Ang 2006

Downside Risk
Andrew Ang
Columbia University and NBER
Joseph Chen
University of Southern California
Economists have long recognized that investors care differently about downside
losses versus upside gains. Agents who place greater weight on downside risk demand
additional compensation for holding stocks with high sensitivities to downside market movements. We show that the cross section of stock returns reflects a downside
risk premium of approximately 6% per annum. Stocks that covary strongly with the
market during market declines have high average returns. The reward for beasring
downside risk is not simply compensation for regular market beta, nor is it explained
by coskewness or liquidity risk, or by size, value, and momentum characteristics.
(JEL C12, C15, C32, G12)
If an asset tends to move downward in a declining market more than it

moves upward in a rising market, it is an unattractive asset to hold,
because it tends to have very low payoffs precisely when the wealth of
investors is low. Investors who are sensitive to downside losses, relative to
upside gains, require a premium for holding assets that covary strongly
with the market when the market declines. Hence, in an economy with
agents placing greater emphasis on downside risk than upside gains,
assets with high sensitivities to downside market movements have high
Portions of this manuscript previously circulated in an earlier article titled Downside Correlation and
Expected Stock Returns. The authors thank Enrique Arzuc Brad Barber, Geert Bekaert, Alon Brav,
John Cochrane, Randy Cohen, Qiang Dai, Kent Daniel, Bob Dittmar, Rob Engle, Wayne Ferson, Eric
Ghysels, John Heaton, David Hirschleifer, N. Jegadeesh, Gautam Kaul, Jonathan Lewellen, Qing Li,
ubos Pastor, Adam Reed, Akhtar Siddique, Rob Stambaugh, and
Terence Lim, Toby Moskowitz, L
Zhenyu Wang. We especially thank Cam Harvey (the editor) and Bob Hodrick for detailed comments.
We also thank two anonymous referees whose comments greatly improved the article. We thank seminar
participants at Columbia University, Koc University, London Business School, London School of
Economics, Morgan Stanley, NYU, PanAgora Asset Management, UNC, USC, the American Finance
Association, the Canadian Investment Review Risk Management Conference, the Conference on Financial Market Risk Premiums at the Federal Reserve Board, the European Finance Association, the Five
Star Conference, an Inquire Europe meeting, a LA Society of Financial Analysts meeting, an NBER
Asset Pricing meeting, a Q-Group meeting, the Texas Finance Festival, the Valuation in Financial
Markets Conference at UC Davis, and the Western Finance Association for helpful discussions. The
authors acknowledge funding from a Q-Group research grant. Address correspondence to Andrew Ang,
Columbia Business School, New York NY 10027 or e-mail: aa610@columbia.edu.
The Author 2006. Published by Oxford University Press on behalf of The Society for Financial Studies. All rights
reserved. For permissions, please email: journals.permissions@oxfordjournals.org.
doi:10.1093/rfs/hhj035
Advance Access publication March 2, 2006
Downloaded from http://rfs.oxfordjournals.org/ at Oakland University on September 30, 2014
Yuhang Xing
Rice University
The Review of Financial Studies / v 19 n 4 2006
Pettengill, Sundaram, and Mathur (1995) and Isakov (1999) estimate the CAPM by splitting the full
sample into two subsamples that consist of observations where the realized excess market return is
positive or negative. Naturally, they estimate a positive (negative) market premium for the subsample
with positive (negative) excess market returns. In contrast, our approach examines premiums for asymmetries in the factor loadings, rather than estimating factor models on different subsamples. Price, Price,
and Nantell (1982) demonstrate that skewness in U.S. equity returns causes downside betas to be different
from unconditional betas but do not relate downside betas to average stock returns.
1192
average returns. In this article, we show that the cross-section of stock

returns reflects a premium for bearing downside risk.
As early as Roy (1952), economists have recognized that investors
care differently about downside losses than they care about upside
gains. Markowitz (1959) advocates using semivariance as a measure of
risk, rather than variance, because semivariance measures downside
losses rather than upside gains. More recently, the behavioral framework of Kahneman and Tverskys (1979) loss aversion preferences and
the axiomatic approach taken by Guls (1991) disappointment aversion
preferences allow agents to place greater weights on losses relative to
gains in their utility functions. Hence, in equilibrium, agents who are
averse to downside losses demand greater compensation, in the form of
higher expected returns, for holding stocks with high downside risk.
According to the Capital Asset Pricing Model (CAPM), a stocks
expected excess return is proportional to its market beta, which is constant across periods of high and low market returns. As Bawa and
Lindenberg (1977) suggest, a natural extension of the CAPM that takes
into account the asymmetric treatment of risk is to specify asymmetric
downside and upside betas. We compute downside (upside) betas over
periods when the excess market return is below (above) its mean. We
show that stocks with high downside betas have, on average, high unconditional average returns. We also find some mixed evidence that there
exists a discount for stocks that have high covariation conditional on
upside movements of the market.
Despite the intuitive appeal of downside risk, which closely corresponds to how individual investors actually perceive risk, there has been
little empirical research into how downside risk is priced in the cross
section of stock returns. Early researchers found little evidence of a
downside risk premium because they did not focus on measuring the
downside risk premium using all individual stocks in the cross section.
For example, Jahankhani (1976) fails to find any improvement over the
traditional CAPM by using downside betas, but his investigation uses
portfolios formed from regular CAPM betas. Similarly, Harlow and Rao
(1989) only evaluate downside risk relative to the CAPM in a maximumlikelihood framework and test whether the return on the zero-beta asset is
the same across all assets. All of these early authors do not directly
estimate a downside risk premium by demonstrating that assets which
covary more when the market declines have higher average returns.1
Downside Risk
1193
Our strategy for finding a premium for bearing downside risk in the
cross section is as follows. First, we directly show at the individual stock
level that stocks with higher downside betas contemporaneously have
higher average returns. Second, we claim that downside beta is a risk
attribute because stocks that have high covariation with the market when
the market declines exhibit high average returns over the same period.
This contemporaneous relationship between factor loadings and risk
premia is the foundation of a cross-sectional risk-return relationship,
and has been exploited from the earliest tests of the CAPM [see, among
others, Black, Jensen, and Scholes (1972) and Gibbons (1982)]. More
recently, Fama and French (1992) also seek, but fail to find, a relationship
between postformation market betas from an unconditional CAPM and
realized average stock returns over the same period. Our study differs
from these earlier tests by examining a series of short one-year samples
using daily data, rather than a single long sample using monthly data.
This strategy provides greater statistical power in an environment where
betas may be time-varying [see comments by Ang and Chen (2005) and
Lewellen and Nagel (2005)].
Third, we differentiate the reward for holding high downside-risk stocks
from other known cross-sectional effects. In particular, Rubinstein (1973),
Friend and Westerfield (1980), Kraus and Litzenberger (1976, 1983), and
Harvey and Siddique (2000) show that agents dislike stocks with negative
coskewness, so that stocks with low coskewness tend to have high average
returns. Downside risk is different from coskewness risk because downside
beta explicitly conditions for market downside movements in a nonlinear
fashion, whereras the coskewness statistic does not explicitly emphasize
asymmetries across down and up markets, even in settings where coskewness
may vary over time [as in Harvey and Siddique (1999)]. Since coskewness
captures some aspects of downside covariation, we are especially careful to
control for coskewness risk in assessing the premium for downside beta. We
also control for the standard list of known cross-sectional effects, including
size and book-to-market factor loadings and characteristics [Fama and
French (1993) and Daniel and Titman (1997)], liquidity risk factor loadings
[Pastor and Stambaugh (2003)], and past return characteristics [Jegadeesh
and Titman (1993)]. Controlling for these and other cross-sectional effects,
we estimate that the cross-sectional premium is approximately 6% per
annum.
Finally, we check if past downside betas predict future expected returns.
We find that, for the majority of the cross section, high past downside beta
predicts high future returns over the next month, similar to the contemporaneous relationship between realized downside beta and realized average
returns. However, this relation breaks down among stocks with very high
volatility. We attribute this to two effects. First, the future downside covariation of very volatile stocks is difficult to predict using past betasthe
1. A Simple Model of Downside Risk

