − 1 φB(N )2 − 2B(N ) = φ B(N)2

=

340

S ECTION 19.5

T HREE LINEAR TERM STRUCTURE MODELS

Then ¶ µ µ ¶ σ2r σ r σΛ φ 2 B(N ) + B(N ) − r ¯ − N − 2B(N ) − (N − B(N )) 2 φ 2φ2 µ ¶ σ2 σr σΛ σ2 − r2 (N − B(N )) . A(N ) = − r B(N)2 − r¯ − 4φ φ 2φ

A(N ) =

We’re done. 19.5.2

Vasicek model by expectation

What if we solve the discount rate forward and take an expectation instead? The Vasicek model is simple enough that we can follow this approach as well, and get the same analytic solution. The same methods work for the other models, but the algebra gets steadily worse. The model is dΛ Λ dr

= −rdt − σΛ dz = φ(¯ r − r)dt + σr dz.

(19.277) (19.278)

The bond price is (N) P0

= E0

µ

ΛN Λ0

¶

(279)

I use 0 and N rather than t and t + N to save a little bit on notation. To find the expectation in (19.279), we have to solve the system (19.277)-(19.278) forward. The steps are simple, though the algebra is a bit daunting. First, we solve r forward. Then, we solve Λ forward. ln Λt turns out to be conditionally normal, so the expectation in (19.279) is the expectation of a lognormal. Collecting terms in the resulting expectation that depend on r0 as the B(N ) term, and the constant term as the A(N) term, we find the same solution as (19.273)-(19.274). The interest rate is just an AR(1). By analogy with a discrete time AR(1) you can guess that its solution is rt =

Z

t

s=0

e−φ(t−s) σr dzs + e−φt r0 + (1 − e−φt )¯ r.

To derive this solution, define r˜ by r˜t = eφt (rt − r¯).

341

(280)

CHAPTER 19

T ERM STRUCTURE OF INTEREST RATES

Then, d˜ rt d˜ rt d˜ rt d˜ rt

= = = =

φ˜ rt dt + eφt drt φ˜ rt dt + eφt φ(¯ r − r)dt + eφt σ r dzt φ˜ rt dt − eφt φe−φt r˜t dt + eφt σr dzt eφt σr dzt .

This equation is easy to solve, r˜t − r˜0

= σr

Z

t

eφs dzs

s=0 t

φt

e (rt − r¯) − (r0 − r¯) = σr

Z

s=0

−φt

rt − r¯ = e

eφs dzs (r0 − r¯) + σ r

Z

t

e−φ(t−s) dzs . s=0

And we have (19.280). Now, we solve the discount factor process forward. It isn’t pretty, but it is straightforward. d ln Λt ln Λt − ln Λ0

dΛ 1 dΛ2 1 − = −(rt + σ 2Λ )dt − σ Λ dzt Λ 2 Λ2 2 Z t Z t 1 2 = − (rs + σΛ )ds − σΛ dzs . 2 s=0 s=0 =

Plugging in the interest rate solution (19.280), ¶ ¸ Z t Z t ·µZ s 1 ln Λt − ln Λ0 = − e−φ(s−u) σr dzu + e−φs (r0 − r¯) + r¯ + σ 2Λ ds − σΛ dzs 2 s=0 u=0 s=0 Interchanging the order of the first integral, evaluating the easy ds integrals and rearranging, ¸ ·µ ¸ ¶ Z t Z t ·Z t Z t 1 = −σΛ dzs − σr e−φ(s−u) ds dzu − r¯ + σ2Λ t + (r0 − r¯) e−φs ds 2 s=0 u=0 s=u s=0 µ ¶ ¸ Z t · ´ ³ −φt σr 1 1 − e = − 1 − e−φ(t−u) dzu − r¯ + σ 2Λ t − (r0 − r¯) . (19.281) σΛ + φ 2 φ u=0 The first integral has a deterministic function of time u. This gives rise to a normally distributed random variable – it’s just a weighted sum of independent normals dzu : µ Z t ¶ Z t 2 f(u)dzu ∼ N 0, f (u)du . u=0

u=0

342

S ECTION 19.5

T HREE LINEAR TERM STRUCTURE MODELS

Thus, ln Λt − ln Λ0 is normally distributed with mean given by the second set of terms in (19.281) and variance var0 (ln Λt − ln Λ0 ) = · ´¸2 σr ³ 1 − e−φ(t−u) du σΛ + φ u=0 # ¶2 µ ¶ Z t "µ σr σr σr σ 2r −2φ(t−u) −φ(t−u) = −2 + 2e σΛ + du σΛ + e φ φ φ φ u=0 µ µ ¶2 ¶ ¢ ¢ σ2 ¡ σr σr σr ¡ 1 − e−φt + r3 1 − e−2φt . (19.282) t − 2 2 σΛ + = σΛ + φ φ φ 2φ =

Z

t

Since we have the distribution of ΛN we are ready to take the expectation. ¡ ¢ 1 ln P (N, 0) = ln E0 eln ΛN −ln Λ0 = E0 (ln ΛN − ln Λ0 ) + σ 20 (ln ΛN − ln Λ0 ) . 2

Plugging in the mean from (19.281) and the variance from (19.282) (N)

ln P0

·µ ¶ ¸ 1 1 − e−φN r¯ + σ2Λ N + (r0 − r¯) (19.283) 2 φ ¶2 µ ¶ µ ¡ ¢ ¢ σ2 ¡ 1 σr σr σr −2φN + σΛ N − 2 + σ Λ 1 − e−φN + r3 1 − e(19.284) + 2 φ φ φ 4φ

= −

All that remains is to make it pretty. To compare it with our previous result, we want to express it in the form ln P (N, r0 ) = A(N ) − B(N )r0 . The coefficient on r0 (19.283) is B(N) =

1 − e−φN , φ

the same expression we derived from the partial differential equation. To simplify the constant term, recall that (19.285) implies 1 − e−2φN = −φB(N )2 + 2B(N ). φ

343

(285)

CHAPTER 19

T ERM STRUCTURE OF INTEREST RATES

Thus, the constant term (the terms that do not multiply r0 ) in (19.283) is ·µ ¶ ¸ 1 2 1 − e−φN A(N ) = − r¯ + σΛ N − r¯ 2 φ ¶2 µ ¶ µ ¡ ¢ ¢ σ2 ¡ 1 σr σr σr + σΛ N − 2 + σ Λ 1 − e−φN + r3 1 − e−2φN + 2 φ φ φ 4φ ·µ ¶ ¸ 1 A(N ) = − r¯ + σ2Λ N − r¯B(N ) 2 ¶2 ¶ µ µ ¢ 1 σr σr σr σ2 ¡ + + σΛ N − + σ Λ B(N) − r2 φB(N )2 − 2B(N) 2 φ φ φ 4φ ¶ µ 2 σ 2r σr 1 σr − r¯ (N − B(N )) − 2 φB(N )2 . + σΛ A(N ) = 2 φ2 φ 4φ Again, this is the same expression we derived from the partial differential equation. This integration is usually expressed under the risk-neutral measure. If we write the riskneutral process dΛ Λ dr

= −rdt = [φ(¯ r − r) − σ r σ Λ ] dt + σr dz.

Then the bond price is (N)

P0

= Ee−

RN

s=0

rs ds

.

The result is the same, of course. 19.5.3

Cox Ingersoll Ross Model

For the Cox-Ingersoll-Ross (1985) model dΛ Λ dr

√ = −rdt − σΛ rdz √ = φ(¯ r − r)dt + σr rdz

our differential equation (19.269) becomes 1 ∂ 2P 2 ∂P ∂P ∂P φ(¯ r − r) + − rP = σr σ Λ r. σr r − 2 ∂r 2 ∂r ∂N ∂r

(286)

Guess again that log prices are a linear function of the short rate, P (N, r) = eA(N)−B(N)r .

344

(287)

S ECTION 19.5

T HREE LINEAR TERM STRUCTURE MODELS

Substituting the derivatives of (19.287) into (19.286), 1 −B(N)φ(¯ r − r) + B(N )2 σ 2r r − A0 (N ) + B 0 (N )r − r = −B(N )σ r σ Λ r. 2

Again, the coefficients on the constant and on the terms in r must separately be zero, 1 B 0 (N ) = 1 − σ2r B(N )2 − (σ r σ Λ + φ) B(N ) 2 A0 (N ) = −B(N)φ¯ r.

(19.288)

The ordinary differential equations (19.288) are quite similar to the Vasicek case, (19.275). However, now the variance terms multiply an r, so the B(N ) differential equation has the extra B(N)2 term. We can still solve both differential equations, though the algebra is a little bit more complicated. The result is ¢ ¡ 2 1 − eγN B(N) = (γ + φ + σr σΛ )(eγN − 1) + 2γ µ µ ¶ ¶ φ¯ r 2γ A(N) = 2 ln + ψN σ 2r ψ(eγN − 1) + 2γ where γ ψ

q = (φ + σ r σ Λ )2 + 2σ2r = φ + σΛ σr + γ.

The CIR model can also be solved by expectation. In fact, this is how Cox Ingersoll and Ross (1985) actually solve it – their marginal value of wealth JW is the same thing as the discount factor. However, where the interest rate in the Vasicek model was a simple conditional normal, the interest rate now has a non-central χ2 distribution, so taking the integral is a little messier. 19.5.4

Multifactor affine models

The Vasicek and CIR models are special cases of the affine class of term structure models (Duffie and Kan 1996, Dai and Singleton 1999). These models allow multiple factors, meaning all bond yields are not just a function of the short rate. Affine models maintain the convenient form that log bond prices are linear functions of the state variables. This means that we can take K bond yields themselves as the state variables, and the yields will reveal anything of interest in the hidden state variables. The short rate and its volatility will be forecast by lagged short rates but also by lagged long rates or interest rate spreads. My presentation and notation is similar to Dai and Singleton’s, but as usual I add the discount factor explicitly. 345

CHAPTER 19

T ERM STRUCTURE OF INTEREST RATES

Here is the affine model setup: dy r dΛ Λ dwi

= φ (¯ y − y) dt + Σdw = δ0 + δ0y

(19.289) (19.290)

= −rdt − b0Λ dw q = αi + β 0i ydzi ; E(dzi dzj ) = 0.

(19.291) (19.292)

Equation (19.289) describes the evolution of the state variables. In the end, yields will be linear functions of the state variables, so we can take the state variables to be yields; thus we use the letter y . y denotes a K− dimensional vector of state variables. φ is now a K × K matrix, y¯ is a K− dimensional vector, Σ is a K × K matrix. Equation (19.290) describes the mean of the discount factor or short rate as a linear function of the state variables. Equation (19.291) is the discount factor. bΛ is a K−dimensional vector that describes how the discount factor responds to the K shocks. The more Λ responds to a shock, the higher the market price of risk of that shock. Equation (19.292) describes the shocks dw. The functional form nests the CIR square root type models if αi = 0 and the Vasicek type Gaussian process if β i = 0. You can’t pick αi and β i arbitrarily, as you have to make sure that αi + β 0i y > 0 for all values of y that the process can attain. Dai and Singleton characterize this “admissibility” criterion. We find bond prices in the affine setup following exactly the same steps as for the Vasicek and CIR models. Again, we guess that prices are linear functions of the state variables y . 0

P (N, y) = eA(N)−B(N) y .

We apply Ito’s lemma to this guess, and substitute in the basic bond pricing equation (19.268). We obtain ordinary differential equations that A(N ) and B(N) must satisfy, ¶ Xµ 1 0 ∂B(N ) 2 0 = −φ B(N ) − B(N)i bΛi + [Σ B(N )]i β i + δ (19.293) ∂N 2 i µ ¶ X 1 ∂A(N ) 2 = y − δ 0 . (19.294) B(N )i bΛi + [Σ0 B(N )]i αi − B(N )0 φ¯ ∂N 2 i I use the notation [x]i to denote the ith element of a vector x. As with the CIR and Vasicek models, these are ordinary differential equations that can be solved by integration starting with A(0) = 0, B(0) = 0. While they do not always have analytical solutions, they are quick to solve numerically – much quicker than solving a partial differential equation. Derivation

To derive (19.294) and (19.293), we start with the basic bond pricing equation (19.268), which I repeat here, ¶ µ µ ¶ µ ¶ dP dP dΛ 1 ∂P + r dt = −Et Et (295) − . P P ∂N P Λ 346

S ECTION 19.5

T HREE LINEAR TERM STRUCTURE MODELS

We need dP/P, 1 ∂P 0 1 1 0 ∂ 2P dP = dy dy + dy. P P ∂y 2P ∂y∂y0

The derivatives are 1 ∂P P ∂y 1 ∂ 2P P ∂y∂y0 1 ∂P P ∂N

= −B(N ) = B(N)B 0 (N ) =

∂A(N ) ∂B(N ) 0 − y. ∂N ∂N

Thus, the first term (19.295) is µ ¶ 1 dP Et y − y) dt + Et (dw0 Σ0 B(N )B 0 (N )Σdw) = −B(N )0 φ (¯ P 2 Et (dwi dwj ) = 0, which allows us to simplify the last term. If w1 w2 = 0, then, ¸· ¸ · X ¤ b1 b1 b1 b2 £ w1 (w0 bb0 w) = w1 w2 b2i wi2 . = b21 w12 + b22 w22 = b2 b1 b2 b2 w2

Applying the same algebra to our case, X X ¢ 2 2¡ Et (dw0 Σ0 B(N)B 0 (N )Σdw) = [Σ0 B(N )]i dwi2 = [Σ0 B(N )]i αi + β 0i y dt. i

i

I use the notation [x]i to denote the ith element of the K−dimensional vector x. In sum, we have µ ¶ ¢ 1X 0 dP 2¡ Et y − y) dt + [Σ B(N)]i αi + β 0i y dt. (296) = −B(N )0 φ (¯ P 2 i The right hand side term in (19.295) is µ ¶ dP dΛ −Et = −B(N)0 dwdw0 bΛ P Λ ¡ ¢ dwdw0 is a diagonal matrix with elements αi + β 0i y . Thus, µ ¶ X ¡ ¢ dP dΛ −Et B(N )i bΛi αi + β 0i y =− P Λ i 347

(297)

CHAPTER 19

T ERM STRUCTURE OF INTEREST RATES

Now, substituting (19.296) and (19.297) in (19.295), along with the easier ∂P/∂N central term, we get µ ¶ ¢ 1X 0 ∂A(N ) ∂B(N ) 0 2¡ 0 0 0 − −B(N ) φ (¯ y − y) + [Σ B(N)]i αi + β i y − y + δ0 + δ y 2 i ∂N ∂N X ¡ ¢ = − B(N )i bΛi αi + β 0i y . i

Once again, the terms on the constant and each yi must separately be zero. The constant term: X 1X 0 ∂A(N) 2 − δ0 = − −B(N )0 φ¯ y+ [Σ B(N)]i αi − B(N)i bΛi αi . 2 i ∂N i µ ¶ ∂A(N ) X 1 2 = y − δ0 B(N )i bΛi + [Σ0 B(N )]i αi − B(N)0 φ¯ ∂N 2 i

The terms multiplying y : B(N )0 φy +

X 1X 0 ∂B(N ) 0 2 [Σ B(N )]i β 0i y + y − δ0y = − B(N )i bΛi β 0i y. 2 i ∂N i

Taking the transpose and solving,

¶ Xµ 1 ∂B(N) 2 = −φ0 B(N ) − B(N )i bΛi + [Σ0 B(N)]i β i + δ. ∂N 2 i

19.6

Bibliography and comments

The choice of discrete vs. continuous time is really one of convenience. Campbell, Lo and MacKinlay (1997) give a discrete-time treatment, showing that bond prices are linear functions of the state variables even in a discrete-time two-parameter square root model. Models also don’t have to be affine. Constantinides (1992) is a nice discrete time model; its discount factor is driven by the squared value of AR(1) state variables. It gives closed form solutions for bond prices. The bond prices are not linear functions of the state variables, but it is the existence of closed forms rather than linearity that makes affine models so attractive. It allows for both signs of the term premium, as we seem to see in the data. So far most of the term structure literature has emphasized the risk-neutral probabilities, rarely making any reference to the separation between drifts and market prices of risk. This was not a serious shortcoming for option pricing uses, for which modeling the volatilities is much more important than for modeling the drifts, and to draw smooth yield curves across 348

S ECTION 19.6

BIBLIOGRAPHY AND COMMENTS

maturities. However, it makes the models unsuitable for bond portfolio analysis and other uses. Many models imply high and time-varying market prices of risk or conditional Sharpe ratios. Recently, Duffee (1999) and Duarte (2000) have started the important task of specifying term structure models that fit the empirical facts about expected returns in term structure models. In particular, they try to fit the Fama-Bliss (1986) and Campbell and Shiller (1991) regressions that relate expected returns to the slope of the term structure (see Chapter 20), while maintaining the tractability of affine models. Term structure models used in finance amount to regressions of interest rates on lagged interest rates. Macroeconomists also run regressions of interest rates on a wide variety of variables, including lagged interest rates, but also lagged inflation, output, unemployment, exchange rates, and so forth. They often interpret these equations as the Federal Reserve’s policy-making rule for setting short rates as a function of macroeconomic conditions. This interpretation is particularly clear in the Taylor rule literature (Taylor 1999) and monetary VAR literature, see Christiano Eichenbaum and Evans (1999), Cochrane (1994) for surveys. Someone, it would seem, is missing important right hand variables. The criticism of finance models is stinging when we only use the short rate as a state variable. Multifactor models are more subtle. If any variable forecasts future interest rates, then it becomes a state variable, and it should be revealed by bond yields. Thus, bond yields should completely drive out any other macroeconomic state variables as interest rate forecasters. They don’t, which is an interesting observation. In addition, there is an extensive literature that studies yields from a purely statistical point of view, Gallant and Tauchen (1997) for example, and a literature that studies high frequency behavior in the federal funds market, for example Hamilton (1996). Obviously, these three literatures need to become integrated. Balduzzi, Bertola and Foresi (1996) consider a model based on the Federal Funds target, and Piazzesi (2000) integrated a careful specification of high-frequency moves in the Federal funds rate into a term structure model. The models studied here are all based on diffusions with rather slow-moving state variables. These models generate one-day ahead densities that are almost exactly normal. In fact, as Johannes (2000) points out, one-day ahead densities have much fatter tails than normal distributions predict. This behavior could be modeled by fast-moving state variables. However, it is more natural to think of this behavior as generated by a jump process, and Johannes nicely fits a combined jump-diffusion for yields. This specification can change pricing and hedging characteristics of term structure models significantly. All of the term structure models in this chapter describe many bond yields as a function of a few state variables. This is a reasonable approximation to the data. Almost all of the variance of yields can be described in terms of a few factors, typically a “level” “slope” and “hump” factor. Knez, Litterman and Scheinkman (1994) make the point with a formal maximum likelihood factor analysis, but you can see the point with a simple eigenvalue decomposition of log yields. 349

CHAPTER 19

σ 6.36 0.61 0.10 0.08 0.07

1 0.45 -0.75 0.47 0.10 0.07

T ERM STRUCTURE OF INTEREST RATES

2 0.45 -0.21 -0.62 -0.49 -0.36

Maturity 3 0.45 0.12 -0.41 0.39 0.68

4 0.44 0.36 0.11 0.55 -0.60

5 0.44 0.50 0.46 -0.55 0.21

“Level” “Slope” “Hump”

Eigenvalue decomposition of the covariance matrix of zero coupon bond yields, 1952-1997. The first column gives the square root of the eigenvalues. The columns marked 1-5 give the eigenvectors corresponding to 1-5 year zero coupon bond yields. I decomposed the covariance matrix as Σ = QΛQ0 ; σ2 gives the diagonal entries in Λ and the rest of the table gives the entries of Q. With this decomposition, we can say that bond yields are generated by y = QΛ1/2 ε; E(εε0 ) = I , thus Q give “loadings” on the shocks ε. Not only is the variance of yields well described by a factor model, but the information in current yields about future yields – the expected changes in yields and the conditional volatility of yields – is well captured by one level and a few spreads as well. It is a good approximation, but it is an approximation. Actual bond prices do not exactly follow any smooth yield curve, and the covariance matrix of actual bond yields does not have an exact K factor structure – the remaining eigenvalues are not zero. Hence you cannot estimate a term structure model directly by maximum likelihood; you either have to estimate the models by GMM, forcing the estimate to ignore the stochastic singularity, or you have to add distasteful measurement errors. As always, the importance of an approximation depends on how you use the model. If you take the model literally, a bond whose price deviates by one basis point is an arbitrage opportunity. In fact, it is at best a good Sharpe ratio, but a K factor model will not tell you how good – it won’t quantify the risk involved in using the model for trading purposes. Hedging strategies calculated from K-factor models may be sensitive to small deviations as well. One solution has been to pick different parameters at each point in time (Ho and Lee 1986). This approach is useful for derivative pricing, but is obviously not a satisfactory solution. Models in which the whole yield curve is a state variable, Kennedy (1994), Santa Clara and Sornette (1999), are another interesting response to the problem, and potentially provide a realistic description of the data. The market price of interest rate risk reflects the market price of real interest rate changes and the market price of inflation – or whatever real factors are correlated with inflation. The relative contributions of inflation and real rates in interest rate changes is very important for the nature of the risks that bondholders face. For example, if real rates are constant and nominal rates change on inflation news, then short term bonds are the safest real long 350

S ECTION 19.7

P ROBLEMS

term investment. If inflation is constant and nominal rates change on real rate news, then long term bonds are the safest long term investment. The data seem to suggest a change in regime between the 1970s and 1990s: in the 70’s, most interest rate changes were due to inflation, while the opposite seems true now. Despite all these provocative thoughts, though, little empirical work has been done that usefully separates interest rate risk premia into real and inflation premium components. Buraschi and Jiltsov (1999) is one recent effort in this direction, but a lot more remains to be done.

19.7 1.

2.

3.

4. 5.

Problems

Complete the proof that each of the three statements of the expectations hypothesis implies the other. Is this also true if we add a constant risk premium? Are the risk premia in each of the three statements of the yield curve of the same sign? Under the expectations hypothesis, if long-term yields are higher than short term yields, does this mean that future long term rates should go up, down, or stay the same? (Hint: a plot of the expected log bond prices over time will really help here.) (N) (1) Start by assuming risk neutrality, E(HP Rt+1 ) = Yt for all maturities N . Try to derive the other representations of the expectations hypothesis. Now you see why we specify that the expected log returns are equal. Look at (19.266) and show that adding orthogonal dw to the discount factor has no effect on bond pricing formulas. Look at (19.266) and show that P = e−rT if interest rates are constant, i.e. if dΛ/Λ = −rdt + σΛ dz .

351

PART IV Empirical survey

352

S ECTION 19.7

P ROBLEMS

This part is a brief attempt to survey some of the central empirical issues that have driven recent thinking about financial economics for, and which are driving the development of our theoretical understanding of the nature of risk and risk premia. This part draws heavily on two previous review articles, Cochrane (1998), (1999a) and on Cochrane and Hansen (1992). Fama’s (1970) and (1991) efficient market reviews are classic and detailed reviews of much of the underlying empirical literature, focusing on crosssectional questions. Campbell (1999, 2000) and Kocheralkota (1996) are good surveys of the equity premium literature.

353

Chapter 20. Expected returns in the time-series and cross-section The first revolution in finance started the modern field. Peaking in the early 1970s, this revolution established the CAPM, random walk, efficient markets, portfolio based view of the world. The pillars of this view are: 1.

2.

The CAPM is a good measure of risk and thus a good explanation why some stocks, portfolios, strategies or funds (assets, generically) earn higher average returns than others. Returns are unpredictable. In particular, (a)

3.

Stock returns are close to unpredictable. Prices are close to random walks; expected returns do not vary greatly through time. “Technical analysis” that tries to divine future returns from past price and volume data is nearly useless. Any apparent predictability is either a statistical artifact which will quickly vanish out of sample, or cannot be exploited after transactions costs. The near unpredictably of stock returns is simply stated, but its implications are many and subtle. (Malkiel 1990 is a classic and easily readable introduction.) It also remains widely ignored, and therefore is the source of lots of wasted trading activity. (b) Bond returns are nearly unpredictable. This is the expectations model of the term structure. If long term bond yields are higher than short term yields – if the yield curve is upward sloping – this does not mean that expected long-term bond returns are any higher than those on short term bonds. Rather, it means that short term interest rates are expected to rise in the future, so you expect to earn about the same amount on short term or long term bonds at any horizon. (c) Foreign exchange bets are not predictable. If a country has higher interest rates than are available in the U.S. for bonds of a similar risk class, its exchange rate is expected to depreciate. After you convert your investment back to dollars, you expect to make the same amount of money holding foreign or domestic bonds. (d) Stock market volatility does not change much through time. Not only are returns close to unpredictable, they are nearly identically distributed as well. Professional managers do not reliably outperform simple indices and passive portfolios once one corrects for risk (beta). While some do better than the market in any given year, some do worse, and the outcomes look very much like good and bad luck. Managers who do well in one year are not more likely to do better than average the next year. The average actively-managed fund does about 1% worse than the market index. The more actively a fund trades, the lower returns to investors. Together, these views reflected a guiding principle that asset markets are, to a good approximation, informationally efficient. (Fama 1970, 1991.) This statement means that market prices already contain most information about fundamental value. Informational efficiency 354

in turn derives from competition. The business of discovering information about the value of traded assets is extremely competitive, so there are no easy quick profits to be made, as there are not in every other well-established and competitive industry. The only way to earn large returns is by taking on additional risk. These statements are not doctrinaire beliefs. Rather, they summarize the findings of a quarter-century of extensive and careful empirical work. However, every single one of them has now been extensively revised by a new generation of empirical research. Now, it seems that : 1.

2.

There are assets, portfolios, funds, and strategies whose average returns cannot be explained by their market betas. Multifactor models dominate the empirical description, performance attribution, and explanation of average returns. Returns are predictable. In particular, (a)

3.

Variables including the dividend/price ratio and term premium can in fact predict substantial amounts of stock return variation. This phenomenon occurs over business-cycle and longer horizons. Daily, weekly and monthly stock returns are still close to unpredictable, and “technical” systems for predicting such movements are still close to useless after transactions costs. (b) Bond returns are predictable. Though the expectations model works well in the long run, a steeply upward sloping yield curve means that expected returns on long term bonds are higher than on short term bonds for the next year. (c) Foreign exchange returns are predictable. If you buy bonds in country whose interest rates are unusually higher than those in the U.S., you expect a greater return, even after converting back to dollars. (d) Stock market volatility does in fact change through time. Conditional second moments vary through time as well as first moments. Means and variances do not seem to move in lockstep, so conditional Sharpe ratios vary through time. Some funds seem to outperform simple indices, even after controlling for risk through market betas. Fund returns are also slightly predictable: past winning funds seem to do better in the future, and past losing funds seem to do worse than average in the future. For a while, this seemed to indicate that there is some persistent skill in active management. However, we now see that multifactor performance attribution models explain most fund persistence: funds earn persistent returns by following fairly mechanical “styles,” not by persistent skill at stock selection (Carhart 1997). Again, these views summarize a large body of empirical work. The strength and interpretation of many results are hotly debated. This new view of the facts need not overturn the view that markets are reasonable competitive and therefore reasonably efficient. It does substantially enlarge our view of what activities provide rewards for holding risks, and it challenges our economic understanding of those risk premia. As of the early 1970s, asset pricing theory anticipated the possibility and 355

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

even probability that expected returns should vary over time and that covariances past market betas would be important for understanding cross-sectional variation in expected returns. What took another 15 to 20 years was to see how important these long-anticipated theoretical possibilities are in the data.

