Empirical Cross-Sectional Asset Pricing

Stefan Nagel
Stanford University, NBER and CEPR

November 13, 2012

Abstract
I review recent research efforts in the area of empirical cross-sectional asset pricing. I
start by summarizing the evidence on cross-sectional return predictability and the failure
of standard (consumption) CAPM models and their conditional versions to explain these
predictability patterns. One response in part of the recent literature is to focus on ad-
hoc factor models, which summarize the cross-section of expected returns in parsimonious
form, or on production-based approaches, which suggest links between firm characteristics
and expected returns. Without imposing restrictions on investor preferences and beliefs,
neither one of these two approaches can answer the question why investors price assets the
way they do. Within the rational expectations paradigm, recent research that imposes
such restrictions has focused on the ICAPM, long-run risks models, as well as frictions and
liquidity risk. Approaches based on investor sentiment have focused on the development
of empirical proxies for sentiment and for the limits to arbitrage that allow sentiment
to affect prices. Empirical work that considers learning and adaptation of investors has
worked with out-of-sample tests of cross-sectional predictability.

I am grateful for comments to Frederico Belo, John Campbell, Marco Giacoletti, Sebastian Infante, Bob
Hodrick, Arthur Korteweg, Ian Martin, Dimitris Papanikolaou, Lasse Pedersen, Jeff Pontiff, Mike Schwert,
and Jeff Wurgler.

1
1 Introduction

It is a fundamental question of financial economics whether assets with different risk exposures

and characteristics are priced to earn different rates of return. I survey studies in empirical

cross-sectional asset pricing that provide evidence on this question and relate the evidence

to models of investor preferences and beliefs. I begin by laying out a framework that will be

helpful for categorizing and interpreting empirical work in this area.

Consider an economy whose state at time t is summarized by a state vector Xt . A

vector of parameters pins down the law of motion of Xt . Assume that Xt is observable

to investors, and let Jt denote the sigma algebra generated by Xt . The economy features

N assets that produce a vector of payoffs Dt (Xt ). Let Mt (Xt , ) be the stochastic discount

factor (SDF) that reflects preferences of a representative investor with preference parameters

. This investor is endowed with knowledge of , but not necessarily of .

Most empirical studies in cross-sectional asset pricing rely on rational expectations asset-

pricing theory in the tradition of Lucas (1978) to derive model predictions and a null hypoth-

esis. Under rational expectations (Muth (1961)), investors are endowed with knowledge of

the parameters . Defining (gross) returns as Rt+1 (Pt+1 + Dt+1 )/Pt , investors price assets

according to

E[Rt+1 (Xt+1 )Mt+1 (Xt+1 , )|Jt ] = 1, (1)

where the conditional expectation E[.|Jt ] is taken under the density f (Xt+1 |Jt , ) with knowl-

edge of . Now define Ht+1 Rt+1 Mt+1 1. Continuing to assume knowledge of , we can

apply the law of iterated expectations and condition down to the econometricians information

set, At , where At Jt . The resulting conditional moment restrictions imply orthogonality

conditions

E[Ht+1 (Xt+1 , ) Zt ] = 0 (2)

for any At -measurable instruments Zt . Studies in cross-sectional asset pricing often focus on

2
the special case

E[Ht+1 (Xt+1 , )0 Zt ] = 0 (3)

of (2), where Zt is a vector of portfolio weights that combines the N assets into a portfolio.

Evaluating (3) is akin to asking whether the elements Zit of Zt can predict differences in

discounted returns Hit+1 of the N assets. The cross-sectional asset pricing literature has

produced a large body of evidence of such cross-sectional return predictability. I start the

survey with a review of this evidence in Section 2. The economic interpretation of these

findings centers around the following key issues:

Stochastic discount factor. The law of one price implies that a stochastic discount factor

(SDF) exists such that (1) holds, and it can be constructed as a linear combination of the asset

payoffs (Hansen and Richard (1987)). The economic content of pricing models is therefore in

the restrictions that they impose on the SDF. Restrictions can be motivated by theory (e.g.,

the CAPM) or have an ad-hoc nature (e.g., restricting a factor model to a small number of

ad-hoc factors). Researchers often look to enrich the specification of investors preferences or

the dynamics of risks to enhance the pricing performance of models. I review these issues in

Sections 3 and 4.

Learning. The standard approach to estimating and evaluating asset-pricing models is

based on the rational expectations paradigm, where it is presumed that investors know the

parameters of the law of motion of Xt . This is an extremely demanding notion. More

realistically, an investor would have to learn by observing data. A Bayesian investor learning
R
would form expectations based on the predictive density f (Xt+1 |Jt ) = f (Xt+1 , |Jt )d,

viewing as a random variable, and pricing assets according to

E B [Rt+1 (Xt+1 )Mt+1 (Xt+1 , )|Jt ] = 1, (4)

where the B superscript of the expectations operator indicates that the expectation is now

taken under the Bayesian predictive density f (Xt+1 |Jt ). For the econometrician, this has the

3
unpleasant consequence that the usual approach of evaluating a sample version of (2) to test

the pricing model is not valid: return predictability that is evident to an econometrician ex-

post may not exist under the predictive distribution that investors perceive at the time they

make decisions. There is only a small empirical literature so far that looks at empirical cross-

sectional asset pricing from the perspective of learning. Out-of-sample tests are onealbeit

imperfectway to study learning effects, and I cover this line of work in Section 6.

Sentiment. The probability distribution of Xt+1 perceived by investors at time t may

deviate from the rational-expectations and Bayesian-learning benchmarks. In this case

E S [Rt+1 (Xt+1 )Mt+1 (Xt+1 , )|Jt ] = 1 (5)

holds with E S [.|Jt ] evaluated under the subjective density perceived by investors. I use

the label sentiment as a catch-all term to describe expectations-formation of this kind.

Section 5 covers empirical studies that take this perspective. Since the boundary between

rational theories and sentiment-approaches is fuzzy,1 I refrain from using the somewhat sim-

plistic categorization of cross-sectional asset-pricing studies into rational and irrational,

or behavioral, approaches.

Frictions. Frictions can affect the SDF in several ways. Frictions may prevent participa-

tion of some classes of investors in the market. Frictions can also give rise to risks that are

of concern to investors, such as liquidity risk. I cover work on these issues in parts of Section

4. Frictions also play an important role in the sentiment-based approaches in Section 5 as an

explanation as to why sophisticated investorswhose beliefs may be closer to the Bayesian-

learning or rational-expectations benchmarksmay be unable to exploit and eliminate the

effects of sentiment on asset prices.

Data snooping. One problem that plagues cross-sectional asset pricing is that much of
1
For example, macroeconomists have worked with models of boundedly rational learning (Bray (1982),
Marcet and Sargent (1989)) in which agents use simplified forecasting rules rather than a full Bayesian ap-
proach. Sims (2003) argues that agents with limited information-processing capacity may find it rational to
be inattentive to some types of information.

4
the empirical work in this area is inherently a search for anomalies. When researchers search

over a large number of candidate predictors for predictability that is unexplained by standard

models of the SDF, it is inevitable that some turn up as significant just by pure chance.

Conventional procedures of statistical inference do not account for this search over many

candidate predictors. The out-of-sample tests reviewed in Section 6 can help shed light on

this problem.

Market efficiency. Empirical work on cross-sectional return predictability is often dis-

cussed as the study of market efficiency, but the meaning of this concept is not clearly defined,

which limits its usefulness as an organizing principle. Fama (1970) defines a (semi-strongly)

efficient market as one that always fully reflects all available (public) information, but it is

not clear whether fully reflect refers to the demanding notion of rational expectations (this

is the view taken in Jensen (1978)), or whether the notion of market efficiency should allow for

learning and adaptation (Schwert (2003) and Timmermann and Granger (2004) propose this

interpretation). Frictions and risks further obscure the precise meaning of market efficiency.

Jensen (1978) augments his rational-expectations definition of market efficiency by suggest-

ing that market efficiency should be defined as the absence of profit opportunities from

trading on public information. Yet, in a market in which sophisticated and sentiment-driven

investors trade, it is possible that sentiment affects prices, yet sophisticated investors do not

perceive any abnormal profit opportunities given the frictions and risks that they face.2 If one

defines such a market as efficient, efficiency becomes a largely empty concept because it does

not distinguish between interesting alternative hypotheses (rational expectations, learning,

sentiment) of how investors price assets. For these reasons, I prefer to explicitly spell out the

belief-formation hypothesis that researchers work with, rather than relying on the imprecise

notion of market efficiency.

While I aim to cover the main conceptual issues in empirical cross-sectional asset pricing,

the survey is necessarily selective in its scope. I focus on recent work, mostly on equity
2
See, for example, Dumas, Kurshev, and Uppal (2009) for an illustration of this point in a general equilib-
rium model.

5
markets. On the econometric side, I emphasize the importance of the choice of moment

conditions when estimating and evaluating asset-pricing models, but I mostly skip a discussion

of methods for estimation and testing.3

2 Cross-Sectional Return Predictability

In this section, I provide a brief survey of the vast literature on cross-sectional return pre-

dictability. I focus on the most common categories of return predictors. Most studies in this

literature assess predictability relative to standard linear factor models of the SDF, such as

the CAPM, consumption CAPM, or ad-hoc factor models as benchmark models. I discuss

these models later in Section 3.

2.1 Technical Predictors

The history of returns and trading volume of a stock is a natural place to look for return

predictors, and the data are easily accessible. In the practice of investment management,

predictors of this kind are in the realm of technical analysis.

One of the most studied predictability patterns in this area is the momentum effect

that entered the academic literature with Jegadeesh and Titman (1993). Jegadeesh and

Titman show that stocks with high returns over the past three to twelve months (winners)

outperform stocks with low recent returns (losers) over the next three to twelve months.

A related phenomenon is the post-earnings-announcement drift. Stocks with unexpectedly

good earnings outperform those with unexpectedly bad earnings over the next six months

(Bernard and Thomas (1989)). Chan, Jegadeesh, and Lakonishok (1996) show that this

earnings momentum is related to, but partly distinct from, the price momentum effect. Lee

and Swaminathan (2000) show that stocks with high trading volume have low future returns

and stronger momentum effects.

3
Recent surveys that devote more attention to the methodology for estimation and testing include Jagan-
nathan, Skoulakis, and Wang (2009) and Goyal (2012).

6
 Over longer horizons, in contrast, returns have a tendency to mean-revert, as shown in

DeBondt and Thaler (1985) and DeBondt and Thaler (1987): Winners over the past three to

five years underperform losers. Jegadeesh and Titman (2001) find that the momentum effect

dies out after a holding period of about 12 months, and returns then start to revert at longer

holding periods.

2.2 Valuation Ratios and Profitability

A second category of predictors can be motivated within a simple present-value framework.

Vuolteenaho (2000) and Cohen, Polk, and Vuolteenaho (2003) build on the log-linear present

value model of Campbell and Shiller (1988) to formulate a present-value relationship in terms

of the book-to-market (B/M ) ratio. They start with the clean-surplus accounting relation

Dt = Yt Bt , where Dt is dividends, Yt is earnings, and Bt is the change in book value

of equity. Assuming no equity issues or repurchases and a covariance-stationary B/M , they

obtain a log B/M ratio

X
bmt Et j (rt+j yt+j ), (6)
j=0

where rt denotes the log stock return, and yt log(1 + Yt /Bt1 ) is the log return on equity

(ROE), and is a constant close to one.4

Thus, holding fixed expectations of future ROE, expected returns must be low if bm is

low. Holding fixed current bm, high expected ROE implies high expected returns. This

motivates the use of bm and expected profitability, preferably jointly, as return predictors.

Of course, expected returns do not necessarily have to vary across stocks. If they do not,

variation across stocks in bm reflects either future profitability or the approximation error

in (6). But if they do vary, bm and/or expected ROE should be correlated with expected

returns.

As (6) is simply an approximate accounting identity, it does not shed light on the struc-
4
If the firm also engages in equity issues or repurchases, yt must be modified to include growth in per-share
book value of equity due to the issuance or repurchase activity (Ohlson (2005)). Equity issues at B/M < 1
or repurchases at B/M > 1 both raise the per-share book value of equity.

7
tural drivers of cross-sectional variation in expected returns. It does not matter whether

priced risk under rational expectations, learning effects, or sentiment drive the variation in

expected returns. The conclusion that bm and expected profitability should predict returns

applies in each one of these cases.

The use of valuation ratios like B/M in return prediction has a long history in finance.

Fama and French (1992) reexamine the evidence and confirm that stocks with low market

capitalization (small firm premium) and high B/M (value premium) earn higher expected

returns. They find that firm size and B/M absorb the predictive role of the earnings-to-price

ratio. Other researchers examine B/M jointly with current profitability measures as proxies

for expected future profitability. Haugen and Baker (1996) document that ROE positively

predicts returns controlling for B/M and various other return predictors. Vuolteenaho (2002)

shows that recent returns, ROE, and B/M jointly predict returns. Given the approximate

accounting identity, it is not all that surprising that valuation ratios like B/M and profitabil-

ity variables are positively related to future returns. The more interesting bit is the failure

of standard models of risk and return to explain this cross-sectional variation in returns.

Current ROE may not be the best predictor of future ROE. For example, some sub-

components of current profitability may help predict future profitability better than others.