In this section, we show how downside risk may be priced cross-sectionally
in an equilibrium setting. Specifically, we work with a rational disappointment aversion (DA) utility function that embeds downside risk following
Gul (1991). Our goal is to provide a simple motivating example of how a
representative agent with a larger aversion to losses, relative to his attraction to gains, gives rise to cross-sectional prices that embed compensation
for downside risk.2
However, our simple approach does not rule out other possible ways in
which downside risk may be priced in the cross section. For example,
Shumway (1997) develops an equilibrium behavioral model based on loss
averse investors. Barberis and Huang (2001) use a loss aversion utility
function, combined with mental accounting, to construct a cross-sectional
equilibrium. However, they do not relate expected stock returns to direct
measures of downside risk. Aversion to downside risk also arises in
models with constraints that bind only in one direction, for example,
binding short-sales constraints [Chen, Hong, and Stein (2001) and Hong
and Stein (2003)] or wealth constraints [Kyle and Xiong (2001)].
While standard power, or CRRA, utility also produces aversion to downside risk, the order of magnitude
of a downside risk premium, relative to upside potential, is economically negligible because CRRA
preferences are locally mean-variance.
1194
average one-year autocorrelation of one-year betas for very volatile stocks is

only 17.3% compared to 43.5% for a typical stock. This is not surprising,
because high volatility increases measurement error. Second, stocks with
very high volatility exhibit anomalously low returns [see Ang, Hodrick,
Xing, and Zhang (2006)]. Fortunately, the proportion of the market
where past downside beta fails to predict future returns is small (less than
4% in terms of market capitalization). Confirming Harvey and Siddique
(2000), we find that past coskewness predicts future returns, but the predictive power of past coskewness is not because past coskewness captures
future exposure to downside risk. Hence, past downside beta and past
coskewness are different risk loadings.
The rest of this article is organized as follows. In Section 1, we present a
simple model to show how a downside risk premium may arise in a crosssectional equilibrium. The framework uses a representative agent with the
kinked disappointment aversion utility function of Gul (1991), which places
larger weight on downside outcomes. Section 2 demonstrates that stocks
with high downside betas have high average returns over the same period
that they strongly covary with declining market returns. In Section 3, we
examine the predictive ability of past downside risk loadings. Section 4
concludes.
Downside Risk
where UW is the felicity function over end-of-period wealth W , which

we choose to be power utility, that is UW W 1 =1 . The
parameter 0 < A 1 is the coefficient of disappointment aversion, F
is the cumulative distribution function for wealth, W is the certainty
equivalent (the certain level of wealth that generates the same utility as
the portfolio allocation determining W ), and K is a scalar given by:
K PrW W APrW > W :
Outcomes above (below) the certainty equivalent W are termed elating (disappointing) outcomes. If 0 < A < 1, then the utility function
(1) down-weights elating outcomes relative to disappointing outcomes.
Put another way, the disappointment-averse investor cares more about
downside versus upside risk. If A 1, disappointment utility reduces to
the special case of standard CRRA utility, which is closely approximated
by mean-variance utility.
To illustrate the effect of downside risk on the cross section of stock
returns, we work with two assets x and y. Asset x has three possible
payoffs ux , mx and dx , and asset y has two possible payoffs uy and dy .
These payoffs are in excess of the risk-free payoff. Our setup has the
minimum number of assets and states required to examine cross-sectional
1195
Rather than considering models with one-sided constraints or

agents with behavioral biases, we treat asymmetries in risk in a rational
representative-agent framework that abstracts from additional interactions from one-sided constraints. The advantage of treating asymmetric
risk in a rational framework is that the disappointment utility function is
globally concave and provides solvable portfolio-allocation problems,
whereas optimal finite portfolio allocations for loss aversion utility may
not exist [see Ang, Bekaert, and Liu (2005)]. Our example with disappointment utility differs from previous studies, because existing work
with Guls (1991) first-order risk aversion utility concentrates on the
equilibrium pricing of downside risk for only the aggregate market,
usually in a consumption setting [see, for example, Bekaert, Hodrick,
and Marshall (1997), Epstein and Zin (1990, 2001) and Routledge and
Zin (2003)]. While a full equilibrium analysis of downside risk would
entail using consumption data, in our simple example and in our empirical work, we measure aggregate wealth by the market portfolio, similar to
a CAPM setting.
Guls (1991) disappointment utility is implicitly defined by the following equation:
Z W

Z 1
1
UW
UW dF W A
UW dF W ;
1
K
W
1
pricing (the expected returns of x and y relative to each other and to the
market portfolio, which consists of x and y), and to incorporate higher
moments (through the three states of x). The full set of payoffs and states
is given by:
Probability
ux ; uy
mx ; uy
dx ; uy
ux ; dy
mx ; dy
dx ; dy
p1
p2
p3
p4
p5
p6
The optimal portfolio weight for a DA investor is given by the solution

to:
max UW ;
wx ;wy
where the certainty equivalent is defined in equation (1), wx (wy ) is the

portfolio weight of asset x (y), end of period wealth W is given by:
W Rf wx x wy y;
and Rf is the gross risk-free rate. An equilibrium is characterized by a set

of asset payoffs, corresponding probabilities, and a set of portfolio weights
so that Equation (3) is maximized, and the representative agent holds the
market portfolio wx wy 1 with 0 < wx < 1 and 0 < wy < 1.
The equilibrium solution even for this simple case is computationally
nontrivial, because the solution to the asset allocation problem [Equation
(3)] entails simultaneously solving for both the certainty equivalent W and
for the portfolio weights wx and wy . In contrast, a standard portfolio
allocation problem for CRRA utility only requires solving the first-order
conditions for the optimal wx and wy . We extend a solution algorithm for
the optimization of Equation (3) developed by Ang, Bekaert, and Liu
(2005) to multiple assets. Appendix A describes our solution method and
details the values used in the calibration. Computing the solution is challenging, because for certain parameter values, equilibrium cannot exist because
nonparticipation may be optimal for low A under DA utility. This is unlike
the asset-allocation problem under standard CRRA utility, where agents
always optimally hold risky assets that have strictly positive risk premia.
In this simple model, the regular beta with respect to the market
portfolio (denoted by ) is not a sufficient statistic to describe the risk
return relationship of an individual stock. In our calibration, an assets
1196
Payoff
Downside Risk
expected returns increase with , but does not fully reflect all risk. This
is because the representative agent cares in particular about downside
risk, through A < 1. Hence, measures of downside risk have explanatory
power for describing the cross section of expected returns. One measure
of downside risk introduced by Bawa and Lindenberg (1977) is the downside beta (denoted by ):

covri ; rm jrm < m

;
varrm jrm < m
coskew
Eri i rm m 2
p
;
varri varrm
where i is the average excess return of asset i. Harvey and Siddique

(2000) predict that lower coskewness should be associated with higher
expected returns. The coskewness measure can be motivated by a thirdorder Taylor expansion of a general Euler equation:
0

U Wt1
Et
r
7
i;t1 0;
U 0 Wt
0
where W is the total wealth of the representative agent, and U can be

approximated by:
1
0
00
000
U 1 WU rm W 2 U r2m :
2
1197
where ri (rm ) is security is (the markets) excess return, and m is the

average market excess return. We also compute a relative downside beta
(denoted by ), relative to the regular CAPM beta, where
covri ; rm =varrm .
Figure 1 shows various riskreturn relationships holding in our DA
cross-sectional equilibrium. In the left figure of the top row, mean excess
returns increase with . To make sure that is not merely reflecting
the regular CAPM beta, we define the CAPM alpha as the excess return
of an asset not accounted for by the regular CAPM beta,
Eri Erm . The figure in the right column in the top row of
Figure 1 shows that the CAPM alpha is increasing with . In the second
row, we also find that CAPM alpha is also increasing with relative
downside beta, . Hence, higher downside risk is remunerated
by higher expected returns not captured by the CAPM.
The right-hand figure in the middle row of Figure 1 plots the CAPM
alpha versus coskewness, where coskewness is defined as:
0.038
0.14
0.036
0.13
0.034
Relations
in
CAPM Alpha
Mean Return
0.12
0.11
0.032
v ersion
Cross-Sectional
0.03
0.1
0.028
0.09
0.026
1.1
1.2
1.3
1.4
1.5
Downside Beta
1.6
1.7
1.8
1.1
1.4
1.5
Downside Beta
1.6
1.7
1.8
0.036
0.036
0.034
CAPM Alpha
0.034
CAPM Alpha
1.3
0.038
0.038
0.032
0.03
0.032
0.03
0.028
0.028
0.026
0.026
0.024
0.024
0.5
0.55
0.6
0.65
0.7
0.75
0.8
0.85
0.9
0.95
0.36
0.038
0.34
0.32
0.3
0.28
Coskew
0.26
0.24
0.22
0.038
0.036
0.036
0.034
0.034
CAPM Alpha
CAPM Alpha
1.2
0.032
0.03
0.032
0.03
0.028
0.028
0.026
0.026
0.024
0.024
1.08
1.09
1.1
1.11
1.12
1.13 1.14
+