20.1

Time-series predictability

I start by looking at patterns in expected returns over time in large market indices, and then look at patterns in expected returns across stocks. 20.1.1

Stocks

Dividend-price ratios forecast excess returns on stocks. Regression coefficients and R2 rise with the forecast horizon. This is a result of the fact that the forecasting variable is persistent. Table 1 gives a simple example of market return predictability. “Low” prices relative to dividends forecast higher subsequent returns. The one-year horizon 0.17 R2 is not particularly remarkable. However, at longer and longer horizons larger and larger fractions of return variation are forecastable. At a 5 year horizon 60% of the variation in stock returns is forecastable ahead of time from the price/divided ratio! (Fama and French 1988.) Horizon k (years) 1 2 3 5

Rt→t+k = a + b(Dt /Pt ) b σ(b) R2 5.3 (2.0) 0.15 10 (3.1) 0.23 15 (4.0) 0.37 33 (5.8) 0.60

Dt+k /Dt = a + b(Dt /Pt ) b σ(b) R2 2.0 (1.1) 0.06 2.5 (2.1) 0.06 2.4 (2.1) 0.06 4.7 (2.4) 0.12

Table 1. OLS regressions of percent excess returns (value weighted NYSE - treasury bill rate) and real dividend growth on the percent VW dividend/price ratio. Rt→t+k indicates the k year return. Standard errors in parenthesis use GMM to correct for heteroskedasticity and serial correlation. Sample 1947-1996. One can object to dividends as the divisor for prices. However, ratios formed with just about any sensible divisor works about as well, including earnings, book value, and moving averages of past prices. Many other variables forecast excess returns, including the term spread between long and short term bonds, the default spread, the level of the T-bill rate, (Fama and French 1989,) 356

S ECTION 20.1

T IME - SERIES PREDICTABILITY

the detrended T-bill rate, and the earnings/dividend ratio (Lamont 1998). Macro variables forecast stock returns as well, including the investment/capital ratio (Cochrane 1991) and the consumption/wealth ratio (Lettau and Ludvigson 2000). Most of these variables are correlated with each other and correlated with or forecast business cycles. This fact suggests a natural explanation, emphasized by Fama and French (1999): Expected returns vary over business cycles; it takes a higher risk premium to get people to hold stocks at the bottom of a recession. When expected returns go up, prices go down. We see the low prices, followed by the higher returns expected and required by the market. (Regressions do not have to have causes on the right and effects on the left. You run regressions with the variable orthogonal to the error on the right, and that is the case here since the error is a forecasting error. This is like a regression of actual weather on a weather forecast.) Table LL, adapted from Lettau and Ludvigson (2000) compares several of these variables. At a one year horizon, both the consumption/wealth ratio and the detrended T bill rate forecast returns, with R2 of 0.18 and 0.10 respectively. At the one year horizon, these variables are more important than the dividend/price and dividend/earnings ratios, and their presence cuts the dividend ratio coefficients in half. However, the d/p and d/e ratios are slower moving than the t bill rate and consumption/wealth ratio. They track decade-to-to decade movements more than business cycle movements. This means that their importance builds with horizon. By six years, the bulk of the return forecastability again comes from the dividend ratios, and it is their turn to cut down the cay and t-bill regression coefficients. The cay and d/e variables have not been that affected by the late 90s, while it has substantially cut down dividend yield forecastability. Horizon(years) 1 1 1 1 6 6 6 6

cay 6.7

5.4 12.4 5.9

d−p

d−e

0.14

0.08

0.07

-0.05

0.95

0.68

0.89

0.65

rrel

-4.5 -3.8 -5.10 1.36

R2 0.18 0.04 0.10 0.23 0.16 0.39 0.03 0.42

Table LL. Long-horizon return forecasts. The return variable is log excess returns on the S&P composite index. cay is Lettau and Ludvigson’s consumption to wealth ratio. d − p is the log dividend yield and e − p is the log earnings yield. rrel is a detrended short term interest rate. Sample 1952:4-1998:3. Source: Lettau and Ludvigson (2000) Table 5. I emphasize that excess returns are forecastable. We have to understand this as timevariation in the reward for risk, not time-varying interest rates. One naturally slips in to non-risk explanations for price variation; for example that the current stock market boom is due to life-cycle savings of the baby boomers. A factor like this does not reference risks; it 357

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

predicts that interest rates should move just as much as stock returns. Persistent d/p; Long horizons are not a separate phenomenon

The results at different horizons are not separate facts, but reflections of a single underlying phenomenon. If daily returns are very slightly predictable by a slow-moving variable, that predictability adds up over long horizons. For example, you can predict that the temperature in Chicago will rise about 1/3 degree per day in the springtime. This forecast explains very little of the day to day variation in temperature, but tracks almost all of the rise in temperature from January to July. Thus, the R2 rises with horizon. Thus, a central fact driving the predictability of returns is that the dividend price ratio is very persistent. Figure 37 plots the d/p ratio and you can see directly that it is extremely slow-moving. Below, I will estimate an AR(1) coefficient around 0.9 in annual data.

Figure 37. To see more precisely how the results at various horizons are linked, and how they result from the persistence of the d/p ratio, suppose that we forecast returns with a forecasting variable x, according to rt+1 xt+1

= axt + εt+1 = ρxt + δ t+1 .

(20.298) (20.299)

(0bviously, you demean the variables or put constants in the regressions.) Small values of b and R2 in (20.298) and a large coefficient ρ in (20.299) imply mathematically that the 358

S ECTION 20.1

T IME - SERIES PREDICTABILITY

long-horizon regression has a large regression coefficient b and large R2 . To see this, write rt+1 + rt+2 rt+1 + rt+2 + rt+3

= a(1 + ρ)xt + aδ t+1 + εt+1 + εt+2 = a(1 + ρ + ρ2 )xt + aρδ t+1 + aδ t+2 + εt+1 + εt+2 + εt+3 .

You can see that with ρ near one, the coefficients increase with horizon, almost linearly at first and then at a declining rate. The R2 are a little messier to work out, but also rise with horizon. The numerator in the long-horizon regression coefficient is E [(rt+1 + rt+2 + ... + rt+k ) xt ]

(300)

where the symbols represent deviations from their means. With stationary r and x, E(rt+j xt ) = E(rt+1 xt−j ), so this is the same moment as E [rt+1 (xt + xt−1 + xt−2 + ...)] ,

(301)

the numerator of a regression coefficient of one year returns on many lags of price dividend ratios. Of course, if you run a multiple regression of returns on lags of p/d, you quickly find that most lags past the first do not help the forecast power. (That statement would be exact in the AR(1) example.) This observation shows once again that one-year and multi-year forecastability are two sides of the same coin. It also suggests that on a purely statistical basis, there will not be a huge difference one-year return forecasts and multi-year return forecasts (correcting the latter for the serial correlation of the error term due to overlap). Hodrick (1991) comes to this conclusion in a careful Monte Carlo experiment, comparing moments of the form (20.300), (20.301) and E(rt+1 xt ). The multi-year regressions, or the implied multi year regressions from one-year forecasts with a slow moving right hand variable are thus mostly useful for illustrating the dramatic economic implications of forecastability, rather than as clever statistical tools that enhance power and allow us to distinguish previously foggy hypotheses. The slow movement of the price-dividend ratio means that on a purely statistical basis, return forecastability is a very open question. What we really know (see Figure 37) is that low prices relative to dividends and earnings in the 50’s preceded the boom market of the early 60’s; that the high price-dividend ratios of the mid-60’s preceded the poor returns of the 70’s; that the low price ratios of the mid-70’s preceded the current boom. We really have three data postwar data points; a once per generation change in expected returns. In addition, the last half of the 1990s has seen a historically unprecedented rise in stock prices and price/dividend ratios (or any other ratio). This rise has cut the postwar return forecasting regression coefficient in half. On the other hand, another crash or even just a decade of poor returns will restore the regression. Data back to the 1600s show the same pattern, but we are often uncomfortable making inferences from centuries-old data. 359

C HAPTER 20 20.1.2

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

Volatility

Price dividend ratios can only move at all if they forecast future returns, if they forecast future dividend growth, or if there is a bubble – if the price-dividend ratio is nonstationary and is expected to grow explosively. In the data, most variation in price-dividend ratios results from varying expected returns. “Excess volatility” – relative to constant discount rate present value models – is thus exactly the same phenomenon as forecastable long-horizon returns. I also derive the very useful price-dividend and return linearizations. Ignoring constants (means), pt − dt rt − Et−1 rt rt+1

= Et

∞ X j=1

ρj−1 (∆dt+j − rt+j )

∞ ∞ X X = (Et − Et−1 ) ρj ∆dt+j − ρj rt+j j=0

j=1

= ∆dt+1 − ρ(dt+1 − pt+1 ) + (dt − pt ).

The volatility test literature starting with Shiller (1981) and LeRoy and Porter (1981) (See Cochrane 1991 for a review) started out trying to make a completely different point. Predictability seems like a sideshow. The stunning fact about the stock market is its extraordinary volatility. On a typical day, the value of the U.S. capital stock changes by a full percentage point, and days of 2 or 3 percentage point changes are not uncommon. In a typical year it changes by 16 percentage points, and 30 percentage point changes are not uncommon. Worse, most of that volatility seems not to be accompanied by any important news about future returns and discount rates. 30% of the capital stock of the United States vanished in a year and nobody noticed? Surely, this observation shows directly that markets are “not efficient” – that prices do not correspond to the value of capital – without worrying about predictability? It turns out however, that “excess volatility” is exactly the same thing as return predictability. Any story you tell about prices that are “too high” or “too low” necessarily imply that subsequent returns will be too low or too high as prices rebound to their correct levels. When prices are high relative to dividends (or earnings, cashflow, book value or some other divisor), one of three things must be true: 1) Investors expect dividends to rise in the future. 2) Investors expect returns to be low in the future. Future cashflows are discounted at a lower than usual rate, leading to higher prices. 3) Investors expect prices to rise forever, giving an adequate return even if there are no dividends. This statement is not a theory, it is an identity: If the price-dividend ratio is high, either dividends must rise, prices must decline, or the price-dividend ratio must grow explosively The open question is, which option holds for 360

S ECTION 20.1

T IME - SERIES PREDICTABILITY

our stock market? Are prices high now because investors expect future earnings, dividends etc. to rise, because they expect low returns in the future, or because they expect prices to go on rising forever? Historically, we find that virtually all variation in price-dividend ratios has reflected varying expected excess returns. Exact present value identity

To document this statement, we need to relate current prices to future dividends and returns. Start with the identity −1 1 = Rt+1 Rt+1 = R−1 t+1

Pt+1 + Dt+1 Pt

(302)

and hence µ ¶ Pt Pt+1 Dt+1 −1 = Rt+1 1 + . Dt Dt+1 Dt

We can iterate this identity forward and take conditional expectations to obtain the identity Ã j ! ∞ X Y Pt −1 = Et Rt+k ∆Dt+k (303) Dt j=1 k=1

where ∆Dt ≡ Dt /Dt−1 . (We could iterate (20.302) forward to Ã j ! ∞ X Y −1 Pt = Rt+k Dt+j , j=1

k=1

but prices are not stationary, so we can’t find the variance of prices from a time-series average. Much of the early volatility test literature concerned stationarity problems. Equation (20.303) also requires a limiting ³Q condition´ that the price dividend ratio cannot explode faster than j −1 returns, limj→∞ Et k=1 Rt+k Pt+j /Dt+j . I come back to this condition below) Equation (??) shows that high prices must, mechanically, come from high future dividend growth or low future returns.

Approximate identity

The nonlinearity of (20.303) makes it hard to handle, and means that we cannot use simple time-series tools. You can linearize (20.303) directly with a Taylor expansion ( Cochrane 1991 takes this approach.) Campbell and Shiller (1988) approximate the one period return identity before iterating, which is algebraically simpler and is the most popular linearization. Start again from the obvious, −1 1 = R−1 t+1 Rt+1 = Rt+1

361

Pt+1 + Dt+1 . Pt

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

Multiplying both sides by Pt /Dt and massaging the result, µ ¶ Pt Pt+1 Dt+1 −1 = Rt+1 . 1+ Dt Dt+1 Dt Taking logs, and with lowercase letters denoting logs of uppercase letters, ¢ ¡ pt − dt = −rt+1 + ∆dt+1 + ln 1 + ept+1 −dt+1

Taking a Taylor expansion of the last term about a point P/D = ep−d µ ¶ P P pt − dt = −rt+1 + ∆dt+1 + ln 1 + + D P [pt+1 − dt+1 − (p − d)] D 1+ D (20.304) pt − dt = −rt+1 + ∆dt+1 + k + ρ (pt+1 − dt+1 ) . Since the average dividend yield is about 4% and average price/dividend ratio is about 25, ρ is a number very near one. I will use ρ = 0.96 for calculations, ρ=

1 P/D = ≈ 1 − D/P = 0.96. 1 + P/D 1 + D/P

Without the constant k, the equation can also apply to deviations from means or any other point. Now, iterating forward is easy, and results in the approximate identity pt − dt = const. +

∞ X j=1

ρj−1 (∆dt+j − rt+j ).

(305)

(Again, we need a condition that pt − dt does not explode faster than ρ−t , limj→∞ ρj (pt+j − dt+j ) = 0. I return to this condition below.) Since (20.305) holds ex-post, we can take conditional expectations and relate price-dividend ratios to ex-ante dividend growth and return forecasts pt − dt = const. + Et

∞ X j=1

ρj−1 (∆dt+j − rt+j ).

(306)

Now it is really easy to see that a high price-dividend ratio must be followed by high dividend growth ∆d, or low returns r. Which is it? Decomposing the variance of price-dividend ratios

To address this issue, equation (20.305) implies ∞ ∞ X X var(pt − dt ) = cov pt − dt , ρj−1 ∆dt+j − cov pt − dt , ρj−1 rt+j (307) j=1

j=1

362

S ECTION 20.1

T IME - SERIES PREDICTABILITY

In words, price-dividend ratios can only vary if they forecast changing dividend growth or of they forecast changing returns. (To derive 20.307 from (20.305), multiply both sides by (pt − dt ) − E(pt − dt ) and take expectations.) Notice that both terms on the right hand side of (20.307) are the numerators of exponentially weighted long-run regression coefficients. This is a powerful equation. At first glance, it would seem a reasonable approximation that returns are unforecastable (the “random walk” hypothesis) and that dividend growth is not forecastable either. But if this were the case, the price/dividend ratio would have to be a constant. Thus the fact that the price/dividend ratio varies at all means that either dividend growth or returns must be forecastable – that the world is not i.i.d. At a simple level, Table 1 includes regressions of long-horizon dividend growth on dividend/price ratios to match the return regressions. The coefficients in the dividend growth case are much smaller, typically one standard error from zero, and the R2 are tiny. Worse, the signs are wrong. To the extent that a high price-dividend ratio forecasts any change in dividends, it seems to forecast a small decline in dividends! Having seen equation (20.307), one is hungry for estimates. Table 2 presents some, taken from Cochrane (1991b). As one might suspect from Table 1, Table 2 shows that in the past almost all variation in price-dividend ratios is due to changing return forecasts. The elements do not have to be between 0 and 100%. For example, -34, 138 occurs because high prices seem to forecast lower real dividend growth (though this number is not statistically significant). Therefore they must and do forecast really low returns, and returns must account for more than 100% of price-dividend variation.

Real std. error Nominal std. error

Dividends -34 10 30 41

Returns 138 32 85 19

Table 2. Variance decomposition of value-weighted NYSE price-dividend ratio. Table entries are the percent of the variance of theP price-dividend ratio attributable j−1 ∆dt+j )/var(pt −dt ) to dividend and return forecasts, 100×cov(pt −dt , 15 j=1 ρ and similarly for returns. This observation solidifies one’s belief in price-dividend ratio forecasts of returns. Yes, the statistical evidence that price-dividend ratios forecast returns is weak, and many return forecasting variables have been tried and discarded, so selection bias is a big worry in forecasting regressions. But the price-dividend ratio (or price-earning, market to book, etc.) has a special status since it must forecast something. To believe that the price-dividend ratio is stationary and varies, but does not forecast returns, you have to believe that the price-dividend ratio does forecast dividends. Given this choice and Table 1, it seems a much firmer conclusion that it forecasts returns. 363

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

It is nonetheless an uncomfortable fact that almost all variation in price-dividend ratios is due to variation in expected excess returns. How nice it would be if high prices reflected expectations of higher future cashflows. Alas, that seems not to be the case. If not, it would be nice if high prices reflected lower interest rates. Again, that seems not to be the case. High prices reflect low risk premia, lower expected excess returns. Campbell’s return decomposition.

Campbell (1991) provides a similar decomposition for unexpected returns, ∞ ∞ X X rt − Et−1 rt = (Et − Et−1 ) ρj ∆dt+j − ρj rt+j . j=0

(308)

j=1

A positive shock to returns must come from a positive shock to forecast dividend growth, or to a negative shock to forecast returns. Since a positive shock to time t dividends is directly paid as a return, (the first sum starts at j = 0), Campbell finds some fraction of return variation is due to current dividends. However, once again, the bulk of index return variation comes from shocks to future returns, i.e. discount rates. To derive (20.308), start with the approximate identity (20.305), and move it back one period pt−1 − dt−1 = const. +

∞ X j=0

ρj (∆dt+j − rt+j ).

Now take innovations of both sides, 0 = (Et − Et−1 )

∞ X j=0

ρj (∆dt+j − rt+j ).

Pulling rt over to the left hand side, you obtain (20.308). (Problem 3 at the end of the chapter guides you through an alternative and more constructive derivation.) Cross-section

So far, we have concentrated on the index. One can apply the same analysis to firms. What causes the variation in price-dividend ratios, or, better book/market ratios (since dividends can be zero) across firms, or over time for a given firm? Vuolteenaho (2000) applies the same sort of analysis to individual stock data. He finds that as much as half of the variation in individual firm book/market ratios reflect expectations of future cashflows. Much of the expected cashflow variation is idiosyncratic, while the expected return variation is common, which is why variation in the index book/market ratio, like variation in the index dividend/price ratio, is almost all due to varying expected excess returns. Bubbles

364

S ECTION 20.1

T IME - SERIES PREDICTABILITY

In deriving the exact and linearized present value identities, I assumed an extra condition that the price-dividend ratio does not explode. Without that condition, and taking expectations of both sides, the exact identity reads ∞ X Pt = Et Dt j=1

Ã

j Y

k=1

−1 Rt+k ∆Dt+k

!

+ lim Et j→∞

Ã

j Y

k=1

R−1 t+k

!

Pt+j Dt+j

(309)

and the linearized identity reads pt − dt = const. + Et

∞ X j=1

ρj−1 (∆dt+j − rt+j ) + Et lim ρj (pt+j − dt+j ). j→∞

(310)

As you can see, the limits in the right hand sides of (20.309) and (20.310) are zero if the price-dividend ratio is stationary, or even bounded. For these terms not to be zero, the price dividend ratio must be expected to grow explosively, and faster than R or ρ−1 . Especially in the linearized form 20.310 you can see that stationary r, ∆d and implies stationary p − d if the last term is zero, and p − d is not stationary if the last term is not zero. Thus, you might want to rule out these terms just based on the view that price dividend ratios do not and are not expected to explode in this way. You can also invoke economic theory to rule them out. The last terms must be zero in an equilibrium of infinitely lived agents or altruistically linked generations. If wealth explodes, optimizing long-lived agents will consume more. Technically, this limiting condition is a first order condition for optimality just like the period to period first order condition. The presence of the last term also presents an arbitrage opportunity in complete markets, as you can short a security whose price contains the last term, buy the dividends separately and eat the difference right away. On the other hand, there are economic theories that permit the limiting terms – overlapping generations models, and they capture the interesting possibility of “rational bubbles” that many observers think they see in markets, and that have sparked a huge literature and a lot of controversy. An investor holds a security with a rational bubble not for any dividends, but on the expectation that someone else will pay even more for that security in the future. This does seem to capture the psychology of investors from the tulip bubble of 17th century Holland to the dot-com bubble of the millenial United States – why else would anyone buy Cisco systems at a price-earnings ratio of 217 and market capitalization 10 times that of General Motors in early 2000? A “rational bubble” imposes a little discipline on this centuries old psychological description, however, by insisting that the person who is expected to buy the security in the future also makes the same calculation. He must expect the price to rise even further. Continuing recursively, the price of a rational bubble must be expected to rise forever. A Ponzi scheme, in which everyone knows the game will end at some time, cannot rationally get off the ground. The expectation that prices will grow at more than a required rate of return forever does not 365

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

mean that sample paths do so. For example, consider the bubble process ( ) Pt R−1 γRPt prob = γP t R−1 Pt+1 = . R 1 prob = γPPttR−1 Figure 38 plots a realization of this process with γ = 1.2. This process yields an expected return R, and the dashed line graphs this expectation as of the first date. Its price is positive though it never pays dividends. It repeatedly grows with a high return γR for a while and then bursts back to one. The expected price always grows, though almost all sample paths do not do so.

Figure 38. Sample path from a simple bubble process. The solid line gives the bubble. The dashed line gives the expected value of the bubble as of time zero, i.e. pRt 0 . Infinity is a long time. It’s really hard to believe that prices will rise forever. The solar system will end at some point; any look at the geological and evolutionary history of the earth suggests that our species will be around a lot less than that. Thus, the infinity in the bubble must really be a parable for “a really long time.” But then the “rational” part of the bubble pops – it must hinge on the expectation that someone will be around to hold the bag; to buy a security without the expectation of dividends or further price increases. (The forever part of usual present value formulas is not similarly worrying because 99.99% of the value comes from the first few hundred years of dividends.) 366

S ECTION 20.1

T IME - SERIES PREDICTABILITY

Empirically, bubbles do not appear to be the reason for historical price-dividend ratio variation. First, price-dividend ratios do seem stationary. (Craine 1993 runs a unit root test with this conclusion.) Even if statistical tests are not decisive, as is expected for a slow moving series, or a series such as that plotted in Figure 38, it is hard to believe that pricedividend ratios can explode rather than revert back to their four-century average level of about 20 to 25. Second, Table 2 shows that return and dividend forecastability terms add up to 100% of the variance of price-dividend ratios. In a bubble, we would expect price variation not matched by any variation in expected returns or dividends, as is the case in Figure 38. I close with a warning: The word “bubble” is widely used to mean very different things. Some people seem to mean any large movement in prices. Others mean large movements in prices that do correspond to low or perhaps negative expected excess returns (I think this is what Shiller 2000 has in mind), i.e. any price movement not explained by a present value model with constant expected returns. 20.1.3

A simple model for digesting predictability

To unite the various predictability and return observations, I construct a simple VAR representation for returns, price growth, dividend growth, dividend price ratio. I start only with a slow moving expected return and unforecastable dividends. This specification implies that d/p ratios reveal expected returns. This specification implies return forecastability. To believe in a lower predictability of returns, you must either believe that dividend growth really is predictable, or that the d/p ratio is really much more persistent than it appears to be. This specification shows that small but persistent changes in expected returns add up to large price changes. We have isolated two important features of the long-horizon forecast phenomenon: dividend/price ratios are highly persistent, and dividend growth is essentially unforecastable. Starting with these two facts, a simple VAR representation can tie together many the predictability and volatility phenomena. Start by specifying a slow-moving state variable xt that drives expected returns, and unforecastable dividend growth, xt rt+1 ∆dt+1

= bxt−1 + δ t = xt + εrt+1 = εdt+1

(20.311) (20.312) (20.313)

All variables are demeaned logs. (The term structure models of Chapter 19 were of this 367

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

form.) From this specification, using the linearized present value identity and return, we can derive a VAR representation for prices, returns, dividends, and the dividend price ratio, (dt+1 − pt+1 ) = b(dt − pt ) + rt+1 ∆pt+1 ∆dt+1

δ t+1 1 − ρb

(20.314)

µ ¶ ρ δ t+1 = (1 − ρb) (dt − pt ) + εdt+1 − 1 − ρb µ ¶ 1 δ t+1 = (1 − b) (dt − pt ) + εdt+1 − 1 − ρb = εdt+1

(20.315) (20.316) (20.317)

Dividend-price ratio: Using the approximate present value identity, we can find the dividend price ratio dt − pt = Et

∞ X j=1

ρj−1 (Et rt+1 − Et dt+j ) =

xt . 1 − ρb

(318)

This equation makes precise my comments that the dividend price ratio reveals expected returns. Obviously, the feature that the dividend price ratio is exactly proportional to the expected return does not generalize. If dividend growth is also forecastable, then the dividendprice ratio is a combination of dividend growth and return forecasts. Actual return forecasting exercises can often benefit from cleaning up the dividend price ratio to focus on the implied return forecast. Returns: Since we know where the dividend/price ratio and dividends are going, we can figure out where returns are going. Use the return linearization (this is equivalent to (20.304)) Rt+1 rt+1

µ ¶ Pt+1 Dt+1 Pt = 1+ / Dt+1 Dt Dt = ρ(pt+1 − dt+1 ) + (dt+1 − dt ) − (pt − dt ).

(20.319)

Now, plug in the from (20.314) and (20.313) to get (20.315). Prices: Write pt+1 − pt = −(dt+1 − pt+1 ) + (dt − pt ) + (dt+1 − dt ).

(320)

Then, plugging in from (20.314) and (20.313), we get (20.316). We can back out parameters from the reduced form return - d/p VAR. (Any two equations carry all the information of this system.) Table RR presents some estimates. 368

S ECTION 20.1 Sample 27-98 48-98 27-92 48-92

a 0.16 0.14 0.28 0.27

T IME - SERIES PREDICTABILITY

a, D/P 4.7 4.0 6.7 6.2

b 0.92 0.97 0.82 0.87

σ(εr ) 19.2 15.0 19.0 14.5

σ(εdp ) 15.2 12.6 15.0 12.4

ρ(εr , εdp ) -0.72 -0.71 -0.69 -0.67

Table RR. Estimates of log excess return and log dividend-price ratio regressions, using annual CRSP data. r is the difference between the log value weighted return and the log treasury bill rate. The estimates are of the system rt+1 dt+1 − pt+1

= a(dt − pt ) + εrt+1 = b(dt − pt ) + εdp,t+1

and rt+1 = (a, D/P )

Dt + εt+1 Pt

I report both the more intuitive coefficients on the actual d/p ratio and the coefficients on the log d/p ratio, which is a more useful specification for our transformations. The two line up; a coefficient of 5 on Dt /Pt implies a coefficient of 5×D/P ≈ 0.25 on (Dt /Pt ) /(D/P ).

You can see that the parameters depend substantially on the sample. In particular, the dramatic returns of the late 1990s, despite low dividend yields, cut the postwar return forecast coefficients in half and the overall sample estimate by about one third. That dramatic decline in the d/p ratio also induces a very high apparent persistence in the d/p ratio, rising to a 0.97 estimate in the 48-98 sample. (Faced with an apparent trend in the data, an autoregression estimates a root near unity.) With these estimates in mind, given the considerations outlined below, I will make calculations using reduced form parameters b ρ σ(εr ) σ(εdp ) ρr,dp

= = = = =

0.9 0.96 15 12.5 −0.7

(20.321)

From these parameters, we can find the underlying parameters of (20.311)-(20.313). I com369

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

ment on each one below as it becomes useful. σ(δ) = σ(εdp )(1 − ρb) = 1.7 (20.322) q σ(εd ) = σ(εr + ρεdp ) = σ2 (εr ) + ρ2 σ 2 (εdp ) + 2ρσ(εr , εdp ) = 10.82

σ(εd , εdp ) = σ(εr εdp ) + ρσ 2 (εdp ) ρ(εr εdp )σ(εr ) + ρσ(εdp ) ρ(εd , εdp ) = = 0.139 σ(εd )

(20.323)

The size of the return forecasting coefficient.