Sloan (1996) decomposes earnings into cash earnings and accruals. He finds that firms with

a high accruals have low future profits and abnormally low future returns. Hirshleifer, Hou,

Teoh, and Zhang (2004) argue that net operating assets summarize the cumulative history

of non-cash earnings and provide a better predictor of future earnings than single-period ac-

cruals. They find that net operating assets strongly predict future returns. Fama and French

(2006) employ a variety of accounting variables, past stock returns, and analyst earnings

forecasts that help forecast future earnings and they show that these variables also predict

returns. Novy-Marx (2012a) argues that gross profits (revenues minus cost of goods sold) are

a good predictor of future earnings, and he finds that gross profitability is strongly positively

related to future stock returns.

8
 Some researchers have used valuation ratios, profitability measures, and other accounting

variables to construct a summary measure of financial distress and find that distressed stocks

have abnormally low returns (Dichev (1998) and Griffin and Lemmon (2002)). Campbell,

Hilscher, and Szilagyi (2008) report a similar finding within a distress-prediction framework

that also employs price-based variables such as lagged returns and volatility.

2.3 Firm Investment and Financing

Theories of firms investment and financing can provide further insights about potential pre-

dictors of returns. According to the q-theory of investment, firm investment is positively

related to expected profitability and negatively related to discount rates and hence future

stock returns. Li, Livdan, and Zhang (2009) provide a recent exposition of this mechanism.

Titman, Wei, and Xie (2004), Fama and French (2006), Anderson and Garcia-Feijoo (2006),

and Li, Livdan, and Zhang (2009) find evidence consistent with this prediction.

High-investment firms are also likely to raise external financing, while firms with low

investment rates distribute capital to investors. As a result, discount rates are also negatively

related to external financing (see, e.g., Li, Livdan, and Zhang (2009)). Consistent with this

prediction, Daniel and Titman (2006), Fama and French (2008), and Pontiff and Woodgate

(2008) construct measures of net equity issuance activity and find a negative relationship to

future stock returns.

High levels of investment and external financing also translate into higher growth. There-

fore, one would expect a negative association between growth and future returns. Lakonishok,

Shleifer, and Vishny (1994) document a negative relation between past sales growth and re-

turns. Cooper, Gulen, and Schill (2008) look at growth in total assets and find a negative

correlation with future returns.

Investment and growth are also related to accruals. Fairfield, Whisenant, and Yohn (2003)

argue that high-growth firms tend to be high-accrual firms, suggesting that the relationship

between accruals and expected returns of Sloan (1996) may be a growth effect. Lewellen

9
and Resutek (2012) however show that investment, external financing, and accruals all have

distinct roles in predicting returns.

While q-theory is useful for identifying potential cross-sectional predictors of returns, the

theory is silent about the reasons why investors price stocks to have different expected rates

of return. The evidence that investment- and financing-related variables predict returns does

not reveal whether priced risk under rational expectations, investor learning, or sentiment are

the drivers of the discount rates that firms face in financial markets and that they respond

to in their investment and financing decisions.

2.4 Idiosyncratic risk

Theories in which idiosyncratic volatility plays a role in pricing typically predict a positive

relation between idiosyncratic volatility and expected returns. For example, in Merton (1987)

a positive relationship arises because investors are imperfectly diversified and demand com-

pensation for bearing idiosyncratic volatility. Empirically, Ang, Hodrick, Xing, and Zhang

(2006) find the opposite result: stocks with high idiosyncratic volatility have extremely low

returns. Huang, Liu, Rhee, and Zhang (2010) caution that this idiosyncratic volatility effect

seems to be driven by the return reversals known to exist in one-month returns (Jegadeesh

(1990)). Idiosyncratic volatility is also one of the variables used in the distress-prediction

model of Campbell, Hilscher, and Szilagyi (2008) and the sign of the effect in their frame-

work is consistent with the findings of Ang et al.

2.5 Seasonality

One puzzling feature of cross-sectional return predicability is that it is subject to strong sea-

sonality. Novy-Marx (2012b) offers an updated view based on recent data. He documents that

many of the cross-sectional return predictors reviewed in this section have a strong January

seasonal: size and value effects, for example, are concentrated in January, while momentum

and profitability-related predictability is weaker in January. The research reviewed in this

10
survey is silent about the source of this seasonality.

3 Linear Factor Models

Most empirical studies in cross-sectional asset pricing employ (log-)linear factor representa-

tions of the SDF. In this section, I review the canonical models that appear in most empirical

work, while Section 4 looks at recent efforts to improve these models. The work reviewed

here and in Section 4 takes a rational expectations approach. To economize on notation, I

use Et [.] as shorthand for E[.|At ].

3.1 Canonical Models

Typically, the SDF specifications in cross-sectional asset pricing studies are special cases of

the SDF that arises in a representative-agent framework in which the representative investor

has the preferences proposed by Epstein and Zin (1989), Epstein and Zin (1991), and Weil

(1990). In this model, the log SDF follows

mt+1 = log ct+1 + ( 1)rw,t+1 , (7)

where rw,t+1 log(Wt+1 /(Wt Ct )) is the log return on wealth, ct+1 log(Ct+1 /Ct ) is

consumption growth, is a time-discount factor, (1 )/[1 (1/)], is the elasticity of

intertemporal substitution, and is a relative risk aversion parameter. With power utility,

= 1/ and = 1, and we get the consumption CAPM

mt+1 = log ct+1 . (8)

As Giovannini and Weil (1989) show, if the state vector is identically and independently

distributed (IID), we get the CAPM,

mt+1 = rw,t+1 . (9)

11
Conditional CAPMs. In (8) and (9), the price of risk, controlled by the parameter , is

constant over time, which is why these models are often referred to as an unconditional

CAPMs. Theories of time-varying risk aversion motivate a generalization to time-varying

coefficients,

mt+1 = 0 (xt ) c (xt )ct+1 . (10)

This is a conditional consumption CAPM. For example, in the external habit model of Camp-

bell and Cochrane (1999), xt is the surplus-consumption ratio and 0t and ct are nonlinear

functions of xt . A conditional CAPM version of (9),

mt+1 = 0 (xt ) w (xt )rw,t+1 , (11)

is also frequently applied in empirical work, but the theoretical motivation is less clear in this

case. One possibility is that investors myopically maximize, period by period, their expected

utility of next-period wealth. In this case, 0 (xt ) = 0 and w (xt ) = (xt ), where (xt ) is a

time-varying relative risk aversion coefficient.

2 its conditional
Beta representation. Let rit+1 denote the log return on asset i, and it

variance. If we combine a log-linear factor model mt+1 = 0t ft 0 ft+1 with the assumption

that rit+1 and ft+1 are conditionally jointly normal, evaluation of the log of the conditional

pricing restriction (1) yields a linear beta representation,

1 2
Et [rit+1 ] + it = 0t + it0 t , (12)
2

where 0t 0t + ft 0 ft 12 ft 0 t ft , it 1 f
t Covt (rit+1 , ft+1 ), t Et [ft+1 ], t is the

conditional covariance matrix of ft+1 with a vector t2 of factor variances on the diagonal, and

t t ft are the factor risk premia. If a conditionally risk-free asset exists with time t + 1

log return rtf , and ft represents log returns on traded assets, we can apply the conditional

12
pricing relation to the risk-free asset and the factors themselves and solve for 0t and ft :

1
0t = rtf + ft 0 ft ft 0 t ft , (13)
2
f 1 f f 1
t = t (t rt + t2 ), (14)
2

where is a vector of ones. Substituting these back into (12), we obtain

1 2 1
Et [rit+1 ] + it = rtf + it0 (ft rtf + t2 ). (15)
2 2

The relations (13) and (14) highlight that 0t and ft have important implications for the

time-series dynamics of returns: they control the conditional risk-free rate and conditional

risk premia. While estimation of linear factor models in cross-sectional asset pricing studies

often focuses on cross-sectional pricing implications, it is important to keep in mind that the

SDF parameters also have time-series implications for conditional moments.

An alternative way of arriving at a linear beta representation is to exponentiate (10) or

(11), carry out a (conditional) first-order Taylor approximation, and normalize the SDF by

its conditional mean,5 which yields

Mt+1 = 1 ft 0 Ft+1 . (16)

Applying this SDF to the pricing restriction for excess returns, Et [Mt+1 (Rt+1 Rtf )] = 0, we

can solve for

Et [Rt+1 ] = Rtf + t (Et [Ft+1 ] Rtf ), (17)

which is approximately identical to (15), especially for returns measured over short horizons.

To provide a standardized treatment, I discuss factor models within the log-linear/log-normal

5
Any constant c, not only c = 1, in (16) would represent a valid normalization of the SDF that produces
identical pricing implications for conditional expected excess returns. In the case of a misspecified SDF,
though, tests of model misspecification are sensitive to the choice of normalization, as pointed out by Kan
and Robotti (2008)

13
framework, even though some of the original studies use the linearized SDF (16) applied to

simple returns.

Mimicking portfolios. The pricing implications of the SDF remain unchanged if we replace

non-traded factors by their (conditional) projection on the test asset returns. More precisely,

in the log-normal setting the projection is on the log returns of test assets, and on simple

returns in case of the linearized SDF (16).

3.2 Ad-hoc Factor Models

The failure of the (consumption) CAPM to explain cross-sectional return predictability has

prompted some researchers to resort to ad-hoc factor models. The motivation for this ap-

proach is the insight that risk factors in the SDF can be replaced by their mimicking portfolios.

The most prominent example is the Fama and French (1993) three-factor model, in which the

SDF is specified as a linear function of a market index return, RM , the difference in returns

between portfolios of small and large stocks, SM B, and the difference in returns between

portfolios of high and low B/M stocks, HM L. Fama and French (1993) show that this model

captures most of the cross-sectional variation in average returns of 25 portfolios sorted by

size and B/M .

SM B and HM L are constructed from portfolios that span the very same expected return

spreads along the size and B/M dimensions that the model is trying to explain. Is the model

therefore tautological? Not entirely. The economic content in Fama and French (1993) is in

the finding that the three factors explain not only cross-sectional, but also most of the time-

series variation in the size and B/M portfolios, with time-series R2 in excess of 90%. This

shows that small firms returns, for example, correlate a lot more strongly with almost-small

firms returns than with large firms returns. However, given this strong co-movement, it is

not surprising that expected returns of small firms are similar to those of almost-small firms

and value stocks expected returns are similar to those of almost-value stocks in the way

captured by the factor model; otherwise a near-arbitrage opportunity would arise. Absence

14
of near-arbitrage opportunities arises in any model under minimal restrictions on preferences

and beliefs (e.g., as in Ross (1976) and Cochrane and Saa-Requejo (2000)), and these weak

restrictions hold in (plausible) models of sentiment and models of learning just as well as in

rational expectations models. Therefore, the fact that the Fama-French model explains size

and B/M returns does not discriminate between these competing explanations.

Fama and French (1993) suggest that SM B and HM L could mimick pervasive priced

macroeconomic risk factors. If so, then SM B and HM L should explain other features of the

cross-section of expected returns, too, not just size and B/M effects. Fama and French (1996)

argue that this is the case because the three-factor model also explains the expected return

variation related to earnings/price, cash-flow/price, sales growth, and long-term reversals.

However, all of these variables are closely related to B/M . The one anomaly in their tests

momentumthat is not, fails to be explained by the three-factor model.

The empirical evidence accumulated since Fama and French (1996) suggests that there

are other important sources of cross-sectional variation in expected returns unrelated to the

Fama-French factors. Most of the cross-sectional return predictors reviewed in Section 2

have predictive power after adjusting for exposure to the Fama-French factors. To name a

few, the idiosyncratic volatility anomaly (Ang, Hodrick, Xing, and Zhang (2006)), the net

equity issuance effect (Daniel and Titman (2006)), and the predictability associated with

gross-profitability (Novy-Marx (2012a)) are all present after risk-adjusting returns with the

Fama-French factors. Lewellen (2011) looks at portfolios sorted on a composite measure

of expected returns that combines many known predictors and his portfolios exhibit large

abnormal returns after adjusting for exposure to the Fama-French factors.

Judging from this evidence, the Fama-French factors seem to be a convenient way of

summarizing the size and value effects, but not more than that. To capture other dimensions

of cross-sectional predictability, too, researchers have proposed alternative factors. Carhart

(1997) adds a momentum factor to the Fama-French model. Novy-Marx (2012a) proposes a

gross-profitability factor. Hou, Xue, and Zhang (2012) construct a four-factor model with a

15
market factor, a size factor, an investment factor, and an ROE factor. The use of profitability-

and investment-related factors is motivated by the present-value relationships and q-theoretic

models discussed above in Sections 2.2 and 2.3. Of course, one must keep in mind that present-

value relationships and q-theory do not answer the question why investors price stocks to have

these differences in expected returns.

3.3 Conditional Factor Models

Much of the early empirical work on size and value premia had been carried out with one-

factor models,

mt+1 = 0t f ft+1 , (18)

in which the price of risk is constant. This raises the question whether the unexplained size

and value premia could simply reflect the fact that researchers ignored time-varying prices of

risk that arise in conditional factor models such as (10) or (11).

To implement tests of conditional factor models, we need to specify the relevant state

variables xt and their functional relationship with the SDF coefficients 0t and ft . Jagan-

nathan and Wang (1996), Cochrane (1996), and Lettau and Ludvigson (2001), and a large

number of subsequent papers (see the references in Lewellen and Nagel (2006)) (LN) examine

models in which the coefficients are linear in the state variables. Applied to our log-linear

SDF here, we get

mt+1 = 0 + 0x xt f ft+1 f x xt ft+1 . (19)

Thus, the conditional one-factor model has been turned into an unconditional multi-factor

model.6 Although intuitively appealing, this route turned out to be less fruitful than it

initially seemed.