1.15
1.16
1.17
1.18
0.1
0.2
0.3
0.4
+
0.5
0.6
0.7
Figure 1
Risk-Return Relations in a Disappointment Aversion Cross-Sectional Equilibrium
This figure shows risk-return relations for an asset in the DA cross-sectional equilibrium. In the top row
we have (i) a plot of the assets mean excess return versus downside beta , and (ii) a plot of the assets
CAPM alpha versus downside beta ; in the second row, (iii) a plot of the assets CAPM alpha versus
relative downside beta, , and (iv) a plot of the assets CAPM alpha versus coskewness; and in the
bottom row, (v) a plot of the assets CAPM alpha versus relative upside beta, , and (vi) a plot of
the assets CAPM alpha versus .
1198
0.024
0.08
Downside Risk
covri ; rm jrm > m

:
varrm jrm > m
Regular, downside, and upside betas are, by construction, not independent of each other. To differentiate the effect of upside risk from downside risk, we introduce two additional measures. Similar to relative
3
Taylor expansions have been used to account for potential skewness and kurtosis preferences in asset
allocation problems by Guidolin and Timmerman (2002), Harvey, Liechty, Liechty, and Muller (2003),
and Jondeau and Rockinger (2006).
1199
The Taylor expansion in Equation (8) is necessarily only an approximation. In particular, since the DA utility function is kinked, polynomial
expansions of U, such as the expansions used by Bansal, Hsieh, and
Viswanathan (1993), may not be good global approximations if the
kink is large (or A is very small).3 Nevertheless, measures like coskewness
based on the Taylor approximation for the utility function should also
have some explanatory power for returns.
Downside beta and coskewness may potentially capture different
effects. Note that, for DA utility, both downside beta and coskewness
are approximations, because the utility function does not have an explicit
form [Equation (1) implicitly defines DA utility]. Since DA utility is
kinked at an endogenous certainty equivalent, skewness and other centered moments may not effectively capture aversion to risk across upside
and downside movements in all situations. This is because they are based
on unconditional approximations to a nonsmooth function. In contrast,
the downside beta in Equation (5) conditions directly on a downside event
that the market return is less than its unconditional mean. In Figure 1,
our model shows that more negative coskewness is compensated by
higher expected returns. However, the Appendix describes a case where
CAPM alphas may increase as coskewness increases, which is the opposite of the relation predicted by the Taylor expansion.
With DA utility, a representative agent is willing to hold stocks with
high upside potential at a discount, all else being equal. A stock with high
upside potential relative to downside risk tends to pay off more when an
investors wealth is already high. Such stocks are not as desirable as
stocks that pay off when the market decreases. Consider two stocks
with the same downside beta, but with different payoffs in up markets.
The stock that covaries more with the market when the market rises has a
larger payoff when the market return is high. This stock does not need as
high an expected return in order for the representative agent to hold it.
Thus, there is a discount for stocks with high upside potential. To measure upside risk, we compute an upside beta (denoted by ) that takes
the same form as Equation (5), except we condition on movements of the
market excess return above its average value:
2. Downside Risk and Realized Returns

In this section, we document that stocks which strongly covary with the
market, conditional on down moves of the market, have contemporaneously high average returns. We document this phenomenon by first
looking at patterns of realized returns for portfolios sorted on downside
risk in Section 2.1. Throughout, we take care in controlling for the regular
beta and emphasize the asymmetry in betas by focusing on relative downside beta in addition to downside beta. In Section 2.2, we examine the
reward to downside risk, controlling for other cross-sectional effects by
using Fama-MacBeth (1973) regressions. We disentangle the different
effects of coskewness risk and downside beta exposure in Section 2.3.
Section 2.4 conducts various robustness tests. In Section 2.5, we show
some additional usefulness of accounting for downside risk by examining
if the commonly used Fama-French (1993) portfolios sorted by size and
book-to-market characteristics exhibit exposure to downside risk.
2.1 Regular, Downside, and Upside Betas
2.1.1 Research design. If there is a cross-sectional relation between risk
and return, then we should observe patterns between average realized
returns and the factor loadings associated with exposure to risk. For
1200
downside beta, we compute a relative upside beta (denoted by ).

We also directly examine the difference between upside beta and downside beta by computing the difference between the two, . In the
last row of Figure 1, our model shows that, controlling for regular beta or
downside beta, higher upside potential is indeed remunerated by lower
expected returns in our model.
Our simple example illustrates one possible mechanism by which compensation to downside risk may arise in equilibrium and how downside
versus upside risk may priced differently. Of course, this example, having
only two assets, is simplistic. Nevertheless, the model provides motivation
to ask if downside risk demands compensation in the cross section of U.S.
stocks, and if such compensation is different in nature from compensation
for risk based on measures of higher moments, such as the HarveySiddique (2000) coskewness measure. As our model shows, the compensation for downside risk is in addition to the reward already incurred in
standard, unconditional risk exposures, such as the regular unconditional
exposure to the market factor reflected in the CAPM beta. In our empirical work, we investigate a premium for downside risk also controlling for
other known cross-sectional effects such as the size and book-to-market
effects explored by Fama and French (1992, 1993), the liquidity effect of
Pastor and Stambaugh (2003), and the momentum effect of Jegadeesh
and Titman (1993).
Downside Risk
1201
example, the CAPM implies that stocks that covary strongly with the
market have contemporaneously high average returns over the same
period. In particular, the CAPM predicts an increasing relationship
between realized average returns and realized factor loadings, or contemporaneous expected returns and market betas. More generally, a
multifactor model implies that we should observe patterns between
average returns and sensitivities to different sources of risk over the
same time period used to compute the average returns and the factor
sensitivities.
Our research design follows Black, Jensen, and Scholes (1972), Fama
and MacBeth (1973), Fama and French (1992), Jagannathan and Wang
(1996), and others and focuses on the contemporaneous relation
between realized factor loadings and realized average returns. More
recently, in testing factor models, Lettau and Ludvigson (2001), Bansal,
Dittmar, and Lundblad (2005), and Lewellen and Nagel (2005), among
others all employ risk measures that are measured contemporaneously
with returns. While both Black, Jensen, and Scholes (1972) and Fama
and French (1992) form portfolios based on preformation factor loadings, they continue to perform their asset pricing tests using postranking
factor loadings, computed using the full sample. In particular, Fama
and French (1992) first form 25 portfolios ranked on the basis of preformation size and market betas. Then, they compute ex post factor
loadings for these 25 portfolios over the full sample. At each month,
they assign the postformation beta of a stock in a Fama-MacBeth (1973)
cross-sectional regression to be the ex post market factor loading of the
appropriate size and book-to-market sorted portfolio to which that
stock belongs during that month. Hence, testing a factor relation entails
demonstrating a contemporaneous relationship between realized covariance between a stock return and a factor with the realized average
return of that stock.
Our work differs from Fama and MacBeth (1973) and Fama and
French (1992) in one important way. Rather than forming portfolios
based on preformation regression criteria and then examining postformation factor loadings, we directly sort stocks on the realized factor loadings
within a period and then compute realized average returns over the same
period for these portfolios. Whereas preformation factor loadings reflect
both actual variation in factor loadings and measurement error effects,
postformation factor dispersion occurs almost exclusively from the actual
covariation of stock returns with risk factors. Moreover, we estimate
factor loadings using higher frequency data over shorter samples, rather
than lower frequency data over longer samples. Hence, our approach has
greater power.
A number of studies, including Fama and MacBeth (1973), Shanken
(1992), Ferson and Harvey (1991), and Pastor and Stambaugh (2003),
2.1.2 Empirical results. We investigate patterns between realized average returns and realized betas. While many cross-sectional asset pricing
studies use a horizon of one month, we work in intervals of 12 months,
from t to t 12, following Kothari, Shanken, and Sloan (1995). Our
choice of an annual horizon is motivated by two concerns. First, we
need a sufficiently large number of observations to condition on periods
of down markets. One month of daily data provides too short a window
for obtaining reliable estimates of downside variation. We check the
robustness of our results to using intervals of 24 months with weekly
frequency data to compute downside betas. Second, Fama and French
(1997), Ang and Chen (2005), and Lewellen and Nagel (2005) show that
market risk exposures are time-varying. Very long time intervals may
cause the estimates of conditional betas to be noisy. Fama and French
(2005) also advocate estimating betas using an annual horizon.
Over every 12-month period, we compute the sample counterparts to
various risk measures using daily data. We calculate a stocks regular
beta, downside beta as described in Equation (5), and upside beta as
described in Equation (9). We also compute a stocks relative downside
beta, , a stocks relative upside beta, , and the difference
between upside beta and downside beta, . Since these risk measures are calculated using realized returns, we refer to them as realized ,
realized , realized , realized relative , realized relative , and
realized .
In our empirical work, we concentrate on presenting the results of
equal-weighted portfolios and equal-weighted Fama-MacBeth (1973)
regressions. While a relationship between factor sensitivities and returns
should hold for both an average stock (equal weighting) or an average
dollar (value weighting), we focus on computing equal-weighted portfolios, because past work on examining nonlinearities in the cross section
has found risk due to asymmetries to be bigger among smaller stocks. For
1202
compute predictive betas formed using conditional information available at time t and then examine returns over the next period. These
studies implicitly assume that risk exposures are constant and not timevarying. Indeed, as noted by Daniel and Titman (1997), in settings
where the covariance matrix is stable over time, preformation factor
loadings are good instruments for the future expected (postformation)
factor loadings. If preformation betas are weak predictors of future
betas, then using preformation betas as instruments will also have low
power to detect ex post covariation between factor loadings and realized returns. We examine the relation between preformation estimates
of factor loadings with postformation realized factor loadings in
Section 3.
Downside Risk
In their article, Harvey and Siddique (2000) state that they use value-weighted portfolios. From personal
correspondence with Cam Harvey, the coskewness effects arise most strongly in equal-weighted portfolios
rather than in value-weighted portfolios.
The theoretical number of lags required to absorb all the moving average error effects is 11, but we
include an additional lag for robustness.
1203
example, the coskewness effect of Harvey and Siddique (2000) is strongest