Does the magnitude of the estimated predictability make sense? Given the statistical uncertainties, do other facts guide us to higher or lower predictability? The coefficient of the one year excess return on the dividend price ratio in Table 1 is about 5, and the estimates in Table RR vary from 4 to 6 depending on the sample. These values are surprisingly large. For example, a naive investor might think that dividend yields move onefor-one with returns; if they pay more dividends, you get more money. Before predictability, we would have explained that high dividend yield means that prices are low in anticipation of lower future dividends, leaving the expected return unchanged. Now we recognize the possibility of time-varying expected returns, but does it make sense that expected returns move even more than dividend yields? Return forecastability follows from the fact that dividends are not forecastable, and that the dividend/price ratio is highly but not completely persistent. We see this in the calculated coefficients of prices and returns on the dividend price ratio in (20.315) and (20.316). We derived rt+1 ∆pt+1

= (1 − ρb) (dt − pt ) + εrt+1 = (1 − b) (dt − pt ) + εpt+1

Since dividends are not forecastable, it is no surprise that the formulas for price growth and return are so similar. The return formula basically just adjusts for the fact that a higher dividend yield directly contributes to return by paying more dividends. To transform units to regressions on D/P, multiply by 25, e.g. rt+1 =

1 − ρb Dt + εrt+1 . D/P Pt

Suppose the d/p ratio were not persistent at all–b = 0. Then both return and price growth coefficients should be 1 in logs or about 25 in levels! If the d/p ratio is one percentage point above its average, we must forecast enough of a rise in prices to restore the d/p ratio to its average in one year. The average d/p ratio is about 4%, though, so prices and hence returns must rise by 25% to change the d/p ratio by one percentage point. d(D/P ) = −D/P d(P )/P .

Suppose instead that the d/p ratio were completely persistent i.e. a random walk with 370

S ECTION 20.1

T IME - SERIES PREDICTABILITY

b = 1. Then the return coefficient is 1 − ρ = 0.04, and about 1.0 in levels, while the price coefficient is 0. If the d/p ratio is one percent above average and expected to stay there, and dividends are not forecastable, then prices must not be forecast to change either. The return is one percentage point higher, because you get the higher dividends as well. Thus, the naive investor who expects dividend yield to move one for one with returns not only implicitly assumes that dividends are not forecastable – which turns out to be true – but also that the d/p ratio will stay put forever.

A persistence parameter b = 0.90 implies price and return regression coefficients of

1 − b = 0.10 1 − ρb = 1 − 0.96 × 0.90 = 0.14

or about 2.5 and 3.4 in levels. If the dividend yield is one percentage point high, and is expected to be 0.9 percentage points high in one year, then prices must increase by P/D × 0.1 percentage points in the next year. The return gets the additional dividend. This, fundamentally, is how zero forecastability of dividends implies that returns move more than one for one with the dividend yield. This is a little below the sample estimates in Table 1 and Table RR of 4-6. That is because in the sample, a high price seems to forecast even lower dividend growth – the wrong sign, which is hard to believe. To continue with a calibration that consistently captures the facts with no dividend forecastability, we either have to lower the persistence coefficient or lower the return forecasting coefficient from the values reported in Table RR. A persistence b = 0.8 implies a return coefficient (1− ρb) = (1− 0.96× 0.8) = 0.23 or in levels 0.23 ×25 = 5.75. However, given the uncertainties of dividend/price forecastability, it seems more sensible to continue calculations with b = 0.9 and corresponding return coefficient of 0.14,equal to the estimate in the 48-98 sample. Going in the other direction, statistical uncertainty, the recent runup in stocks despite low dividend yields, and the dramatic portfolio implications of time-varying returns for investors whose risks or risk aversion do not change over time all lead one to consider lower predictability. As we see from these calculations though, there are only two ways to make sense of lower predictability. You could follow the “new economy” advocates, and believe that this time, prices really are rising on advance news of dividend growth, even though prices have not forecast dividend growth in the past. If not, you have to believe that dividend price ratios are substantially more persistent than they have seemed in the postwar data. Much more persistent d/p is a tough road to follow, since D/P ratios already move incredibly slowly. Now, they basically change sign once a generation; high in the 50’s, low in the 60’s, high in the mid-70’s, and decreasing ever since (see Figure 37.) As a quantitative example, suppose the D/P ratio had an AR(1) coefficient of 0.96 in annual data. This means a half life of ln 0.5/ ln 0.96 = 17 years. In this case, the price coefficient would be coefficient 371

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

would be 1 − 0.96 1−b = =1 D/P 0.04

and the return coefficient would be 1 − 0.962 1 − ρb = ≈2 D/P 0.04

A one percentage point higher d/p ratio means that prices must rise 1 percentage point next year, so returns must be about 2 percentage points higher. A two for one movement of expected returns with the dividend yield thus seems about the lower bound for return predictability, so long as dividend growth remains unforecastable. Persistence, price volatility and expected returns

From the dividend-price ratio equation (20.314) we can find the volatility of the dividend price ratio and related it to the volatility and persistence of expected returns. σ(dt − pt ) =

1 σ (xt ) 1 − ρb

With b = 0.9, 1/(1 − ρb) = 1/(1 − 0.96 × 0.9) = 7. 4. Thus, the high persistence of expected returns means that a small expected return variation translates into a potentially very large price variation; or equivalently that very large price variations, unaccounted for by forecasts of dividend variation, can be explained by small variation in expected returns. Translating to levels, a one percentage point change in expected returns with persistence b = 0.9 corresponds to a 7.4% increase in price. The Gordon growth model is a classic and even simpler way to see this point. With constant dividend growth g and return r, the present value identity becomes P =

D . r−g

A price-dividend ratio of 25 means r − g = 0.04. Then, a one percentage point permanent change in expected return translates into a 25 percentage point change in price! This is an overstatement, since expected returns are not this persistent, but it allows you to clearly see the point. This point also shows that small market imperfections in expected returns can translate into substantial market imperfections in prices, if those expected return changes are persistent. We know markets cannot be perfectly efficient (Grossman and Stiglitz 1980). If they were perfectly efficient, there would be no traders around to make them efficient. Especially in situations where short sales or arbitrage are constrained by market frictions, prices of similar assets can be substantially different, while the expected returns of those assets are almost the same. For example the “closed end fund” puzzle (Thompson 1978) noted that baskets of securities sold for substantial price discounts relative to the sum of the individual securi372

S ECTION 20.1

T IME - SERIES PREDICTABILITY

ties. However, these price differentials persist for a long time. You can’t short the closed end funds to buy the securities and keep that short position on for years. 20.1.4

Mean-reversion

I introduce long-horizon return regressions and variance ratios. I show that they are related: each one picks up a string of small negative return autocorrelations. I show though that the direct evidence for mean reversion and Sharpe ratios that rise with horizon is weak. Long run regressions and variance ratios

The first evidence of long-run forecastability in the stock market did not come from d/p regressions, but rather from clever ways of looking at the long-run univariate properties of returns. Fama and French (1988a) ran regressions of long-horizon returns on past longhorizon returns, rt→t+k = a + bk rt−k→t + εt+k ,

(324)

basically updating classic autocorrelation tests from the 60s to long horizon data. They found negative and significant b coefficients: a string of good past returns forecasts bad future returns. Poterba and Summers (1988) considered a related “variance ratio” statistic. If stock returns are i.i.d., then the variance of long horizon returns should grow with the horizon var(rt→t+k ) = var(rt+1 + rt+2 + .. + rt+k ) = kvar(rt+1 ).

(325)

They computed the variance ratio statistic vk =

1 var(rt→t+k ) . k var(rt+1 )

They found variance ratios below one. Stocks, it would seem, really are safer for “long-run investors” who can “afford to wait out the ups and downs of the market,” common Wall Street advice, long maligned by academics. These two statistics are closely related, and reveal the same basic fact: stock returns have a string of small negative autocorrelations. To see this relation, write the variance ratio statistic

vk =

1 var k

³P k

j=1 rt+j

var(rt+1 )

´

=

k k X X |k − j| |k − j| ρj = 1 + 2 ρj , k k j=1

j=−k

373

(326)

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

and the regression coefficient in (20.324) k k X X 1 cov bk = rt+j , rt−j+1 var(rt→t+k ) j=1 j=1 =

k k 1 X |k − j| k var(rt+1 ) X |k − j| ρk+j = ρk+j . var(rt→t+k ) k vk k j=−k

j=−k

Both statistics are based on tent-shaped sums of autocorrelations, as illustrated by Figure 39. If there are many small negative autocorrelations which bring returns back slowly after a shock, these autocorrelations might be individually insignificant. Their sum might be economically and statistically significant, however, and these two statistics will reveal that fact by focusing on the sum of autocorrelations. The long-horizon regression weights emphasize the middle of the autocorrelation function, so a k year horizon long-horizon regression is comparable to a somewhat longer variance ratio.

Variance ratio weights Long horizon regression weights

Return autocorrelations Figure 39. Long horizon regression and variance ratio weights on autocorrelations. Moving average representation and mean reversion

The “mean-reversion” description of these statistics comes from their implications for where values go at long horizons following a shock. We can show that the square root of the variance ratio measures the long-horizon impact of a shock relative to its instantaneous impact – the extent to which values revert back towards their mean following a shock. You can always write returns as a moving average of their own shocks. From a regression of returns on past returns a(L)rt = εt

374

(327)

S ECTION 20.1

T IME - SERIES PREDICTABILITY

you can find the θj in rt =

∞ X

θj εt−j = θ(L)εt = a(L)−1 εt .

j=0

(Most simply, just simulate (20.327) forward.) The θj are the moving average representation or impulse-response function – they tell you the path of expected returns following a shock. Let vt represent the cumulative returns, or the log value of a dollar invested, ∆vt = rt . P Then, the partial sum kj=1 θj tells you the effect on invested wealth vt+k of a univariate return shock εt

Relating variance ratios, long-horizon regressions and moving averages for finite k is possible but not pretty. However, we can nicely relate the limiting response – where limk→∞ Et vt+k ends up after a shock – to the autocorrelations, and thus to the limit of the variance ratio statistic very simply as

1+2

∞ X j=1

2 ∞ X ρj = θ j /σ2ε .

(328)

j=0

If returns are i.i.d., the variance ratio is one at all horizons; all autocorrelations are zero, and all θ past the first are zero so the long-run price moves one for one with the shock. A longP string of small negative autocorrelations means a variance ratio less than one, and means ∞ j=0 θj < 1 so the long-run effect on price is lower than the impact effect - this is “mean-reversion.” The right hand equality of (20.328) follows by just taking the k → ∞ in (20.326). For the second equality, you can recognize in both expressions the spectral density of r at frequency zero. (Cochrane 1986 discusses these and other properties of variance ratios.)

Numbers

Table A1 presents an estimate of the variance of long-horizon returns and long-horizon return regressions. The long-horizon regressions do show some interesting mean reversion, especially in the 3-5 year range. However, that turns around at year 7 and disappears by year 10. The variance ratios do show some long-horizon stabilization. At year 10, the variance ratio is (16.3/19.8)2 = 0.68, and the long-run price impact of a shock is 16.8/19.8 = 0.85. The mean log return grows linearly with horizon whether returns are autocorrelated or not – E(r1 + r2 ) = 2E(r). If the variance also grows linearly with the horizon, as it does for non-autocorrelated returns, then the Sharpe ratio grows with the square root of horizon. If the variance grows more slowly than horizon, then the Sharpe ratio grows faster than the square root of the horizon. This is the fundamental question for whether stocks are (unconditionally) “safer for the long run.” Table A1 includes the long-horizon Sharpe ratios, and you can see that they do increase. 375

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

logs, 1926-1996. √ σ (rk ) / k βk √ Sharpe/ k

1 19.8 0.08 0.31

2 20.6 -0.15 0.30

3 19.7 -0.22 0.30

5 18.2 -0.04 0.31

7 16.5 0.24 0.36

10 16.3 0.08 0.39

Table A1. Mean reversion using logs, 1926-1996. r denotes the difference between the log value weighted NYSE return and the log treasury bill return. σ(rk ) = σ(rt→t+k ) is the variance of long-horizon returns. β k is the regression coefficient in rt→t+k = α + β k rt−k→t + εt+k . The Sharpe ratio is E(rt→t+k )/σ(rt→t+k ) You would not be to blame if you thought that the evidence of Table A1 was rather weak, especially compared with the dramatic dividend/price regressions. It is, and it is for this reason that most current evidence for predictability focuses on other variables such as the d/p ratio. In addition, Table A2 shows that the change from log returns to levels of returns, while having a small effect on long-horizon regressions, destroys any evidence for higher Sharpe ratios at long horizons. Table A3 shows the same results in the postwar period. Some of the negative long-horizon regression coefficients are negative and significant, but there are just as large positive coefficients, and no clear pattern. The variance ratios are flat or even rising with horizons, and the Sharpe ratios are flat or even declining with horizon. 1926-1996 √ levels σ (rk ) / k βk √ Sharpe/ k

1 20.6 0.02 0.41

2 22.3 -0.21 0.41

3 22.5 -0.22 0.41

5 24.9 -0.03 0.40

7 28.9 0.22 0.40

10 39.5 -0.63 0.38

Table A2. r denotes the difference between the gross (not log) long-horizon valueweighted NYSE return and the gross treasury bill return. 1947-1996 √ logs σ (rk ) / k βk √ Sharpe/ k 1947-1996 √ levels σ (rk ) / k βk √ Sharpe/ k

1 15.6 -0.10 0.44 1 17.1 -0.13 0.50

2 14.9 -0.29* 0.46 2 17.9 -0.33* 0.51

3 13.0 0.30* 0.51 3 16.8 0.30 0.55

5 13.9 0.30 0.46 5 21.9 0.25 0.48

7 15.0 0.17 0.41 7 29.3 0.13 0.41

10 15.6 -0.18 0.36 10 39.8 -0.25 0.37

Table A3. Mean-reversion in postwar data. In sum, the direct evidence for mean-reversion in index returns seems quite weak. I consider next whether indirect evidence, values of these statistics implied by other estimation 376

S ECTION 20.1

T IME - SERIES PREDICTABILITY

techniques, still indicate mean-reversion. (The mean-reversion of individual stock returns as examined by Fama and French (1988a) is somewhat stronger, and results in the stronger cross-sectional “reversal” effect described in section 2.5 below.) Keep in mind also that the unconditional Sharpe ratio does not in the end, drive investment decisions. Investment decisions are driven by the conditional moments of asset returns at any moment in time, using every information variable that there is. 20.1.5

Mean-reversion and forecastability20.335

I reconcile large forecastability from d/p ratios with a small mean reversion. I calculate the univariate return process implied by the simple VAR, and find that it displays little mean reversion. I show that if dividend shocks are uncorrelated with expected return shocks, there must be some mean reversion. If one rules out the small positive correlation in our samples, one gets a slightly higher estimate of univariate mean-reversion. I tie the strong negative correlation between return and d/p shocks to an essentially zero correlation between expected return and dividend growth shocks. How is it possible that variables such as the dividend price ratio forecast returns strongly, but there seems to be little evidence for mean reversion in stock returns? To answer this question, we have to connect the d/p regressions and the mean-reversion statistics. Forecastability from variables such as the dividend-price ratios is related to, but does not necessarily imply mean-reversion. (Campbell 1991 emphasizes this point.) Mean-reversion is about the univariate properties of the return series, forecasts of rt+j based on {rt , rt−1 , rt−2 ...}. Predictability is about the multivariate properties, forecasts of rt+j based on {xt , xt−1 , xt−2 , ...} as well as {rt , rt−1 , rt−2 ...}. Variables xt can forecast rt+1 , while {rt−j } fail to forecast rt+1 . As a simple example, suppose that returns are i.i.d., but you get to see tomorrow’s newspaper. You forecast returns with a variable xt = rt+1 , rt+1 xt+1

= xt = δ t+1 .

In this example, xt forecasts returns very well, but lagged returns do not forecast returns at all. To examine this issue, continue with the VAR representation built up from a slowly moving expected return and unforecastable dividends, (20.311)-(20.317). We want to find the univariate return process implied by this VAR: what would happen if you took infinite data from the system and ran a regression of returns on lagged returns? The answer, derived below, 377

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

is of the form rt =

1 − γL νt. 1 − bL

(329)

This is just the kind of process that can display slow mean-reversion or momentum. The moving average coefficients are rt = ν t − (γ − b)ν t−1 − b(γ − b)ν t−2 − b2 (γ − b)ν t−3 − b3 (γ − b)ν t−4 − ...

(330)

Thus, if γ > b, a positive return shock sets off a long string of small negative returns, which cumulatively bring the value back towards where it started. If γ < b, a positive shock sets off a string of small positive returns, which add “momentum” to the original increase in value. The long-run statistics are 2 µ ¶2 ∞ X X 1−γ 1+2 ρj = θj /σ2 (ν t ) = . 1−b j=0

Thus, if γ > b, returns will have a variance ratio below one, and if γ < b a variance ratio above one.

Now, what value of γ does our VAR predict? Is there a sensible structure of the VAR that generates substantial predictability but little mean-reversion? The general formula, derived below, is that γ solves ¡ ¢ 1 + b2 σ2 (εd ) + (1 + ρ2 )σ2 (εdp ) − 2(ρ + b)σ(εd , εdp ) 1 + γ2 = = 2q, γ bσ2 (εd ) + ρσ 2 (εdp ) − (ρ + b)σ(εd , εdp )

(331)

and hence, γ=q−

Case 1: No predictability.

p q 2 − 1.

If returns are not predictable in this system; if σ(δ) = 0 so σ(εdp ) = 0; then (20.331) specializes to 1 + b2 1 + γ2 = . γ b γ = b, so returns are not autocorrelated. Sensibly enough.

Case 2: Constant dividend growth.

Next, suppose that the case that dividend growth is constant; σ(εd ) = 0 and variation in expected returns is the only reason that returns vary at all. In this case, (20.331) specializes 378

S ECTION 20.1

T IME - SERIES PREDICTABILITY

quickly to 1 + ρ2 1 + γ2 = , γ ρ

and thus γ = ρ. This is a substantial amount of mean reversion. (γ − b) in (20.330) is then 0.96 − 0.90 = 0.06, so that each year j after a shock returns come back by 6 × bj percent of the original shock. The cumulative impact is that value ends up at (1 − γ)/(1 − b) = (1 − 0.96)/(1 − 0.9) = 0.4 or only 40% of the original shock. Case 3: Dividend growth uncorrelated with expected return shocks.

Pure variation in expected returns is of course not realistic. Dividends do vary. If we add dividend growth uncorrelated with expected return shocks – with σ(εdp , εd ) = 0– (20.331) specializes to 1 + b2 1 + ρ2 1 + γ2 bσ2 (εd ) ρσ2 (εdp ) = + = 2q 2 2 2 γ b bσ (εd ) + ρσ (εdp ) ρ bσ (εd ) + ρσ2 (εdp )

(332)

In this case, b < γ < ρ. There will be some mean reversion in returns – this model cannot generate γ ≤ b. However, the mean reversion in returns will be lower than with no dividend growth, because dividend growth obscures the information in ex-post returns about timevarying expected returns. (See (20.333).) How much lower depends on the parameters. Using the parameters (20.321), I find that (20.332) implies p γ = q − q 2 − 1 = 0.928.

Our baseline VAR with no correlation between dividend growth and expected return shocks thus generates a univariate return process that is slightly on the mean-reversion edge of uncorrelated. The long-run response to a shock is 1 − 0.928 1−γ = = 0.72 1−b 1 − 0.9

This is a lot less mean-reversion than 0.4, but still somewhat more mean reversion than we see in Tables A1-A3. This case is an important baseline worth stressing. If expected returns are positively correlated, realized returns are negatively autocorrelated. If (unchanged) expected dividends are discounted at a higher rate, today’s price falls. You can see this most easily by just looking at the return or its linearization, (20.319) rt+1 = ∆dt+1 − ρ(dt+1 − pt+1 ) + (dt − pt ).

(333)

The d − p ratio is proportional to expected returns. A positive shock to expected returns, uncorrelated with dividend growth, lowers actual returns. A little more deeply, look at the 379

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

return innovation identity (20.308),

rt − Et−1 rt = (Et − Et−1 )

∞ X j=0

ρj ∆dt+j −

∞ X j=1

ρj rt+j .

(334)

P j If expected returns (Et − Et−1 ) ∞ j=1 ρ rt+j increase, with no concurrent news about current or future dividends, then rt − Et−1 rt decreases.

This is the point to remark on a curious feature of the return - dividend/price VAR; the negative correlation between ex-post return shocks and dividend/price ratio shocks. All the estimates were around -0.7. At first glance such a strong correlation between VAR residuals seems strange. At second glance, it is expected. From (20.333) you can see that a positive innovation to the dividend price ratio will correspond to a negative return innovation, unless a striking dividend correlation gets in the way. More deeply, you can see the point in (20.334). Quantitatively, from (20.315), the return shock is related to the dividend growth shock and the expected return shock by εr = εd −

ρ δ = εd − ρεdp 1 − ρb

Thus, a zero correlation between the underlying dividend growth and expected return shocks, ρ(εd , δ) = 0 implies a negative covariance between return shocks and expected return shocks. σ(εr , δ) = −

ρ σ2 (δ) 1 − ρb

The correlation is a perfect −1 if there are no dividend growth shocks. At the parameters (??) σ(εdp ) = 12.5, σ(εr ) = 15, we obtain ρ(εr , δ) = ρ(εr , εdp ) = −

σ(εdp ) 12.5 ρ σ(δ) = −ρ = −0.96 × = −0.8. 1 − ρb σ(ε) σ(ε) 15

The slight 0.1 positive correlation between dividend growth and expected return shocks results (or, actually, results from) a slightly lower −0.7 specification for the correlation of return and d/p shocks. The strong negative correlation between return shocks and expected return shocks, expected from a low correlation between dividend growth shocks and expected return shocks, is crucial to the finding that returns are not particularly correlated despite predictability. Consider what would happen if the correlation ρ(εr , εdp ) = ρ(εr , δ) were zero. The expected return xt is slow moving. If it is high now, it has been high for a while, and there has likely been a series of good past returns. But it also will remain high for a while, leading to a period of high future returns. This is “momentum,” positive return autocorrelation, the opposite of mean-reversion. Case 4: Dividend growth shocks positively correlated with expected return shocks

As we have seen, the VAR with no correlation between expected return and dividend 380

S ECTION 20.1

T IME - SERIES PREDICTABILITY

growth shocks cannot deliver uncorrelated returns or positive “momentum” correlation patterns. At best, volatile dividend growth can obscure an underlying negative correlation pattern. However, looking at (20.333) or (20.334), you can see that adding dividend growth shocks positively correlated with expected return shocks could give us uncorrelated or positively correlated returns. The estimate in Table RR implied a slight positive correlation of dividend growth and expected return shocks, ρεd δ = 0.14 in (20.323). If we use that estimate in (20.331), we recover an estimate γ = 0.923;

1−γ = 0.77 1−ρ

This γ is quite close to b = 0.9, and the small mean reversion is more closely consistent with Tables A1-A3. Recall that point estimates as in Table 1 actually showed that a high d/p ratio forecast higher dividends – the wrong sign. This point estimate means that shocks to the d/p ratio and expected returns are positively correlated with shocks to expected dividend growth. If you generalize the VAR to allow such shocks, along with a richer specification allowing additional lags and variables, you find that VARs give point estimates with slight but very small mean reversion. (See Cochrane 1994 for a plot. The estimated univariate process has slight mean-reversion, with an impulse-response ending up at about 0.8 of its starting value, and no different from the direct estimate. Can we generate unforecastable returns in this system? To do so, we have to increase the between expected return shocks and dividend growth. Equating (20.331) to ¡ correlation ¢ 1 + b2 /b and solving for ρ(εd , εdp ), we obtain ρ(εd , εdp ) =

(1 − ρb) (ρ − b) σ(εdp ) = 0.51. (1 − b)2 (ρ + b) σ(εd )

This is possible, but not likely. Any positive correlation between dividend growth and expected return shocks strikes me as suspect. If anything, I would expect that since expected returns rise in “bad times” when risk or risk aversion increases, we should see a positive shock to expected returns associated with a negative shock to current or future dividend growth. Similarly, if we are going to allow dividend price ratios to forecast dividend growth, a high dividend price ratio should forecast lower dividends. Tying together all these thoughts, I think it’s reasonable to impose zero dividend forecastability and zero correlation between dividend growth and expected return shocks. This specification means that returns are really less forecastable than they seem in some samples. As we have seen, b = 0.9 and no dividend forecastability means that the coefficient of return on D/P is really about 3.4 rather than 5 or 6. This specification means that expected returns really account for 100% rather than 130% of the price-dividend variance. However, it also means that univariate mean reversion is slightly stronger than it seems in our sample. 381

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

This section started with the possibility that the implied mean reversion from a multivariate system could be a lot larger than that revealed by direct estimates. Instead, we end up by reconciling strong predictability and slight mean-reversion. How to find the univariate return representation

To find the implied univariate representation, we have to find a representation rt+1 = a(L)ν t

(335)

in which the a(L) is invertible. The Wold decomposition theorem tells us that there is a unique moving invertible moving average representation in which the ν t are the one-step ahead forecast error shocks, i.e. the errors in a regression model a(L)rt+1 = ν t+1 . Thus, if you find any invertible moving average representation, you know you have the right one. We can’t do it by simply manipulating the systems starting with (20.311), because they are expressed in terms of multivariate shocks, errors in regressions that include x. There are three fundamental representations of a time series: its Wold moving average representation, its autocorrelation function, and its spectral density. To find the univariate representation (20.335), you either calculate the autocorrelations E(rt rt−j ) from (20.311) and then try to recognize what process has that autocorrelation pattern, or you calculate the spectral density and try to recognize what process has that spectral density. In our simple setup, we can write the return-d/p VAR (20.314)-(20.315) as

rt+1 = (1 − ρb) (dt − pt ) + (εdt+1 − ρεdpt+1 ) (dt+1 − pt ) = b(dt − pt ) + εdpt+1

Then, write returns as rt+1 (1 − bL) rt+1 (1 − bL) rt+1

(1 − ρb) εdpt + (εdt+1 − ρεdpt+1 ) 1 − bL = (1 − ρb) εdpt + (εdt+1 − ρεdpt+1 ) − b (εdt − ρεdpt ) = (εdt+1 − ρεdpt+1 ) + (εdpt − bεdt ) (20.336) =

Here, you can see that rt must follow an ARMA(1,1) with one root equal to b and the other root to be determined. Write yt = (1 − bL)rt , and thus yt = (1 − γL)ν t . Then the autocovariances of y from (20.336) are ¢ ¡ 2 E(yt+1 ) = 1 + b2 σ2 (εd ) + (1 + ρ2 )σ2 (εdp ) − 2(ρ + b)σ(εd , εdp ) E(yt+1 yt ) = −bσ 2 (εd ) − ρσ 2 (εdp ) − (ρ + b)σ(εd , εdp )

while yt = (1 − γL)ν t implies 2 E(yt+1 ) =

¢ ¡ 1 + γ 2 σ2ν

E(yt+1 yt ) = −γσ2ν .