To see why, suppose that a conditional one-factor model mt+1 = 0t ft ft+1 is the true
6
The xt term should not be regarded as a risk factor, though, because it is pre-determined. As a conse-
quence, under the restrictions of the model, it plays no role in generating cross-sectional variation in expected
returns.

16
model, but a researcher applies a constant-price-of-risk model like (18). How big is the pricing

error that results from this misspecification? Lewellen and Nagel (2006) (LN) perform these

calculations for a linearized SDF applied to simple returns. Here I show similar calculations

within a log-linear/log-normal framework. The conditional factor model implies

1 2
Et [rit+1 ] rtf + it = Covt (rit+1 , ft+1 )ft . (20)
2

Taking unconditional expectations and noting that the expected conditional covariance equals

the unconditional covariance if f is conditionally de-meaned, we get

1
E[rit+1 ] E[rtf ] + E[it
2
] = Cov(rit+1 , ft+1 ft )E[ft ] + Cov{Covt (rit+1 , ft+1 ), ft } (21)
2

The first term on the RHS is the result that one would obtain with the constant-price-of-risk

model (18). Even if conditional betas and the factor risk premium t = f t2 are time-

varying, an unconditional beta representation holds as long as the price of risk is constant.

Only if the price of risk varies over time and is correlated with the conditional covariance of

returns and the factor, the second term on the RHS introduces a wedge between the pricing

implications for unconditional expected returns of the true conditional factor model and the

econometricians constant-price-of-risk model.

How big could Cov{Covt (rit+1 , ft+1 ), ft } be for a value minus growth portfolio? Con-

sider the case of the conditional CAPM where f is a market index log return. To maxi-

mize time-variation in the price of risk, consider the case in which all time-variation in the

market risk premium is driven by time-variation in ft rather than by time-varying con-

ditional factor volatility, so that t = ft 2 .7 With it = Covt (rit+1 , ft+1 )/ 2 we obtain

Cov{Covt (rit+1 , ft+1 ), ft } = Cov(it , t ). To get an upper bound, consider the case in which

it and t are perfectly correlated. As LN point out, plausible values for the standard de-
7
Evidence from ICAPM and long-run risks models reviewed below in Section 4 suggest that conditional
factor volatility drives at least part of time-variation in factor risk premia. This would imply lower time-
variation in ft and smaller Cov{Covt (rit+1 , ft+1 ), ft }.

17
viation of it for a value-growth spread portfolio are around 0.30. Predictive regressions or

calibrated models of time-varying risk premia like Campbell and Cochrane (1999) suggest

a standard deviation of t of at most 0.5% monthly. Thus, an upper bound for the wedge

amounts to 0.15% monthly, or less than 2% in annualized termsmuch less than the empir-

ically observed value-growth spread of 0.50% monthly. If the correlation of it and t takes

a more realistic value substantially closer to zero, the gap to the observed value premium

becomes even bigger. Therefore, if the constant-price-of-risk version of the CAPM of the

form (18) fails to explain the value premium, it is not plausible that the conditional CAPM

can explain it. A similar logic applies in the case with a non-traded factor, as in the con-

sumption CAPM, but there it is the covariance between conditional consumption betas and

the consumption risk premium that determines the pricing error.

This reasoning leadsseeminglyto a puzzle. Studies like Jagannathan and Wang (1996)

and Lettau and Ludvigson (2001) use an SDF of the form (19) and find a dramatic improve-

ment in explanatory power for the cross-section of stock returns compared with an uncon-

ditional one-factor model. The key to reconciling these findings is to note that LNs anal-

ysis takes into account the models implications for the time-series dynamics of conditional

moments: according to theory, for a significant wedge to exist between the unconditional

expected return predictions of the conditional factor model and its constant-price-of-risk ver-

sion, Cov(it , t ) must be big. In contrast, if the model is fit to a cross-section of expected

returns, without taking into account conditional moment restrictions, this restriction is ig-

nored.8 As a consequence, the cross-sectional estimates of 0t and ft can imply dynamics

of conditional moments, via (13) and (14), that are completely out of line with the data.

Nagel and Singleton (2011) illustrate this gap between implied and empirical dynamics of

conditional moments in the case of the conditional consumption CAPM. They show that the

cross-sectional estimates of ft can be more than an order of magnitude too volatile compared

with models that feature strong, but plausible, time-variation in risk premia, like Campbell
8
In a finite sample, by estimating E[Mt+1 Rt+1 ] = 1 via GMM, or in a two-pass regression where, in the
second stage, average returns of the test portfolios are regressed on their factor betas.

18
and Cochrane (1999).

3.4 Pitfalls with Low-Dimensional Factor Structures in Test Asset Returns

There is a further rather mechanical reason why fitting the unconditional versions (19) of

conditional factor modelsor in fact any multi-factor modelto the cross-section of expected

returns yields small pricing errors for certain types of test portfolios. Most studies that test

models of the form (19) use size and B/M portfolios as test assets. For these portfolios,

three factors explain almost all the time-series and cross-sectional variation in returns: the

market factor, SMB, and HML (Fama and French (1993)). Thus, within the log-normal
e
framework here, we can write the log excess returns, rt+1 rt+1 rtf + 12 E[it
2 ], on N test

assets approximately as

rte = Bpet + et , (22)

where B is a N K factor loading matrix, pet is a K-dimensional vector of log excess returns

on the factors underlying the test asset returns (the Fama-French factors, with K = 3, in the

case of the size- and B/M -sorted test portfolios ), and the residual et is uncorrelated with pet .

Now consider the application of a model like (19), which we can write as mt+1 = 0 0 ft+1 ,

where ft is uncorrelated with et and has dimension J = 3. The N J matrix of betas with

respect to these proposed factors f is

0
= BCov(pet+1 , ft+1 )1 (23)

where is the covariance matrix of f. Note that the betas on the proposed factors are

K = J linear combinations of the factor loadings B. As a consequence, explains the cross-

section of E[rte ] = BE[pet ] as well as B, namely perfectly, irrespective of the nature of the

proposed factors pe .9 That explains the cross-section of expected returns therefore has

little economic content. It just re-discovers the fact that the test asset returns have a low-
9
As long as the correlation between the elements of pet and ft is not exactly zero. Lewellen, Nagel, and
Shanken (2010) also analyze more general settings in which J < K and in which ft can be correlated with et .

19
dimensional factor structure and that the proposed factors f are uncorrelated with residuals

e. This is the point made in Lewellen, Nagel, and Shanken (2010) (LNS) (see also Daniel and

Titman (2012)). In many existing tests of models of the type (19), the number of proposed

factors is J = 3 and so the small pricing errors and high cross-sectional R2 emphasized in

these papers actually carry little economic meaning.

Unlike the cross-sectional R2 , goodness-of-fit specification tests, as in Hansen (1982),

Shanken (1985), and Hansen and Jagannathan (1997), adjust for the low-dimensional fac-

tor structure in test asset returns. These tests are correctly sized under the null hypoth-

esis that the model is correctly specified. However, a low-dimensional factor structure can

create a power problem. When the magnitude of the pricing errors is small due to the

low-dimensionality problem, these tests have little power to reject a false model. Moreover,

researchers often intentionally focus on R2 -type measures of model performance rather than

these specification tests to highlight the magnitude of pricing errors on economically inter-

esting portfolios, rather than linear combinations of portfolios formed based on statistical

considerations (Cochrane (1996)). Alas, if the test asset portfolios have a low-dimensional

factor structure, R2 -type measures are not economically informative. To make the tests more

informative, additional information needs to be brought in through judicious choice of the

moment conditions that are used to estimate and evaluate the model.

3.5 More Informative Moment Conditions

LNS discuss several avenues to enhance the informativeness of the tests. For example, one

can expand the set of test portfolios to include assets that do not share the low-dimensional

factor structure of size and B/M portfolios. LNS add industry portfolios and find a dramatic

deterioration in the cross-sectional fit of several previously proposed models of the type (19)

(see, also, Daniel and Titman (2012)).

If factors of a proposed model are traded, a second possibility is to exploit the fact that

theory links the prices of risk, via (13) and (14), to expected excess returns of the factors.

20
This implies additional moment conditions for the factors themselves. Within our log-normal

framework, the unconditional moment restrictions become

e
E[rt+1 ] = (24)
e
E[ft+1 ] = . (25)

To what extent estimation pays attention to the additional restriction (25) depends on the

estimator that is employed. A two-pass regression estimator where the second stage regresses
e
the sample average of rt+1 e
and ft+1 on first-stage estimates of betas with OLS may pay little

attention to (25). In contrast, maximum likelihood (ML), as in Shanken (1992), enforces (25)

in terms of its sample analog by estimating = (1/T ) Tt=1 ft+1

e .10 A similar logic applies
P

with a GMM estimator of a linearized SDF as in (16) and moment conditions for excess

returns E[Mt+1 (Rt+1 Rtf )] = 0. A GMM estimator with an identity weighting matrix may

put very little weight on the moment condition for the factor excess return, while an efficient

estimator with the optimal weighting matrix of Hansen (1982) enforces perfect pricing of the

factor asymptotically.11 Thus, inclusion of (25) among the moment conditions combined with

the use of efficient estimation methods is likely to enhance the informativeness of the tests.12

Empirical work in cross-sectional asset pricing typically works with unconditional moment

restrictions for estimation and evaluation, but conditional moment restrictions can enhance

the informativeness of tests. This enhancement is especially relevant for conditional factor

models. The time-varying prices of risk in these models imply strong predictions about the

time-variation of conditional risk premia. It seems useful to ask whether these time-series
10
ML is Pasymptotically equivalent to a time-series regression of rte on fte , which automatically imposes
= (1/T ) Tt=1 ft+1
e
(Gibbons, Ross, and Shanken (1989)), or a GLS cross-sectional regression which does
the same (Shanken (1992)).
11
This can be seen by noting that the asymptotic weighting matrix S 1 is proportional to the inverse of the
covariance matrix of the residuals in a regression of returns on factors (see, e.g., the appendix of Jagannathan
and Wang (2002)). If the factor is included among the test asset returns, the corresponding element of the
residual vector has zero variance, and hence S 1 1 puts infinite weight on pricing the factor perfectly.
12
For non-traded factors, this approach is not available. One might think that replacing non-traded factors
with mimicking portfolios might help, but this does not address the problem. One can show that if test asset
returns follow (22), K = J, and ft is uncorrelated with et , the mimicking portfolio always price the test assets
perfectly even if the mimicking factor risk premia are restricted to be equal to their expected excess returns.

21
predictions are in line with the data. With traded factors, Lewellen and Nagel (2006) show

that this can be done by utilizing higher frequency data (e.g., daily returns), and estimating

a factor model as a time-series regression within short time windows (e.g., quarters). The

idea is that conditional betas and the conditional market risk premium (co-)vary very little

within a short time window, so a time-series regression within a short window produces ap-

proximately unbiased estimates of conditional betas and conditional alphas. The conditional

factor model can then be tested by examining whether the average of the (noisy) short-

window alpha estimates is significantly different from zero. Based on this method, LN find

that the conditional CAPM fails to explain the size and value premia. Ang and Kristensen

(2012) refine this approach by using a more general non-parametric estimation methodology

(LNs estimator is a special case of theirs) and they find similar results.

With non-traded factors, we can exploit conditional moment restrictions by specifying

instruments and working with the moment conditions (2). But which instruments should

we use? Even if we fix the set of random variables that we want to condition on, there is

still an infinite number of ways to construct the instruments and to choose the number of

moment conditions, as any function of At -measurable random variables is a valid and feasible

instrument. Some instruments may enhance the informativeness of the moment conditions,

while others may add little information over and above the unconditional moment restrictions.

Nagel and Singleton (2011) build on Hansen (1985) and Chamberlain (1987) to implement

an optimal-instruments estimator that uses the statistically most informative instruments.

Nagel and Singleton examine several scaled consumption CAPMs of the form (19) and find

that these models are unable to fit the cross-sectional and time-series dimension of the data

simultaneously (see, also, Roussanov (2011) for related work).

In summary, the conditional factor models that seem to do so well in explaining the cross-

section of unconditional expected returns on size and B/M portfolios struggle to explain the

data when additional moment conditions are brought inas one would expect, given the

discussion in Section 3.3. More generally, these results highlight that researchers must be

22
careful to choose moment conditions that can be informative about the economic explanatory

power of a model.

4 Recent Advances in Modeling the SDF

As the conditional factor model route proved to be less fruitful than it initially seemed, other

approaches to enhance standard specifications of the SDF have recently received more atten-

tion. A number of recent papers work with the general version of the Epstein-Zin-Weil SDF

(7). They obtain empirically implementable models by making additional assumptions about

the observability of wealth portfolio returns (ICAPM) or about the dynamics of consumption

(long-run risks). I also review work that links asset prices to the first-order conditions of

producers, of financial market participants, and of financial intermediaries.