for equal-weighted portfolios.4 We work with equally-weighted portfolios
to emphasize the differences between downside risk and coskewness. In a
series of robustness checks, we also examine if our findings hold using
value-weighted portfolios or in value-weighted Fama-MacBeth regressions. We concentrate only on NYSE stocks to minimize the illiquidity
effects of small firms, but also consider all stocks on the NYSE, AMEX,
and NASDAQ in robustness tests.
At the beginning of the one-year period at time t, we sort stocks into five
quintiles based on their realized , realized , realized , realized
relative , realized relative , or realized over the next 12
months. In the column labeled Return, Table 1 reports the average
realized excess return from time t to t 12 in each equally weighted
quintile portfolio. The table also reports the average cross-sectional realized , , or of each quintile portfolio. These average returns and
betas are computed over the same 12-month period. Hence, Table 1 shows
relationships between contemporaneous factor loadings and returns.
Although we use a 1-year horizon, we evaluate 12-month returns at a
monthly frequency. This use of overlapping information is more efficient
but induces moving average effects. To adjust for this, we report t-statistics
of differences in average excess returns between quintile portfolio 5 (high
betas) and quintile portfolio 1 (low betas) using 12 Newey-West (1987a)
lags.5 The sample period is from July 1963 to December 2001, with our last
12-month return period starting in January 2001. As part of our robustness
checks (below), we also examine nonoverlapping sample periods.
Panel A of Table 1 shows a monotonically increasing pattern between
realized average returns and realized . Quintile 1 (5) has an average
excess return of 3.5% (13.9%) per annum, and the spread in average
excess returns between quintile portfolios 1 and 5 is 10.4% per annum,
with a corresponding difference in contemporaneous market betas of
1.36. Our results are consistent with the earliest studies testing the
CAPM, like Black, Jensen, and Scholes (1972), who find a reward for
holding higher-beta stocks. However, this evidence per se does not mean
that the CAPM holds, because the CAPM predicts that no other variable
other than beta should explain a firms expected return. Nevertheless, it
demonstrates that bearing high market risk is rewarded with high average
returns. Panel A also reports the positive and negative components of
beta ( and ). By construction, higher or higher must also
mean higher unconditional , so high average returns are accompanied by
Table 1
Returns of Stocks Sorted by Realized Factor Loadings
Panel A: Stocks sorted by realized b
Return
1 Low
2
3
4
5 High
3.52%
6.07%
7.58%
9.48%
13.95%
0.28
0.59
0.82
1.10
1.64
0.36
0.67
0.90
1.18
1.72
High-low
t-stat
10.43%
[4.98]
1.36
1.36
Return
0.19
0.51
0.77
1.06
1.63
1 Low
2
3
4
5 High
2.71%
5.62%
7.63%
10.16%
14.49%
0.40
0.63
0.83
1.06
1.49
0.19
0.61
0.89
1.23
1.92
0.42
0.62
0.79
0.99
1.34
1.44
High-low
t-stat
11.78%
[6.16]
1.09
1.72
0.92
Return
1 Low relative
2
3
4
5 High relative
4.09%
7.69%
8.53%
9.56%
10.73%
0.98
0.83
0.80
0.84
0.98
0.56
0.73
0.86
1.08
1.60
1.12
0.84
0.75
0.72
0.71
High-low
t-stat
6.64%
[7.70]
0.00
1.04
0:41
Panel E: Stocks sorted by realized relative b
Portfolio
Return
1 Low
2
3
4
5 High
5.73%
7.42%
8.29%
9.33%
9.83%
0.44
0.62
0.82
1.05
1.49
0.63
0.73
0.90
1.10
1.46
0:04
0.45
0.76
1.12
1.85
High-low
t-stat
4.11%
[2.62]
1.05
0.83
1.89
Panel F: Stocks sorted by realized b b
Portfolio
Return
0.91
0.82
0.81
0.85
1.03
1.19
0.95
0.88
0.87
0.95
0.25
0.58
0.76
0.99
1.56
1 Low
2
3
4
5 High
11.35%
9.83%
8.53%
7.35%
3.55%
0.12
0:23
1.31
High-low
t-stat
7:81%
[9.03]
Portfolio
Return
1 Low relative
2
3
4
5 High relative
10.48%
9.51%
8.53%
7.72%
4.37%
High-low
t-stat
6:11%
[9.02]
Panel D: Stocks sorted by realized b
Portfolio
Portfolio
Panel C: Stocks sorted by realized relative b

0.94
0.83
0.80
0.84
1.02
1.45
1.03
0.87
0.78
0.70
0.37
0.62
0.75
0.95
1.46
0.08
0:75
1.08
This table lists the equal-weighted average returns and risk characteristics of stocks sorted by realized
betas. For each month, we calculate ; ; , relative (given by ), relative (given by
), and with respect to the market of all stocks listed on the NYSE using daily continuously compounded returns over the next 12 months. For each risk characteristic, we rank stocks into
quintiles (15) and form equal-weighted portfolios at the beginning of each 12-month period. The
number of stocks in each portfolio varies across time from 216 to 317 stocks. The column labeled
Return reports the average return in excess of the one-month Treasury-bill rate over the next 12
months (which is the same period as the period used to compute ; , and ). The row labeled Highlow reports the difference between the returns of portfolio 5 and portfolio 1. The entry labeled t-stat in
square brackets is the t-statistic computed using Newey-West (1987a) heteroskedastic-robust standard
errors with 12 lags for the High-low difference. The columns labeled , and report the timeseries and cross-sectional average of equal-weighted individual stock betas over the 12-month holding
period. The sample period is from July 1963 to December 2001, with the last 12-month period from
January 2001 to December 2001, and observations are at a monthly frequency.
high , , and regular . Note that, for these portfolios sorted by

realized , the spread in realized and is also similar to the spread
in realized . In the remainder of the panels in Table 1, we decompose
1204
Portfolio
Panel B: Stocks sorted by realized b
Downside Risk
1205
the reward for unconditional market risk into downside and upside
components.
Panel B shows that stocks with high contemporaneous have high
average returns. Stocks in the quintile with the lowest (highest) earn
2.7% (14.5%) per annum in excess of the risk-free rate. The average
difference between quintile portfolios 1 and 5 is 11.8% per annum,
which is statistically significant at the 1% level. These results are consistent with agents disliking downside risk and avoiding stocks that covary
strongly when the market dips, such as the DA representative agent
described in Section 1. Stocks with high must carry a premium in
order to entice agents to hold them. An alternative explanation is that
agents have no particular emphasis on downside risk versus upside
potential. High stocks may earn high returns simply because, by
construction, high stocks have high regular . The average spread
between quintile portfolios 1 and 5 is very large (0.19 to 1.92), but sorting
on also produces variation in and . However, the variation in
or is not as disperse as the variation in . Another possible explanation is that sorting on high contemporaneous covariance with the market
mechanically produces high contemporaneous returns. However, this
concern is not applicable to our downside risk measure, since we are
picking out precisely those observations for which stocks already have
very low returns when the market declines. In Panels C and D, we
demonstrate that it is the reward for downside risk alone that is behind
the pattern of high stocks earning high returns.
In panel C of Table 1, we sort stocks by realized relative downside beta
. Relative downside beta focuses on the incremental impact of
downside beta over the regular, unconditional market beta. Panel C
shows that stocks with high realized relative have high average
returns. The difference in average excess returns between portfolios 5
and 1 is 6.6% per annum and is highly significant with a robust t-statistic
of 7.70. We can rule out that this pattern of returns is attributable to
regular beta because the loadings are flat over portfolios 1 to 5. Hence,
the high realized returns from high relative are produced by the
exposure to downside risk, measured by high loadings.
Panel D shows a smaller spread for average realized excess returns for
stocks sorted on realized , relative to the spreads for and in panels
A and B. Since only measures exposure to a rising market, stocks that
rise more when the market return increases should be more attractive
and, on average, earn low returns. We do not observe a discount for
stocks that have attractive upside exposure. We find that low (high)
stocks earn, on average, 5.7% (9.8%) per annum in excess of the risk-free
rate. This pattern of high returns to high loadings seems to be
inconsistent with agents having strong preferences for upside potential;
however, this measure does not control for the effects of regular or for
1206
the effects of . Instead, the increasing pattern of returns in panel D