382

S ECTION 20.1

T IME - SERIES PREDICTABILITY

Hence, we can find γ from the condition ¡ ¢ 1 + b2 σ2 (εd ) + (1 + ρ2 )σ2 (εdp ) − 2(ρ + b)σ(εd , εdp ) 1 + γ2 = = 2q. γ bσ2 (εd ) + ρσ2 (εdp ) − (ρ + b)σ(εd , εdp ) The solution (the root less than one) is γ=q−

p q 2 − 1.

For more general processes, such as computations from an estimated VAR, it is better to approach the problem via the spectral density. This approach allows you to construct the uni¤0 £ variate representation directly without relying on cleverness. If you write yt = rt xt , the VAR is yt = A(L)η t . Then spectral density of returns Sr (z) is given by the top left element of Sy (z) = A(z)E(ηη0 )A(z −1 )0 with z = e−iω . Like the autocorrelation, the spectral density is the same object whether it comes from the univariate or multivariate representation. You can find the autocorrelations by (numerically) inverse-Fourier transforming the spectral density. The autocorrelations and spectral densities are directly revealing: a string of small negative autocorrelations or a dip in the spectral density near frequency zero correspond to mean-reversion; positive autocorrelations or a spectral density higher at frequency zero than elsewhere corresponds to momentum. To find the univariate, invertible moving average representation from the spectral density, you have to factor the spectral density Srr (z) = a(z)a(z) where a(z) is a polynomial with roots outside the unit circle, a(z) = (1 − γ 1 z)(1 − γ 2 z)...γ i < 1. Then, since a(L) is invertible, rt = a(L)εt σ2ε = 1 is the univariate representation of the return process. 20.1.6

Multivariate mean-reversion

I calculate the responses to multivariate rather than univariate shocks. In a multivariate system you can isolate expected return shocks and dividend growth shocks. The price response to expected return shocks is entirely stationary. We are left with a troubling set of facts: high price/dividend ratios strongly forecast low returns, yet high past returns do not seem to forecast low subsequent returns. Surely, there must be some sense in which “high prices” forecast lower subsequent returns? The resolution must involve dividends (or earnings, book value, or a similar divisor for prices). A price rise with no change in dividends results in lower subsequent returns. A price rise that comes with a dividend rise does not result in lower subsequent returns. A high return combines dividend news and price-dividend news, and so obscures the lower expected return message. In a more time-series language, instead of looking at the response to a univariate return shock – a return that was unanticipated based on lagged returns – let us look at the 383

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

responses to multivariate shocks – a return that was unanticipated based on lagged returns and dividends. This is easy to do in our simple VAR. We can simulate (20.314) -(20.317) forward and trace the responses to a dividend growth shock and an expected return (d/p ratio) shock. Figures 40 and 41 present the results of this calculation. (Cochrane 1994 presents a corresponding calculation using an unrestricted VAR, and the results are very similar.)

Figure 40. Responses to a one standard deviation (1.7%) negative expected return shock in the simple VAR. Start with Figure 40. The negative expected return shock raises prices and the p-d ratio immediately. We can identify such a shock in the data as a return shock with no contemporaneous movement in dividends. The p-d ratio then reverts to its mean. Dividends are not forecastable, so they show no immediate or eventual response to the expected return shock. It could be the case that prices move in advance of future dividends; if this were the case we would see dividends rising to meet higher prices after a return shock. Instead, prices show a long and complete reversion back to the level of dividends. This shock looks a lot like a negative yield shock to bonds: such a shock raises prices now so that bonds end up at the same maturity value despite a smaller expected return. The cumulative return “mean-reverts” even more than prices. For given prices, dividends are now smaller (smaller d-p) so returns deviate from their mean by more than price growth. 384

S ECTION 20.1

T IME - SERIES PREDICTABILITY

Figure 41. Responses to a one standard deviation (14%) dividend growth shock in the simple VAR. The cumulative return ends up below its previously expected value. Compare this value response to the univariate value response, which we calculated above ends up at about 0.8 of its time-1 response. The dividend shock raises prices and cumulative returns immediately and proportionally to dividends, so the price-dividend ratio does not change. Expected returns or the discount rate, reflected in any slope of the value line, do not change. If the world were i.i.d., this is the only kind of shock we would see, and dividend-price ratios would always be constant. Figure (40) and (41) plot the responses to “typical,” one standard deviation shocks. Thus you can see that actual returns are typically about half dividend shocks and half expected return shocks. That is why returns alone are a poor indicator of expected returns. In sum, at last we can see some rather dramatic “mean-reversion.” Good past returns by themselves are not a reliable signal of lower subsequent returns, because they contain substantial dividend growth noise. Good returns that do not include good dividends isolate an expected return shock. This does signal low subsequent returns. It sets off a completely transitory variation in prices.

385

C HAPTER 20 20.1.7

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

Cointegration and short vs. long-run volatility

If d − p, ∆p and ∆d then the long-run variance of ∆d and ∆p must be the same, longrun movements in d and p must be perfectly correlated, and d and p must end up in the same place after any shock. Thus, the patterns of predictability, volatility, univariate and multivariate mean-reversion really all just stem from these facts, the persistence of d − p and the near-unforecastability of ∆d. You might think that the facts about predictability depend on the exact structure of the VAR, including parameter estimates. In fact, most of what we have learned about predictability and mean reversion comes down to a few facts: the dividend-price ratio, returns, and dividend growth are all stationary; dividend growth is not (or at best weakly) forecastable, and dividend growth varies less than returns. These facts imply that the dividend and price responses to each shock are eventually equal in Figures (40) and (41). If d − p, ∆p and ∆d are stationary, then d and p must end up in the same place following a shock. The responses of a stationary variable (d − p) must die out. If dividends are not forecastable, then it must be the case that prices do all the adjustment following a price shock that does not affect dividends. Stationary d − p, ∆p and ∆d also implies that the variance of long-horizon ∆p must equal the variance of long-horizon ∆d. 1 1 var (pt+k − pt ) = lim var (dt+k − dt ) , k→∞ k k→∞ k lim

(337)

and the correlation of long-run price and dividend growth must approach one. These facts follow from the fact that the variance ratio of a stationary variable must approach zero, and d − p is stationary. Intuitively, long run price growth cannot be more volatile than long run dividend growth, or the long-run p − d ratio would not be stationary. Now, if dividend growth is not forecastable, its long run volatility is the same as its short run volatility – its variance ratio is one. Short run price growth is more volatile than short run dividend growth, so we conclude that prices must be mean-reverting; their variance ratio must be below one.

Quantitatively, this observation supports the magnitude of univariate mean reversion that we have found so far. Dividend growth has a short run, and thus long-run, standard deviation of about 10% per year, while returns and prices Thus, prices must p have a long-run variance ratio of about 2/3, or a long-run response to univariate shocks of 2/3 = 0.82.

The change in prices is not the same thing as the return, especially at long horizons, since returns include the intervening dividends. One can address this question with a slightly different accounting: define d as the dividend paid to a dollar investment. The resulting dividend series is still not predictable and has roughly the same volatility, so in this case we 386

S ECTION 20.1

T IME - SERIES PREDICTABILITY

get approximately the same result. The work of Lettau and Ludvigson (2000) suggests that we may get much more dramatic implications by including consumption data. The ratio of stock market values to consumption should also be stationary; if wealth were to explode people would surely consume more and vice versa. The ratio of dividends to aggregate consumption should also be stationary. Consumption growth seems independent at all horizons, and consumption growth is very stable, with roughly 1% annual standard deviation. For example, Lettau and Ludvigson 2000 find that none of the variables that forecast returns in Table LL – including d − p and a consumption to wealth ratio – forecast consumption growth at any horizon. These facts suggest that aggregate dividends are forecastable by the consumption/dividend ratio, and strongly so – the long-run volatility of aggregate dividend growth must be the 1% volatility of consumption growth, not the 10% short run volatility of dividend growth. These facts also mean that almost all of the 15% or more variation in annual stock market wealth must be transitory – the long run volatility of stock market value must be no more than the 1% consumption growth volatility! Again, total market value is not the same thing as price, price is not the same thing as cumulated return, and aggregate dividends are not the same thing as the dividend concept we have used so far (dividends paid to a dollar investment with dividends consumed), or dividends paid to a dollar investment with dividends reinvested. Lettau and Ludvigson show that the consumption/wealth ratio does forecast returns, but noone has yet worked out the mean-reversion implications of this fact. My statements about the implications of stationary d − p, ∆d, ∆p, r are developed in detail in Cochrane 1994. They are special cases of the representation theorems for cointegrated variables developed by Engel and Granger (1987). A regression of a difference like ∆p on a ratio like p − d is called the error-correction representation of a cointegrated system. Error correction regressions have subtly and dramatically changed almost all empirical work in finance and macroeconomics. The vast majority of the successful return forecasting regressions in this section, both time-series and cross-section, are error-correction regressions of one sort or another. Corporate finance is being redone with regressions of growth rates on ratios, as is macroeconomic forecasting. For example, the consumption/GDP ratio is a powerful forecaster of GDP growth. 20.1.8

Bonds

The expectations model of the term structure works well on average and for horizons of 4 years or greater. At the one year horizon, however, a forward rate 1 percentage point higher than the spot rate seems entirely to indicate a one percentage point higher expected excess return rather than a one percentage point rise in future interest rates.

387

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

The venerable expectations model of the term structure specifies that long term bond yields are equal to the average of expected future short term bond yields. As with the CAPM and random walk, the expectations model was the workhorse of empirical finance for a generation. And as with those other views, a new round of research has significantly modified the traditional view.

Maturity N 1 2 3 4 5

Avg. Return (N) E(hprt+1 ) 5.83 6.15 6.40 6.40 6.36

Std. error 0.42 0.54 0.69 0.85 0.98

Std. dev. (N) σ(hprt+1 ) 2.83 3.65 4.66 5.71 6.58

Table 4. Average continuously compounded (log) one-year holding period returns on zero-coupon bonds of varying maturity. Annual data from CRSP 1953-1997. Table 4 calculates the average return on bonds of different maturities. The expectations hypothesis seems to do pretty well. Average holding period returns do not seem very different across bond maturities, despite the increasing standard deviation of bond returns as maturity rises. The small increase in returns for long term bonds, equivalent to a slight average upward slope in the yield curve, is usually excused as a small “liquidity premium.” In fact, the curious pattern in Table 4 is that bonds do not share the high Sharpe ratios of stocks. Whatever factors account for the volatility of bond returns, they seem to have very small risk prices. Table 4 is again a tip of an iceberg of an illustrious career for the expectations hypothesis. Especially in times of great inflation and exchange rate instability, the expectations hypothesis does a very good first-order job. However, one can ask a more subtle question. Perhaps there are times when long term bonds can be forecast to do better, and other times when short term bonds are expected to do better. If the times even out, the unconditional averages in Table 4 will show no pattern. Equivalently, we might want to check whether a forward rate that is unusually high forecasts an unusual increase in spot rates. 388

S ECTION 20.1

N 1 2 3 4

T IME - SERIES PREDICTABILITY

Change in yields (1) (1) yt+N − yt = (N→N+1) (1) = a + b(ft − yt ) + εt+N 2 a σ(a) b σ(b) R 0.1 0.3 -0.10 0.36 -0.02 -0.01 0.4 0.37 0.33 0.005 -0.04 0.5 0.41 0.33 0.013 -0.3 0.5 0.77 0.31 0.11

Holding period returns (N+1) (1) hprt+1 − yt = (N→N+1) (1) = a + b(ft − yt ) + εt+1 2 a σ(a) b σ(b) R -0.1 0.3 1.10 0.36 0.16 -0.5 0.5 1.46 0.44 0.19 -0.4 0.8 1.30 0.54 0.10 -0.5 1.0 1.31 0.63 0.07

Table 5. Forecasts based on forward-spot spread. OLS regressions 1953-1997 annual data. Yields and returns in annual percentages. The left hand panel runs the change in the one year yield on the forward-spot spread. The right hand panel runs the one period excess return on the forward-spot spread. Table 5 gets at these issues, updating Fama and Bliss’ (1986) classic regression tests. (Campbell and Shiller 1991 and Campbell 1995 make the same point with regressions of yield changes on yield spreads.) The left hand panel presents a regression of the change in yields on the forward-spot spread. The expectations hypothesis predicts a coefficient of 1.0, since the forward rate should equal the expected future spot rate. At a one-year horizon we see instead coefficients near zero and a negative adjusted R2 . Forward rates one year out seem to have no predictive power whatsoever for changes in the spot rate one year from now. On the other hand, by 4 years out, we see coefficients within one standard error of 1.0. Thus, the expectations hypothesis seems to do poorly at short (1 year) horizons, but much better at longer horizons and on average (Table 4). If the yield expression of the expectations hypothesis does not work at one year horizons, then the expected return expression of the expectations hypothesis must not hold either – one must be able to forecast one year bond returns. To check this fact, the right hand panel of Table 5 runs regressions of the one year excess return on long-term bonds on the forward-spot spread. Here, the expectations hypothesis predicts a coefficient of zero: no signal (including the forward-spot spread) should be able to tell you that this is a particularly good time for long bonds vs. short bonds. As you can see, the coefficients in the right hand panel of Table 5 are all about 1.0. A high forward rate does not indicate that interest rates will be higher one year from now; it seems entirely to indicate that you will earn that much more holding long term bonds (The right hand panel is really not independent evidence, since the coefficients in the right and left hand panels of Table 5 are mechanically linked. For example 1.14 + (-0.14) = 1.0, and this holds as an accounting identity. Fama and Bliss call them “complementary regressions.”) Figures 42 and 43 provide a pictorial version of the results in Table 5. Suppose that the yield curve is upward sloping as in the left panel. What does this mean? A naive investor might think this pattern indicates that long-term bonds give a higher return than short term bonds. The expectations hypothesis denies this conclusion. If the expectations hypothesis were true, the forward rates plotted against maturity in the left hand panel would translate 389

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

one-for-one to the forecast of future spot rates in the right hand panel, as plotted in the line marked “Expectations model.” Rises in future short rates should lower bond prices, cutting off the one-period advantage of long-term bonds. The rising short rates would directly raise the multi-year advantage of short term bonds. We can calculate the actual forecast of future spot rates from the estimates in the left hand panel of Table 5, and these are given by the line market “Estimates” in Figure 43. The essence of the phenomenon is sluggish adjustment of the short rates. The short rates do eventually rise to meet the forward rate forecasts, but not as quickly as the forward rates predict that they should.

Figure 42. If the current yield curve is as plotted here.... As dividend growth should be forecastable so that returns are not forecastable, short-term yields should be forecastable so that returns are not forecastable. In fact, yield changes are almost unforecastable at a one year horizon, so, mechanically, bond returns are. We see this directly in the first row of the left hand panel of Table 5 for the one-period yield. It is an implication of the right hand panel as well. If

(N+1)

hprt+1

(1)

− yt

(N→N+1)

= 0 + 1(ft

390

(1)

− yt ) + εt+1

(338)

S ECTION 20.1

T IME - SERIES PREDICTABILITY

Figure 43. ...this is the forecast of future one year interest rates. The dashed line gives the forecast from the expectations hypothesis. The solid line is constructed from the estimates in Table 4. then, writing out the definition of holding period return and forward rate, (N)

(N+1)

pt+1 − pt

(1)

+ pt

(N) pt+1 (N) yt+1

(N)

= 0 + 1(pt = 0 = 0

(N+1)

− pt

(1)

+ pt ) + εt+1

(20.339)

(N) + 1(pt ) + εt+1 (N) + 1(yt ) − εt+1 /N

A coefficient of 1.0 in (20.338) is equivalent to yields or bond prices that follow random walks; yield changes that are completely unpredictable. Of course yields are stationary and not totally unpredictable. However, they move slowly. Thus, yield changes are very unpredictable at short horizons but much more predictable at long horizons. That is why the coefficients in the right hand panel of Table 5 build with horizon. If we did holding period return regressions at longer horizons, they would gradually approach the expectations hypothesis result. The roughly 1.0 coefficients in the right hand panel of Table 5 mean that a one percentage point increase in forward rate translates into a one percentage point increase in expected return. It seems that old fallacy of confusing bond yields with their expected returns also contains a grain of truth, at least for the first year. However, the one-for-one variation of expected returns with forward rates does not imply a one-for-one variation of expected returns 391

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

with yield spreads. Forward rates are related to the slope of the yield curve, (N→N+1)

ft

(1)

− yt

(N)

= pt

(N+1)

− pt (N)

(1)

− yt

(N+1)

(1)

= −N yt + (N + 1)yt − yt ´ ³ ´ ³ (N+1) (N) (N+1) (1) + yt = N yt − yt − yt

Thus, the forward-spot spread varies a more than the yield spread, so regression coefficients of holding period yields on yield spreads give coefficients greater than one. Expected returns move more than one-for-one with yield spreads. Campbell (1995) reports coefficients of excess returns on yield spreads that rise from one at a 2 month horizon to 5 at a 5 year horizon. The facts are analogous to the dividend/price regression. There, dividends were essentially unforecastable. This implied that a one percentage point change in dividend yield implied a 5 percentage point change in expected excess returns. Of course, there is risk: the R2 are all about 0.1-0.2, about the same values as the R2 from the dividend/price regression at a one year horizon, so this strategy will often go wrong. Still, 0.1-0.2 is not zero, so the strategy does pay off more often than not, in violation of the expectations hypothesis. Furthermore, the forward-spot spread is a slow moving variable, typically reversing sign once per business cycle. Thus, the R2 build with horizon as with the D/P regression, peaking in the 30% range (Fama and French 1989). The fact that the regressions in Table 5 run the change in yield on the forward-spot spread and the excess return on the forward-spot spread is very important. The overall level of (N) interest rates moves up and down a great deal but slowly over time. Thus, if you run yt+j = (N+1)

a + bft + εt+N , you will get a coefficient b almost exactly equal to 1.0 and a stupendous R2 , seemingly a stunning validation of the expectations hypothesis. If you run a regression of tomorrow’s temperature in Chicago on today’s temperature, the regression coefficient will be near 1.0 with a huge R2 as well, since the temperature varies a lot over the year. But today’s temperature is not a useful temperature forecast. To measure a temperature forecast we want to know if the forecast can predict the change in temperature. Is (forecast - today’s temperature) a good measure of (tomorrow’s temperature - today’s temperature)? Table 5 runs this regression.

The decomposition in (20.339) warns us of one of several econometric traps in this kind of regression. Notice that two of the three right hand variables are the same. Thus any mea(N+1) (1) and pt will induce a spurious common movement in left and right surement error in pt hand variables. In addition, since the variables are a triple difference, the difference may eliminate a common signal and isolate measurement error or noise. There are pure measurement errors in the bond data, and we seldom observe pure discount bonds of the exactly desired maturity. In addition, various liquidity and microstructure effects can influence the yields of particular bonds in ways that are not exploitable for typical investors. As an example of what this sort of “measurement error” can do, suppose all bond yields 392

S ECTION 20.1

T IME - SERIES PREDICTABILITY

are 5%, but there is one “error” in the two period bond price at time 1 – rather than being -10 it is -15. The table below tracks the effects of this error. It implies a blip of the one year forward rate in year one, and then a blip in the return from holding this bond from year one to year two. The price and forward rate “error” automatically turns in to a subsequent return when the “error” is corrected. If the price is real, of course, this is just the kind of event we want the regression to tell us about – the forward rate did not correspond to a change in future spot rate, so there was a large return; it was a price that was “out of line” and if you could trade on it, you should. But the regression will also pounce on measurement error in prices and indicate spurious returns. t (1) pt (2) pt (3) pt (i) yt , i 6= 2 (2) yt (1→2) ft (1→2) (1) ft − yt (2→1) (1) hprt − yt

0 -5 -10 -15 5 5 5 0 0

1 -5 -15 -15 5 7.5 10 5 0

2 -5 -10 -15 5 5 5 0 5

3 -5 -10 -15 5 5 5 0 0

Numerical example of the effect of measurement error in yields on yield regressions. 20.1.9

Foreign exchange

The expectations model works well on average. However, a foreign interest rate one percentage point higher than its usual differential with the US rate (equivalently, a one percentage point higher forward-spot spread) seems to indicate even more than one percentage point expected excess return; a further appreciation of the foreign currency. Suppose interest rates are higher in Germany than in the U.S. Does this mean that one can earn more money by investing in German bonds? There are several reasons that the answer might be no. First, of course is default risk. While not a big problem for German government bonds, Russia and other governments have defaulted on bonds in the past and may do so again. Second, and more important, is the risk of devaluation. If German interest rates are 10%, US interest rates are 5%, but the Euro falls 5% relative to the dollar during the year, you make no more money holding the German bonds despite their attractive interest rate. Since lots of investors are making this calculation, it is natural to conclude that an interest rate differential across countries on bonds of similar credit risk should reveal an expectation of currency devaluation. The logic is exactly the same as the “expectations hypothesis” in 393

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

the term structure. Initially attractive yield or interest rate differentials should be met by an offsetting event so that you make no more money on average in one country or another, or in one currency versus another. As with bonds, the expectations hypothesis is slightly different from pure risk neutrality since the expectation of the log is not the log of the expectation. Again, the size of the phenomena we study usually swamps this distinction. As with the expectations hypothesis in the term structure, the expected depreciation view ruled for many years, and still constitutes an important first-order understanding of interest rate differentials and exchange rates. For example, interest rates in east Asian currencies were very high on the eve of the currency collapses of 1997, and many banks were making tidy sums borrowing at 5% in dollars to lend at 20% in local currencies. This situation should lead one to suspect that traders expect a 15% devaluation, and most likely a small chance of a larger devaluation. That is, in this case, exactly what happened. Many observers and policy analysts who ought to know better often attribute high nominal interest rates in troubled countries to “tight monetary policy” that is “strangling the economy” to “defend the currency.” In fact, one’s first order guess should be that such high nominal rates reflect a large probability of inflation and devaluation – loose monetary and fiscal policy – and that they correspond to much lower real rates. Still, does a 5% interest rate differential correspond to an exactly 5% expected depreciation, or does some of it still represent a high expected return from holding debt in that country’s currency? Furthermore, while expected depreciation is clearly a large part of the story for high interest rates in countries that have constant high inflation or that may suffer spectacular depreciation of a pegged exchange rate, how does the story work for, say, the U.S. vs. Germany, where inflation rates diverge little, yet exchange rates fluctuate a surprisingly large amount? Table 6 presents the facts, as summarized by Hodrick (2000) and Engel (1996). The first row of Table 6 presents the average appreciation of the dollar against the indicated currency over the sample period. The dollar fell against DM, yen and Swiss Franc, but appreciated against the pound. The second row gives the average interest rate differential – the amount by which the foreign interest rate exceeds the US interest rate. According to the expectations hypothesis, these two numbers should be equal – interest rates should be higher in countries whose currencies depreciate against the dollar. The second row shows roughly the right pattern. Countries with steady long-term inflation have steadily higher interest rates, and steady depreciation. The numbers in the first and second rows are not exactly the same, but exchange rates are notoriously volatile so these averages are not well measured. Hodrick shows that the difference between the first and second rows is not statistically different from zero. This fact is exactly analogous to the fact of Table 4 that the expectations hypothesis works well “on average” for US bonds and is the tip of an iceberg of empirical successes for the expectations hypothesis as applied to currencies. As in the case of bonds, however, we can also ask whether times of temporarily higher or lower interest rate differentials correspond to times of above and below average depre394

S ECTION 20.1

T IME - SERIES PREDICTABILITY

ciation as they should. The third and fifth rows of Table 6 address this question, updating Hansen and Hodrick’s (1980) and Fama’s (1984) regression tests. The number here should be +1.0 in each case – an extra percentage point interest differential should correspond to one extra percentage point expected depreciation. As you can see, we have exactly the opposite pattern: a higher than usual interest rate abroad seems to lead, if anything to further appreciation. It seems that the old fallacy of confusing interest rate differentials across countries with expected returns, forgetting about depreciation, also contains a grain of truth. This is the “forward discount puzzle,” and takes its place alongside the forecastability of stock and bond returns. Of course it has produced a similar avalanche of academic work dissecting whether it is really there and if so, why. Hodrick (1987), Engel (1996), and Lewis (1995) provide surveys. The R2 shown in Table 6 are quite low. However, like D/P, the interest differential is a slow-moving forecasting variable, so the return forecast R2 build with horizon. Bekaert and Hodrick (1992) report that the R2 rise to the 30-40% range at six month horizons and then decline again. Still, taking advantage of this predictability, like the bond strategies described above, is quite risky. Mean appreciation Mean interest differential b, 1975-1989 R2 b, 1976-1996

DM -1.8 -3.9 -3.1 .026 -0.7

£ 3.6 2.1 -2.0 .033 -1.8

U -5.0 -3.7 -2.1 .034 -2.4

SF -3.0 -5.9 -2.6 .033 -1.3

Table 6. The first row gives the average appreciation of the dollar against the indicated currency, in percent per year. The second row gives the average interest differential – foreign interest rate less domestic interest rate, measured as the forward premium – the 30 day forward rate less the spot exchange rate. The third through fifth rows give the coefficients and R2 in a regression of exchange rate changes on the interest differential = forward premium, st+1 − st = a + b(ft − st ) + εt+1 = a + b(rtf − rtd ) + εt+1

where s = log spot exchange rate, f = forward rate, rf = foreign interest rate, rd = domestic interest rate. Source: Hodrick (1999) and Engel (1996). The puzzle does not say that one earns more by holding bonds from countries with higher interest rates than others. Average inflation, depreciation, and interest rate differentials line up as they should. If you just buy bonds with high interest rates, you end up with debt from Turkey and Brazil, whose currencies inflate and depreciate steadily. The puzzle does say that one earns more by holding bonds from countries whose interest rates are higher than usual relative to U.S. interest rates. However, the fact that the “usual” rate of depreciation and “usual” interest differential 395

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

varies through time, if they are well-defined concepts at all, may diminish if not eliminate the out-of-sample performance of trading rules based on these regressions. The foreign exchange regressions offer a particularly clear-cut case in which “Peso problems” can skew forecasting regressions. Lewis (1995) credits Milton Friedman for coining the term to explain why Mexican interest rates were persistently higher than U.S. interest rates in the early 1970’s even though the currency had been pegged for more than a decade. A small probability of a huge devaluation each period can correspond to a substantial interest differential. You will see long stretches of data in which the expectations hypothesis seems not to be satisfied, because the collapse does not occur in sample. The Peso subsequently collapsed, giving substantial weight to this view. Since “Peso problems” have become a generic term for the effects of small probabilities of large events on empirical work. Rietz (1988) offered a Peso problem explanation for the equity premium that investors are afraid of another great depression which has not happened in sample. Selling out of the money put options and earthquake insurance in Los Angeles are similar strategies whose average returns in a sample will be severely affected by rare events.