4.1 Substituting Out Consumption Using the Intertemporal Budget Con-

straint

The Epstein-Zin-Weil SDF (7) features a consumption growth factor. As consumption is no-

toriously difficult to measure, replacing consumption growth with easier-to-observe variables

could perhaps improve the empirical performance of the model. Campbell (1993) shows that

the intertemporal budget constraint can be used to substitute out consumption. Recent work

by Campbell, Giglio, Polk, and Turley (2012) extends this approach to a setting with con-

ditional heteroskedasticity. The intertemporal budget constraint implies that the log return

on wealth follows
w
rt+1 = at + ct+1 + vt+1 , (26)

where at log((Wt Ct )/Ct ) and vt+1 log(Wt+1 /Ct+1 ). Campbell et al. further approxi-

mate vt+1 + at+1 , where takes a value close to one. Using this approximation in (26)

23
and substituting into the log Epstein-Zin-Weil SDF (7) yields

w
mt+1 = log + at + at+1 rt+1 . (27)

Assuming that asset returns and consumption growth are conditionally jointly log-normal

and applying the conditional pricing restriction Et [Mt+1 Rt+1 ] = 1 with the SDF (27) to the

return on wealth allows solving forward for at . Substituting this result for at+1 in (27) and

absorbing the conditional mean of the SDF in gt , we obtain

cf dr 1 risk
mt+1 = gt Nt+1 + Nt+1 + Nt+1 (28)
2

where

X
dr
Nt+1 (Et+1 Et ) j rt+1+j
w
, (29)
j=1

cf w w dr
Nt+1 rt+1 Et [rt+1 ] + Nt+1 , (30)

X
risk
Nt+1 (Et+1 Et ) j Vart+j (mt+1+j + rt+1+j
w
). (31)
j=1

Thus, one obtains a conditional three-factor model that represents a discrete-time version

of the intertemporal CAPM (ICAPM) of Merton (1973). The N dr and N risk factors reflect

the exposure of the SDF to changes in investment opportunities in the form of news about

expected returns and news about risk, respectively. N cf represents news about expected cash

flows on the wealth portfolio.

Campbell and Vuolteenaho (2004) (CV) investigate the homoskedastic version of (28)
risk is zero, as in Campbell (1993). Empirical implementation of the model requires
where Nt+1

forecasts of returns to compute N dr . CV use a vector-autoregression (VAR) to produce these

forecasts. They find that the high market beta of growth stocks in the post-1963 period arises

largely from a strong negative loading on discount-rate shocks, while value stocks have higher

24
cash-flow betas (see, also, Bansal, Dittmar, and Lundblad (2005)). With > 1, cash-flow

risks carry a higher risk premium than discount-rate risks, resulting in a value premium.

When evaluating the fit of their model, CV take into account the fact that the ICAPM

restricts the risk prices in (28) to be a function of only one parameter (). Incorporating

this restriction in the empirical estimation helps to avoid the problems associated with a

low-dimensional factor structure in test asset returns that I reviewed in Section 3.4.

Campbell, Giglio, Polk, and Turley (2012) investigate the full heteroskedastic version of

(28) (Brennan, Wang, and Xia (2004) empirically analyze a related model). Campbell et

al. assume a one-factor structure of volatility shocks. In this case the conditional variances

that enter N risk are proportional to the common volatility factor. In the post-1963 sample,

they find that value stocks have small positive exposure to these volatility shocks, but the

exposure of growth stocks is much bigger. The positive correlation of equity returns with

volatility shocks is counter-intuitive, but since volatility shocks have a negative price of risk,

this further helps to explain the value premium.

Overall, the ICAPM shows some promise for explaining the cross-section of expected

returns, but some challenges remain. VAR dynamics are subject to substantial estimation

uncertainty, and empirical results differ to some extent across plausible specifications of in-

vestment opportunity dynamics (see Chen and Zhao (2009) and Bansal, Kiku, Shaliastovich,

and Yaron (2012)). A second challenge is that the dynamics of investment opportunities es-

timated by the VAR may not be internally consistent with the model. The Epstein-Zin-Weil

SDF implies equilibrium restrictions on the joint time-series dynamics of risk premia and

volatility that are ignored in an unrestricted VAR. Taking these unrestricted VAR dynamics

as exogenously given, an Epstein-Zin-Weil investor would want to time the market, while

the ICAPM implementation of Campbell et al. assumes that the investor holds 100% of

wealth in the market portfolio of risky assets. Holding a market-timing portfolio would in

turn change the investors view about the riskiness of value and growth stocks. Bansal, Kiku,

Shaliastovich, and Yaron (2012) explore a specification in which the market risk premium is

25
linear in the conditional variance of the market return, which is a step (albeit not a complete

one) towards reconciling the dynamics of investment opportunities with the model.

4.2 Substituting Out the Wealth Return by Modeling Long-Run Con-

sumption Dynamics

While the ICAPM approach is motivated by the notion that returns on wealth may be easier

to observe than consumption, one can also make the argument that the wealth portfolio

return may be difficult to observe because many important components of wealth are traded

in markets subject to severe frictions (e.g., housing), or not traded at all (e.g., human capital).

If one follows this argument, then substituting out the wealth return rather than consumption

growth could be preferable.

Using (26) to substitute out the wealth return from the log Epstein-Zin-Weil SDF (7)

yields

mt+1 = log + ( 1) ct+1 ( 1)at + ( 1)at+1 . (32)

Applying this SDF to the conditional pricing restriction Et [Mt+1 Rt+1 ] = 1 for the wealth

return and substituting for the wealth return in the conditional expectation (but not the

variance) using (26), one can solve forward for at . Substituting the result into (32) for at+1

and absorbing the conditional mean of the SDF in gt , we obtain

1
sc 1 lc 1 risk
mt+1 = gt Nt+1 ( )Nt+1 + N , (33)
2 1 t+1

where

sc
Nt+1 ct+1 Et [ct+1 ], (34)

X
lc
Nt+1 (Et+1 Et ) j ct+1+j , (35)
j=1

and N risk is defined as before in (31). Restoy and Weil (2011) derive the homoskedastic

26
version of (33) without N risk . If > 1/, assets earn a positive risk premium for positive
lc . Bansal and Yaron (2004)
correlation with shocks to long-run consumption growth, Nt+1

label this risk as long-run risk and they propose that consumption growth contains a small

highly persistent component that makes this risk relevant for asset pricing.

Consistent with a premium for long-run risk, Parker and Julliard (2005) find that the

covariance of stock returns with future consumption growth several years ahead is positively

related to B/M and negatively related to size, which helps explain the cross-section of average

returns on size and B/M portfolios. Liew and Vassalou (2000) find that the Fama-French

size and B/M factors help predict GDP growth at an annual horizon, which also points in

this direction. Hansen, Heaton, and Li (2008) find similar results with a different approach

to measuring long-run risk. Bansal, Kiku, and Yaron (2007) assume a parametric model

for consumption growth with a small highly persistent component and stochastic volatility.

They show that N lc and N risk can be constructed empirically by projecting consumption

growth and squared consumption innovations on the markets price-dividend ratio and the

risk-free rate. Their model has a somewhat harder task than the ICAPM, because it links

consumption dynamics and asset returns in an internally consistent fashion. The model

explains part of the value and size premia during the 1930 to 2002 period, but it leaves a

substantial portion unexplained. Standard errors are big because the long-run consumption

dynamics are estimated with low precision.

These problems in estimating long-run consumption dynamics raise some doubts about

the merit of the rational expectations assumption in these models. Similar concerns arise

in the ICAPM, where the long-run dynamics of discount rates are difficult to pin down. As

Hansen, Heaton, and Li (2008) emphasize, if econometricians struggle to pin down these

long-run dynamics, it is a leap of faith to endow agents with knowledge of the parameters of

the data-generating process. Developing an asset-pricing framework that incorporates learn-

ing, parameter uncertainty, and model uncertainty seems particularly relevant for models

whose asset-pricing implications are driven by long-run dynamics. In fact, Collin-Dufresne,

27
Johannes, and Lochstoer (2012) show that learning could be the source of long-run risk: a

Bayesian agent learning about a constant consumption growth rate is aware that her subjec-

tive beliefs about the mean growth rate evolve as a martingale, and thus perceives long-run

risk.

4.3 Production-Based Approaches

While the consumption-based asset-pricing model links the SDF to consumers marginal rate

of substitution state by state, production-based models link the SDF to producers marginal

rate of transformation (Cochrane (1991)). This raises the prospect that one could recover the

SDF from (possibly better measured) macroeconomic data on the production side without

relying on (hard-to-measure) consumption data. Typical production functions, however, do

not allow the marginal rate of transformation to vary across states, which makes recovery of

the SDF impossible without relying on investors first-order condition, too.

The model of Belo (2010) is an exception. Belo specifies a production function that allows

producers to move productivity across states, and he works out how to empirically recover

the SDF as a function of industry-level production variables from the producer first-order

conditions, without any assumptions about consumer preferences. He finds that this SDF

does quite well in pricing size, B/M , and industry portfolios.

Cochrane (1996) uses an alternative approach. If firms have a constant returns-to-scale

technology, access to financial markets, and they trade to remove arbitrage opportunities be-
I ]=1
tween physical investment and financial markets, then the pricing equation Et [Mt+1 Rt+1

holds for physical investment returns RI in the same way it holds for stock returns. Just as

one can construct a pricing kernel by projecting M on securities returns (Hansen and Jagan-

nathan (1991)), one can construct one by projecting M on investment returns. Restricting

the number of physical investment returns to a few common factors, Cochrane arrives at an

ad-hoc factor model where the SDF is linear in two investment returns, approximated by

investment growth rates (private residential and non-residential investment). Li, Vassalou,

28
and Xing (2006) use a similar approach with a more disaggregated set of investment growth

rates. These ad-hoc factor models have some success in fitting the cross-section of stock

returns, but their economic content is limited. A test of these models really just tests the law

of one price under the inclusion of physical investment returns, combined with the ad-hoc

restriction to a small number of factors.

A third approach, pioneered by Berk, Green, and Naik (1999), is to exogenously fix an

SDF as a function of some aggregate shocks, and proceed to model the evolution of firms

asset composition and systematic risk. Berk et al. use a dynamic real options model in which

value firms consist of assets in place and growth options, while growth stocks are pure growth

options. They exogenously specify a log SDF of the form

1
mt+1 = rtf z2 z t+1 , (36)
2

where t+1 is also the only aggregate shock that affects firms assets-in-place cash flows,

while the value of growth options is additionally exposed to interest-rate shocks that have a

component orthogonal to . As a consequence, value stocks have a greater negative covariance

with the SDF, i.e., a higher systematic risk and expected return, than growth stocks. Carlson,

Fisher, and Giammarino (2004) work with a related model with a single source of aggregate

uncertainty. Zhang (2005) endogenizes firms cash flows and systematic risk in a competitive

industry equilibrium. Zhang also uses an SDF like (36), but with a time-varying loading on .

In his model, is the only source of aggregate uncertainty. Novy-Marx (2011) finds support

for the prediction implied by these models that operating leverage should be positively related

to expected returns.

Berk et al., Carlson et al., and Zhang argue that their models provide a rational ex-

planation for the value premium. This conclusion is not warranted. These models surely

produce interesting insights into the origins of firms systematic risk exposures, but these

models cannot shed light on the question whether the empirically observed cross-sectional

variation in expected returns reflects priced risk under rational expectations, the effects of

29
learning, or the effects of sentiment. The exogenous specification of an SDF like (36) does

not embody any economic assumptions other than no arbitrage combined with ad-hoc re-

strictions on the types of shocks that do and do not enter the SDF. That the no-arbitrage

restriction holds does not imply that the SDF reflects rational pricing. For example, in an

economy in which there is a single source of aggregate uncertainty, as in Zhang (2005), any

systematic effects of investor sentiment would also be driven by this single aggregate shock.

Consequently, under the econometricians objective probability measure, the SDF would also

take the form (36). Berk, Green, and Naik (1999) impose the ad-hoc restriction that the part

of the interest-rate shock uncorrelated with does not show up as a shock in the SDF. But

why dont investors care about that part of the shock? Are there rational reasons for them

not to care about it, or are they imperfectly rational in attaching a zero price of risk to this

shock?13 Without a structural model of consumer preferences and beliefs, the model cannot

answer these questions.

A natural fourth approach is therefore to jointly model consumer preferences and the

production side in general equilibrium. In many cases, though, models that take this ap-

proach also imply that a conditional consumption CAPM or even a conditional CAPM holds,

which brings the models into conflict with the empirical evidence in section 3.5. For example,

Gomes, Kogan, and Zhang (2003) have a single aggregate shock in their model, and the mar-

ket portfolio is conditionally a perfectly correlated mimicking portfolio for consumption, and

hence the conditional CAPM holds (the same happens in Zhang (2005) and Carlson, Fisher,

and Giammarino (2004)). To get around these undesirable predictions, some researchers

have added additional sources of uncertainty to the model. For example, Papanikolaou

(2011) works with Epstein-Zin-Weil preferences and a production technology that is sub-

ject to productivity shocks and technological innovation shocks. He constructs a long-short

factor portfolio that mimicks the technological innovation shocks and which helps explain the
13
A similar assumption is made in Lettau and Wachter (2007). As Lettau and Wachter and Santos and
Veronesi (2010) point out, with external habit preferences like in Campbell and Cochrane (1999), the interest-
rate shock would appear in the SDF and lead to a growth premium rather than a value premium.

30
cross-section of average returns in the data.