may be due to the patterns of or , which increase from quintile
portfolios 1 to 5. The spread in regular is 1.05, while the spread in is
0.83. From the CAPM, high implies high returns, and if agents dislike
downside risk, high also implies high returns. Hence, we now turn to
measures that control for these confounding effects when we look for an
upside risk premium.
In panel E, we examine the effect of while controlling for regular
by sorting stocks according to realized relative upside beta, .
Examining relative helps us to disentangle the effect of from the
effects of regular and . Panel E shows that stocks with high realized
relative have low returns. We find that high (low) relative stocks
earn, on average, 4.4% (10.5%) per annum in excess of the risk-free rate.
Furthermore, stocks sorted by relative upside beta produce a spread in
while keeping the spread in regular and relatively flat. The
differences in regular and across the highest quintile and the lowest
quintile portfolios are relatively low at 0.12 and 0.23, respectively. In contrast, we obtain a wide spread in of 1.31 between the
highest quintile and the lowest quintile portfolios. This pattern of low
returns to high relative stocks is consistent with agents accepting a
discount for holding stocks with high upside potential, which would
result from the DA agent equilibrium in Section 1.
Finally, we sort stocks by the realized difference between upside beta
and downside beta in Table 1, panel F. We look at this
measure to gauge the effect of upside risk relative to downside risk. In
panel F, we observe a decreasing pattern in average realized excess returns
with increasing . On average, we find that stocks with high
(low) earn 3.6% (11.4%) per annum in excess of the risk-free
rate. While this direction is consistent with a premium for downside risk
and a discount for upside potential, it is hard to separate the effects of
downside risk independently from upside risk using these
portfolio sorts. The quintile portfolios sorted by show little
variation in regular , but they show a decreasing pattern in and an
increasing pattern in . Hence, it is difficult to separate whether the
patterns in realized returns arise because of exposure to downside losses
or exposure to upside gains. Thus, our preferred measures to examine
downside or upside risk are relative and relative in panels C and E,
which control for the effect of .
In summary, Table 1 demonstrates that downside risk is rewarded in
the cross section of stock returns. Stocks with high loadings earn high
average returns over the same period that are not mechanically driven by
high regular, unconditional betas. Stocks that covary strongly with the
market conditional on positive moves of the market command significant
Downside Risk
discounts. However, these relations do not control for other known

patterns in the cross section of stock returns, which we now investigate.
Eri i rm m 3
cokurt p
;
varri varrm 3=2
10
where ri is the firm excess return, rm is the market excess return, i is the
average excess stock return, and m is the average market excess returns.
Finally, we also include the Pastor and Stambaugh (2003) historical
liquidity beta at time t to proxy for liquidity exposure.
1207
2.2 Fama-MacBeth Regressions

A long literature from Banz (1981) onward has shown that various firm
characteristics also have explanatory power in the cross section. The size
effect [Banz (1981)], the book-to-market effect [Basu (1983)], the momentum effect [Jegadeesh and Titman (1993)], the volatility effect [Ang et al.
(2006)], exposure to coskewness risk [Harvey and Siddique (2000)], exposure to cokurtosis risk [Scott and Horvarth (1980) and Dittmar (2002)],
and exposure to aggregate liquidity risk [Pastor and Stambaugh (2003)],
all imply different patterns for the cross section of expected returns. We
now demonstrate that downside risk is different from all of these effects
by performing a series of cross-sectional Fama and MacBeth (1973)
regressions at the firm level, over the sample period from July 1963 to
December 2001.
We run Fama-MacBeth regressions of excess returns on firm characteristics and realized betas with respect to various sources of risk. We
use a 12-month horizon for excess returns to correspond to the contemporaneous period over which our risk measures are calculated. Since
the regressions are run using a 12-month horizon, but at the overlapping
monthly frequency, we compute the standard errors of the coefficients
by using 12 Newey-West (1987a) lags. Table 2 reports the results listed
by various sets of independent variables in Regressions IVI. We also
report means and standard deviations to help gauge economic significance. We regress realized firm returns over a 12-month horizon (t to
t 12) on realized market beta, downside beta, and upside beta, (, ,
and ) computed over the same period. Hence, these regressions
capture a contemporaneous relationship between average returns and
factor loadings or characteristics. We control for log-size, book-tomarket ratio, and past 12-month excess returns of the firm at the
beginning of the period t. We also include realized standard deviation
of the firm excess returns, realized coskewness, and realized cokurtosis
as control variables. All of these are also computed over the period
from t to t 12. We define cokurtosis in a similar manner to coskewness in Equation (6):
Table 2
Fama-MacBeth Regressions
Model
Intercept

Bk-Mkt
Past Ret
Std Dev
Coskewness
Cokurtosis
L
II
III
IV
VI
0.300
[9.35]
0.177
[8.19]
0.044
[3.39]
0.054
[1.66]
0.046
[1.42]
0.246
[7.62]
0.257
[7.79]
0.069
[7.17]
0:029
[4.85]
0:039
[8.82]
0.017
[3.87]
0.017
[1.91]
8:433
[10.7]
0:229
[10.7]
0.015
[1.57]
0.064
[7.44]
0:025
[4.15]
0:007
[1.47]
0.024
[5.17]
0.063
[6.32]
0.028
[2.68]
0.003
[0.22]
0:013
[3.03]
0.023
[5.03]
0.053
[5.40]
0:181
[4.31]
0.062
[6.00]
0.020
[2.33]
0:034
[7.77]
0.017
[3.67]
0.020
[2.12]
5:781
[6.41]
0:196
[5.07]
0.045
[4.40]
0.056
[5.25]
0.017
[1.91]
0:034
[7.39]
0.018
[3.76]
0.015
[1.50]
6:459
[7.04]
0:188
[4.59]
0.047
[4.52]
0:008
[0.93]
Mean
(Std Dev)
0.828
(0.550)
0.882
(0.739)
0.722
(0.842)
5.614
(1.523)
0.768
(0.700)
0.085
(0.370)
0.355
(0.174)
0.179
(0.188)
2.240
(1.353)
0.166
(0.456)
This table shows the results of Fama-MacBeth (1973) regressions of 12-month excess returns on firm
characteristics and realized-risk characteristics. The sample period is from July 1963 to December 2001,
with the last 12-month period from January 2001 to December 2001, and observations are at a monthly
frequency (451 months) for all stocks listed on the NYSE. For regression VI, the sample period is from
January 1967 to December 2001 (397 months). The number of stocks in each regression varies across time
from 1080 to 1582 stocks. The t-statistics in square brackets are computed using Newey-West (1987a)
heteroskedastic-robust standard errors with 12 lags. The firm characteristics are log of market capitalizations (Log-Size), book-to-market ratios (Bk-mkt), and past 12-month excess returns (Past ret),
all computed at the beginning of each period. The realized risk characteristics are ; ; , standard
deviations (Std Dev), coskewness and cokurtosis are all calculated over the following 12-month period
using daily continuously compounded returns. We also include the Pastor-Stambaugh (2003) liquidity
beta, L , for January 1967 to January 2001. All independent variables are Winsorized at the 1% level and
at the 99% within each month. We report time-series averages of the cross-sectional mean and standard
deviation (in parentheses) of each independent variable in the last column.
In order to avoid putting too much weight on extreme observations,