20.2

The Cross-section: CAPM and Multifactor Models

Having studied how average returns change over time, now we study how average returns change across different stocks or portfolios. 20.2.1

The CAPM

For a generation, portfolios with high average returns also had high betas. I illustrate with the size-based portfolios. The first tests of the CAPM such as Lintner (1965) were not a great success. If you plot or regress the average returns versus betas of individual stocks, you find a lot of dispersion, and the slope of the line is much too flat – it does not go through any plausible riskfree rate. Miller and Scholes (1972) diagnosed the problem. Betas are measured with error, and measurement error in right hand variables biases down regression coefficients. Fama and MacBeth (1973) and Black, Jensen and Scholes (1972) addressed the problem by grouping stocks into portfolios. Portfolio betas are better measured because the portfolio has lower residual variance. Also, individual stock betas vary over time as the size, leverage, and risks of the business change. Portfolio betas may be more stable over time, and hence easier to measure accurately. There is a second reason for portfolios. Individual stock returns are √ so volatile that you cannot reject the hypothesis that all average returns are the same. σ/ T is big when σ = 396

S ECTION 20.2

T HE C ROSS - SECTION : CAPM AND MULTIFACTOR MODELS

40 − 80%. By grouping stocks into portfolios based on some characteristic (other than firm name) related to average returns, you reduce the portfolio variance and thus make it possible to see average return deferences. Finally, I think much of the attachment to portfolios comes from a desire to more closely mimic what actual investors would do rather than simply form a test.

Fama and MacBeth and Black Jensen and Scholes formed their portfolios on betas. They found individual stock betas, formed stocks into portfolios based on their betas, and then estimated the portfolio’s beta in the following period. More recently, size, book/market, industry, and many other characteristics have been used to form portfolios. Ever since, the business of testing asset pricing models has been conducted in a simple loop: 1.

2. 3.

Find a characteristic that you think is associated with average returns. Sort stocks into portfolios based on the characteristic, and check that there is a difference in average returns between portfolios. Worry here about measurement, survival bias, fishing bias, and all the other things that can ruin a pretty picture out of sample. Compute betas for the portfolios, and check whether the average return spread is accounted for by the spread in betas. If not, you have an anomaly. Consider multiple betas. This is the traditional procedure, but econometrics textbooks urge you not to group data in this way. They urge you to use the characteristic as an instrument for the poorly measured right hand variable instead. It is an interesting and unexplored idea whether this instrumental variables approach could fruitfully bring us back to the examination of individual securities rather than portfolios. The CAPM proved stunningly successful in empirical work. Time after time, every strategy or characteristic that seemed to give high average returns turned out to also have high betas. Strategies that one might have thought gave high average returns (such as holding very volatile stocks) turned out not to have high average returns when they did not have high betas. To give some sense of that empirical work, Figure 44 presents a typical evaluation of the Capital Asset Pricing Model. (Chapter 15 presented some of the methodological issues surrounding this evaluation; here I focus on the facts.) I examine 10 portfolios of NYSE stocks sorted by size (total market capitalization), along with a portfolio of corporate bonds and long-term government bonds. As the spread along the vertical axis shows, there is a sizeable spread in average returns between large stocks (lower average return) and small stocks (higher average return), and also a large spread between stocks and bonds. The figure plots these average returns against market betas. You can see how the CAPM prediction fits: portfolios with higher average returns have higher betas. In particular, notice that the long term and corporate bonds have mean returns in line with their low betas, despite their standard deviations nearly as high as those of stocks. Comparing this graph with the similar Figure 5 of the consumption-based model back in Chapter 2, the CAPM fits very well. 397

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

Figure 44. The CAPM. Average returns vs. betas on the NYSE value-weighted portfolio for 10 size-sorted stock portfolios, government bonds, and corporate bonds, 1947-1996. The solid line draws the CAPM prediction by fitting the market proxy and treasury bill rates exactly (a time-series test). The dashed line draws the CAPM prediction by fitting an OLS cross-sectional regression to the displayed data points. The small firm portfolios are at the top right. The points far down and to the left are the government bond and treasury bill returns. In fact, Figure 44 captures one of the first significant failures of the CAPM. The smallest firms (the far right portfolio) seem to earn an average return a few percent too high given their betas. This is the celebrated “small-firm effect” (Banz 1981). Would that all failed economic theories worked so well! It is also atypical in that the estimated market line through the stock portfolios is steeper than predicted, while measurement error in betas usually means that the estimated market line is too flat. 20.2.2

Fama-French 3 factors

Book to market sorted portfolios show a large variation in average returns that is unrelated to market betas. The Fama and French 3 factor model successfully explains the average returns of the 25 size and book to market sorted portfolios with a 3 factor model, consisting of the market, a small minus big (SMB) portfolio and a high minus low (HML) portfolio. 398

S ECTION 20.2

T HE C ROSS - SECTION : CAPM AND MULTIFACTOR MODELS

In retrospect, it is surprising that the CAPM worked so well for so long. The assumptions on which it is built are very stylized and simplified. Asset pricing theory recognized at least since Merton (1971a,b) the theoretical possibility, indeed probability, that we should need factors, state variables or sources of priced risk beyond movements in the market portfolio in order to explain why some average returns are higher than others. The Fama - French model is one of the most popular multi-factor models that now dominate empirical research. Fama and French (1993) presents the model; Fama and French (1996) gives an excellent summary, and also shows how the 3 factor model performs in evaluating expected return puzzles beyond the size and value effects that motivated it. “Value” stocks have market values that are small relative to the accountant’s book value. (Book values essentially track past investment expenditures.) This category of stocks has given large average returns. “Growth” stocks are the opposite of value and have had low average returns. Since low prices relative to dividends, earnings or book value forecast times when the market return will be high, it is natural to suppose that these same signals forecast categories of stocks that will do well; the “value effect” is the cross-sectional analogy to price-ratio predictability in the time-series. High average returns are consistent with the CAPM, if these categories of stocks have high sensitivities to the market, high betas. However, small and especially value stocks seem to have abnormally high returns even after accounting for market beta. Conversely “growth” stocks seem to do systematically worse than their CAPM betas suggest. Figure 45 shows this value-size puzzle. It is just like Figure 44, except that the stocks are sorted into portfolios based on size and book-market ratio9 rather than size alone. As you can see, the highest portfolios have three times the average excess return of the lowest portfolios, and this variation has nothing at all to do with market betas. Figures 46 and 47 dig a little deeper to diagnose the problem, by connecting portfolios that have different size within the same book/market category, and different book/market within size category. As you can see, variation in size produces a variation in average returns that is positively related to variation in market betas, as we had in Figure 45. Variation in book/market ratio produces a variation in average return is negatively related to market beta. Because of this value effect, the CAPM is a disaster when confronted with these portfolios. (Since the size effect disappeared in 1980, it is likely that almost the whole story can be told with book/market effects alone.) To explain these patterns in average returns, Fama and French advocate a multifactor model with the market return, the return of small less big stocks (SMB) and the return of high book/market minus low book/market stocks (HML) as three factors. They show that variation in average returns of the 25 size and book/market portfolios can be explained by varying loadings (betas) on the latter two factors. (All their portfolios have betas close to one on the market portfolio. Thus, market beta explains the average return difference between 9

I thank Gene Fama for providing me with these data.

399

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

Figure 45. Average returns vs. market beta for 25 stock portfolios sorted on the basis of size and book/market ratio. stocks and bonds, but not across categories of stocks.) Figures 48 and 49 illustrate Fama and French’s results. The vertical axis is still the average return of the 25 size and book/market portfolios. Now, the horizontal axis is the predicted values from the Fama-French three factor model. The points should all lie on a 45◦ line if the model is correct. The points lie much closer to this prediction than they do in Figures 46 and 47. The worst fit is for the growth stocks (lowest line, left hand panel), for which there is little variation in average return despite large variation in size beta as one moves from small to large firms. 20.2.3

What are the size and value factors?

What are the macroeconomic risks for which the Fama-French factors are proxies or mimicking portfolios? There are hints of some sort of “distress” or “recession” factor at work. A central part of the Fama French model is the fact that these three pricing factors also explain a large part of the ex-post variation in the 25 portfolios – the R2 in time-series regressions are very high. In this sense, one can regard it as an APT rather than a macroeconomic 400

S ECTION 20.2

T HE C ROSS - SECTION : CAPM AND MULTIFACTOR MODELS

Figure 46. Average excess returns vs. market beta. Lines connect portfolios with different size category within book to market categories. factor model. The Fama-French model is not a tautology, despite the fact that factors and test portfolios are based on the same set of characteristics. We would like to understand the real, macroeconomic, aggregate, nondiversifiable risk that is proxied by the returns of the HML and SMB portfolios. Why are investors so concerned about holding stocks that do badly at the times that the HML (value less growth) and SMB (small-cap less large-cap) portfolios do badly, even though the market does not fall? Fama and French (1995) note that the typical “value” firm has a price that has been driven down from a long string of bad news, and is now in or near financial distress. Stocks bought on the verge of bankruptcy have come back more often than not, which generates the high average returns of this strategy. This observation suggests a natural interpretation of the value premium: If a credit crunch, liquidity crunch, flight to quality or similar financial event comes along, stocks in financial distress will do very badly, and this is just the sort of time at which one particularly does not want to hear that one’s stocks have become worthless! (One cannot count the “distress” of the individual firm as a “risk factor.” Such distress is idiosyncratic and can be diversified away. Only aggregate events that average investors care about can result in a risk premium.) Unfortunately, empirical support for this theory is weak, since the HML 401

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

Figure 47. Average excess returns vs. market beta. Lines connect portfolios with different book to market categories within size categories. portfolio does not covary strongly with other measures of aggregate financial distress. Still, it is a possible and not totally tested interpretation, since we have so few events of actual systematic financial stress in recent history. Heaton and Lucas’ (1997) results add to this story for the value effect. They note that the typical stockholder is the proprietor of a small, privately held business. Such an investor’s income is of course particularly sensitive to the kinds of financial events that cause distress among small firms and distressed value firms. Such an investor would therefore demand a substantial premium to hold value stocks, and would hold growth stocks despite a low premium. Lettau and Ludvigson (2000) (also discussed in the next section) document that HML has a time-varying beta on both the market return and on consumption. Thus, though there is (unfortunately) very little unconditional correlation between HML and recession measures, Lettau and Ludvigson document that HML is sensitive to bad news in bad times. Liew and Vassalou (1999) are an example of current attempts to link value and small firm returns to macroeconomic events. They find that in many countries counterparts to HML and SMB contain information above and beyond that in the market return for forecasting GDP growth. For example, they report a regression GDPt→t+1 = a + 0.065 M KTt−1→t + 0.058 HM Lt−1→t + εt+1

402

S ECTION 20.2

T HE C ROSS - SECTION : CAPM AND MULTIFACTOR MODELS

Figure 48. Average excess return vs. prediction of the Fama-French 3 factor model. Lines connect portfolios of different size categories within book to market category. GDPt→t+1 denotes the next year’s GDP growth and MKT, HML denote the previous year’s return on the market index and HML portfolio. Thus, a 10% HML return reflects a 1/2 percentage point rise in the GDP forecast.

On the other hand, one can ignore Fama and French’s motivation and regard the model as an arbitrage pricing theory. If the returns of the 25 size and book/market portfolios could be perfectly replicated by the returns of the 3 factor portfolios – if the R2 in the time-series regressions were 100% – then the multifactor model would have to hold exactly, in order to preclude arbitrage opportunities. In fact the R2 of Fama and French’s time-series regressions are all in the 90%-95% range, so extremely high Sharpe ratios for the residuals (which are portfolios) would have to be invoked for the model not to fit well. Equivalently, given the average returns from HML and SMB and the failure of the CAPM to explain those returns, there would be near-arbitrage opportunities if value and small stocks did not move together in the way described by the Fama-French model. One way to assess whether the three factors proxy for real macroeconomic risks is by checking whether the multifactor model prices additional portfolios, and especially portfolios that do not have high R2 values. Fama and French (1996) extend their analysis in this direction: They find that the SMB and HML portfolios comfortably explain strategies based on alternative price multiples (P/E, B/M), strategies based on 5 year sales growth (this is especially interesting since it is the only strategy that does not form portfolios based on price variables) and the tendency of 5 year returns to reverse. All of these strategies are not ex403

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

Figure 49. Average excess returns vs. predictions of the Fama-French 3 factor model. Lines connect portfolios of different book to market category within the same size category. plained by CAPM betas. However they all also produce portfolios with high R2 values in a time-series regression on the HML and SMB portfolios! This is good and bad news. It might mean that the model is a good APT; that the size and book/market characteristics describe the major sources of priced variation in all stocks. On the other hand it might mean that these extra sorts just haven’t identified other sources of priced variation in stock returns. (Fama and French also find that HML and SMB do not explain “momentum,” despite large R2 values. More on momentum later.) One’s first reaction may be that explaining portfolios sorted on the basis of size and book to market by factors sorted on the same basis is a tautology. This is not the case. For example, suppose that average returns were higher for stocks whose ticker symbols start later in the alphabet. (Maybe investors search for stocks alphabetically, so the later stocks are “overlooked.”) This need not trouble us if Z stocks happened to have higher betas. If not – if letter of the alphabet were a CAPM anomaly like book to market – however, it would not necessarily follow that letter based stock portfolios move together. Adding A-L and M-Z portfolios to the right hand side of a regression of the 26 A,B,C, etc. portfolios on the market portfolio need not (and probably does not) increase the R2 at all. The size and book to market premia are hard to measure, and seem to have declined substantially in recent years. But even if they decline back to CAPM values, Fama and French will still have found a surprisingly large source of common movement in stock returns. More to the point, in testing a model It is exactly the right thing to do to sort stocks 404

S ECTION 20.2

T HE C ROSS - SECTION : CAPM AND MULTIFACTOR MODELS

into portfolios based on characteristics related to expected returns. When Black Jensen and Scholes and Fama and MacBeth first tested the CAPM, they sorted stocks into portfolios based on betas, because betas are a good characteristic for sorting stocks into portfolios that have a spread in average returns. If your portfolios have no spread in average returns – if you just choose 25 random portfolios – then there will be nothing for the asset pricing model to test. In fact, despite the popularity of the Fama French 25, there is really no fundamental reason to sort portfolios based on 2 way or larger sorts of individual characteristics. You should use all the characteristics at hand that (believably!) indicate high or low average returns and simply sort stocks according to a one-dimensional measure of expected returns. The argument over the status of size and book/market factors continues, but the important point is that it does so. Faced with the spectacular failure of the CAPM documented in Figures and 4, one might have thought that any hope for a rational asset pricing theory was over. Now we are back where we were, examining small anomalies and arguing over refinements and interpretations of the theory. That is quite an accomplishment! 20.2.4

Macroeconomic factors

Labor income, industrial production, news variables and conditional asset pricing models have also all had some successes as multifactor models. I have focused on the size and value factors since they provide the most empirically successful multifactor model to date, and have therefore attracted much attention. Several authors have used macroeconomic variables as factors in order to examine directly the story that stock performance during bad macroeconomic times determines average returns. Jagannathan and Wang (1996) and Reyfman (1997) use labor income; Chen Roll and Ross (1986) use industrial production and inflation among other variables. Cochrane (1996) uses investment growth. All these authors find that average returns line up against betas calculated using these macroeconomic indicators. The factors are theoretically easier to motivate, but none explains the value and size portfolios as well as the (theoretically less solid, so far) size and value factors. Lettau and Ludvigson (2000) specify a macroeconomic model that does just as well as the Fama-French factors in explaining the 25 Fama-French portfolios. Their plots of actual average returns vs. model predictions show a relation as strong as those of Figures 48 and 49. Their model is mt+1 = a + b(cawt )∆ct+1

where caw is a measure of the consumption-wealth ratio. This is a “scaled factor model” of 405

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

the sort advocated in Chapter 8. You can think of it as capturing a time-varying risk aversion. This is a stunning result. Though Merton’s (1971a,b) theory says that variables which predict market returns should show up as factors which explain cross-sectional variation in average returns, surprisingly few papers have actually tried to see whether this is true, now that we do have variables that we think forecast the market return. Campbell (1996) and Ferson and Harvey (1999) are among the few exceptions. 20.2.5

Momentum and reversal

Sorting stocks based on past performance, you find that a portfolio that buys long-term losers and sells long-term winners does better than the opposite – individual stock long-term returns mean-revert. This “reversal” effect makes sense given return predictability and meanreversion, and is explained by the Fama-French 3 factor model. However, a portfolio that buys short-term winners and sells short-term losers also does well – “momentum.” This effect is a puzzle. Since a string of good returns gives a high price, it is not surprising that stocks that do well for a long time (and hence build up a high price) subsequently do poorly, and stocks that do poorly for a long time (and hence dwindle down to a low price, market value, or market to book ratio) subsequently do well Table 3, taken from Fama and French (1996) reveals that this is in fact the case. (As usual, this table is the tip of an iceberg of research on these effects, starting with DeBont and Thaler 1985 and Jagadeesh and Titman 1993.)

Strategy Reversal Momentum Reversal Momentum

Period 6307-9312 6307-9312 3101-6302 3101-6302

Portfolio Formation Months 60-13 12-2 60-13 12-2

Average Return, 10-1 (Monthly %) -0.74 +1.31 -1.61 +0.38

Table 3. Average monthly returns from reversal and momentum strategies. Each month, allocate all NYSE firms on CRSP to 10 portfolios based on their performance during the “portfolio formation months” interval. For example, 60-13 forms portfolios based on returns from 5 years ago to 1 year, 1 month ago. Then buy the best-performing decile portfolio and short the worst-performing decile portfolio. Source: Fama and French (1996) Table VI. Reversal

406

S ECTION 20.2

T HE C ROSS - SECTION : CAPM AND MULTIFACTOR MODELS

Here is the “reversal” strategy. Each month, allocate all stocks to 10 portfolios based on performance in year -5 to year -1. Then, buy the best-performing portfolio and short the worst-performing portfolio. The first row of Table 3 shows that this strategy earns a hefty -0.74% monthly return10 . Past long-term losers come back and past winners do badly. This is a cross-sectional counterpart to the mean-reversion that we studied in section 1.4. Fama and French (1998a) already found substantial mean-reversion – negative long-horizon return autocorrelations – in disaggregated stock portfolios, so one would expect this phenomenon. Spreads in average returns should correspond to spreads in betas. Fama and French verify that these portfolio returns are explained by their 3 factor model. Past losers have a high HML beta; they move together with value stocks, and so inherit the value stock premium. Momentum

The second row of Table 3 tracks the average monthly return from a “momentum” strategy. Each month, allocate all stocks to 10 portfolios based on performance in the last year. Now, quite surprisingly, the winners continue to win, and the losers continue to lose, so that buying the winners and shorting the losers generates a positive 1.31% monthly return. At every moment there is a most-studied anomaly, and momentum is that anomaly as I write. It is not explained by the Fama French 3 factor model. The past losers have low prices and tend to move with value stocks. Hence the model predicts they should have high average returns, not low average returns. Momentum stocks move together, as do value and small stocks so a “momentum factor” works to “explain” momentum portfolio returns. This is so obviously ad-hoc (i.e. an APT factor that will only explain returns of portfolios organized on the same characteristic as the factor) that nobody wants to add it as a risk factor. A momentum factor is more palatable as a performance attribution factor – to say that a fund did well by following a momentum strategy rather than by stock picking ability, leaving aside why a momentum strategy should work. Carhart (1997) uses it in this way to show that similar momentum behavior in fund returns is due to momentum in the underlying stocks rather than persistent stock-picking skill. Momentum may be explained as just a new way of looking at an old phenomenon, the small apparent predictability of monthly individual stock returns. A tiny regression R2 for forecasting monthly returns of 0.0025 (1/4%) is more than adequate to generate the momentum results of Table 3. The key is the large standard deviation of individual stock returns, typically 40% or more at an annual basis. The average return of the best performing decile of a normal distribution is 1.76 standard deviations above the mean11 , so the winning moFama and French do not provide direct measures of standard deviations for these portfolios. One can infer however from the betas, R2 values and standard deviation of market and factor portfolios that the standard deviations are roughly 1-2 times that of the market return, so that Sharpe ratios of these strategies are comparable to that of the market return. 11 We’re looking for 10

R∞ rf (r)dr E(r|r ≥ x) = Rx∞ x f (r)dr

407

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

mentum portfolio typically went up about 80% in the previous year, and the typical losing portfolio went down about 60% per year. Only a small amount of continuation will give a 1% monthly return when multiplied by such large√past returns. To be precise, the monthly indi2 vidual stock standard deviation is about 40%/ √ 12 ≈ 12% . If the R is 0.0025, the standard deviation of the predictable part of returns is 0.0025 × 12% = 0.6%. Hence, the decile predicted to perform best will earn 1.76 × 0.6% ≈ 1% above the mean. Since the strategy buys the winners and shorts the losers, an R2 of 0.0025 implies that one should earn a 2% monthly return by the momentum strategy – more even than the 1.3% shown in Table 3. Lewellen (2000) offers a related explanation for momentum coming from small cross-correlations of returns. We have known at least since Fama (1965) that monthly and higher frequency stock returns have slight, statistically significant predictability with R2 in the 0.01 range. However, such small though statistically significant high frequency predictability, especially in small stock returns, has also since the 1960s always failed to yield exploitable profits after one accounts for transactions costs, thin trading, high short sale costs and other microstructure issues. Hence, one naturally worries whether momentum is really exploitable after transactions costs. Momentum does require frequent trading. The portfolios in Table 3 are reformed every month. Annual winners and losers will not change that often, but the winning and losing portfolio must still be turned over at least once per year. Carhart (1996) calculates transactions costs and concludes that momentum is not exploitable after those costs are taken into account. Moskowitz and Grinblatt (1999) note that most of the apparent gains come from short positions in small, illiquid stocks, positions that also have high transactions costs. They also find that a large part of momentum profits come from short positions taken November, anticipating tax-loss selling in December. This sounds a lot more like a small microstructure glitch rather than a central parable for risk and return in asset markets. Table 3 already shows that the momentum effect essentially disappears in the earlier data sample, while reversal is even stronger in that sample. Ahn, Boudoukh, Richardson, and Whitelaw (1999) show that apparent momentum in international index returns is missing from the futures markets, also suggesting a microstructure explanation. Of course, it is possible that a small positive autocorrelation is there and related to some risk. However, it is hard to generate real positive autocorrelation in realized returns. As we saw extensively in section 20.335, a slow and persistent variation in expected returns most naturally generates negative autocorrelation in realized returns. News that expected returns are higher means future dividends are discounted at a higher rate, so today’s price and return declines. The only way to overturn this prediction is to suppose that expected return where x is defined as the top 10th cutoff, Z

∞

f (r)dr =

x

1 . 10

With a normal distribution, x = 1.2816σ and E(r|r ≥ x) = 1.755σ.

408

S ECTION 20.3

S UMMARY AND INTERPRETATION

shocks are positively correlated with shocks to current or expected future dividend growth. A convincing story for such correlation has not yet been constructed. On the other hand, the required positive correlation is very small and not very persistent.

20.3

Summary and interpretation

While the list of new facts appears long, similar patters show up in every case. Prices reveal slow-moving market expectations of subsequent excess returns, because potential offsetting events seem sluggish or absent. The patterns suggest that there are substantial expected return premia for taking on risks of recession and financial stress unrelated to the market return. Magnifying glasses

The effects are not completely new. We knew since the 1960s that high frequency returns are slightly predictable, with R2 of 0.01 to 0.1 in daily to monthly returns. These effects were dismissed because there didn’t seem to be much that one could do about them. A 51/49 bet is not very attractive, especially if there is any transactions cost. Also, the increased Sharpe ratio one can obtain by exploiting predictability is directly related to the forecast R2 , so tiny R2 , even if exploitable, did not seem like an important phenomenon. Many of the new facts amount to clever magnifying glasses, ways of making small facts economically interesting. For forecasting market returns, we now realize that R2 rise with horizon when the forecasting variables are slow-moving. Hence small R2 at high frequency can mean really substantial R2 , in the 30-50% range, at longer horizons. Equivalently, we realize that small expected return variation can add up to striking price variation if the expected return variation is persistent. For momentum effects, the ability to sort stocks and funds into momentum-based portfolios means that incredibly small predictability times portfolios with huge past returns gives important subsequent returns. Dogs that did not bark

In each case, an apparent difference in yield should give rise to an offsetting movement, but seems not to do so. Something should be predictable so that returns are not predictable, and it isn’t. The d/p forecasts of the market return were driven by the fact that dividends should be predictable, so that returns are not. Instead, dividend growth seems nearly unpredictable. As we saw, this fact and the speed of the d/p mean reversion imply the observed magnitude of return predictability. The term structure forecasts of bond returns were driven by the fact that bond yields should be predictable, so that returns are not. Instead, yields seem nearly unpredictable at the one year horizon. This fact means that the forward rate moves one for one with expected returns, and that a one percentage point increase in yield spread signals as much as a 5 percentage point increase in expected return. Exchange rates should be forecastable so that foreign exchange returns are not. Instead, 409

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

a one percentage point increase in interest rate abroad seems to signal a greater than one percentage point increase in expected return. Prices reveal expected returns

If expected returns rise, prices are driven down, since future dividends or other cash flows are discounted at a higher rate. A “low” price, then, can reveal a market expectation of a high expected or required return. Most of our results come from this effect. Low price/dividend, price/earnings, price/book values signal times when the market as a whole will have high average returns. Low market value (price times shares) relative to book value signals securities or portfolios that earn high average returns. The “small firm” effect derives from low prices – other measures of size such as number of employees or book value alone have no predictive power for returns (Berk 1997). The “5 year reversal” effect derives from the fact that 5 years of poor returns lead to a low price. A high long-term bond yield means that the price of long term bonds is “low,” and this seems to signal a time of good long-term bonds returns. A high foreign interest rate means a low price on foreign bonds, and this seems to indicate good returns on the foreign bonds. The most natural intepretatation of all these effects is that the expected or required return – the risk premium – on individual securities as well as the market as a whole varies slowly over time. Thus we can track market expectations of returns by watching price/dividend, price/earnings or book/market ratios. Macroeconomic risks

The price-based patterns in time-series and cross-sectional expected returns suggest a premium for holding risks related to recession and economy-wide financial distress. All of the forecasting variables are connected to macroeconomic activity (Fama and French 1989). The dividend price ratio is highly correlated with the default spread and rises in bad times. The term spread forecasts bond and stock returns, and is also one of the best recession forecasters. It rises steeply at the bottoms of recessions, and is inverted at the top of a boom. Thus, return forecasts are high at the bottom of business cycles and low at the top of booms. “Value” and “small-cap” stocks are typically distressed. Formal quantitative and empirically successful economic models of the recession and distress premia are still in their infancy (I think Campbell and Cochrane 1999 is a good start), but the story is at least plausible, and the effects have been expected by theorists for a generation. To make this point come to life, think concretely about what you have to do to take advantage of the value or predictability strategies. You have to buy stocks or long-term bonds at the bottom, when stock prices are low after a long and depressing bear market; in the bottom of a recession or financial panic; a time when long-term bond prices and corporate bond prices are unusually low. This is a time when few people have the guts (the risktolerance) or the wallet to buy risky stocks or risky long-term bonds. Looking across stocks rather than over time, you have to invest in “value” or small market capitalization companies, dogs by any standards. These are companies with years of poor past returns, years of poor 410

S ECTION 20.3

S UMMARY AND INTERPRETATION

sales, companies on the edge of bankruptcy, far off of any list of popular stocks to buy. Then, you have to sell stocks and long term bonds in good times, when stock prices are high relative to dividends, earnings and other multiples, when the yield curve is flat or inverted so that long term bond prices are high. You have to sell the popular “growth” stocks with good past returns, good sales and earnings growth. I’m going on a bit here to counter the widespread impression, best crystallized by Shiller (2000) that high price earnings ratios must signal “irrational exuberance.” Perhaps, but is it just a coincidence that this exuberance comes at the top of an unprecedented economic expansion, a time when the average investor is surely feeling less risk averse than ever, and willing to hold stocks despite historically low risk premia? I don’t know the answer, but the rational explanation is surely not totally impossible! Is it just a coincidence that we are finding premia just where a generation of theorists said we ought to – in recessions, credit crunches, bad labor markets, investment opportunity set variables, and so forth? This line of explanation for the foreign exchange puzzle is still a bit farther off, though there are recent attempts to make economic sense of the puzzle (See Engel’s 1996 survey; Atkeson, Alvarez and Kehoe 1999 is a recent example.) At a verbal level, the strategy leads you to invest in countries with high interest rates. High interest rates are often a sign of monetary instability or other economic trouble, and thus may mean that the investments are be more exposed to the risks of global financial stress or a global recession than are investments in the bonds of countries with low interest rates, who are typically enjoying better times. Overall, the new view of finance amounts to a profound change. We have to get used to the fact that most returns and price variation come from variation in risk premia, not variation in expected cash flows, interest rates, etc. Most interesting variation in priced risk comes from non-market factors. These are easy to say, but profoundly change our view of the world. Doubts

Momentum is, so far, unlike all the other results. The underlying phenomenon is a small predictability of high frequency returns. However, the price-based phenomena make this predictability important by noting that, with a slow-moving forecasting variable, the R2 build over horizon. Momentum is based on a fast-moving forecast variable – the last year’s return. Therefore the R2 decline with horizon. Instead, momentum makes the tiny autocorrelation of high frequency returns significant by forming portfolios of extreme winners and losers, so a small continuation of huge past returns gives a large current return. All the other results are easily digestible as a slow, business-cycle related time-varying expected return. This specification gives negative autocorrelation (unless we add a distasteful positive correlation of expected return and dividend shocks) and so does not explain momentum. Momentum returns have also not yet been linked to business cycles or financial distress in even the informal way that I suggested for the price-based strategies. Thus, it still lacks much of a plausible economic interpretation. To me, this adds weight to the view that it isn’t there, it isn’t exploitable, or it represents a small illiquidity (tax-loss selling of small illiquid stocks) that will be quickly remedied once a few traders understand it. In the entire history of finance there has always been an anomaly-du-jour, and momentum is it right now. We will have to wait to 411

C HAPTER 20

E XPECTED RETURNS IN THE TIME - SERIES AND CROSS - SECTION

see how it is resolved. Many of the anomalous risk premia seem to be declining over time. The small firm effect completely disappeared in 1980; you can date this as the publication of the first small firm effect papers or the founding of small firm mutual funds that made diversified portfolios of small stocks available to average investors. To emphasize this point, Figure 50 plots size portfolio average returns vs. beta in the period since 1979. You can see that not only has the small firm premium disappeared, the size-related variation in beta and expected return has disappeared.