4.4 Rare Disasters and Higher Moments

The vast majority of empirical studies in cross-sectional asset pricing work within a log-

normal framework, either explicitly, by combining a log-linear SDF with the assumption of

joint log-normality of returns and the SDF, or implicitly, by applying a first-order Taylor

approximation to a non-linear SDF. This approach may fail to account for higher moments

and tail risks that investors care about. Recent work by Barro (2006) revived interest in the

pricing implications of disaster risks.

Empirical work on the cross-sectional pricing implications of disaster risks is still in its

infancy. In principle, one can account for these in empirical work by applying non-linear esti-

mation methods (e.g., GMM) to Et [Mt+1 Rt+1 ] = 1 without linearizing the SDF or assuming

log-normality. Alternatively, one can include higher-order polynomial terms in the approxi-

mation. This is the approach taken by Harvey and Siddique (2000) and Dittmar (2002). To

the extent that tail risks are important, sample selection and the finite-sample distribution

of estimators are of particular relevance.

4.5 Participation Constraints and Liquidity

In typical empirical applications of representative-agent models, consumption is measured as

the (per-capita) aggregate consumption of all consumers in the economy. In reality, not all

consumers participate in financial markets. Asset prices might reflect the consumption risks

of financial market participants, but not necessarily the consumption risk of non-participants.

To explore this idea, Mankiw and Zeldes (1991) construct a series of per-capita stockholder

consumption, and they show that with this series it is easier to explain the equity premium in

a consumption-based asset pricing model than with overall per-capita consumption because

stockholder consumption has a higher covariance with stock returns. Malloy, Moskowitz,

and Vissing-Jrgensen (2009) apply this idea to the cross-section of stock returns within a

31
long-run risks model in which the log SDF follows a homoskedastic version of (28). They

construct a stockholder consumption growth series from Consumer Expenditure Survey data

from 1982 to 2004. They are able to explain cross-sectional variation across size and B/M

portfolios with a much lower coefficient of relative risk aversion than with standard measures

of per-capita consumption growth.

Much of the trading activity in financial markets is undertaken by financial institutions,

not households. The idealized notion that agents optimally, frequently, and timely make

decisions according to their first-order conditions seems more likely to hold for financial

institutions than for households. Transaction costs, information costs, limited attention, and

behavioral biases on the part of households may reduce the informativeness of aggregate

consumption data about priced risks in financial markets. It therefore seems worthwhile to

explore the first-order conditions of financial institutions.

From a financial institutions viewpoint, the definition of good and bad states of the world

likely depends on factors that are not well captured by aggregate consumption. One important

concern for financial institutions seems to be liquidity. This reasoning motivated Holmstrom

and Tirole (2001), Acharya and Pedersen (2005), and Brunnermeier and Pedersen (2009) to

propose models in which the SDF is liquidity-state-dependent. The notions of liquidity are

different in each of these models (the marginal value of liquidity services of tradeable assets in

Holmstrom-Tirole; transaction costs in Acharya-Pedersen; funding liquidity in Brunnermeier-

c
Pedersen), but a general feature is that the consumption-based SDF Mt+1 is augmented with

a factor t that captures the state of aggregate liquidity

c t
Mt+1 = Mt+1 . (37)
t+1

Augmenting the standard Epstein-Zin-Weil SDF (7) in this way yields, in logs,

mt+1 = log ct+1 + ( 1)rw,t+1 t+1 , (38)

32
where t+1 log(t+1 /t ). In this model, assets that pay off well in states in which

liquidity is scarce (low t+1 ) trade at a price premium and have a low expected return.

Pastor and Stambaugh (2003) construct an empirical liquidity factor by measuring the

(illiquidity-driven) tendency of individual stock returns to revert at daily frequency. They find

a substantial spread in average returns between stocks that load differently on their liquidity

factor and a liquidity-factor-mimicking portfolio helps explain the momentum effect (see,

also, Sadka (2006)). Acharya and Pedersen (2005) construct a liquidity factor by aggregating

individual stocks Amihud (2002) illiquidity ratios. Adrian, Etula, and Muir (2012) propose

broker-dealer leverage as a proxy for t+1 . Their rationale is that de-leveraging indicates

deteriorating funding conditions and high marginal utility for financial intermediaries. A

single-factor SDF with this variable as the only factor does remarkably well in explaining the

joint cross-section of average returns on size, B/M , momentum, and bond portfolios.

Frazzini and Pedersen (2011) consider an economy in which some investors are leverage-

constrained (see, also, Baker, Bradley, and Wurgler (2011)). In this model, t reflects

the weighted-average Lagrange multiplier on the leverage constraints. In their model, all

securities have the same covariance with t+1 . Applied to the CAPM special case of (38)

in our log-normal framework, this yields

1 2 1 2
Et [rit+1 ] + it = rtf + t + it (Et [rt+1
w
] rtf t + it ) (39)
2 2

Thus, the zero-beta rate rtf + t is higher than rtf and the slope Et [rt+1
w ] r f + 1 2 of
t t 2 it

the risk-return tradeoff is flattened by t , as in Black (1972).14 Frazzini and Pedersen find

support for this prediction in a variety of international markets and asset classes.

With regards to leverage constraints, it is also worthwhile recalling the discussion of the

ICAPM in section 4.1. One unresolved problem in this literature is the assumption that the

investor whose Euler equation is used to price assets is assumed to be always fully invested
14
There is a wedge between the zero-beta rate and the risk-free rate in this model, because leverage-
constrained investors refrain from holding the risk-free asset and it is priced by unconstrained investors.

33
in the market portfolio, without timing the market, even though (exogenously specified)

investment opportunities are time-varying. Campbell, Giglio, Polk, and Turley (2012) point

out that one potential resolution could be to interpret the Euler equation as the first-order

condition of a leverage-constrained investor. The discussion here shows that this would likely

entail an additional factor in the SDF driven by the Lagrange multiplier on the constraint.

5 Sentiment

I now turn to empirical work on cross-sectional asset pricing that departs from the rational

expectations and Bayesian learning paradigms. Differences in opinion,15 excessive extrapola-

tion of growth, underreaction, limited attention are examples of such departures. Common

to these approaches is that the SDF Mt+1 that reflects sentiment-driven investors marginal

utilities prices assets under subjective beliefs, i.e., EtS [Mt+1 Rt+1 ] = 1. Under objective beliefs,

the pricing restriction Et [Mt+1 Rt+1 ] = 1 holds with

t+1
Mt+1 = Mt+1 , (40)
t

where t is the Radon-Nikodym derivative of the subjective probability measure with respect

to the objective one. Thus, for the econometricianwho works with data generated under

the objective measurethe challenge is still the same as in rational expectations theories:

specify and test specifications of the SDF. The only difference is that with sentiment effects

present, the econometrician has to entertain the possibility that subjective beliefs enter Mt+1

through t+1 /t . These relationships between pricing under the subjective and objective

measure have important implications for the interpretation of empirical tests.

Part of the empirical literature has focused on ad-hoc factor models (as in the debate

between Daniel and Titman (1997) and Davis, Fama, and French (2000) about charac-

teristics versus covariances) or on theoretical models with reduced-form SDFs (see the
15
Differences in opinion could also arise under Bayesian learning with non-common priors

34
discussion in Section 4.3) to discriminate between risk-based and sentiment explanations

of cross-sectional return predictability. Yet, without taking a stand on some theoretical re-

strictions on M , these approaches cannot shed light on this question. For example, suppose

that mt+1 log(Mt+1 ) follows a linear one-factor model, and log(t+1 /t ) has one-factor

representation where the single factor reflects aggregate sentiment fluctuations. In this case

mt+1 log(Mt+1 ) follows a linear two-factor model. As a consequence, it is possible that an

ad-hoc factor model perfectly explains expected returns, even though pricing is influenced by

sentiment.

Furthermore, risk- and sentiment-based explanations could be true simultaneously. Con-

sider an economy with two investors. Investor A holds subjective beliefs. As first-order con-

dition implies that EtS [Mt+1

A R A
t+1 ] = 1 holds under As subjective beliefs where Mt+1 reflects

the risk exposures taken on by A (e.g., high exposure to assets that A is optimistic about).
B R
Investor B is a sophisticated arbitrageur with rational expectations, and Et [Mt+1 t+1 ] = 1

B reflects the risk exposures taken on by B (e.g.,

holds under Bs objective beliefs where Mt+1

low exposure to assets that A is optimistic about). An econometricians examination of the

A would reveal that it misprices assets under the objective measure,
explanatory power of Mt+1
B , in
suggestive of a sentiment explanation. An examination of the explanatory power of Mt+1

contrast, would reveal that it prices assets correctly under the objective measure, consistent

with a risk-based explanation. Of course, both the sentiment and risk-based explanations are

true simultaneously in this example as both investors first-order conditions hold.

5.1 Limits to Arbitrage

For sentiment to have price impact, some limits to arbitrage must exist that prevent so-

phisticated investors from completely neutralizing the effect of sentiment-driven investors.16

If sentiment is correlated across investors, and if it affects groups of stocks or an entire

asset class in the same direction, there is a natural explanation: trading against sentiment
16
Gromb and Vayanos (2010) survey the theoretical literature on limits to arbitrage.

35
requires taking risky positions with undiversifiable risk. Sophisticated investors want to

be compensated for taking on such risk, which limits their willingness to trade aggressively

against sentiment. Many cross-sectional return predictability patterns offer fairly high Sharpe

Ratios, though, so this is unlikely to be a full explanation.

Frictions could constitute an additional impediment. Pontiff (1996) shows that the market

values of closed-end fund shares are more likely to deviate from their net asset value if an

arbitrage strategy is subject to greater residual risk and to greater costs for initiating and

holding the position. In a similar vein, a number of recent papers find that various cross-

sectional return predictability patterns are stronger among stocks subject to such frictions.

Short-sale constraints are one example. They are most likely to be relevant among stocks

with low institutional ownership, because institutional investors, and especially passive ones,

are the most active lenders of stocks to short sellers. Based on this idea, Nagel (2005)

shows that cross-sectional return predictability based on several predictors, including B/M ,

is stronger among stocks with low institutional ownership (controlling for size).

Small stocks and stocks without analyst coverage are typically not attractive investments

for professional investors with big portfolios. Griffin and Lemmon (2002) find that the value

premium is bigger among these stocks, and Hong, Lim, and Stein (2000) report stronger

momentum effects. Campbell, Hilscher, and Szilagyi (2008) find that the distress anomaly is

stronger for stocks with low analyst coverage, institutional ownership, and liquidity.

Cross-sectional predictability is typically stronger among stocks with high idiosyncratic

volatility; see for example Ali, Hwang, and Trombley (2003) for B/M , Lipson, Mortal, and

Schill (2011) and Lam and Wei (2011) for the asset growth anomaly, Mendenhall (2004) for

the post-earnings announcement drift, and Mashruwala, Rajgopal, and Shevlin (2006) for

the accrual anomaly. These results could indicate that idiosyncratic risk acts as a deterrent

to more aggressive exploitation of predictability by sophisticated investors (see, e.g., Pontiff

(2006)), but it is not clear whether this theory is quantitatively plausible. After all, strategies

based on these predictors offer relatively high Sharpe Ratios.

36
5.2 Predicting Returns with Cross-Sectional Sentiment Proxies

Some researchers have developed proxy measures for investors subjective beliefs, with the

goal of using these to predict returns in the cross-section.

Diether, Malloy, and Scherbina (2011) and Chen, Hong, and Stein (2002) build on the idea

in Miller (1977) that if investors with divergent beliefs face short-sale constraints, prices are

set by optimists, while pessimists sit on the sidelines with zero positions. Greater divergence

of beliefs then implies greater overpricing and lower future returns. Diether et al. proxy for

investors belief dispersion with analyst forecast dispersion, while Chen et al. proxy for it

with the (lack of) breadth of mutual fund ownership of a stock. Both find support for the

prediction that expected returns are low when dispersion is high.

La Porta (1996) uses analyst forecasts as a sentiment proxy and finds that stocks with

high forecasted growth perform poorly subsequently, especially around future earnings an-

nouncements, supporting the hypothesis of Lakonishok, Shleifer, and Vishny (1994) that the

value premium could arise from excessive extrapolation of firms past performance.

Teo and Woo (2004) proxy for investor sentiment with the returns of mutual funds in

small/large and growth/value styles. They find that high style returns in the past years pre-

dict low returns for stocks associated with these styles, consistent with the model of Barberis

and Shleifer (2003). Frazzini and Lamont (2008) identify stocks with positive sentiment as

those that are held by mutual funds that have recently received a lot of inflows. They show

that these stocks perform poorly in the future. Furthermore, positive-sentiment stocks tend

to be growth stocks and stocks with high equity issuance activity.

5.3 Cross-Sectional Implications of Time-Varying Aggregate Sentiment

Fluctuations in aggregate investor sentiment may have cross-sectional implications if different

types of stocks have different sensitivity to variations in aggregate sentiment.

Lemmon and Portniaguina (2006) find that expected returns of small stocks and stocks

with low institutional ownership are low when the University of Michigan consumer senti-

37
ment index is high. Baker and Wurgler (2006) construct an index from several aggregate

sentiment proxies from the prior literature. They find that expected returns of stocks that

are unattractive to arbitrageurs (young, small, unprofitable, extreme-growth, distressed, and

high volatility stocks) are most sensitive to sentiment. Stambaugh, Yu, and Yuan (2012) use

the Baker-Wurgler index to examine time-variation in a broad set of existing cross-sectional

predictability patterns (most of those discussed in sections 2.2 and 2.3). They find that long-

short strategies constructed from these predictors have low returns following high aggregate

sentiment, and most of this time-variation comes from the short component of the portfolio,

consistent with a short-sale constraints explanation. It is noteworthy that none of these pa-

pers finds that the value premium can be predicted with aggregate sentiment proxies. An

exception is Ben-Rephael, Kandel, and Wohl (2012). They construct an aggregate sentiment

index from a measure of investor reallocation between equity and bond funds and find that

high sentiment predicts low returns most strongly among small stocks and growth stocks.