each month we Winsorize all independent variables at the 1% and 99%
levels.6 Winsorization has been performed in cross-sectional regressions
by Knez and Ready (1997), among others, and ensures that extreme
outliers do not drive the results. It is particularly valuable for dealing
with the book-to-market ratio, because extremely large book-to-market
values are sometimes observed due to low prices, particularly before a
firm delists.
For example, if an observation for the firms book-to-market ratio is extremely large and above the 99th
percentile of all the firms book-to-market ratios that month, we replace that firms book-to-market ratio
with the book-to-market ratio corresponding to the 99th percentile. The same is done for firms whose
book-to-market ratios lie below the 1%-tile of all firms book-to-market ratios that month.
1208
Log-size
Downside Risk
By construction, is a weighted average of and . If we place both and on the right-hand side
of regressions IIVI and omit , the coefficients on are the same to three decimal places as the
coefficients reported in Table 2. Similarly, if we specify both and to be regressors, the coefficients on
are almost unchanged.
1209
We begin with regression I in Table 2 to show the familiar, standard set

of cross-sectional return patterns. While the regular market beta carries a
positive coefficient, the one-factor CAPM is rejected because is not a
sufficient statistic to explain the cross section of stock returns. The results
of the regression confirm several CAPM anomalies found in the literature. For example, small stocks and stocks with high book-to-market
ratios have high average returns [see Fama and French (1992)], while
stocks with high past returns also continue to have high returns [see
Jegadeesh and Titman, (1993)]. The very large and highly significant
negative coefficient (8.43) on the firms realized volatility of excess
returns confirms the anomalous finding of Ang et al. (2006), who find
that stocks with high return volatilities have low average returns. Consistent with Harvey and Siddique (2000), stocks with high coskewness
have low returns. Finally, stocks with positive cokurtosis tend to have
high returns, consistent with Dittmar (2002).
In regressions IIVI, we separately examine the downside and upside
components of beta and show that downside risk is priced.7 We turn first
to regression II, which reveals that downside risk and upside risk are
priced asymmetrically. The coefficient on downside risk is positive (0.069)
and highly significant, confirming the portfolio sorts in Table 1. The
coefficient on is negative (0.029), but lower in magnitude than the
coefficient on . These results are consistent with the positive premium
on relative and the discount on relative reported in panels C and E
of Table 1.
Regression III shows that the reward for both downside and upside risk
is robust to controlling for size, book-to-market, and momentum effects.
Note that asymmetric beta risk does not remove the book-to-market or
momentum effects. But, importantly, the Fama-MacBeth coefficients for
and remain almost unchanged from their regression II estimates at
0.064 and 0.025, respectively. While neither nor is a sufficient
statistic to explain the cross section of stock returns, Table 2 demonstrates a robust reward for holding stocks with high (low) ( ) loadings controlling for the standard size, book-to-market, and past return
effects.
In regressions IVVI, accounting for additional measures of risk does
not drive out the significance of but drives out the significance of .
Once we account for coskewness risk in regression IV, the coefficient on
becomes very small (0.003) and becomes statistically insignificant,
with a t-statistic of 0.22. Controlling for coskewness also brings down
the coefficient on from 0.064 in regression III (without the coskewness
1210
risk control) to 0.028 in regression IV (with the coskewness risk control).

Nevertheless, the premium for downside risk remains positive and statistically significant. In regression V, where we add controls for realized
firm volatility and realized firm cokurtosis, the coefficient on downside
risk remains consistently positive at 0.062 and also remains highly statistically significant, with a robust t-statistic of 6.00. The preference function of Section 1 that weights downside outcomes more than upside
outcomes also produces a discount for upside potential. We observe
these patterns in data only for regressions II and III, where the upside
discount is about half the size in magnitude of the premium for downside
risk. After the additional controls beyond size, book-to-market, and
momentum are included in regressions IV and V, the discount for potential upside becomes fragile. Thus, the premium for downside risk dominates in the cross section.
Finally, regression VI investigates the reward for downside and upside
risk, controlling for the full list of firm characteristics and realized factor
loadings. We lose five years of data in constructing the Pastor-Stambaugh
historical liquidity betas, so this regression is run from January 1967 to
December 2001. The coefficient on is 0.056, with a robust t-statistic of
5.25. In contrast, the coefficient on is statistically insignificant,
whereas the premium for coskewness is significantly negative, at 0.188.
Since both and coskewness risk measure downside risk, and the
coefficients on both risk measures are statistically significant, we carefully disentangle the and coskewness effects in Section 2.4.
To help interpret the economic magnitudes of the risk premia reported
in the Fama-MacBeth regressions, the last column of Table 2 reports the
time-series average of the cross-sectional mean and standard deviation of
each of the factor loadings or characteristics. The average market beta is
less than one (0.83) because we are focusing on NYSE firms, which tend
to be skewed towards large firms with relatively low betas. The average
downside beta is 0.88, with a cross-sectional standard deviation of 0.74.
This implies that, for a downside risk premium of 6.9% per annum, a twostandard-deviation move across stocks in terms of corresponds to a
change in expected returns of 2 0:069 0:74 10:2% per annum.
While the premium on coskewness appears much larger in magnitude,
at approximately 19% per annum, coskewness is not a beta and must be
carefully interpreted. A two-standard-deviation movement across coskewness changes expected returns by 2 0:188 0:19 7:1% per
annum, which is slightly less than, but of the same order of magnitude
as, the effect of downside risk.
The consistent message from the regressions in Table 2 is that reward
for downside risk is always positive at approximately 6% per annum
and statistically significant. High downside beta is compensated for by
high average returns, and this result is robust to controlling for other
Downside Risk
firm characteristics and risk characteristics, including upside beta. Moreover, downside beta risk remains significantly positive in the presence of
coskewness risk controls. On the other hand, the reward for upside risk
( ) is fragile. A priori, we expect the coefficient on to be negative,
but in data, it often flips sign and is insignificant when we control for
other cross-sectional risk attributes. Thus, aversion to downside risk is
priced more strongly, and more robustly, in the cross section than
investors attraction to upside potential.
1211
2.3 Downside Beta Risk and Coskewness Risk

The Fama-MacBeth (1973) regressions in Table 2 demonstrate that
both downside beta and coskewness have predictive power for the
cross-section of stock returns. Since both and coskewness capture
the effect of asymmetric higher moments and downside risk, we now
measure the magnitude of the reward for exposure to downside beta,
while explicitly controlling for the effect of coskewness. Table 3 presents the results of this exercise.
To control for the effect of coskewness, we first form quintile portfolios
sorted on coskewness. Then, within each coskewness quintile, we sort
stocks into five equally weighted portfolios based on . Both coskewness
and are computed over the same 12-month horizon for which we
examine realized excess returns. After forming the 5 5 coskewness and
portfolios, we average the realized excess returns of each quintile
over the five coskewness portfolios. This characteristic control procedure
creates a set of quintile portfolios with near-identical levels of coskewness risk. Thus, these quintile portfolios control for differences in
coskewness.
Panel A of Table 3 reports average excess returns of the 25 coskewness
portfolios. The column labeled Average reports the average
12-month excess returns of the quintiles, controlling for coskewness
risk. The row labeled High-low reports the differences in average
returns between the first and fifth quintile portfolios within each
coskewness quintile. The last row reports the 51 quintile difference for
the quintiles that control for the effect of coskewness exposure. The
average excess return of 7.6% per annum in the bottom right entry of
Panel A is the difference in average returns between the fifth and first
quintile portfolios that control for coskewness risk. This difference has a
robust t-statistic of 4.16. Hence, coskewness risk cannot account for the
reward for bearing downside beta risk.
In Panel A, the patterns within each coskewness quintile moving from
low to high stocks (reading down each column) are very interesting. As coskewness increases, the differences in average excess returns
due to different loadings decrease. The effect is quite pronounced. In
the first coskewness quintile, the difference in average returns between
Table 3
Returns of Stocks Sorted by Realized Downside Beta and Coskewness
Panel A: b Sorts controlling for coskewness
Coskewness Quintiles
Portfolio
5 High
Average
1 Low
2
3
4
5 High
7.21%
10.55%
13.63%
15.63%
21.84%
5.74%
8.40%
11.30%
12.82%
15.85%
4.03%
6.94%
8.30%
9.35%
11.51%
3.40%
5.59%
6.08%
6.74%
6.81%
0.22%
2.61%
3.76%
2.56%
2.32%
4.21%
6.82%
8.61%
9.42%
11.67%
High-low
t-stat
14.64%
[5.62]
10.11%
[5.22]
7.48%
[3.91]
3.41%
[1.87]
2.10%
[1.32]
7.55%
[4.16]
5 High
Average
Panel B: Coskewness sorts controlling for b
Quintiles
Portfolio
1 Low coskew
2
3
4
5 High coskew
High-low
t-stat
1 Low
4.69%
4.17%
2.74%
1.50%
0.41%
7.15%
6.19%
6.51%
5.24%
2.96%
9.30%
9.61%
8.68%
6.68%
3.86%
12.59%
12.33%
11.31%
9.16%
5.37%
17.61%
18.21%
16.07%
12.83%
7.65%
10.27%
10.10%
9.06%
7.08%
4.05%
4:28%
[4.23]
4:18%
[5.64]
5:45%
[7.12]
7:22%
[8.09]
9:96%
[7.94]
6:22%
[8.17]
This table examines the relation between and coskewness. For each month, we compute and
coskewness with respect to the market of all stocks listed on the NYSE using daily continuously
compounded returns over the next 12 months. In panel A, we first rank stocks into quintiles (15) at
the beginning of each 12-month period based on coskewness over the next 12 months. Then, we rank
stocks within each first-sort quintile into additional quintiles according to computed over the next 12
months. For each 5 5 grouping, we form an equal-weighted portfolio. In panel B, we reverse the order
so that we first sort on and then on coskewness. The sample period is from July 1963 to December
2001, and the number of stocks in each portfolio varies across time from 43 to 64 stocks. We report the
average return in excess of the one-month Treasury bill rate over the next 12 months. For the column
labeled Average, we report the average return of stocks in each second-sort quintile. This controls for
coskewness in panel A (B). The row labeled High-low reports the difference between the returns of
portfolio 1 and portfolio 5. The entry labeled t-stat in square brackets is the t-statistic computed using
Newey-West (1987a) heteroskedastic-robust standard errors with 12 lags for the High-low difference.
the low and high quintiles is 14.6% per annum. The average return
difference in the low and high portfolios decreases to 2.1% per annum
for the quintile of stocks with the highest coskewness.
The reason for this pattern is as follows. As defined in Equation (6),
coskewness is effectively the covariance of a stocks return with the square
of the market return, or with the volatility of the market. A stock with
negative coskewness tends to have low returns when market volatility is
high. These are also usually, but not always, periods of low market
returns. Volatility of the market treats upside and downside risk symmetrically, so both extreme upside and extreme downside movements of the
market have the same volatility. Hence, the prices of stocks with large
1212
1 Low
Downside Risk
2.4 Robustness Checks