Figure 50. Average returns vs. market betas. CRSP size portfolios less treasury bill rate, monthly data 1979-1998. The value premium has been cut roughly in half in the 1990s, √ and 1990 is roughly the date of widespread popularization of the value effect, though σ/ T leaves a lot of room for error here. As you saw in Table RR, the last 5 years of high market returns have cut the estimated return predictability from the dividend-price ratio in half. These facts suggest an uncomfortable implication: that at least some of the premium the new strategies yielded in the past was due to the fact that they were simply overlooked or are artifacts of data-dredging. Since they are hard to measure, one is tempted to put less emphasis on these average 412

S ECTION 20.4

P ROBLEMS

returns. However, they are crucial to our interpretation of the facts. The CAPM is perfectly consistent with the fact that there are additional sources of common variation. For example, it was long understood that stocks in the same industry move together; the fact that value or small stocks also move together need not cause a ripple. The surprise is that investors seem to earn an average return premium for holding these additional sources of common movement, whereas the CAPM predicts that (given beta) they should have no effect on a portfolio’s average returns.

20.4 1.

2. 3.

Problems

Does equation (20.308) condition down to information sets coarser than that observed by agents? Or must we assume that whatever VAR is used by the econometrician contains all information seen by agents? Show that the two regressions in Table 5 are complementary – that the coefficients add up to one, mechanically, in sample. Derive the return innovation decomposition (20.319), directly. Write the return rt = ∆dt + ρ (pt − dt ) − (pt−1 − dt−1 )

Apply Et − Et−1 to both sides, rt − Et−1 rt = (Et − Et−1 ) ∆dt + ρ (Et − Et−1 ) (pt − dt ) .

4. 5.

(340)

Use the price-dividend identity and iterate forward to obtain (20.308). Find the univariate representation and mean-reversion statistics for prices implied by the simple VAR and the three dividend examples. Find the univariate return representation from a general return forecasting VAR. rt+1 = axt + εrt+1 xt+1 = bxt + εxt+1

6.

7.

Find the correlation between return and x shocks necessary to generate uncorrelated returns. Show that stationary xt − yt , ∆xt , ∆yt imply that xt and yt must have the same variance ratio and long-run differences must become perfectly correlated. Start by showing that the long run variance limk→∞ var(xt+k − xt )/k for any stationary variable must be zero. Apply that fact to xt − yt . Compute the long-horizon regression coefficients and R2 in the VAR (20.311)-(20.317). Show that the R2 do indeed rise with horizon. Do coefficients and R2 rise forever, or do they turn around at some point?

413

Chapter 21. Equity premium puzzle and consumption-based models The original specification of the consumption-based model was not a great success, as we saw in Chapter 1. Still, it is in some sense the only model we have. The central task of financial economics is to figure out what are the real risks that drive asset prices and expected returns. Something like the consumption-based model – investors’ first order conditions for savings and portfolio choice – has to be the starting point. Rather than dream up models, test them and reject them, financial economists since the work of Mehra and Prescott (1986) and Hansen and Jagannathan (1991) have been able to work backwards to some extent, characterizing the properties that discount factors must have in order to explain asset return data. Among other things, we learned that the discount factor had to be extremely volatile, while not too conditionally volatile; the riskfree rate or conditional mean had to be pretty steady. This knowledge is now leading to a much more successful set of variations on the consumption-based model.

21.1 21.1.1

Equity premium puzzles The basic equity premium/riskfree rate puzzle

The postwar US market Sharpe ratio is about 0.5 – an 8% return and 16% standard deviation. The basic Hansen-Jagannathan bound

σ(m) E(Re ) ≤ ≈ γσ(∆c) σ(Re ) E(m)

implies σ(m) ≥ 50% on an annual basis, requiring huge risk aversion or consumption growth volatility. The average risk free rate is about 1%, so E(m) ≈ 0.99. High risk aversion with power utility implies a very high riskfree rate, or requires a negative subjective discount factor. Interest rates are quite stable over time and across countries, so Et (m) varies little. High risk aversion with power utility implies that interest rates are very volatile. In Chapter 1, we derived the basic Hansen-Jagannathan (1991) bounds. These are characterizations of the discount factors that price a given set of asset returns. Manipulating 414

S ECTION 21.1

E QUITY PREMIUM PUZZLES

0 = E(mRe ) we found |E(Re )| σ(m) ≥ . E(m) σ(Re )

(341)

In continuous time, or as an approximation in discrete time, we found that time-separable utility implies γσ(∆c) ≥

|E(Re )| σ(Re )

(342)

where γ = −cu00 /u0 is the local curvature of the utility function, and risk aversion coefficient for the power case. Equity premium puzzle

The postwar mean value weighted NYSE is about 8% per year over the T-bill rate, with a standard deviation of about 16%. Thus, the market Sharpe ratio E(Re )/σ(Re ) is about 0.5 for an annual investment horizon. If there were a constant risk free rate, E(m) = 1/Rf

would nail down E(m). The T-bill rate is not very risky, so E(m) is not far from the inverse of the mean T-bill rate, or about E(m) ≈ 0.99. Thus, these basic facts about the mean and variance of stocks and bonds imply σ(m) > 0.5. The volatility of the discount factor must be about 50% of its level in annual data! Per capita consumption growth has standard deviation about 1% per year. With log utility, that implies σ(m) = 0.01 = 1% which off by a factor of 50. To match the equity premium we need γ > 50,which seems a huge level of risk aversion. Equivalently, a log utility investor with consumption growth of 1% and facing a 0.5 Sharpe ratio should be investing dramatically more in the stock market, borrowing to do so. He should invest so much that his wealth and hence consumption growth does vary by 50% each year. Correlation puzzle

The bound takes the extreme possibility that consumption and stock returns are perfectly correlated. They are not, in the data. Correlations are hard to measure, since they are sensitive to data definition, timing, time-aggregation, and so forth. Still, the correlation of annual stock returns and nondurable plus services consumption growth in postwar U.S. data is no more than about 0.2. If we use this information as well – if we characterize the mean and standard deviation of all discount factors that have correlation less than 0.2 with the market return – the calculation becomes |E(Re )| 1 1 σ(m) ¯ ≥ ¯¯ = 0.5 = 2.5 e ¯ E(m) 0.2 ρm,Re σ(R )

with σ(m) ≈ γσ(∆c), we now need a risk aversion coefficient of 250! 415

C HAPTER 21

E QUITY PREMIUM PUZZLE AND CONSUMPTION - BASED MODELS

Here is a classier way to state the correlation puzzle. Remember that proj(m|X) should price assets just as well as m itself. Now, m = proj(m|X)+ε and σ 2 (m) = σ2 (proj (m|X))+ σ2 (ε). Some of the early resolutions of the equity premium puzzle ended up adding noise uncorrelated with asset payoffs to the discount factor. This modification increased discount factor volatility and satisfied the bound. But as you can see, adding ε increases σ 2 (m) with no effect whatsoever on the model’s ability to price assets. As you add ε, the correlation between m and asset returns declines. A bound with correlation, or equivalently comparing σ2 (proj(m|X)) rather than σ 2 (m) to the bound avoids this trap. Average interest rates and subjective discount factors

It has been traditional to use risk aversion numbers of 1 to 5 or so, but perhaps this is tradition, not fact. What’s wrong with γ = 50 to 250? The most basic piece of evidence for low γ comes from the relation between consumption growth and interest rates. " µ ¶−γ # Ct+1 f Rt = Et (mt+1 ) = Et β Ct or, in continuous time, rtf = δ + γEt (∆c) .

(343)

We can take unconditional expectations to compare these equations with average interest rates and consumption growth. Average real interest rates are also about 1% Thus, γ = 50 to 250 with a typical δ such as δ = 0.01 implies a very high riskfree rate, of 50 − 250%. To get a reasonable interest rate, we have to use a subjective discount factor δ = −0.5 to −2.5, or −50% to −250%. That’s not impossible – present values can converge with negative discount rates (Kocherlakota 1990) – but it does not seem reasonable. People prefer earlier consumption, not later consumption. Interest rate variation and the conditional mean of the discount factor

Again, however, maybe we’re being too doctrinaire. What evidence is there against γ = 50 − 250 with corresponding δ = −0.5 to −2.5?

Real interest rates are not only low on average, they are also relatively stable over time and across countries. γ = 50 in equation (21.343) means that a country or a boom time with consumption growth 1 percentage point higher than normal must have real interest rates 50 percentage points higher than normal, and consumption 1 percentage point lower than normal should be accompanied by real interest rates of 50 percentage points lower than normal– you pay them 48% to keep your money. We don’t see anything like this. γ = 50 to 250 in a time-separable utility function implies that consumers are essentially unwilling to substitute (expected) consumption over time, so huge interest rate variation must force them to make the small variations in consumption growth that we do see. This level of aversion to intertemporal substitution is too large. For example, think about what interest

416

S ECTION 21.1

E QUITY PREMIUM PUZZLES

rate you need to convince someone to skip a vacation. Take a family with $50,000 per year consumption, and which spends $2,500 (5%) on an annual vacation. If interest rates are good enough, though, the family can be persuaded to skip this year’s vacation and go on a much more lavish vacation next year. The required interest rate is ($52, 500/$47, 500)γ − 1. For γ = 250 that is an interest rate of 3 × 1011 ! For γ = 50, we still need an interest rate of 14, 800%. I think most of us would give in and defer the vacation for somewhat lower interest rates! A reasonable willingness to substitute intertemporally is central to most macroeconomic models that try to capture output, investment, consumption, etc. dynamics. As always, we can express the observation as a desired characteristic of the discount factor. Though mt+1 must vary a lot, its conditional mean Et (mt+1 ) = 1/Rtf must not vary much. You can get variance in two ways – variance in the conditional mean and variance in the unexpected component; var(x) = var [Et (x)] + var [x − Et (x)]. The fact that interest rates are stable means that almost all of the 50% or more unconditional variance must come from the second term. The power functional form is really not an issue. To get past the equity premium and these related puzzles, we will have to introduce other arguments to the marginal utility function – some non-separability. One important key will be to introduce some non-separability that distinguishes intertemporal substitution from risk aversion. 21.1.2

Variations

Just raising the interest rate will not help, as all-stock portfolios have high Sharpe ratios too. Uninsured individual risk is not an obvious solution. Individual consumption is not volatile enough to satisfy the bounds, and is less correlated with stock returns than aggregate consumption. The average return in postwar data may overstate the true expected return; a target of 3-4% is not unreasonable. Is the interest rate “too low”?

A large literature has tried to explain the equity premium puzzle by introducing frictions that make treasury bills “money-like” and so argue that the short-term interest rate is artificially low. (Aiyagari and Gertler 1991 is an example). However, high Sharpe ratios are pervasive in financial markets. Portfolios long small stocks and short big stocks, or long value (high book/market) and short growth stocks, give Sharpe ratios of 0.5 or more as well. Individual shocks

Maybe we should abandon the representative agent assumption. Individual income shocks 417

C HAPTER 21

E QUITY PREMIUM PUZZLE AND CONSUMPTION - BASED MODELS

are not perfectly insured, so individual income and consumption is much more volatile than aggregate consumption. Furthermore, through most of the sample, only a small portion of the population held any stocks at all. This line of argument faces a steep uphill battle. The basic pricing equation applies to each consumer. Individual income growth may be more volatile than the aggregate, but it’s not credible that any individual’s consumption growth varies by 50% -250% per year! Keep in mind, this is nondurable and services consumption and the flow of services from durables, not durables purchases. Furthermore, individual consumption growth is likely to be less correlated with stock returns than is aggregate consumption growth, and the more volatile it is, the less correlated. As a simple example, write individual consumption equal to aggregate consumption plus an idiosyncratic shock, uncorrelated with economywide variables, ∆cit = ∆cat + εit .

Hence, ¢ ¡ cov(∆cit , rt ) = cov ∆cat + εit , rt = cov (∆cat , rt ) .

As we add more idiosyncratic variation, the correlation of consumption with the any aggregate such as stock returns declines in exact proportion so that the asset pricing implications are completely unaffected. Luck and a lower target

One nagging doubt is that a large part of the U.S. postwar average stock return may represent good luck rather than ex-ante expected return. First of all, the standard deviation of stock returns √ is so high that standard errors are surprisingly large. Using the standard√formula σ/ T , the standard error of average stock returns in 50 years of data is about 16/ 50 ≈ 2.3. This fact means that a two-standard error confidence interval for the expected return extends from about 3% to about 13%! This is a pervasive, simple, but√surprisingly under-appreciated problem in empirical asset pricing. In 20 years of data, 16/ 20 = 3.6 so we can barely say that an 8% average return is above zero. 5 year performance averages √ of something like a stock return are close to meaningless on a statistical basis, since 16/ 5 = 7. 2. (This is one reason that many funds are held to tracking error limits relative to a benchmark. You may be able to measure performance relative to a benchmark, even if your return and the benchmark are both very volatile. √ If σ(Ri − Rm ) is small, then σ(Ri − Rm )/ T can be small, even if σ(Ri ) and σ(Rm ) are large.) However, large standard errors can argue that the equity premium is really higher than the postwar return. Several other arguments suggest a bias – that a substantial part of the 8% average excess return of the last 50 years was good luck, and that the true equity premium is more like 3-4%. 418

S ECTION 21.1

E QUITY PREMIUM PUZZLES

Brown, Goetzmann and Ross (1995) suggest that the U.S. data suffer from selection bias. One of the reasons that I write this book in the U.S., and that the data has been collected from the U.S., is precisely because U.S. stock returns and growth have been so good for the last 50 - 100 years. One way to address this question is to look at other samples. Average returns were a lot lower in the U.S. before WWII. In Shiller’s (1989) annual data from 1871-1940, the S&P500 average excess return was only 4.1% However, Campbell (1999, table 1) looks across countries for which we have stock market data from 1970-1995, and finds the average equity premium practically the same as that for the U.S. in that period. The other countries averaged a 4.6% excess return while the U.S. had a 4.4% average excess return in that period. On the other hand, Campbell’s countries are Canada, Japan, Australia and Western Europe. These probably shared a lot of the U.S. “good luck” in the postwar period. There are lots of countries for which we don’t have data, and usually because returns were very low in those countries. As Brown, Goetzmann and Ross (1995) put it, “Looking back over the history of the London or the New York stock markets can be extraordinarily comforting to an investor – equities appear to have provided a substantial premium over bonds, and markets appear to have recovered nicely after huge crashes. ... Less comforting is the past history of other major markets: Russia, China, Germany and Japan. Each of these markets has had one or more major interruptions that prevent their inclusion in long term studies” [my emphasis]. Think of the things that didn’t happen in the last 50 years. We had no banking panics, and no depressions; no civil wars, no constitutional crises; we did not lose the cold war, no missiles were fired over Berlin, Cuba, Korea or Vietnam. If any of these things had happened, we might well have seen a calamitous decline in stock values, and I would not be writing about the equity premium puzzle. A view that stocks are subject to occasional and highly non-normal crashes – world wars, great depressions, etc. – makes sampling uncertainty even larger, and means that the average return from any sample that does not include a crash will be larger than the actual average return – the Peso Problem again (Reitz 1988). Fama and French (2000) notice that the price/dividend ratio is low at the beginning of the sample and high at the end. Much of that is luck–the dividend yield is stationary in the very long run, with slow-moving variation through good and bad times. We can understand their alternative calculation most easily using the return linearization, rt+1 = ∆dt+1 + (dt − pt ) − ρ(dt+1 − pt+1 ).

Then, imposing the view that the dividend price ratio is stationary, we can estimate the average return as E (rt+1 ) = E (∆dt+1 ) + (1 − ρ)E(dt − pt ).

The right hand expression gives an estimate of the unconditional average return on stocks equal to 3.4%. This differes from the sample average return of 9% because, the d/p ratio 419

C HAPTER 21

E QUITY PREMIUM PUZZLE AND CONSUMPTION - BASED MODELS

declined dramatically in the postwar sample. Here is the fundamental issue: Was it clear to people in 1947 (or 1871, or whenever one starts the sample) and throughout the period that the average return on stocks would be 8% greater than that of bonds, subject only to the 16% year to year variation? Given that knowledge, would investors have changed their portfolios, or would they have stayed pat, patiently explaining that these average returns are earned in exchange for risk that they are not prepared to take? If people expected these mean returns, then we face a tremendous challenge of explaining why people did not buy more stocks. This is the basic assumption and challenge of the equity premium puzzle. But phrased this way, the answer is not so clear. I don’t think it was obvious in 1947 that the United States would not slip back into depression, or another world war, but would instead experience a half century of economic growth and stock returns never before seen in human history. 8% seems like an extremely – maybe even irrationally – exuberant expectation for stock returns as of 1947, or 1871. (You can ask the same question, by the way, about value effects, market timing, or other puzzles we try to explain. Only if you can reasonably believe that people understood the average returns and shied away because of the risks does it make sense to explain the puzzles by risk rather than luck. Only in that case with the return premia continue anyway!) This consideration mitigates, but cannot totally solve the equity premium puzzle. Even a 3% equity premium is tough to understand with 1% consumption volatility. If the premium is 3%, the Sharpe ratio is 3/16 ≈ 0.2, so we still need risk aversion of 20, and 100 if we include correlation. 20-100 is a lot better than 50-250, but is still quite a challenge. 21.1.3

Predictability and the equity premium

The Sharpe ratio varies over time. This means that discount factor volatility must vary over time. Since consumption volatility does not seem to vary over time, this suggests that risk aversion must vary over time – a conditional equity premium puzzle. Conventional portfolio calculations suggest that people are not terribly risk averse. These calculations implicitly assume that consumption moves proportionally to wealth, and inherits the large wealth volatility. If stock returns mean-revert, E(Re )/σ(Re ) and hence σ(m)/E(m) rises faster than the square root of the horizon. Consumption growth is roughly i.i.d., so σ(∆c) rises about with the square root of horizon. Thus, mean-reversion means that the equity premium puzzle is even worse for long-horizon investors and long-horizon returns. We have traced the implications of the unconditional Sharpe ratio, and of low and relatively constant interest rates. The predictability of stock returns also has important implications for discount factors. 420

S ECTION 21.1

E QUITY PREMIUM PUZZLES

Heteroskedasticity in the discount factor–conditional equity premium puzzle

The Hansen-Jagannathan bound applies conditionally of course, ¡ e ¢ Et Rt+1 1 σ t (mt+1 ) . e ) = − ρ (Re , m σt (Rt+1 ) E t+1 t (mt+1 ) t t+1 Mean returns are predictable, and the standard deviation of returns varies over time. So far, however, the two moments are forecasted by different sets of variables and at different horizons – d/p, term premium, etc. forecast the mean at long horizons; past squared returns and implied volatility forecast the variance at shorter horizons – and these variables move at different times. Hence, it seems that the conditional Sharpe ratio on the left hand side moves over time. (Glosten, Jagannathan and Runkle 1993, French Schwert and Stambaugh 1987, Yan 2000 find some co-movements in conditional mean and variance, but do not find that all movement in one moment is matched by movement in the other.) On the right hand side, the conditional mean discount factor equals the risk free rate and so must be relatively stable over time. Time-varying conditional correlations are a possibility, but hard to interpret. Thus, the predictability of returns strongly suggests that the discount factor must be conditionally heteroskedastic – σt (mt+1 ) must vary through time. Certainly the discount factors on the volatility bound, or the mimicking portfolios for discount factors, both of which have ρ = 1, must have time-varying volatility. In the standard time-separable model, σt (mt+1 ) = γ t σ t (∆ct+1 ). Thus, we need either time-varying consumption risk or time-varying curvature; loosely speaking a time-varying risk aversion. The data don’t show much evidence of conditional heteroskedasticity in consumption growth, leading one to favor a time-varying risk aversion. However, this is a case in which high risk aversion helps: if γ is sufficiently high, a small and perhaps statistically hard to measure amount of consumption heteroskedasticity can generate a lot of discount factor heteroskedasticity. (Kandel and Stambaugh 1990 follow this approach to explain predictability.) Capm, portfolios and consumption

The equity premium puzzle is centrally about the smoothness of consumption. This is why it was not noticed as a major puzzle in the early development of financial theory. In turn, the smoothness of consumption is centrally related to the predictability of returns. In standard portfolio analyses, there is no puzzle that people with normal levels of risk aversion do not want to hold far more stocks. From the usual first order condition and with Λ = VW (W ) we can also write the Hansen-Jagannathan bound in terms of wealth, analogously to (21.342), ¯ ¯ ¯E(r) − rf ¯ −W VW W ≤ σ (∆w) (344) σ(r) VW The quantity −W WW W /VW is in fact the measure of risk aversion corresponding to most survey and introspection evidence, since it represents aversion to bets on wealth rather than 421

C HAPTER 21

E QUITY PREMIUM PUZZLE AND CONSUMPTION - BASED MODELS

to bets on consumption. (They can be the same for power utility, but not in general.) For an investor who holds the market, σ (∆w) is the standard deviation of the stock return, about 16%. With a market Sharpe ratio of 0.5, we find the lower bound on risk aversion, 0.5 −W VW W ≈ 3. = VW 0.16

Furthermore, the correlation between wealth and the stock market is one in this calculation, so no correlation puzzle crops up to raise the required risk aversion. This is the heart of the oft-cited Friend and Blume (1975) calculation of risk aversion, one source of the idea that 3-5 is about the right level of risk aversion rather than 50 or 250. The Achilles heel is the hidden simplifying assumption that returns are independent over time, and the investor has no other source of income, so no variables other than wealth show up in its marginal value VW . In such an i.i.d. world, consumption moves one-for-one with wealth, and σ (∆c) = σ (∆w). If your wealth doubles and nothing else has changed, you double consumption. This calculation thus hides a consumption-based “model,” and the model has the drastically counterfactual implication that consumption growth has a 16% standard deviation! All this calculation has done is say that “in a model in which consumption has a 16% volatility like stock returns, we don’t need high risk aversion to explain the equity premium.” Hence the central point – the equity premium is about consumption smoothness. Just looking at wealth and portfolios, you do not notice anything unusual. In the same way, retreating to the CAPM or factor models doesn’t solve the puzzle either. The CAPM is a specialization of the consumption-based model, not alternatives, and thus hide an equity premium puzzle. For example, I derived the CAPM above as a consequence of log utility. With log utility, you have to believe that properly measured consumption growth has a 50% per year standard deviation! That testable implication is right there in the model, though often ignored. Most implementations of the CAPM take the market premium as given (ignoring the link to consumption in the model’s derivation) and estimate the market premium as a free parameter. The equity premium puzzle asks whether the market premium itself makes any sense. The long-run equity premium puzzle

The fact that annual consumption is much smoother than wealth is an important piece of information. In the long-run, consumption must move one-for-one with wealth, so consumption and wealth volatility must be the same. Therefore, we know that the world is very far from i.i.d., so predictability will be an important issue in understanding risk premia. Predictability can imply mean reversion and Sharpe ratios that rise faster than the square root of horizon. Thus, E(Ret→t+k ) σ(mt→t+k ) ¢≤ ¡ e ≈ γσ(∆ct→t+k ). E(mt→t+k ) σ Rt→t+k

422

S ECTION 21.2

N EW MODELS

If stocks do mean-revert, then discount factor volatility must increase faster than the square root of the horizon. Consumption growth is close to i.i.d.,so the volatility of consumption growth only increases with the square root of horizon. Thus mean-reversion implies that the equity premium puzzle is even worse at long investment horizons.