The above studies focus on directional measures of sentiment, but some researchers have

also explored aggregate measures of belief dispersion. Cross-sectional implications arise if

belief dispersion for individual stocks is highly sensitive to fluctuations in aggregate belief

dispersion for some stocks, and less for others, or if some stocks are more likely to be affected

by short-sale constraints than others. Yu (2011) constructs an aggregate measure of belief

dispersion from individual stock analyst forecasts and finds that high dispersion forecasts a

strong value premium. Hong and Sraer (2012) argue that high-beta stocks have the highest

exposure to aggregate belief dispersion. As a consequence, a positive relationship between risk

(beta) and return arises only in times of low belief dispersion. In times of high dispersion, this

relationship follows an inverted U-shape. They find empirical support for these predictions.

This theory is an alternative (based on heterogeneity in beliefs) to the theory of Frazzini and

Pedersen (2011) (based on heterogeneity in risk aversion and leverage constraints) to explain

the low average returns of high beta stocks discussed in section 4.5.

38
5.4 Limited Attention

A different line of work examines whether cross-sectional return predictability could arise

from investors failure to pay timely attention to valuation-relevant pieces of information. Sims

(2003) suggested limited attention as an explanation of inertia in macroeconomic variables,

and Peng and Xiong (2006) and Hirshleifer, Lim, and Teoh (2011) apply these ideas in models

of cross-sectional asset pricing. Sloan (1996) and Hirshleifer, Hou, Teoh, and Zhang (2004))

suggest limited attention as an explanation of the accrual anomaly: investors seem to pay

insufficient attention to the heterogeneity in the time-series properties of different earnings

components.

Evidence in a number of empirical studies is consistent with investors paying limited

attention to relatively complex information. DellaVigna and Pollet (2007) show that de-

mographic information can predict industry profitability far in the future, but the market

seems to be slow in incorporating this information, resulting in industry-return predictabil-

ity. Cohen and Frazzini (2008) examine economically linked firms and find that the market

does not fully incorporate these economic links when responding to news. Following a simi-

lar logic, Menzly and Ozbas (2010) find cross-industry return predictability for economically

linked industries. Cohen and Lou (2011) empirically categorize firms into complicated and

easy-to-analyze ones and find that the returns of easy-to-analyze firms predict the returns of

their more complicated peers. Belo, Gala, and Li (2012) document predictable variation in

cash-flows and expected returns related to political cycles.

Hou, Peng, and Xiong (2009) look at aggregate proxies of investor attention. They argue

that high trading volume and being in an up-market proxies for high investor attention,

and they find that earnings momentum (interpreted as underreaction) weakens with investor

attention while price momentum (interpreted as overreaction) strengthens.

39
6 Out-of-Sample Tests

The asset-pricing approaches that I have discussed so far may be too stark in their treatment

of beliefs. Rational expectations theories assume that investors act in equilibrium as if they

knew the specification and parameters of the equilibrium law of motion that generates the

data. This may be a useful benchmark, but real-world investors have to learn the data-

generating process. To the extent regimes change, parameters drift, and memory may get

lost, it is not clear how fast, if at all, this eventual convergence to rational expectations

happens. Sentiment theories on the other hand rest on the notion that investors have some

fixed biases and attention deficits. Real-world investors may indeed have biases and attention

deficits, but they are probably not be completely immune to learning either.

If investors learn and adapt, this has implications for cross-sectional return predictability

(see Lo (2004) for an exposition of this view). There are still many open questions about how

learning affects asset prices, and the body of empirical literature that deals with these issues

is still quite small. Nevertheless, some progress has been made. Much of this work employs

out-of-sample testing methods.

Out-of-sample tests can also help address data-snooping concerns. I review empirical

work that speaks to the data-snooping issue after the discussion of learning.

6.1 Learning and Out-of-Sample Evaluation

Intuitively, one might think that if learning is driving return predictability, this predictability

should not exist out of sample (where out of sample here is meant to be forward in time,

not backwards or cross-country). But it is not as simple as that. As an illustration of the

problem, consider the model of Lewellen and Shanken (2002) (LS). LS set up a model in which

a risk-neutral Bayesian investor observes dividends drawn from an IID normal distribution

with mean . The investor initially starts with a diffuse prior. The gross risk-free rate is

constant at Rf . After observing t observations of dividends, the investors posterior mean

e
is the sample mean Dt . Now define the excess dollar return Rt+1 Pt+1 + Dt+1 Rf Pt .

40
Combining this definition with the valuation relation Pt = Dt /(Rf 1) yields

 
e 1
Rt+1 = 1+ (Dt+1 Dt ). (41)
(t + 1)(Rf 1)

From the investors viewpoint, EtB [Dt+1 ] = Dt and so EtB [Rt+1

e ] = 0, i.e., excess returns

are completely unforecastable in real time. However, an econometrician examining a trading

strategy that buys Dt units of the stock at t, financed at Rf , finds that it has an expected

return of
2
 
e 1
E[Rt+1 Dt ] = 1+ <0 (42)
(t + 1)(Rf 1) t

This trading strategy uses only information available in real time, and hence it is imple-

mentable out of sample, yet it earns non-zero expected returns. The reason is that the

econometricians evaluation of the population moment in (42) (or its large sample approx-

imation) uses information about that the investor does not have in real time (at time t

the investors best guess of is Dt ). Lack of out-of-sample forecastability is therefore not

generally a suitable null hypothesis for testing whether return predictability is generated by

learning.

As the expression (42) shows, the magnitude of the trading strategys expected excess

return declines with the number of periods that the investor has learned the dividend process.

This time-decay could potentially provide a signature of learning that one could look for in

the data. The form of the time-decay will typically depend on the specifics of the process that

the investors are learning about, but the prediction that return predictability in out-of-sample

tests is weaker than in original in-sample tests may be a robust one.

The learning problem faced by real-world investors may be of a somewhat different nature,

though. In Bayesian learning models, the agent can costlessly pay attention to any number

of variables that might be relevant (within the set of models entertained by the agent).

In practice, even sophisticated investors and their computer algorithms can pay attention

only selectively to a limited set of variables. Over time they may discover new ones, and they

41
form an opinion about their relevance. Schwartzstein (2012) provides a model that formalizes

this type of learning under selective attention. In his model, an agent pays attention to a

predictor z only with a certain probability that depends on the agents belief that z is an

important predictor. Eventually, as the agent updates the belief on the importance of z,

the agent will learn to pay attention to z, but the learning process may take a long time.

Applied to cross-sectional asset-pricing, learning under selective attention would mean that

(abnormal) return predictability associated with a predictor z can exist as long as investors

pay insufficient attention to z. Once attention picks up (for example, when an academic

study publicizes the predictive power of z), investors quickly incorporate z into their return-

forecasting models and the predictability associated with z disappears. Out-of-sample tests

that check for post-publication predictability can shed light on this type of learning.

A number of studies have examined whether the evidence for cross-sectional return pre-

dictability is weaker out of sample, subsequent to the publication of academic studies. Dimson

and Marsh (1999) find that the small-firm premium in the UK was not evident from 1989-

1997 after studies publicized the UK small firm premium. Schwert (2003) finds that the size

and value effects in the US disappeared during the 1994-2002 period. Green, Hand, and

Soliman (2011) find that the accrual anomaly has disappeared post publication. In contrast,

Jegadeesh and Titman (2001) find that momentum profits continued to exist in the years

following the publication of Jegadeesh and Titman (1993).

One problem with these analyses is that the out-of-sample period is short, so it is difficult

to distinguish true disappearance of predictability from disappearance by chance. One possi-

ble remedy is to consider many return predictors simultaneously. McLean and Pontiff (2012)

study the post-publication return predictability associated with 82 characteristics that aca-

demic studies had identified. They estimate that the average decay in return predictability

is about 35%.

The studies above are true out-of-sample studies. They use data samples that were not

yet available to the researchers that published the original research. Decay in predictability

42
can also be studied with pseudo out-of-sample tests in which the researcher splits a data

sample into a training sample (used to estimate a predictive relationship) and a forecasting

sample (used to evaluate the predictive performance of the forecasting model). Haugen and

Baker (1996) use this approach and find substantial pseudo out-of-sample predictability for

a variety of predictors. Lewellen (2011) constructs a summary measure for expected returns

from 15 predictors. Using recursively expanding windows as training samples and a one-

month ahead horizon for forecasting samples, he finds a decay in predictive ability of about

20-30% compared with in-sample estimates. Cooper, Gutierrez, and Marcum (2005) find only

small out-of-sample predictability with a strategy that picks the best performing size, B/M ,

and momentum-sorted portfolios in 10-year training samples. Chordia, Subrahmanyam, and

Qing (2011) find that momentum and turnover-related predictability is weaker among liquid

stocks in the second half of the 1976-2009 period.

Overall, the evidence indicates that predictability tends to be somewhat weaker out of

sample, but, for most cross-sectional predictors, the predictability is still substantial, even

out of sample. To what extent the magnitude of the decay is consistent with learning is still

an open question.

6.2 Data Snooping and Out-of-Sample Evaluation

Weaker predictability in true out-of-sample tests can also be an indication of data-snooping

problems in the original studies. Correlations between a predictor z found in sample could

reflect an in-sample correlation between z and noise rather than correlation of z with the

true predictable component of returns. If econometricians undertake specification searches

looking for predictors with significant predictive power, conventional procedures of statisti-

cal inference do not account for this specification search and hence overstate the statistical

significance. One approach to this problem is to adjust critical values of test statistics for

the specification search, as proposed in Lo and MacKinlay (1990) and White (2000), but im-

plementing these methods may be difficult when specification searches occur in an informal

43
manner or through sequential investigations by different researchers.

True out-of-sample tests can help address data-snooping concerns. If researchers found an

in-sample relationship that happens to be just an in-sample correlation between a predictor

z and noise, there is no reason for the predictive relationship to exist in a data sample that

was not available to the researchers that published the initial results. In contrast, pseudo

out-of-sample tests are not a solution to the data-snooping problem, because researchers can

mine the data for significant pseudo out-of-sample predictability in the same way as it can

be mined for in-sample predictability.

In addition to the post-publication evidence discussed above, such true out-of-sample

evidence exists for a number of cross-sectional predictors. Davis, Fama, and French (2000)

collect new accounting data and to construct B/M sorted portfolios prior to the 1960s and

they find a robust value premium. Dimson, Nagel, and Quigley (2003) do a similar exercise

to extend UK data back into the 1950s and they find a strong value premium. Other studies

have examined data from non-U.S. stock markets and confirmed the existence of a value

premium (Fama and French (1998)), momentum (Rouwenhorst (1998)), the idiosyncratic

volatility effect (Ang, Hodrick, Xing, and Zhang (2009)), the equity issuance effect (McLean,

Pontiff, and Watanabe (2009)), the asset growth anomaly (Watanabe, Yu, and Xu (2011))

and value and momentum effects in various asset classes internationally (Asness, Moskowitz,

and Pedersen (2012)). In contrast, Leippold and Lohre (2012) find that the accrual anomaly

is not a globally robust feature of stock returns. Thus, with few exceptions, the evidence

indicates that cross-sectional return predictability uncovered in earlier studies is unlikely to

be simply the result of data snooping.

7 Concluding Remarks

I conclude by highlighting a few key issues that emerged in this survey.

1. A substantial body of evidence has now moved us beyond Fama and French (1996).

44
 Empirical studies have found a number of robust cross-sectional return predictability

patterns that are not subsumed by size and B/M and by the Fama-French three-

factor model. For researchers using ad-hoc factor models as benchmarks in abnormal

return calculations, there is now little empirical justification for relying on the Fama-

French factors, rather than a factor model that incorporates these other sources of

cross-sectional return predictability as well.

2. Ad-hoc factor models are useful tools for summarizing the main dimensions of cross-

sectional return predictability, but they cannot be used to test risk-based rational ex-

pectations explanations of this predictability against sentiment explanations. Without

restrictions on investor preferences, and hence on the risks that are relevant to investors,

empirical tests cannot discriminate between these competing explanations. The same

comment applies to production-based models with an exogenously specified reduced-

form SDF. Recent work on the ICAPM and long-run risk models that employs such

restrictions shows more promise to make progress on this question than the ad-hoc

factor model literature.

3. The view of sentiment- and risk-based explanations as a dichotomy may be false to

begin with. First, in a market in which sentiment-driven investors trade with rational

investors, systematic sentiment can affect prices, yet rational investors first-order con-

dition holds as well. Thus, a sentiment- and a risk-based explanation of cross-sectional

return predictability can be true simultaneously. It is an interesting question whether

recent empirical successes of the ICAPM could perhaps be interpreted in this way.

Second, theories of learning and adaptation fall neither into the rational expectations

nor the sentiment category. As real-world investors learn from data in similar ways

as econometricians do, these theories may deserve greater attention in the study of

cross-sectional asset pricing.