We now show that our results do not depend on the way we have
measured asymmetries in betas or the design of our empirical tests. In
particular, we show that our results are robust to measuring asymmetries
1213
negative coskewness tend to decrease when the market falls, but the prices
of these stocks may also decrease when the market rises. In contrast,
downside beta concentrates only on the former effect by explicitly considering only the downside case. When coskewness is low, there is a wide
spread in because there is large scope for market volatility to represent
both large negative and large positive changes. This explains the large
spread in average returns across the quintiles for stocks with low
coskewness.
The small 2.1% per annum 51 spread for the quintiles for the
highest coskewness stocks is due to the highest coskewness stocks exhibiting little asymmetry. The distribution of coskewness across stocks is
skewed towards the negative side and is negative on average. Across the
low to the high coskewness quintiles in panel A, the average coskewness
ranges from 0.41 to 0.09. Hence, the quintile of the highest coskewness
stocks have little coskewness. This means that high coskewness stocks
essentially do not change their behavior across periods where market
returns are stable or volatile. Furthermore, the range of in the highest
coskewness quintile is also smaller. The small range of for the highest
coskewness stocks explains the low 2.1% spread for the quintiles in
the second last column of panel A.
Panel B of Table 3 repeats the same exercise as panel A, except we
examine the reward for coskewness controlling for different levels of .
Panel B first sorts stocks on coskewness before sorting on , and then
averages across the quintiles. This exercise examines the coskewness
premium controlling for downside exposure. Controlling for , the 51
difference in average returns for coskewness portfolios is 6.2%, which is
highly statistically significant with a t-statistic of 8.17. Moreover, there
are large and highly statistically significant spreads for coskewness in
every quintile. Coskewness is able to maintain a high range within
each portfolio, unlike the diminishing range for within each
coskewness quintile in Panel A.
In summary, downside beta risk and coskewness risk are different. The
high returns to high stocks are robust to controlling for coskewness
risk and vice versa. Downside beta risk is strongest for stocks with low
coskewness. Coskewness does not differentiate between large market
movements on the upside or the downside. For stocks with low coskewness, downside beta is better able to capture the downside risk premium
associated only with market declines than an unconditional coskewness
measure.
rf
covri ; rm jrm < rf

varrm jrm < rf
and
rf
covri ; rm jrm > rf

:
varrm jrm > rf
11
We define downside and upside beta relative to the zero rate of return as:
0
covri ; rm jrm < 0

varrm jrm < 0
and
0
covri ; rm jrm > 0

:
varrm jrm > 0
12
We show the correlations among these risk measures in Table 4, which

reports the time-series averages of the cross-sectional correlations of regular beta and the various downside and upside risk measures. Table 4
shows that , rf , and 0 are all highly correlated with each other with
correlations greater than 0.96. Similarly, we find that , rf , and 0 are
also highly correlated with each other. Given these correlations, it is not
surprising that reproducing Table 1 and Table 2 using either one of these
alternative cutoff points yields almost identical results.8 Therefore, the
finding of a downside risk premium is indeed being driven by emphasizing
losses versus gains, rather than by using a particular cutoff point for the
benchmark.
Table 4 also shows that the regular measure of beta is quite different
from measures of downside beta and upside beta. The correlation of
regular beta with downside beta or upside beta is 0.78 and 0.76, respectively. Therefore, downside beta and upside beta capture different
aspects of risk, and are not simply reflective of the regular, unconditional market beta. Interestingly, the correlation between downside beta
and upside beta is only around 0.46. Thus, a high downside risk exposure does not necessarily imply a high upside risk exposure. We examine
8
These results are available upon request.
1214
with respect to different cutoff points across up markets and down

markets. We also show that our results are robust to using longerfrequency data. Finally, we show that our results are not driven by
using equal weighting, concentrating on NYSE stocks, or using overlapping portfolios by checking robustness with respect to value weighting,
including all stocks listed on NYSE, AMEX, and NASDAQ, and using
nonoverlapping portfolios.
We begin by using other cutoff points to determine up markets and
down markets. Our measures of and use returns relative to realized
average market excess return. Naturally, realized average market returns
vary across time and may have particularly low or high realizations.
Alternatively, rather than using the average market excess return as the
cutoff point between up markets and down markets, we can also use the
risk-free rate or the zero rate of return as the cutoff point. We define
downside and upside beta relative to the risk-free rate as:
Downside Risk
Table 4
Correlations of Beta Measures

rf
z

rf
z
rf
0
rf
0
1.000
0.779
1.000
0.770
0.971
1.000
0.762
0.967
0.990
1.000
0.760
0.464
0.447
0.447
1.000
0.769
0.444
0.467
0.461
0.972
1.000
0.776
0.439
0.452
0.464
0.969
0.991
1.000
further the cross-sectional determinants of future downside risk exposure in Section 3.

We now turn to additional robustness checks to make sure our findings
are being driven by variation in downside risk rather than by some
statistical bias introduced by our testing method. In Table 5, we subject
our results to a battery of additional robustness checks. Here, we check to
see if our results are robust to excluding small stocks, using longerfrequency data to compute and relative , creating value-weighted
portfolios, using all stocks, and using nonoverlapping annual observations.9 We report the robustness checks for realized in panel A and for
realized relative in panel B. In each panel, we report average 12month (or 24-month) excess returns of quintile portfolios sorted by
realized , or realized relative , over the same period. The table
also reports the differences in average excess returns between quintile
portfolios 1 and 5 with robust t-statistics.
One possible worry is that our use of daily returns introduces a bias due
to nonsynchronous trading. Indeed, one of the reasons for our focus on
just the NYSE is to minimize these effects. To further check the influence
of very small stocks, the first column of Table 5 excludes from our sample
stocks that fall within the lowest size quintile. When small stocks are
9
In addition, we conduct further robustness checks that are available upon request. In particular, to
control for the influence of nonsynchronous trading, we also repeat our exercise using control for
nonsynchronous trading in a manner analogous to using a Scholes-Williams (1977) correction to compute
the downside betas. Although this method is ad hoc, using this correction does not change our results. We
also find that the point estimates of the premiums are almost unchanged when we exclude stocks that fall
into the highest volatility quintile, but statistical significance increases dramatically.
1215
This table reports the time-series averages of cross-sectional correlations of various risk characteristics for
stocks listed on the NYSE. The risk characteristics are regular beta , downside beta , downside
beta defined relative to the risk-free rate rf ), downside beta defined relative to a zero return 0 ,
upside beta , upside beta relative to the risk-free rate rf , and upside beta relative to a zero return
0 . The regular downside and upside beta, and , respectively, are calculated relative to the
sample mean market return. All risk characteristics are computed using daily returns over the past 12
months. The sample period is from July 1963 to December 2001, with the last 12-month period from
January 2001 to December 2001. There are a total of 451 observations at a monthly frequency.
Table 5
Robustness Checks of Realized Downside Beta Portfolios
Portfolio
Exclude small
Two-year weekly
Value-weighted
Panel A: Sorts by realized b
All stocks
Nonoverlapping
4.48%
6.25%
7.69%
10.33%
12.81%
7.08%
12.54%
16.86%
21.34%
29.35%
2.62%
4.23%
6.04%
9.32%
9.76%
4.14%
7.23%
8.84%
11.35%
19.37%
3.69%
5.74%
8.33%
10.68%
16.15%
High-low
t-stat
8.34%
[4.54]
22.27%
[4.96]
7.14%
[3.30]
15.24%
[5.57]
12.46%
[3.51]
Panel B: Sorts by realized relative b