21.2

New models

We want to end up with a model that explains a high market Sharpe ratio, and the high level and volatility of stock returns, with low and relatively constant interest rates, roughly i.i.d. consumption growth with small volatility, and that explains the predictability of excess returns – the fact that high prices today correspond to low excess returns in the future. Eventually, we would like the model to explain the predictability of bond and foreign exchange returns as well, the time-varying volatility of stock returns and the cross-sectional variation of expected returns, and it would be nice if in addition to fitting all of the facts, people in the models did not display unusually high aversion to wealth bets. I start with a general outline of the features that most models that address these puzzles share. Then, I focus on two models, the Campbell-Cochrane (1999) habit persistence model and the Constantinides and Duffie (1996) model with uninsured idiosyncratic risks. The mechanisms we uncover in these models apply to a large class. The Campbell-Cochrane model is a representative from the literature that attacks the equity premium by modifying the representative agent’s preferences. The Constantinides and Duffie model is a representative of the literature that attacks the equity premium by modeling uninsured idiosyncratic risks, market frictions, and limited participation. 21.2.1

Outlines of new models

Additional state variables are the natural route to solving the empirical puzzles. Investors must not be particularly scared of the wealth or consumption effects of holding stocks, but of the fact that stocks do badly at particular times, or in particular states of nature. Broadly speaking, most solutions introduce something like a “recession” state variable. This fact makes stocks different, and more feared, than pure wealth bets, whose risk is unrelated to the state of the economy. In the ICAPM way of looking at things, we get models of this sort by specifying things so there is an additional recession state variables z in the value function V (W, z). Then, expected returns are

−W VW W E(r) − r = cov VW f

µ

423

¶ zVW z dW ,r + cov(z, r). W VW

(345)

C HAPTER 21

E QUITY PREMIUM PUZZLE AND CONSUMPTION - BASED MODELS

In a utility framework, we add other arguments to the utility function u(C, z), so −CuCC E(r) − r = cov uC f

µ

¶ zuCz dC ,r + cov(z, r). C uC

(346)

The extra utility function arguments must enter non-separably. If u(C, z) = f (C) + g(z), then uCz = 0. All utility function modifications are of this sort – they add extra goods like leisure, nonseparability over time in the form of habit persistence, or nonseparability across states of nature so that consumption if it rains affects marginal utility if it shines. The lesson of the equity premium literature is that the second term must account for essentially all of the market premium. Since the cross-sectional work surveyed in Chapter 20 seemed to point to something like a recession factor as the primary determinant of cross-sectional variation in expected returns, a gratifying unity seems close at hand – and a fundamental revision of the CAPM-i.i.d. view of the source of risk prices. The predictability of returns – emphasized by the dramatic contrast between consumption and wealth volatility at short horizons – suggests a natural source of state variables. Unfortunately, the sign is wrong. The fact that stocks go up when their expected subsequent returns are low means that stocks, like bonds, are good hedges for shocks to their own opportunity sets. Therefore, adding the effects of predictability typically lowers expected returns. (The “typically” in this sentence is important. The sign of this effect – the sign of zVW z – does depend on the utility function and environment. For example, there is no risk premium for log utility.) Thus, we need an additional state variable, and one strong enough to not only explain the equity premium, given that the first terms in (21.345) and (21.346) are not up to the job, but one stronger still to overcome the effects of predictability. Recessions are times of low prices and high expected returns. We want a model in which recessions are bad times, so that investors fear bad stock returns in recessions. But high expected returns are good times for a pure Merton investor. Thus, the other state variable(s) that describe a recession – high risk aversion, low labor income, high labor income uncertainty, liquidity, etc. – must overcome the “good times” of high expected returns and indicate that times really are bad after all. 21.2.2

Habits

A natural explanation for the predictability of returns from price/dividend ratios is that people get less risk averse as consumption and wealth increase in a boom, and more risk averse as consumption and wealth decrease in a recession. We can’t tie risk aversion to the level of consumption and wealth, since that increases over time while equity premia have not declined. Thus, to pursue this idea, we must specify a model in which risk aversion depends on the level of consumption or wealth relative to some “trend” or the recent past. Following this idea, Campbell and Cochrane (1999) specify that people slowly develop habits for higher or lower consumption. Thus, the “habits” form the “trend” in consumption. 424

S ECTION 21.2

N EW MODELS

The idea is not implausible. Anyone who has had a large pizza dinner or smoked a cigarette knows that what you consumed yesterday can have an impact on how you feel about more consumption today. Might a similar mechanism apply for consumption in general and at a longer time horizon? Perhaps we get used to an accustomed standard of living, so a fall in consumption hurts after a few years of good times, even though the same level of consumption might have seemed very pleasant if it arrived after years of bad times. This thought can at least explain the perception that recessions are awful events, even though a recession year may be just the second or third best year in human history rather than the absolute best. Law, custom and social insurance also insure against falls in consumption as much as low levels of consumption. The Model

We model an endowment economy with i.i.d. consumption growth. ∆ct+1 = g + vt+1 ; vt+1 ∼ i.i.d. N (0, σ 2 ).

We replace the utility function u(C) with u(C − X) where X denotes the level of habits. E

∞ X t=0

δt

(Ct − Xt )1−γ − 1 . 1−γ

Habits should move slowly in response to consumption, something like xt ≈ λ

∞ X

φj ct−j

(347)

j=0

or, equivalently xt = φxt−1 + λct .

(348)

(Small letters denote the logs of large letters throughout this section, ct = ln Ct , etc.) Rather than letting habit itself follow an AR(1) we let the “surplus consumption ratio” of consumption to habit follow an AR(1): St =

Ct − Xt Ct

st+1 = (1 − φ)¯ s + φst + λ (st ) (ct+1 − ct − g)

(349)

Since s contains c and x, this equation also specifies how x responds to c, and it is locally the same as (21.347). We also allow consumption to affect habit differently in different states by 425

C HAPTER 21

E QUITY PREMIUM PUZZLE AND CONSUMPTION - BASED MODELS

specifying an square root type process rather than a simple AR(1), 1p 1 − 2 (st − s¯) − 1 S¯r γ S¯ = σ 1−φ

λ(st ) =

(21.350) (21.351)

The extra complication of (21.349) rather than (21.347) means consumption is always above habit, since S = es > 0. Other habit models can give consumption below habit which leads to infinite or imaginary marginal utility. St becomes the single state variable in this economy. Time-varying expected returns, price-dividend ratios, etc. are all functions of this state variable.

Marginal utility is uc (Ct , Xt ) = (Ct − Xt )−γ = St−γ Ct−γ .

The model assumes an external habit – each individual’s habit is determined by everyone else’s consumption, as in Abel’s (1990) “keeping up with the Joneses” specification. This is mostly a technical simplification, since it allows us to ignore terms by which current consumption affect future habits; the opposite specification gives very similar results (see problem 2). With marginal utility, we now have a discount factor. µ ¶−γ St+1 Ct+1 uc (Ct+1 , Xt+1 ) =δ Mt+1 ≡ δ . uc (Ct , Xt ) St Ct Since we have a stochastic process for S and C , and each is lognormal, we can evaluate the conditional mean of the discount factor to evaluate the riskfree rate 1 rtf = − ln Et (Mt+1 ) = − ln(δ) + γg − γ(1 − φ). 2

(352)

We gave up on analytic solutions and evaluated the price-dividend ratio as a function of the state variable by iteration on a grid: ¶¸ · µ Pt Ct+1 Pt+1 (st ) = Et Mt+1 (st+1 ) 1+ Ct Ct Ct+1 With price-dividend ratios, we can calculate returns, expected returns, etc. How does it work – equity premium and predictability

We choose parameters, simulate 100,000 artificial data points, and report standard statistics and tests in artificial data. The parameters g = 1.89, σ = 1.50, rf = 0.94 match their values in postwar data. The parameter φ = 0.87 matches the autocorrelation of the pricedividend ratio and the choice γ = 2.00 matches the postwar Sharpe ratio. δ = 0.89, S¯ = 0.57 follow from the model. 426

S ECTION 21.2

N EW MODELS

Table 2cc presents means and standard deviations predicted by the model. The model replicates the postwar Sharpe ratio, with a constant 0.94% risk free rate and a reasonable subjective discount factor δ < 1. Of course, we picked the parameters to do this, but given the above equity premium discussion it’s already an achievement that we are able to pick any parameters to hit these moments. Some models can replicate the Sharpe ratio, but do not replicate the level of expected returns and return volatility. E = 1% and σ = 2% will give the right Sharpe ratio, but this model predicts the right levels as well. The model also gets the level of the price-dividend ratio about right. Table 2cc. Means and standard deviations of simulated and historical data. Statistic E(R − Rf )/σ(R − Rf ) E(r − rf ) σ(r − rf ) exp[E(p − d)] σ(p − d)

Consumption claim 0.50 6.64 15.2 18.3 0.27

Dividend claim 6.52 20.0 18.7 0.29

Postwar data 0.50 6.69 15.7 24.7 0.26

The model is simulated at a monthly frequency; statistics are calculated from artificial time-averaged data at an annual frequency. Asterisks (*) denote statistics that model parameters were chosen to replicate. All returns are annual percentages. Table 5cc shows how the artificial data match the predictability of returns from pricedividend ratios. The paper goes on, and shows how the model matches the volatility test result that almost all return variation is due to variation in expected excess returns, the “leverage effect” of higher volatility after a big price decline, and several related phenomena. Table 5cc. Long-horizon return regressions Horizon (Years) 1 2 3 5 7

Cons. claim 10×coef. R2 -2.0 0.13 -3.7 0.23 -5.1 0.32 -7.5 0.46 -9.4 0.55

Postwar data 10×coef. R2 -2.6 0.18 -4.3 0.27 -5.4 0.37 -9.0 0.55 -12.1 0.65

How does it work?

How does this model get around all the equity premium - riskfree rate difficulties described above, and explain predictability as well? When a consumer has a habit, local curvature depends on how far consumption is above 427

C HAPTER 21

E QUITY PREMIUM PUZZLE AND CONSUMPTION - BASED MODELS

the habit, as well as the power γ, ηt ≡

γ −Ct ucc (Ct − Xt ) = . uc (Ct − Xt ) St

As consumption falls toward habit, people become much less willing to tolerate further falls in consumption; they become very risk averse. Thus a low power coefficient γ can still mean a high, and time-varying curvature. Recall our fundamental equation for the Sharpe ratio, Et (r) − rtf = η t σt (∆c)corrt (∆c, r). σ t (r)

High curvature ηt means that the model can explain the equity premium, and curvature ηt that varies over time as consumption rises in booms and falls toward habit in recessions means that the model can explain a time-varying and countercyclical (high in recessions, low in booms) Sharpe ratio, despite constant consumption volatility σt (∆c) and correlation corrt (∆c, r). So far so good, but didn’t we just learn that raising curvature implies high and timevarying interest rates? This model gets around interest rate problems with precautionary saving. Suppose we are in a bad time, in which consumption is low relative to habit. People want to borrow against future, higher, consumption, and this force should drive up interest rates. (In fact, many habit models have very volatile interest rates.) However, people are also much more risk averse when consumption is low. This consideration induces them to save more, in order to build up assets against the event that tomorrow might be even worse. This desire to save drives down interest rates. Our λ(s) specification makes these two forces exactly offset, leading to constant real rates. The precautionary saving motive also makes the model more plausibly consistent with variation in consumption growth across time and countries. Adding (21.351) to (21.352), we can write 1 ³ γ ´2 2 rf = ρ + γg − σ 2 S¯

The power coefficient γ = 2 controls the relation between consumption growth and interest rates, while the curvature coefficient γ/St controls the risk premium. Thus this habit model allows high “risk aversion” with low “aversion to intertemporal substitution,” and it is consistent with the consumption and interest rate data. As advertised, this model explains the equity premium and predictability by fundamentally changing the story for why consumers are afraid of holding stocks. The k− period stochastic discount factor is µ ¶−γ St+k Ct+k k Mt→t+k = δ . St Ct

covariances with S shocks now drive average returns as well as covariances with C shocks. S = (C − X)/C is a recession indicator – it is low after several quarters of consumption 428

S ECTION 21.2

N EW MODELS

declines and high in booms. While (Ct+k /Ct )−γ and (St+k /St )−γ enter symmetrically in the formula, the volatility of (Ct+k /Ct )−γ with γ = 2 is so low that it accounts for essentially no risk premia. Therefore, it must be true, and it is, that variation in (St+k /St )−γ is much larger, and accounts for nearly all risk premia. In the Merton language of (21.345) and (21.346), variation across assets in expected returns is driven by variation across assets in covariances with recessions far more than by variation across assets in covariances with consumption growth. At short horizons, shocks to St+1 and Ct+1 move together, so the distinction between a recession state variable and consumption risk is minor; one can regard S as an amplification mechanism for consumption risks in marginal utility. dS/∂C ≈ 50, so this amplification generates the required volatility of the discount factor. At long horizons, however, St+k becomes less and less conditionally correlated with Ct+k . St+k depends on Ct+k relative to its recent past, but the overall level of consumption may be high or low. Therefore, investors fear stocks because they do badly in occasional serious recessions, times of recent belt-tightening. These risks are at the long run unrelated to the risks of long-run average consumption growth. As another way to digest how this model works, we can substitute in the s process from (21.349) and write the marginal rate of substitution as Mt+1

=

ln Mt+1

= = =

ln Mt+1

µ

¶−γ St+1 Ct+1 δ St Ct ln δ − γ (st+1 − st ) − γ(ct+1 − ct ) {ln δ − γ(1 − φ)¯ s} + {γ (1 − φ) st + γgλ (st )} − γ [λ (st ) + 1] (ct+1 − ct ) a + b(st ) + d(st )(ct+1 − ct )

Up to the question of logs vs. levels, this is a “scaled factor model” of the form we studied in Chapter 8. It still is a consumption-based model, but the sensitivity of the discount factor to consumption changes over time. The long-run equity premium is even more of a puzzle. Most recession state variables, such as recessions, labor, and instruments for time-varying expected returns (“shifts in the investment opportunity set”) are stationary. Hence, the standard deviation of their growth rates eventually stops growing with horizon. At a long enough horizon, the standard deviation of the discount factor is dominated by the standard deviation of the consumption growth term, and we return to the equity premium puzzle at a long enough run. Since this model produces predictability of the right sign, it produces a long run equity premium puzzle. How it manages this feat with a stationary state variable St is subtle (and we didn’t notice it until the penultimate draft!) The answer is that while St is stationary, St−γ −γ is not. St has a fat tail approaching zero so the conditional variance of St+k grows without bound. While the distinction between stationary S and nonstationary S −γ seems initially minor, it 429

C HAPTER 21

E QUITY PREMIUM PUZZLE AND CONSUMPTION - BASED MODELS

is in fact central. Any model that wishes to explain the equity premium at long and short runs by means of an additional, stationary state variable must find some similar transformation so that the volatility of the stochastic discount factor remains high at long horizons. This model does have high risk aversion. The utility curvature and value function curvature are both high. Many authors require that a “solution” of the equity premium puzzle display low risk aversion. This is a laudable goal, and no current model has attained it. No current model generates the equity premium with a low and relatively constant interest rate, low risk aversion, and the right pattern of predictability – high prices forecast low returns, not high returns, and consumption is roughly a random walk. Constantinides (1990) and Boldrin, Christiano and Fisher (1997) are habit models with a large equity premium and low risk aversion, but they don’t get the pattern of predictability right. Boldrin, Christiano and Fisher have highly variable interest rates to keep consumption from being predictable. Constantinides (1990) has a constant interest rate, but consumption growth that is serially correlated, so consumption rises to meet i.i.d. wealth growth. The long-run equity premium is solved with counterfactually high long-run consumption volatility. To get a high equity premium with low risk aversion, we need to find some crucial characteristic that separates stock returns from wealth bets. This is a difficult task. After all, what are stocks if not a bet? The answer must be some additional state variable. Stocks must pay off badly in particularly unfortunate states of the world. Again, the trouble with predictability is that stocks pay off well in particularly bad states of the world – states with low future returns. This makes stocks even more desirable, requiring even higher risk aversion to explain the equity premium. The alternative, not yet found, is to find some measure of the state of the world that is particularly bad when stocks pay off badly, enough to explain not only the standard equity premium, but the long run equity premium resulting from the fact that stocks are less risky at longer horizons. I write this not to say that such a model is impossible. The point is to show the hurdle that must be overcome, in the hope that someone will overcome it. 21.2.3

Heterogeneous agents and idiosyncratic risks

A long, increasing, and important literature in the equity premium attacks the problem instead with relatively standard preferences, but instead adds uninsured idiosyncratic risk. As with the preference literature, this literature is interesting beyond the equity premium. We are learning a lot about who holds stocks and why, what risks they face. We are challenged to think of new assets and creative ways of using existing assets to share risks better. Constantinides and Duffie (1996) provide a very clever and simple model in which idiosyncratic risk can be tailored to generate any pattern of aggregate consumption and asset prices. It can generate the equity premium, predictability, relatively constant interest rates, smooth and unpredictable aggregate consumption growth and so forth. Furthermore, it requires no transactions costs, borrowing constraints or other frictions, and the individual con430

S ECTION 21.2

N EW MODELS

sumers can have any nonzero value of risk aversion. Of course, we still have to evaluate whether the idiosyncratic risk process we construct to explain asset pricing phenomena are reasonable and consistent with microeconomic data. A simple version of the model

I start with a very simplified version of the Constantinides-Duffie model. Each consumer i has power utility, U =E

X

1−γ e−δt Cit

t

Individual consumption growth Cit+1 is determined by an independent, idiosyncratic normal (0,1) shock ηit , ln

µ

Cit+1 Ci,t

¶

1 2 = ηit+1 yt+1 − yt+1 2

(353)

where yt+1 is, by construction since it multiplies the shock ηit , the cross-sectional standard deviation of consumption growth. yt+1 is dated t + 1 since it is the cross-sectional standard deviation given aggregates at t+1. The aggregates are determined first, and then the shocks ηit are handed out. Now, yt+1 is specified so that people suffer a high cross-sectional variance of consumption growth on dates of a low market return Rt+1 , ·

yt+1 = σ ln

µ

¶¯ ¸ s p Cit+1 ¯¯ 2 Rt+1 = δ − ln Rt+1 . ¯ Cit γ(γ + 1)

(354)

Given this structure, the individual is exactly happy to consume {Cit } without further trading in the stock. (We can call Cit income Iit , and prove the optimal decision rule is to consume income Cit = Iit .) His first-order condition for an optimal consumption-portfolio decision " # µ ¶−γ Cit+1 −δ 1 = Et e Rt+1 Cit holds, exactly. To prove this assertion, just substitute in for Cit+1 /Cit and take the expectation: · ¸ 1 2 1 = Et exp −δ − γη it+1 yt+1 + γyt+1 + ln Rt+1 2

Since η is independent of everything else, we can use E [f (ηy)] = E [E(f (ηy|y)] . Now, 431

C HAPTER 21

E QUITY PREMIUM PUZZLE AND CONSUMPTION - BASED MODELS

with η normal (0,1),

Therefore, we have

· ¸ ¤ ¢ ¡ £ 1 2 2 E exp −γηit+1 yt+1 | yt+1 = exp γ yt+1 . 2 · ¸ 1 2 2 1 2 1 = Et exp −δ + γ yt+1 + γyt+1 + ln Rt+1 . 2 2

Substituting in from (21.354), µ · ¸ ¶ 2 1 1 = Et exp −δ + γ(γ + 1) (δ − ln Rt+1 ) + ln Rt+1 2 γ(γ + 1) 1 = Et 1!

The general model

In the general model, Constantinides and Duffie define s r Ct+1 2 yt+1 = ln mt+1 + δ + γ ln γ(γ + 1) Ct

(355)

where mt is a strictly positive discount factor that prices all assets under consideration, pt = Et [mt+1 xt+1 ] for all xt+1 ∈ X.

(356)

By starting with a discount factor that can price a large collection of assets, where I used the −1 to price the single return Rt+1 in (21.354), idiosyncratic risk can be discount factor Rt+1 constructed to price exactly a large collection of assets. We can exactly match the Sharpe ratio, return forecastability, and other features of the data. Then, they let ln

µ

vit+1 vit

¶

1 2 = ηit+1 yt+1 − yt+1 2

Cit+1 = vit+1 Ct+1 . yt+1 is still the conditional standard deviation of consumption growth, given aggregates – returns and aggregate consumption. This variation allows uncertainty in aggregate consumption. We can tailor the idiosyncratic risk to and consumption-interest rate facts as well.

432

S ECTION 21.2

N EW MODELS

Following exactly the same argument as before, we can now show that " # µ ¶−γ Cit+1 −δ 1 = Et e Rt+1 Cit for all the assets priced by m. A technical assumption

Astute readers will notice the possibility that the square root term in (21.354) and (21.355) might be negative. Constantinides and Duffie rule out this possibility by assuming that the discount factor m satisfies ln mt+1 ≥ δ + γ ln

Ct+1 Ct

(357)

in every state of nature, so that the square root term is positive. We ³ construct ´γ i such discount factors by picking parameters a, b in mt+1 = h can sometimes 0 δ Ct+1 max a + b xt+1 , e to satisfy (21.356). However, neither this construction nor a Ct discount factor satisfying (21.357) is guaranteed to exist for any set of assets. The restriction (21.357) is a tighter form of the familiar restriction that mt+1 ≥ 0 that is equivalent to the absence of arbitrage in the assets under consideration. Ledoit and Bernardo (1997) show that the restriction m > a is equivalent to restrictions on the maximum gain/loss ratio available from the set of assets under consideration. Thus, the theorem really does not apply to any set of arbitrage-free payoffs. The example m = 1/R is a positive discount factor that prices a single asset return 1 = E(R−1 R), but does not necessarily satisfy restriction (21.357). For high R, we can have very negative ln 1/R. This example only works if the distribution of R is limited to R ≤ eδ . How the model works

As the Campbell-Cochrane model is blatantly (and proudly) reverse-engineered to surmount (and here, to illustrate) the known pitfalls of representative consumer models, the Constantinides-Duffie model is reverse engineered to surmount the known pitfalls of idiosyncratic risk models. Idiosyncratic risk stories face two severe challenges, as explained in section 1.2. First, the basic pricing equation applies to each individual. If we are to have low risk aversion and power utility, the required huge volatility of consumption is implausible for any individual. Second, if you add idiosyncratic risk uncorrelated with asset returns, it has no effect on pricing implications. Constantinides and Duffie’s central contribution is to very cleverly solve the second problem. In idiosyncratic risk models, we cannot specify individual consumption directly as we do in representative agent endowment economies, and go straight to finding prices. The endowment economy structure says that aggregate consumption is fixed, and prices have to adjust so that consumers are happy consuming the given aggregate consumption stream. 433

C HAPTER 21

E QUITY PREMIUM PUZZLE AND CONSUMPTION - BASED MODELS

However, individuals can always trade consumption with each other. The whole point of assets is that one individual can sell another some consumption, in exchange for the promise of some consumption in return in the next period. We have to give individuals idiosyncratic income shocks, and then either check that they do not want to trade away the idiosyncratic shock, or find the equilibrium consumption after they do so. Early idiosyncratic risk papers found quickly how clever the consumers could be in getting rid of the idiosyncratic risks by trading the existing set of assets. Telmer (1993) and Lucas (1994) found that if you give people transitory but uninsured income shocks, they respond borrowing and lending or by building up a stock of savings. As in the classic permanent income model, consumption only responds by the interest rate times the change in permanent income, and at low enough interest rates, not at all. “Self-insurance through storage” removes the extra income volatility and we are back to smooth individual consumption and an equity premium puzzle. Constantinides and Duffie get around this problem by making the idiosyncratic shocks permanent. The normal ηit shocks determine consumption growth. In an evaluation in microeconomic data, this makes us look for sources of permanent shocks. This, at a deeper level, is why idiosyncratic consumption shocks have to be uncorrelated with the market. We can give individuals idiosyncratic income shocks that are correlated with the market. Say, agent A gets more income when the market is high, and agent B gets more income when it is low. But then A will short the market, B will go long, and they will trade away any component of the shock that is correlated with the returns on available assets. I argued above that this effect made idiosyncratic shocks hopeless as candidates to explain the equity premium puzzle. Shocks uncorrelated with asset returns have no effect on asset pricing, and shocks correlated with asset returns are quickly traded away. The only way out is to exploit the nonlinearity of marginal utility. We can give people income shocks that are uncorrelated with returns, so they can’t be traded away. Then we have a nonlinear marginal utility function turn these shocks into marginal utility shocks that are correlated with asset returns, and hence can affect pricing implications. This is why Constantinides and Duffie specify that the variance of idiosyncratic risk rises when the market declines. If marginal utility were linear, an increase in variance would have no effect on the average level of marginal utility. Therefore, Constantinides and Duffie specify power utility, and the interaction of nonlinear marginal utility and changing conditional variance produces an equity premium. As a simple calculation that shows the basic idea, start with individuals i with power utility so

0=E

"µ

i Ct+1 Cti

¶−γ

434

Ret+1

#

S ECTION 21.2

N EW MODELS

Now aggregate across people by summing over i, with EN = "

0 = E EN

Ãµ

i Ct+1 Cti

¶−γ !

1 N

#

PN

i=1

Ret+1 .

If the cross-sectional variation of consumption growth is lognormally distributed, ·µ ¶ ¸ 2 −γEN ∆cit+1 + γ2 σ 2N ∆cit+1 e 0=E e Rt+1 As you see, the economy displays more risk aversion than would a “representative agent” with aggregate consumption ∆cat+1 = EN ∆cit+1 . That risk aversion can also vary over time if σ N varies over time, and this variation can generate risk premia. Microeconomic evaluation and risk aversion

Like the Campbell-Cochrane model, this could be either a new view of stock market (and macroeconomic) risk, or just a clever existence proof for a heretofore troubling class of models. The first question is whether the microeconomic picture painted by this model is correct, or even plausible. Is idiosyncratic risk large enough? Does idiosyncratic risk really rise when the market falls, and enough to account for the equity premium? Are there enough permanent idiosyncratic shocks? Do people really shy away from stocks because of stock returns are low at times of high labor market risk? This model does not change the first puzzle. To get power utility consumers to shun stocks, they still must have tremendously volatile consumption growth or high risk aversion. The point of this model is to show how consumers can get stuck with high consumption volatility in equilibrium, already a difficult task. More seriously than volatility itself, consumption growth variance also represents the amount by which the distribution of individual consumption and income spreads out over time, since the shocks must be permanent and independent across people. The 50% or larger consumption growth volatility that we require to reconcile the Sharpe ratio with risk aversion of one means that the distribution of consumption (and income) must also spread out by 50% per year. The distribution of consumption does spread out, but not this much. For example, Deaton and Paxson (1994) report that the cross-sectional variance of log consumption within an age cohort rises from about 0.2 at age 20 to 0.6 at age 60. √ This 0.2 = estimate means that the cross sectional standard deviation of consumption rises from √ . 45 or 45% at age 20 to 0.6 = . 77 or 77% at age 60. (77% means that an individual one standard deviation better off than the mean consumes 77% more than the mean consumer.) We are back to about 1% per year. Finally, and most crucially, the cross-sectional uncertainty about individual income must not only be large, it must be higher when the market is lower. This risk-factor is after all the central element of Constantinides and Duffie’s explanation for the market premium. Figure 51 shows how the cross-sectional standard deviation of consumption growth varies with the 435

C HAPTER 21

E QUITY PREMIUM PUZZLE AND CONSUMPTION - BASED MODELS

market return and risk aversion in my simple version of Constantinides and Duffie’s model. If we insist on low (γ = 1 to 2) risk aversion, the cross-sectional standard deviation of consumption growth must be extremely sensitive to the level of the market return. Looking at the γ = 2 line for example, is it plausible that a year with 5% market return would show a 10% cross-sectional variation in consumption growth, while a mild 5% decline in the market is associated with a 25% cross-sectional variation?

Figure 51. Cross-sectional standard deviation of individual consumption growth as a function of the market return in q the simple q version of the Constantinides-Duffie model. 2 t ln R1t + δ + γ ln CCt−1 . Parameter values are The plot is the variable yt = γ(γ+1) ρ = 0.05, ln Ct /Ct−1 = 0.01.

All of these empirical problems are avoided if we allow high risk aversion rather than a large risk to drive the equity premium. The γ = 25 line in Figure 51 looks possible; a γ = 50 line would look even better. With high risk aversion we do not need to specify highly volatile individual consumption growth, spreading out of the income distribution, or dramatic sensitivity of the cross-sectional variance to the market return. As in any model, a high equity premium must come from a large risk, or from large risk aversion. Labor market risk correlated with the stock market does not seem large enough to account for the equity premium without high risk aversion.