4. The choice of moment conditions is crucial for econometric evaluation to be informative

45
 about the performance of a model. In particular, fitting SDFs with time-varying prices

of risk to the cross-section of expected returns only, without attention to conditional

moment restrictions, can produce implied time-variation in conditional moments that

is grossly inconsistent with the data.

References

Acharya, V., and L. H. Pedersen. 2005. Asset Pricing with Liquidity Risk. Journal of

Financial Economics 77:374410.

Adrian, T., E. Etula, and T. Muir. 2012. Financial Intermediaries and the Cross-Section

of Asset Returns. Working Paper, Northwestern University and Federal Reserve Bank

of New York.

Ali, A., L. S. Hwang, and M. A. Trombley. 2003. Arbitrage Risk and the Book-to Market

Anomaly. Journal of Financial Economics 69:355373.

Amihud, Y. 2002. Illiquidity and Stock Returns: Cross-Section and Time-Series Effects.

Journal of Financial Markets 5:3156.

Anderson, C. W., and L. Garcia-Feijoo. 2006. Empirical Evidence on Capital Investment,

Growth Options and Security Returns. Journal of Finance 61:171194.

Ang, A., R. J. Hodrick, Y. Xing, and X. Zhang. 2006. The Cross-Section of Volatility and

Expected Returns. Journal of Finance 61:259299.

Ang, A., R. J. Hodrick, Y. Xing, and X. Zhang. 2009. High Idiosyncratic Volatility and Low

Returns: International and Further U.S. EvidenceJournal of Financial Economics 91:1

23.

Ang, A., and D. Kristensen. 2012. Testing conditional factor modelsJournal of Financial

Economics 106:132 156.

Asness, C., T. Moskowitz, and L. H. Pedersen. 2012. Value and Momentum Everywhere.

46
 Journal of Finance:forthcoming.

Baker, M., B. Bradley, and J. Wurgler. 2011. Benchmarks as Limits to Arbitrage: Under-

standing the Low-Volatility AnomalyFinancial Analysts Journal 67:4054.

Baker, M., and J. Wurgler. 2006. Investor Sentiment and the Cross-Section of Stock Re-

turns. Journal of Finance 61:16451680.

Bansal, R., R. Dittmar, and C. Lundblad. 2005. Consumption, Dividends, and the Cross

Section of Equity Returns. Journal of Finance 60:16391672.

Bansal, R., D. Kiku, I. Shaliastovich, and A. Yaron. 2012. Volatility, the Macro- economy

and Asset Prices. Working Paper, Duke University and University of Pennsylvania.

Bansal, R., D. Kiku, and A. Yaron. 2007. Risks for the Long Run: Estimation and Inference.

Working Paper, Duke University.

Bansal, R., and A. Yaron. 2004. Risks for the Long Run: A Potential Resolution of Asset

Pricing Puzzles. Journal of Finance 59:14811509.

Barberis, N., and A. Shleifer. 2003. Style InvestingJournal of Financial Economics 68:161

199.

Barro, R. J. 2006. Rare Disasters and Asset Markets in the Twentieth CenturyQuarterly

Journal of Economics 121:823866.

Belo, F. 2010. Production-Based Measures of Risk for Asset Pricing. Journal of Monetary

Economics 57:146163.

Belo, F., V. D. Gala, and J. Li. 2012. Government Spending, Political Cycles and the

Cross-Section of Stock Returns. Journal of Financial Economics:forthcoming.

Ben-Rephael, A., S. Kandel, and A. Wohl. 2012. Measuring Investor Sentiment with Mutual

Fund FlowsJournal of Financial Economics 104:363 382.

Berk, J. B., R. C. Green, and V. Naik. 1999. Optimal Investment, Growth Options and

Security Returns. Journal of Finance 54:15531607.

47
Bernard, V. L., and J. K. Thomas. 1989. Post-Earnings-Announcement Drift: Delayed

Price Response or Risk Premium? Journal of Accounting Research 27:136.

Black, F. 1972. Capital Market Equilibrium with Restricted Borrowing. Journal of Busi-

ness 45:444455.

Bray, M. 1982. Learning, Estimation and the Stability of Rational Expectations Equilibria.

Journal of Economic Theory 26:318339.

Brennan, M. J., A. W. Wang, and Y. Xia. 2004. Estimation and Test of a Simple Model

of Intertemporal Capital Asset Pricing. Journal of Finance 59:17431775.

Brunnermeier, M. K., and L. H. Pedersen. 2009. Market Liquidity and Funding Liquidity.

Review of Financial Studies 22:22012238.

Campbell, J. Y. 1993. Intertemporal Asset Pricing Without Consumption Data. American

Economic Review 83:487512.

Campbell, J. Y., and J. H. Cochrane. 1999. By Force of Habit: A Consumption-Based

Explanation of Aggregate Stock Market BehaviorJournal of Political Economy 107:205

251.

Campbell, J. Y., S. Giglio, C. Polk, and R. Turley. 2012. An Intertemporal CAPM with

Stochastic Volatility. Working Paper, Harvard University.

Campbell, J. Y., J. Hilscher, and J. Szilagyi. 2008. In Search of Distress Risk. Journal of

Finance 63:28992939.

Campbell, J. Y., and R. J. Shiller. 1988. The Dividend-Price Ratio and Expectations of

Future Dividends and Discount Factors. Review of Financial Studies 1:195228.

Campbell, J. Y., and T. Vuolteenaho. 2004. Bad Beta, Good Beta. American Economic

Review 94:12491275.

Carhart, M. M. 1997. On Persistence of Mutual Fund Performance. Journal of Fi-

nance 52:5782.

48
Carlson, M., A. Fisher, and R. Giammarino. 2004. Corporate Investment and Asset

Price Dynamics: Implications for the Cross-section of ReturnsThe Journal of Fi-

nance 59:25772603.

Chamberlain, G. 1987. Asymptotic Efficiency in Estimation with Conditional Moment

Restrictions. Journal of Econometrics 34:305344.

Chan, L. K., N. Jegadeesh, and J. Lakonishok. 1996. Momentum Strategies. Journal of

Finance 51:16811713.

Chen, J., H. Hong, and J. C. Stein. 2002. Breadth of Ownership and Stock Returns. Journal

of Financial Economics 66:171205.

Chen, L., and X. Zhao. 2009. Return Decomposition. Review of Financial Studies 22:5213

5249.

Chordia, T., A. Subrahmanyam, and T. Qing. 2011. Trends in the Cross-Section of Ex-

pected Returns. Working Paper, UCLA.

Cochrane, J. H. 1991. Production-Based Asset Pricing and the Link Between Stock Returns

and Economic Fluctuations. Journal of Finance 46:209237.

Cochrane, J. H. 1996. A Cross-Sectional Test of an Investment-Based Asset Pricing Model.

Journal of Political Economy 104:572621.

Cochrane, J. H., and J. Saa-Requejo. 2000. Beyond Arbitrage: Good-Deal Asset Price

Bounds in Incomplete MarketsJournal of Political Economy 108:79119.

Cohen, L., and A. Frazzini. 2008. Economic Links and Predictable Returns. Journal of

Finance 63:19772011.

Cohen, L., and D. Lou. 2011. Complicated Firms. Journal of Financial Eco-

nomics:forthcoming.

Cohen, R. B., C. Polk, and T. Vuolteenaho. 2003. The Value Spread. Journal of Fi-

nance 58:609641.

49
Collin-Dufresne, P., M. Johannes, and L. A. Lochstoer. 2012. Parameter Learning in Gen-

eral Equilibrium: The Asset Pricing Implications. Working Paper, Columbia Business

School.

Cooper, M., H. Gulen, and M. Schill. 2008. Asset Growth and the Cross-Section of Stock

Returns. Journal of Business 63:16091652.

Cooper, M., R. C. Gutierrez, and B. Marcum. 2005. On the Predictability of Stock Returns

in Real Time. Journal of Business 78:469500.

Daniel, K., and S. Titman. 1997. Evidence on the Characteristics of Cross Sectional Vari-

ation in Stock Returns. Journal of Finance 52:133.

Daniel, K., and S. Titman. 2006. Market Reactions to Tangible and Intangible Information.

Journal of Finance 61:16051643.

Daniel, K., and S. Titman. 2012. Testing Factor-Model Explanations of Market Anomalies.

Critical Finance Review 1:103139.

Davis, J., E. Fama, and K. French. 2000. Characteristics, Covariances, and Average Re-

turns: 1929 to 1997. Journal of Finance 55:389406.

DeBondt, W. F., and R. Thaler. 1985. Does the Stock Market Overreact? Journal of

Finance 40:793805.

DeBondt, W. F., and R. Thaler. 1987. Further Evidence on Investor Overreaction and

Stock Market Seasonality. Journal of Finance 42:557581.

DellaVigna, S., and J. M. Pollet. 2007. Demographics and Industry Returns. American

Economic Review 97:16671702.

Dichev, I. D. 1998. Is the Risk of Bankruptcy a Systematic Risk? Journal of Fi-

nance 53:11311147.

Diether, K. B., C. J. Malloy, and A. Scherbina. 2011. Differences in Opinion and the Cross

Section of Stock Returns. Journal of Finance 57:21132140.

50
Dimson, E., and P. Marsh. 1999. Winter. Murphys Law and Market Anomalies. Journal

of Portfolio Management 26:5369.

Dimson, E., S. Nagel, and G. Quigley. 2003. November/December. Capturing the Value

Premium in the United Kingdom. Financial Analysts Journal 59:3545.

Dittmar, R. F. 2002. Nonlinear Pricing Kernels, Kurtosis Preference, and Evidence from

the Cross Section of Equity ReturnsThe Journal of Finance 57:pp. 369403.

Dumas, B., A. Kurshev, and R. Uppal. 2009. Equilibrium Portfolio Strategies in the Pres-

ence of Sentiment Risk and Excess Volatility. Journal of Finance 64:579629.

Epstein, L., and S. Zin. 1989. Substitution, Risk Aversion, and the Temporal Behavior of

Consumption and Asset Returns: An Theoretical Framework. Econometrica 57:937

969.

Epstein, L., and S. Zin. 1991. Substitution, Risk Aversion, and the Temporal Behavior

of Consumption and Asset Returns: An Empirical Investigation. Journal of Political

Economy 99:263286.

Fairfield, P., J. S. Whisenant, and T. Yohn. 2003. Accrued Earnings and Growth: Implica-

tions for Future Profitability and Market Mispricing. Accounting Review 78:353371.

Fama, E. 1970. Efficient Capital Markets: A Review of Theory and Empirical Work. Jour-

nal of Finance 25:383417.

Fama, E. F., and K. R. French. 1992. The Cross-Section of Expected Stock Returns. Journal

of Finance 47:427465.

Fama, E. F., and K. R. French. 1993. Common Risk Factors in the Returns on Stocks and

Bonds. Journal of Financial Economics 33:2349.

Fama, E. F., and K. R. French. 1996. Mulitifactor Explanations of Asset Pricing Anomalies.

Journal of Finance 51:5587.

Fama, E. F., and K. R. French. 1998. Value versus Growth: The International Evidence.

51
 Journal of Finance 53:19751999.

Fama, E. F., and K. R. French. 2006. Profitability, Investment and Average Returns.

Journal of Financial Economics 82:491518.

Fama, E. F., and K. R. French. 2008. Average Returns, B/M and Share Issues. Journal of

Finance 63:29712995.

Frazzini, A., and O. A. Lamont. 2008. Dumb Money: Mutual Fund Flows and the Cross-

Section of Stock Returns. Journal of Financial Economics 88:299322.

Frazzini, A., and L. H. Pedersen. 2011. Betting Against Beta. Working Paper, NYU.

Gibbons, M., S. Ross, and J. Shanken. 1989. A Test of the Efficiency of a Given Portfolio.

Econometrica 57:11211152.

Giovannini, A., and P. Weil. 1989. Risk Aversion and Intertemporal Substitution. Working

Paper, NBER.

Gomes, J., L. Kogan, and L. Zhang. 2003. Equilibrium Cross-Section of Returns. Journal

of Political Economy 111:693732.

Goyal, A. 2012. Empirical Cross-Sectional Asset Pricing: A SurveyFinancial Markets and

Portfolio Management 26:338.

Green, J., J. R. M. Hand, and M. T. Soliman. 2011. Going, Going, Gone? The Apparent

Demise of the Accruals AnomalyManagement Science 57:797816.

Griffin, J. M., and M. L. Lemmon. 2002. Book-to-Market Equity, Distress Risk and Stock

Returns. Journal of Finance 57:23172336.

Gromb, D., and D. Vayanos. 2010. Limits to Arbitrage. Annual Review of Financial Eco-

nomics 2:251275.

Hansen, L. P. 1982. Large Sample Properties of Generalized Method of Moments Estima-

tors. Econometrica 50:10291054.

52
Hansen, L. P. 1985. A Method for Calculating Bounds on the Asymptotic Covariance Ma-

trices of Generalized Method of Moments Estimators. Journal of Econometrics 30:203

238.

Hansen, L. P., J. C. Heaton, and N. Li. 2008. Consumption Strikes Back? Measuring

Long-Run Risk. Journal of Political Economy 116:260302.

Hansen, L. P., and R. Jagannathan. 1991. Implications of Security Market Data for Models

of Dynamic Economies. Journal of Political Economy 99:225262.

Hansen, L. P., and R. Jagannathan. 1997. Assessing Specification Errors in Stochastic

Discount Factor Models. Journal of Finance 52:557590.