436

S ECTION 21.3

B IBLIOGRAPHY

The larger set of asset pricing facts has not yet been studied in this model. It is clearly able to generate return predictability, but that requires a pattern of variation in idiosyncratic risk that remains to be characterized and evaluated. It can generate cross-sectional patterns such as value premia if value stocks decline at times of higher cross-sectional volatility; that too remains to be studied. Summary

In the end, the Constantinides-Duffie model and the Campbell-Cochrane model are quite similar in spirit. First, both models make a similar, fundamental change in the description of stock market risk. Consumers do not fear much the loss of wealth of a bad market return per se. They fear that loss of wealth because it tends to come in recessions, in one case defined as times of heightened labor market risk, and in the other case defined as a fall of consumption relative to its recent past. This recession state-variable or risk-factor drives most variation in expected returns. Second, both models require high risk aversion. While Constantinides and Duffie’s proof shows that one can dream up a labor income process to rationalize the equity premium for any risk aversion coefficient, we see that even vaguely plausible characterizations of actual labor income uncertainty require high risk aversion to explain the historical equity premium. Third, both models provide long-sought demonstrations that it is possible to rationalize the equity premium in their respective class of models. This existence proof is particularly stunning in Constantinides and Duffie’s case. Many authors (myself included) had come to the conclusion that the effort to generate an equity premium from idiosyncratic risk was hopeless because any idiosyncratic risk that would affect asset prices would be traded away.

21.3

Bibliography

Shiller (1982) made the first calculation that showed either a large risk aversion coefficient or counterfactually large consumption variability was required to explain means and variances of asset returns. Mehra and Prescott (1985) labeled this fact the “equity premium puzzle.” However, they described these puzzles in the context of a two-state Markov model for consumption growth, identifying a stock as a claim to consumption and a risk free bond. Weil (1989) emphasized the interaction between equity premium and risk-free rate puzzles. Hansen and Jagannathan (1991) sparked the kind of calculations I report here in a simplified manner. Cochrane and Hansen (1992) derived many of the extra discount factor moment restrictions I surveyed here, calculating bounds in each case. Luttmer (1996), (1999) tackled the important extension to transactions costs. Kocherlakota (1996) is a nice summary of equity premium facts and models. Much of the material in this Chapter is adapted from a survey in Cochrane (1997). Campbell (1999) and (2000) are two excellent recent surveys. Ferson (1995) is a nice survey of consumption-based model variations as well as some of the beta pricing models discussed in the last chapter. The general picture of all solutions based on changing preferences is that they introduce 437

C HAPTER 21

E QUITY PREMIUM PUZZLE AND CONSUMPTION - BASED MODELS

non-separabilities. If the marginal utility of consumption depends on z as well as c, uc (c, z), then expected returns depend on covariances with z as well. In turn, this happens if we cannot write the utility function as u(c, z) = v(c) + w(z), the separable form. Habit persistence introduces a non-time-separable utility function, since P u(c, x) and x depends on past c–you can’t write a utility function with habits as a sum t vt (ct ), so one period’s consumption affects another period’s marginal utility. The Campbell-Cochrane model I presented here is a tip of an iceberg of habit research, including prominent contributions by Constantinides (1990), Ferson and Constantinides (1991), Heaton (1995), Abel (1990). Models can be nonseparable across goods as well. Leisure is the most natural extra variable to add to a utility function. It’s not clear a priori whether more leisure enhances the marginal utility of consumption (why bother buying a boat if you’re at the office all day and can’t use it) or vice versa (if you have to work all day, it’s more important to come home to a really nice big TV). However, we can let the data speak on this matter. Explicit versions of this approach have not been very successful to date. (Eichenbaum, Hansen and Singleton 1989). On the other hand, recent research has found that adding labor income as an extra adhoc “factor” can be useful in explaining the cross section of average stock returns, especially if it is scaled by a conditioning variable (Jagannathan and Wang 1996, Reyfman 1997, Lettau and Ludvigson 2000). The non-state separable utility functions following Epstein and Zin (1989) are a major omission of this presentation. The expectation E in the standard utility function sums over states of nature, e.g. U = prob(rain) × u(C if it rains) + prob(shine) × u(C if it shines).

“Separability” means one adds across states, so the marginal utility of consumption in one state is unaffected by what happens in another state. But perhaps the marginal utility of a little more consumption in the sunny state of the world is affected by the level of consumption in the rainy state of the world. Epstein and Zin and Hansen, Sargent and Tallarini (1997) propose recursive utility functions of the form £ ¤ Ut = Ct1−γ + βf Et f −1 (Ut+1 ) .

If f (x) = x this expression reduces to power utility. These utility functions are not stateseparable. As with habits, these utility functions distinguish risk aversion from intertemporal substitution–one coefficient can be set to capture the consumption-interest rate facts, and a completely separate coefficient can be set to capture the equity premium. So far, this style of model as in Epstein and Zin (1989), Weil (1989), Kandel and Stambaugh (1991), and Campbell (1996) does not generate time-varying risk aversion, but that modification should not be too difficult, and could lead to a model that works very much like the habit model I surveyed here. Habit persistence is the opposite of durability. If you buy a durable good yesterday, that lowers your marginal utility of an additional purchase today, while buying a habit-forming good raises your marginal utility of an additional purchase today. Thus the durability of goods 438

S ECTION 21.3

B IBLIOGRAPHY

should introduce a non-time-separability of the form u(ct + θxt ), xt = f(ct−1 , ct−2 , ...) rather than the habit persistence form u(ct − θxt ). Since goods are durable, and we have a lot of data on durables purchases, it would be good to include both durability and habit persistence in our models. (In fact, even “nondurables” contain items like clothing; the truly nondurable purchases are such a small fraction of total consumption that we rely on very little data.) One must be careful with the time horizon in such a specification. At a sufficiently small time horizon, all goods are durable. A pizza eaten at noon lowers marginal utility of more pizza at 12:05. Thus, our common continuous time, time separable assumption really cannot be taken literally. Hindy and Huang (1992) argue that consumption should be “locally substitutable” in continuous time models. Heaton (1993) found that at monthly horizons, consumption growth displays the negative autocorrelation suggestive of durability with constant interest rates, while at longer horizons consumption is nearly unforecastable after accounting for time-aggregation. There is also a production first order condition that must be solved, relating asset prices to marginal rates of transformation. The standard here is the q theory of investment, which is based on an adjustment cost. If the stock market is really high, you issue stock and make new investments. The trouble with this view is that f 0 (K) declines very slowly, so the observed price volatility implies huge investment volatility. The q theory adds adjustment costs to damp the investment volatility. The q theory has had as much trouble fitting the data as the consumption-based model. Cochrane (1991d) reports one success when you transform the data to returns – high stock returns are associated with high investment growth. The more recent investment literature has focused on specifying the adjustment cost problem with asymmetries and irreversibilities, for example Abel and Eberly (1996). There is an important literature that puts new utility functions together with production functions, to construct complete explicit economic models that replicate the asset pricing facts. Such efforts should also at least preserve if not enhance our ability to understand the broad range of dynamic microeconomic, macroeconomic, international and growth facts that the standard models were constructed around. Jermann (1998) tried putting habit persistence consumers in a model with a standard technology Y = θf (K, L) from real business cycle models. The easy opportunities for intertemporal transformation provided by that technology meant that the consumers used it to dramatically smooth consumption, destroying the prediction of a high equity premium. To generate the equity premium, Jermann added an adjustment cost technology, as the production-side literature had found necessary. This modification resulted in a high equity premium, but also large variation in riskfree rates. Boldrin, Christiano and Fisher (1997) also added habit-persistence preferences to real business cycle models with frictions in the allocation of resources to two sectors. They generate about 1/2 the historical Sharpe ratio. They find some quantity dynamics are improved over the standard model. However, they still predict highly volatile interest rates and persistent consumption growth. To avoid the implications of highly volatile interest rates, I suspect we will need representations of technology that allow easy transformation across time but not across states of 439

C HAPTER 21

E QUITY PREMIUM PUZZLE AND CONSUMPTION - BASED MODELS

nature, analogous to the need for easy intertemporal substitution but high risk aversion in preferences. Alternatively, the Campbell-Cochrane model above already produces the equity premium with constant interest rates, which can be interpreted as a linear production function f(K). Models with this kind of precautionary savings motive may not be as severely affected. Tallarini (1999) uses non-state separable preferences similar to those of Epstein and Zin in a general equilibrium model with production. He shows a beautiful observational equivalence result: A model with standard preferences and a model with non-state-separable preferences can predict the same path of quantity variables (output, investment, consumption, etc.) but differ dramatically on asset prices. This result offers one explanation of how the real business cycle and growth literature could go on for 25 years examining quantity data in detail and miss all the modifications to preferences that we seem to need to explain asset pricing data. It also means that asset price information is crucial to identifying preferences and calculating welfare costs of policy experiments. Finally, it offers hope that adding the deep modifications necessary to explain asset pricing phenomena will not demolish the success of standard models at describing the movements of quantities. The Constantinides and Duffie model has roots in a calculation by Mankiw (1986) that idiosyncratic risk could make the representative consumer seem more risk averse than the individuals. Work on evaluating the mechanisms in this model in microeconomic data is starting. Heaton and Lucas (1996) calibrate idiosyncratic risk from the PSID, but their model explains at best 1/2 of the sample average stock return, and less still if they allow a net supply of bonds with which people can smooth transitory shocks. More direct tests of these features in microeconomic consumption data are underway, for example Brav, Constantinides and Geczy (1999), Storesletten, Telmer and Yaron (1999). Kandel and Stambaugh (1986) present a model in which a small amount of time-varying consumption volatility and a high risk aversion coefficient generate the large time-varying discount factor volatility we need to generate returns predictability. Aiyagari and Gertler (1991), though aimed at the point that the equity premium might be explained by a “too low” riskless rate, nonetheless was an important paper in specifying and solving models with uninsured individual risks and transactions costs to keep people from trading them away.

21.4 1.

2.

Problems

Derive the analogue to the Hansen-Jagannathan bound in continuous time for an “excess return,” i.e. considering a self-financing portfolio, rather than a single return less the risk free rate. Suppose habit accumulation is linear, and there is a constant riskfree rate or linear

440

S ECTION 21.4

P ROBLEMS

technology equal to the discount rate, Rf = 1/δ. The consumer’s problem is then max

∞ X t=0

3.

δt

(Ct − Xt )1−γ 1−γ

s.t.

X t

δ t Ct =

X

δ t et + W0 ; Xt = θ

t

∞ X

φj Ct−j

j=1

where et is a stochastic endowment. In an internal habit specification, the consumer considers all the effects that current consumption has on future utility through Xt+j . In an external habit specification, the consumer ignores such terms. Show that the two specifications give identical asset pricing predictions in this simple model, by showing that internal-habit marginal utility is proportional to external-habit marginal utility, state by state. Suppose a consumer has quadratic utility with a constant interest rate equal to the subjective discount rate, but a habit or durable consumption good, so that utility is 1 u(ct − θct−1 ) = − (c∗ − ct + θct−1 ). 2

4.

Show that external habit persistence θ > 0 implies positive serial correlation in consumption changes. Show that the same solution holds for internal habits, or durability. Show that durability leads to negative serial correlation in consumption changes. Many models predict too much variation in the conditional mean discount factor, or too much interest rate variation. This problem guides you through a simple example. Introduce a simple form of external habit formation, u = (Ct − θCt−1 )1−γ

5.

and suppose consumption growth Ct+1 /Ct is i.i.d. Show that interest rates still vary despite i.i.d. consumption growth. We showed that if m satisfies the Hansen-Jagannathan bound, then proj(m|X) should also do so. Hansen and Jagannathan also compute bounds with positivity, solutions to min σ(m) s.t. p = E(mx), m ≥ 0, E(m) = µ.

6.

Does proj(m|X) also lie in the same bound? One most often compares consumption-based models to Hansen-Jagannathan bounds. Can you compare the CAPM discount factor m = a − bRem to the bound? To the bound with positivity?

441

Chapter 22.

References

Abel, Andrew B. 1988, “Stock Prices under Time-Varying Dividend Risk: An Exact Solution in an Infinite-Horizon General Equilibrium Model,” Journal of Monetary Economics 22, 375-93. Abel, Andrew B., 1990, “Asset Prices Under Habit Formation and Catching Up With the Jones,” American Economic Review 80, 38-42. Abel, Andrew B., 1994, “Exact Solutions for Expected Rates of Return under Markov Regime Switching: Implications for the Equity Premium Puzzle,” Journal of Money, Credit, and Banking 26, 345-61. Abel, Andrew B., 1999, “Risk Premia and Term Premia in General Equilibrium,” Journal of Monetary Economics 43, 3-33. Abel, Andrew B. and Janice C. Eberly, 1996, “Optimal Investment with Costly Reversibility.” Review of Economic Studies 63, 581-593. Abel, Andrew B. and Janice C. Eberly, 1999, “The Effects of Irreversibility and Uncertainty on Capital Accumulation,” Journal of Monetary Economics 44, 339-77. Aiyagari, S. Rao and Mark Gertler, 1991, “Asset Returns with Transactions Costs and Uninsured Individual Risk: A Stage III exercise,” Journal of Monetary Economics 27, 309-331. Andrews, Donald W. K., 1991, “Heteroskedasticity and Autocorrelation Consistent Covariance Matrix Estimation,” Econometrica 59, 817-58. Atkeson, Andrew, Fernando Alvarez, and Patrick Kehoe, 1999, “Volatile Exchange Rates and the Forward Premium Anomaly: A Segmented Asset Market View,” Working paper, University of Chicago. Bachelier, L. 1900, “Theory of Speculation,” in Cootner, P. (ed), The Random Character of Stock Prices, Cambridge, MA: MIT press 1964. Balduzzi, Pierluigi., Giuseppe Bertola and Silverio Foresi, 1996, “A Model of Target Changes and the Term Structure of Interest Rates,” Journal of Monetary Economics 39, 223, 49. Banz, Rolf W. 1981, “The Relationship Between Return and Market Value of Common Stocks,” Journal of Financial Economics, 9, 3-18. Barsky, Robert and Bradford J. DeLong 1993, “Why Does the Stock Market Fluctuate?” Quarterly Journal of Economics 108, 291-311. Bekaert, Geert and Robert J. Hodrick, 1992, “Characterizing Predictable Components in Excess Returns on Equity and Foreign Exchange Markets,”Journal of Finance 47, 467-509. Becker, Connie, Wayne E. Ferson, Michael Schill, and David Myers 1999, “Conditional Market Timing with Benchmark Investors,” Journal of Financial Economics 52, 119-148. Berk, Jonathan, 1997, “Does Size Really Matter?” Financial Analysts Journal, September/October 1997, 12-18. 442

Bernardo, Antonio and Olivier Ledoit. “Gain Loss and Asset Pricing.” (1999) Journal of Political Economy 108, 144-172. Black, Fischer, Michael Jensen and Myron Scholes, 1972, “The Capital Asset Pricing Model: Some empirical Tests,” in Michael Jensen, Ed., Studies in the Theory of Capital Markets (Praeger, New York NY) Black, Fischer and Myron Scholes, 1973, “The Valuation of Options and Corporate Liabilities.” Journal of Political Economy 81, 637-654. Bollerslev, Tim, R. Chou and K. Kroner 1992, “ARCH Modeling in Finance: A Review of Theory and empirical Evidence,” Journal of Econometrics,52, 5-59. Boudoukh, Jacob, Matthew Richardson 1994 “The Statistics of Long-Horizon Regressions Revisited,” Mathematical Finance 4, 103-119. Boudoukh, Jacob, Matthew Richardson, Robert Stanton and Robert Whitelaw, 1998, “The Stochastic Behavior of Interest Rates: Implications from a Nonlinear, Continuous-time, Multifactor Model,” Manuscript, University of California at Berkeley. Brav, Alon, George Constantinides and Christopher Geczy, 1999, “Asset Pricing with Heterogeneous Consumers and Limited Participation: empirical Evidence,” Manuscript Duke University. Breeden, Douglas T., 1979, “An Intertemporal Asset Pricing Model with Stochastic Consumption and Investment Opportunities,” Journal of Financial Economics 7, 265-296. Breeden, Douglas T., Michael R. Gibbons and Robert H. Litzenberger, 1989, “empirical Tests of the Consumption-Oriented CAPM,” Journal of Finance 44, 231-262. Brown, Stephen, William Goetzmann and Stephen A. Ross, 1995, “Survival” Journal of Finance 50, 853-873. Buraschi, Andrea and Alexei Jiltsov, 1999, “How Large is the Inflation Risk Premium in the U.S. Nominal Term Structure,” Manuscript London Business School. Burnside, Craig, Martin Eichenbaum, and Sergio Rebelo, 1993, “Labor Hoarding and the Business Cycle,” Journal of Political Economy 101, 245-73. Campbell, John Y. 1991, “A Variance Decomposition for Stock Returns,” Economic Journal 101, 157-179. Campbell, John Y. 1995, “Some Lessons from the Yield Curve,” Journal of Economic Perspectives, 9, 129-152. Campbell, John Y. 1996. “Understanding Risk and Return.” Journal of Political Economy 104, 298-345. Campbell, John Y. 1999. “Asset Prices, Consumption, and the Business Cycle.” in John B. Taylor and Michael Woodford eds. Handbook of Macroeconomics. Amsterdam: North-Holland. Campbell, John Y., 2000, “Asset Pricing at the Millennium,” Journal of Finance August Campbell, John Y., and John H. Cochrane, 1999, “By Force of Habit: A Consumption-Based Explanation of Aggregate Stock Market Behavior” Journal of 443

CHAPTER 22

REFERENCES

Political Economy, 107, 205-251. Campbell, John Y., Andrew W. Lo and A. Craig MacKinlay 1997, The Econometrics of Financial Markets Princeton NJ: Princeton University Press. Campbell, John Y. and Robert J. Shiller. 1988a. “The Dividend-Price Ratio and Expectations of Future Dividends and Discount Factors,” Review of Financial Studies 1, 195-227. Campbell, John Y. and Robert J. Shiller. 1988b. “Stock Prices, Earnings, and Expected Dividends.” Journal of Finance 43: 661-676. Campbell, John Y. and Robert J. Shiller 1991, “Yield Spreads and Interest Rates: A Bird’s Eye View” Review of Economic Studies 58, 495-514. Carhart, Mark M. 1997, “On Persistence in Mutual Fund Performance,” Journal of Finance 52, 57-82. Chamberlain, Gary and Michael Rothschild, 1983, “Arbitrage, Factor Structure, and Mean-Variance Analysis on Large Asset Markets,” Econometrica 51, 1281-1304 Chen, Nai-Fu, Richard Roll and Steven Stephen A. Ross, 1986, “Economic Forces and the Stock Market,” Journal of Business 59, 383-403. Christiano, Lawrence, Martin Eichenbaum, and Charles Evans, 1999, “Monetary Policy Shocks: What Have we Learned and to What End?” Forthcoming in John Taylor, ed., Handbook of Monetary Economics Cochrane, John H 1988.,“How Big is the Random Walk in GNP?”, Journal of Political Economy 96, 893-920. Cochrane, John H., 1991a, ”Explaining the Variance of Price-Dividend Ratios” 5, 243-280. Cochrane, John H., 1991b, “A Simple Test of Consumption Insurance,” Journal of Political Economy 99, 957-976. Cochrane, John H., 1991c “Volatility Tests and Efficient Markets: A Review Essay” Journal of Monetary Economics 27, 463-485. Cochrane, John H., 1991d “Production-Based Asset Pricing and the Link Between Stock Returns and Economic Fluctuations,” Journal of Finance 46, 207-234. Cochrane, John H., 1994, “Permanent and Transitory Components of GNP and Stock Prices” Quarterly Journal of Economics 109, 241-266. Cochrane John H. 1994, “Shocks,” Carnegie-Rochester Conference Series on Public Policy 41, 295-364. Cochrane, John H., 1996, “A Cross-Sectional Test of an Investment-Based Asset Pricing Model,” Journal of Political Economy 104, 572-621. Cochrane, John H., 1997, “Where is the Market Going? Uncertain Facts and Novel Theories,” Economic Perspectives 21: 6 (November/December 1997) Federal Reserve Bank of Chicago. Cochrane, John H., 1999a, “New Facts in Finance,” Economic Perspectives Federal Reserve Bank of Chicago 23 (3) 36-58. Cochrane, John H., 1999b, “Portfolio Advice for a Multifactor World” Economic

444

Perspectives Federal Reserve Bank of Chicago 23 (3) 59-78. Cochrane, John H., 2000, “A Resurrection of the Stochastic Discount Factor/GMM Methodology,” Manuscript, University of Chicago. Cochrane, John H. and Lars Peter Hansen, 1992, “Asset Pricing Explorations for Macroeconomics” In Olivier Blanchard and Stanley Fisher, Eds.,1992 NBER Macroeconomics Annual 115-165. Cochrane, John H. and Jesús Saá-Requejo 2000, “Beyond Arbitrage: Good Deal Asset Price Bounds in Incomplete Markets” Journal of Political Economy 108, 79-119. Cochrane, John H. and Argia M. Sbordone, 1988, “Multivariate Estimates of the Permanent Components in GNP and Stock Prices” Journal of Economic Dynamics and Control, 12, 255-296. Constantinides, George M., 1989, “Theory of Valuation: Overview and Recent Developments,” in Sudipto Bhattacharya and George M. Constantinides, eds., Theory of Valuation Totwa NJ: Rowman & Littlefield Constantinides, George M. 1990. “Habit Formation: A Resolution of the Equity Premium Puzzle.” Journal of Political Economy 98: 519-543. Constantinides, George M., 1992, “A Theory of the Nominal Term Structure of Interest Rates,” Review of Financial Studies 5, 531-52. Constantinides, George M. 1998, “Transactions Costs and the Volatility Implied by Option Prices.” Manuscript, Graduate School of Business, University of Chicago. Constantinides, George M. and Darrell Duffie. 1996, “Asset Pricing with Heterogeneous Consumers.” Journal of Political Economy 104, 219–240.. Constantinides, George M. and Thaleia Zariphopoulou, 1997, “Bounds on Option Prices in an Intertemporal Setting with Proportional Transaction Costs and Multiple Securities,” Manuscript, Graduate School of Business, University of Chicago. Cox, John C. and Chi-fu Huang, 1989, “Optimal Consumption and Portfolio Policies when Asset Prices Follow a Diffusion Process,” Journal of Economic Theory 39, 33-83. Cox, John C., Jonathan E. Ingersoll, and Stephen A. Ross. 1985. “A Theory of the Term Structure of Interest Rates.” Econometrica 53: 385-408. Cox, John C., Stephen A. Ross, and Mark Rubinstein, 1979, “Option Pricing: A Simplified Approach,” Journal of Financial Economics 7, 229-63. Cox, John C., and Mark Rubinstein, 1985, Options Markets Englewood Cliffs, NJ: Prentice Hall. Dai, Qiang and Kenneth J. Singleton, 1999, “Specification Analysis of Affine Term Structure Models,” forthcoming, Journal of Finance Daniel, Kent, David Hirshleifer and Avanidhar Subrahmanyam, 1998, “Investor Psychology and Security Market Under- and Overreactions,” Journal of Finance 53, 1839-1885. Debreu, Gerard 1959, The Theory of Value New York: Wiley and Sons. DeBondt, Werner F.M. and Thaler, Richard H., 1985, “Does the Stock Market

445

CHAPTER 22

REFERENCES

Overreact?” Journal of Finance 40, 793-805. Dixit, Avinash and R. Pindyck. Investment Under Uncertainty. Princeton NJ: Princeton University Press, 1994. Duarte, Jefferson 2000 “The Relevance of the Price of Risk in Affine Term-Structure models” Manuscript, University of Chicago Duffee, Gregory, 1999, “Forecasting future interest rates: Are affine models failures?” Manuscript, University of California at Berkeley. Duffie, J. Darrel and Rui Kan, 1996, “A Yield Factor Model of the Term structure of interest rates” Mathematical Finance 6, 379-406 Dybvig, Philip H. and Jonathan E. Ingersoll Jr., 1982, “Mean-Variance Theory in Complete Markets,” Journal of Business 55, 233-51. Dybvig, P., and Stephen Ross, 1985, “Yes, the APT is testable,” Journal of Finance 40, 1173-1188. Dybvig, P., J. Ingersoll Jr. and Stephen Ross 1996, “Long Forward and Zero-Coupon Rates Can Never Fall,” Journal of Business, 69, 1-25. Eichenbaum, Martin, Lars Peter Hansen and Kenneth Singleton, 1988, “A Time-Series Analysis of Representative Agent Models of Consumption and Leisure Choice under Uncertainty,” Quarterly Journal of Economics 103, 51-78. Engel, Charles, 1996, “The Forward Discount Anomaly and the Risk Premium: a Survey of Recent Evidence,” Journal of empirical Finance 3, 123-192. Engle, Robert F. and Clive W. J. Granger, 1987, “Cointegration and Error Correction: Representation, Estimation, and Testing,” Econometrica 55, 251-276. Epstein, Larry G. and Stanley E. Zin. 1989. “Substitution, Risk Aversion and the Temporal Behavior of Asset Returns.” Journal of Political Economy 99: 263-286. Fama, Eugene F., 1965, “The Behavior of Stock Market Prices, Journal of Business 38, 34-105. Fama, Eugene F., 1970, “Efficient Capital Markets: A review of Theory and empirical Work,” Journal of Finance 25, 383-417. Fama, Eugene F., 1984, “Forward and Spot Exchange Rates,” Journal of Monetary Economics 14, 319-338. Fama, Eugene F. 1991, “Efficient Markets II,” Journal of Finance 46, 1575-1618. Fama, Eugene F. and Robert R. Bliss, 1987, “The information in Long-Maturity Forward Rates,” American Economic Review, 77, 680-92. Fama, Eugene F. and Kenneth R. French. 1988a. “Permanent and Temporary Components of Stock Prices.” Journal of Political Economy 96: 246-273. Fama, Eugene F. and Kenneth R. French. 1988b. “Dividend Yields and Expected Stock Returns.” Journal of Financial Economics 22: 3-27. Fama, Eugene F. and Kenneth R. French, 1989, “Business Conditions and Expected Returns on Stocks and Bonds,” Journal of Financial Economics 25, 23-49. Fama, Eugene F. and Kenneth R. French, 1993, “Common Risk Factors in the Returns on Stocks and Bonds,” Journal of Financial Economics 33, 3-56. 446

Fama, Eugene F., and Kenneth R. French, 1996, “Multifactor Explanations of Asset-Pricing Anomalies,” Journal of Finance 47, 426-465. Fama, Eugene F., and Kenneth R. French, 1997a, “Size and Book-to-Market Factors in Earnings and Returns,”Journal of Finance 50, 131-55. Fama, Eugene F., and Kenneth R. French, 1997b, “Industry Costs of Equity,” Journal of Financial Economics 43, 153-193. Fama, Eugene F., and Kenneth R. French, 2000, “The Equity Premium,” Working paper, University of Chicago. Fama, Eugene F., and James D. MacBeth, 1973, “Risk Return and Equilibrium: empirical Tests,” Journal of Financial Political Economy 71, 607-636. Ferson, Wayne E., 1995, “Theory and Empirical Testing of Asset Pricing Models,” in R. A. Jarrow, V. Maksimovic, and W. T. Ziemba, eds., Handbooks in OR & MS, Volume 9, Finance Amsterdam: Elsevier Science B.V. Ferson, Wayne E. and George Constantinides. 1991. “Habit Persistence and Durability in Aggregate Consumption: empirical Tests.” Journal of