Hansen, L. P., and S. F. Richard. 1987. The Role of Conditioning Information in Deducing

Testable Restrictions Implied by Dynamic Asset Pricing Models. Econometrica 55:587

613.

Harvey, C. R., and A. Siddique. 2000. Conditional Skewness in Asset Pricing TestsJournal

of Finance 60:12631295.

Haugen, R. A., and L. Baker, Nardin. 1996. Commonality in the Determinants of Expected

Stock Returns. Journal of Financial Economics 41:401439.

Hirshleifer, D., K. Hou, S. H. Teoh, and Y. Zhang. 2004. Do Investors Overvalue Firms

with Bloated Balance Sheets. Journal of Accounting and Economics 38:297331.

Hirshleifer, D., S. S. Lim, and S. H. Teoh. 2011. Limited Investor Attention and Stock

Market Misreactions to Accounting Information. Review of Asset Pricing Studies 1:35

73.

Holmstrom, B., and J. Tirole. 2001. LAPM: A Liquidity-Based Asset Pricing Model. Jour-

nal of Finance 56:18371867.

Hong, H., T. Lim, and J. C. Stein. 2000. Bad News Travels Slowly: Size, Analyst Coverage,

and the Profitability of Momentum StrategiesJournal of Finance 55:265295.

53
Hong, H., and D. A. Sraer. 2012. Speculative Betas. Working Paper, Princeton University.

Hou, K., L. Peng, and W. Xiong. 2009. A Tale of Two Anomalies: The Implications

of Investor Attention for Price and Earnings Momentum. Working Paper, Princeton

University.

Hou, K., C. Xue, and L. Zhang. 2012. Digesting Anomalies. Working Paper, The Ohio

State University.

Huang, W., Q. Liu, S. G. Rhee, and L. Zhang. 2010. Return Reversals, Idiosyncratic Risk

and Expected Returns. Review of Financial Studies 23:147168.

Jagannathan, R., G. Skoulakis, and Z. Wang. 2009. The Analysis of the Cross Section of

Security Returns. In Handbook of Financial Econometrics, ed. Y. At-Sahalia and L. P.

Hansen. Volume 2, pp. 73134.

Jagannathan, R., and Z. Wang. 1996. The Conditional CAPM and the Cross-section of

Expected Returns. Journal of Finance 51:354.

Jagannathan, R., and Z. Wang. 2002. Empirical Evaluation of Asset-Pricing Models: A

Comparison of the SDF and Beta MethodsThe Journal of Finance 57:pp. 23372367.

Jegadeesh, N. 1990. Evidence of Predictable Behavior of Security Returns. Journal of

Finance 45:881898.

Jegadeesh, N., and S. Titman. 1993. Returns to Buying Winners and Selling Losers: Im-

plications for Market Efficiency. Journal of Finance 48:6591.

Jegadeesh, N., and S. Titman. 2001. Profitability of Momentum Strategies: An Evaluation

of Alternative Explanations. Journal of Finance 56:699720.

Jensen, M. C. 1978. Some Anomalous Evidence Regarding Market Efficiency. Journal of

Financial Economics 6:95101.

Kan, R., and C. Robotti. 2008. Specification Tests of Asset Pricing Models Using Ex-

cess Returnscation Tests of Asset Pricing Models Using Excess Returns. Journal of

54
 Empirical Finance 15:816838.

La Porta, R. 1996. Expectations and the Cross-Section of Stock ReturnsThe Journal of

Finance 51:17151742.

Lakonishok, J., A. Shleifer, and R. W. Vishny. 1994. Contrarian Investment, Extrapolation

and Risk. Journal of Finance 49:15411578.

Lam, F. E. C., and K. J. Wei. 2011. Limits-to-Arbitrage, Investment Frictions, and the

Asset Growth AnomalyJournal of Financial Economics 102:127149.

Lee, C. M., and B. Swaminathan. 2000. Price Momentum and Trading Volume. Journal of

Finance 55:20172069.

Leippold, M., and H. Lohre. 2012. Data Snooping and the Global Accrual Anomaly. Applied

Financial Economics 22:509535.

Lemmon, M., and E. Portniaguina. Winter 2006. Consumer Confidence and Asset Prices:

Some Empirical EvidenceReview of Financial Studies 19:14991529.

Lettau, M., and S. C. Ludvigson. 2001. Resurrecting the (C)CAPM: A Cross-Sectional Test

When Risk Premia Are Time-Varying. Journal of Political Economy 109:12381287.

Lettau, M., and J. A. Wachter. 2007. Why Is Long-Horizon Equity Less Risky? A Duration-

Based Explanation of the Value PremiumJournal of Finance 62:5592.

Lewellen, J. 2011. The Cross Section of Expected Stock Returns. Working Paper, Dart-

mouth College.

Lewellen, J., and S. Nagel. 2006. The Conditional CAPM Does Not Explain Asset Pricing

Anomalies. Journal of Financial Economics 79:289314.

Lewellen, J., S. Nagel, and J. Shanken. 2010. A Skeptical Appraisal of Asset-Pricing Tests.

Journal of Financial Economics 96:175194.

Lewellen, J., and R. J. Resutek. 2012. The Predictive Power of Investments and Accruals.

Working Paper, Dartmouth College.

55
Lewellen, J., and J. Shanken. 2002. Learning, Asset-Pricing Tests and Market Efficiency.

Journal of Finance 57:11131145.

Li, E. X., D. Livdan, and L. Zhang. 2009. Anomalies. Review of Financial Studies 22:4301

4334.

Li, Q., M. Vassalou, and Y. Xing. 2006. Sector Investment Growth Rates and the Cross

Section of Equity Returns. Journal of Busines:16371665.

Liew, J., and M. Vassalou. 2000. Can Book-to-Market, Size and Momentum be Risk Factors

that Predict Economic Growth? Journal of Financial Economics 57:221245.

Lipson, M. L., S. Mortal, and M. Schill. 2011. On the Scope and Drivers of the Asset

Growth EffectJournal of Financial and Quantitative Analysis 46:16511682.

Lo, A. W. 2004. The Adaptive Markets Hypothesis. Journal of Portfolio Manage-

ment 30:1529.

Lo, A. W., and A. C. MacKinlay. 1990. Data-Snooping Biases in Tests of Financial Asset

Pricing Models. Review of Financial Studies 3:431467.

Lucas, R. 1978. Asset Prices in an Exchange Economy. Econometrica 46:14291445.

Malloy, C. J., T. J. Moskowitz, and A. Vissing-Jrgensen. 2009. Long-Run Stockholder

Consumption Risk and Asset Returns. Journal of Finance 64:24272479.

Mankiw, N. G., and S. Zeldes. 1991. The Consumption of Stockholders and Nonstockhold-

ersJournal of Financial Economics 29:97112.

Marcet, A., and J. Sargent, Thomas. 1989. Convergence of Least Squares Learning

Mechanisms in Self-Referential Linear Stochastic Models. Journal of Economic The-

ory 48:337368.

Mashruwala, C., S. Rajgopal, and T. Shevlin. 2006. Why is the accrual anomaly not arbi-

traged away? The role of idiosyncratic risk and transaction costs. Journal of Accounting

and Economics 42:3 33.

56
McLean, D. R., and J. Pontiff. 2012. Does Academic Research Destroy Stock Return Pre-

dictability? Working Paper, Boston College.

McLean, D. R., J. Pontiff, and A. Watanabe. 2009. Share Issuance and Cross-Sectional

Returns: International Evidence. Journal of Financial Economics 94:117.

Mendenhall, R. 2004. Arbitrage risk and post-earnings announcement drift. Journal of

Business 77:875894.

Menzly, L., and O. Ozbas. 2010. Market Segmentation and Cross-predictability of Re-

turnsThe Journal of Finance 65:15551580.

Merton, R. C. 1973. An Intertemporal Capital Asset Pricing ModelEconometrica 41:867

887.

Merton, R. C. 1987. A Simple Model of Capital Market Equilibrium with Incomplete

Information. Journal of Finance 42:483510.

Miller, E. M. 1977. Risk, Uncertainty and Divergence of Opinion. Journal of Fi-

nance 32:11511168.

Muth, J. F. 1961. Rational Expectations and the Theory of Price Movements. Economet-

rica 29:123.

Nagel, S. 2005. Short Sales, Institutional Investors and the Cross-Section of Stock Returns.

Journal of Financial Economics 78:277309.

Nagel, S., and K. J. Singleton. 2011. Estimation and Evaluation of Conditional Asset

Pricing Models. Journal of Finance 66:873909.

Novy-Marx, R. 2011. Operating LeverageReview of Finance 15:103134.

Novy-Marx, R. 2012a. The Other Side of Value: The Gross Profitability Premium. Journal

of Financial Economics:forthcoming.

Novy-Marx, R. 2012b. Pseudo-Predictability in Conditional Asset Pricing Tests: Explain-

ing Anomaly Performance with Politics, the Weather, Global Warming, Sunspots, and

57
 the Stars. Working Paper, NBER.

Ohlson, J. A. 2005. On Accounting-Based Valuation Formulae. Review of Accounting Stud-

ies 10:323347.

Papanikolaou, D. 2011. Investment Shocks and Asset Prices. Journal of Political Econ-

omy 119:639685.

Parker, J. A., and C. Julliard. 2005. Consumption Risk and the Cross Section of Expected

Returns. Journal of Political Economy 113:185222.

Pastor, L., and R. F. Stambaugh. 2003. Liquidity Risk and Expeced Stock Returns. Journal

of Political Economy 111:642685.

Peng, L., and W. Xiong. 2006. Investor Attention, Overconfidence and Category Learn-

ingJournal of Financial Economics 80:563 602.

Pontiff, J. 1996. Costly Arbitrage: Evidence from Closed-End FundsQuarterly Journal of

Economics 111:11351151.

Pontiff, J. 2006. Costly Arbitrage and the Myth of Idiosyncratic Risk. Journal of Accounting

and Economics 42:3552.

Pontiff, J., and A. Woodgate. 2008. Share Issuance and Cross-Sectional Returns. Journal

of Finance 63:921945.

Restoy, F., and P. Weil. 2011. Approximate Equilibrium Asset PricesReview of Fi-

nance 15:128.

Ross, S. A. 1976. The Arbitrage Theory of Capital Asset Pricing. Journal of Economic

Theory 13:34160.

Roussanov, N. 2011. Composition of Wealth, Conditioning Information, and the Cross-

Section of Stock Returns. Working Paper, University of Pennsylvania.

Rouwenhorst, K. G. 1998. International Momentum Strategies. Journal of Finance 53:267

284.

58
Sadka, R. 2006. Momentum and Post-Earnings-Announcement Drift Anomalies: The Role

of Liquidity Risk. Journal of Financial Economics 80:309349.

Santos, T., and P. Veronesi. 2010. Habit formation, the cross section of stock returns and

the cash-flow risk puzzleJournal of Financial Economics 98:385 413.

Schwartzstein, J. 2012. Selective Attention and Learning. Working Paper, Dartmouth Col-

lege.

Schwert, G. W. 2003. Anomalies and Market Efficiency. In Handbook of the Economics

of Finance, ed. G. M. Constantinides, M. Harris, G. M. Stulz, Rene Constantinides,

M. Harris, and R. Stulz, pp. 939972. Amsterdam: Elsevier North Holland.

Shanken, J. 1985. Multivariate Tests of the Zero-Beta CAPM. Journal of Financial Eco-

nomics 14:327 348.

Shanken, J. 1992. On the Estimation of Beta-Pricing ModelsReview of Financial Stud-

ies 5:133.

Sims, C. A. 2003. Implications of Rational InattentionJournal of Monetary Eco-

nomics 50:665 690.

Sloan, R. 1996. Do Stock Prices Fully Reflect Information in Accruals and Cash Flows

About Future Earnings? Accounting Review 71:289315.

Stambaugh, R. F., J. Yu, and Y. Yuan. 2012. The Short of It: Investor Sentiment and

AnomaliesJournal of Financial Economics 104:288 302.

Teo, M., and S.-J. Woo. 2004. Style effects in the cross-section of stock returnsJournal of

Financial Economics 74:367 398.

Timmermann, A., and C. W. Granger. 2004. Efficient Market Hypothesis and Forecasting.

International Journal of Forecasting 20:1527.

Titman, S., K. J. Wei, and F. Xie. 2004. Capital Investments and Stock Returns. Journal

of Financial and Quantitative Analysis 39:677700.

59
Vuolteenaho, T. 2000. Understanding the Aggregate Book-to-Market Ratio and Its Impli-

cations to Current Equity-Premium Expectations. Working Paper, Harvard University.

Vuolteenaho, T. 2002. What Drives Firm-Level Stock Returns? Journal of Finance 52:233

264.

Watanabe, A., T. Yu, and Y. Xu. 2011. The Asset Growth Effect and Market Efficiency:

Insights from International Stock Markets. Working Paper, University of Alberta.

Weil, P. 1990. Nonexpected Utility in Macroeconomics. Quarterly Journal of Eco-

nomics 105:2942.

White, H. 2000. A Reality Check for Data SnoopingEconometrica 68:10971126.

Yu, J. 2011. Disagreement and Return Predictability of Stock PwortfoliosJournal of Fi-

nancial Economics 99:162 183.

Zhang, L. 2005. The Value PremiumThe Journal of Finance 60:pp. 67103.

60