1 Low
2
3
4
5 High
5.07%
8.01%
8.66%
9.44%
10.37%
11.46%
15.78%
18.06%
20.48%
21.37%
3.19%
6.86%
7.07%
7.56%
7.18%
5.58%
8.66%
9.41%
10.86%
14.21%
5.67%
8.78%
8.42%
9.67%
12.05%
High-low
t-stat
5.30%
[6.46]
9.91%
[4.29]
3.99%
[3.06]
8.63%
[7.02]
6.38%
[3.87]
We perform robustness checks of the results in Table 1. For each month, we calculate and relative
using weekly (Wednesday to Tuesday) continuously compounded returns over each 24-month period or
daily continuously compounded returns over each 12-month period. We report the results using realized
in panel A and the results using realized relative (given by ) in panel B. For each risk
characteristic, we rank stocks into quintiles (15). In the first column of each panel, we form equalweighted portfolios among NYSE stocks, but exclude stocks that fall in the lowest size quintile. In the
second column, we form equal-weighted portfolios among NYSE stocks formed by ranking on or
computed using two years of weekly data. In the third column of each panel, we form valueweighted portfolios using stocks listed on the NYSE at the beginning of each 12-month period. In the
fourth column, we use all stocks listed on the NYSE, AMEX, and NASDAQ and form equal-weighted
portfolios at the beginning of each period, using quintile breakpoints based on NYSE stocks. In the last
column, we compute the risk characteristics using stocks listed on the NYSE and form equal-weighted
portfolios at the beginning of each January using nonoverlapping 12-month horizon observations. We
report the average return in excess of the one-month T-bill rate over the next 12 months. The row labeled
High-low reports the difference between the returns of portfolio 5 and portfolio 1. For the columns
labeled Exclude small, Value-weighted, and All stocks, the sample period is from July 1963 to
December 2001, with the last 12-month period from January 2001 to December 2001, with observations
at a monthly frequency. For the column labeled Two-year weekly, the sample period is from July 1963
to December 2001, with the last 24-month period spanning January 2000 to December 2001. For the
column labeled Nonoverlapping, the sample period is from January 1964 to December 2001, with the
last 12-month period lasting from January 2001 to December 2001. The number of stocks in each
portfolio varies across time from 216 to 317 stocks, except for All stocks, where it varies from 289
to 2330 stocks. The entry labeled t-stat in square brackets is the t-statistic computed using Newey-West
(1987a) heteroskedastic-robust standard errors with 24 lags for the second column, 1 lag for the last
column, and 12 lags for all other columns for the High-low difference.
removed, the difference between quintiles 5 and 1 for the stocks sorted by
realized remains strongly statistically significant (with a robust t-statistic
of 4.54) at 8.34% per annum, but is slightly reduced from the 51 difference
of 11.8% per annum when small stocks are included in Table 1. Similarly,
the 51 difference in average returns for relative also remains highly
significant.
1216
1 Low
2
3
4
5 High
Downside Risk
10
We have also reproduced the Fama-MacBeth regressions using risk measures calculated at the weekly
frequency and found virtually identical results to Table 2. In addition, when we examine realized betas
and realized returns over a 60-month horizon using monthly frequency returns, we find the same
qualitative patterns that are as statistically significant as using a 12-month horizon.
1217
The second column computes and ( ) using two years of

weekly data, rather than one year of daily data. We compute weekly
returns from Wednesdays to Tuesdays, and use two years of data to
ensure that we have a sufficient number of observations to compute the
factor loadings. We report the contemporaneous 24-month realized
return over the same 24-month period used to measure the downside
risk loadings. The table shows that there is no change in our basic
message: there exists a reward for exposures to downside risk and relative
downside risk.10
In the third column of Table 5, we examine the impact of constructing
value-weighted portfolios rather than equal-weighted portfolios. Using
value weighting preserves the large spreads in average excess returns for
sorts by and relative . In particular, the 51 spread of valueweighted quintile portfolios in realized returns from sorting on realized
is 7.1% per annum. Although this has reduced from 11.8% per annum
using equal-weighted portfolios in Table 1, the difference remains statistically significant at the 1% level. Similarly, the 51 spread in relative
portfolios in panel B reduces from 6.6% per annum using equal weighting
to 4.0% per annum with value weighting. This difference is also significant at the 1% level.
In the fourth column, labeled All stocks, we use all stocks listed on
the NYSE, AMEX, and NASDAQ, rather than restricting ourselves to
stocks listed on the NYSE. We form equal-weighted quintile portfolios at
the beginning of the period based on realized beta rankings. To keep our
results comparable with our earlier results, we use breakpoints calculated
over just the NYSE stocks. Using all stocks increases the average excess
returns, so our main results using only the NYSE universe are conservative. The 51 spreads in average excess returns increase substantially using
all stocks. For the (relative ) quintile portfolios, the 51 difference
increases to 15.2% (8.6%) per annum, compared to 11.8% (6.6%) per
annum using only NYSE stocks. By limiting our universe to NYSE
stocks, we deliberately understate our results to avoid confounding
influences of illiquidity and nonsynchronous trading.
In our last robustness check in Table 5, we use nonoverlapping
observations. While the use of the overlapping 12-month horizon in
Tables 1 and 2 is statistically efficient, we examine the effect of using
nonoverlapping 12-month periods in the last column of Table 5.
Our 12-month periods start at the beginning of January and end in
December of each calendar year. With nonoverlapping samples, it is not
necessary to control for the moving average errors with robust t-statistics,
2.5 Downside Risk in Size and Book-to-Market Portfolios

While we have demonstrated that exposures to high or high relative
loadings are compensated by high average returns and this effect is
consistent with investors placing greater weight on downside risk, we
have not demonstrated that downside risk is useful in pricing portfolios
sorted on other attributes. We now examine if portfolios of stocks sorted
by other stock characteristics also exhibit contemporaneous exposure to
downside risk. We focus on the Fama and French (1993) set of 25
portfolios sorted by size and book-to-market. To price these portfolios,
Fama and French (1993) develop a linear asset pricing model that augments the excess market return factor, rm , with size and value factors
(SMB and HML, respectively). Table 6 examines if these portfolios
exhibit exposure to downside risk, even after controlling for the standard
Fama-French model.
In Table 6, we examine linear factor models of the form:
m a bm rm bm r
m bSMB SMB bHML HML;
13
where r
m minrm ; m equals rm if the excess market return is below its
sample mean, or its sample mean otherwise. We estimate the coefficients
bm , bm , bSMB , and bHML by GMM using the moment conditions:
Emr 0;
1218
14
but we have fewer observations. Nevertheless, the results show that the
point estimates of the 51 spreads are still statistically significant at the 1%
level. Not surprisingly, the point estimates remain roughly unchanged
from Table 1.
In unreported results, we also conduct additional robustness checks
to value weighting and using all stocks in a Fama-MacBeth regression
setting. First, we run a set of value-weighted Fama-MacBeth regressions to make sure that small stocks are not driving our results. We
do this by running a cross-sectional weighted least squares regression
for each period, where the weights are the market capitalization of a
firm at the beginning of each period. Using value-weighted regressions
continues to produce a strong, statistically significant, positive relation between downside risk and contemporaneous returns with or
without any additional controls. Similar to the results of using all
stocks in the portfolio formations of Table 5, using all stocks in the
Fama-MacBeth regressions only increases the magnitude of the downside risk premium, which remains overwhelmingly statistically significant.
Downside Risk
Table 6
GMM Specification Tests
Linear factor model coefficients
a
bm
bm
bSMB
bHML
J 2 test
Specification I
1.02
5:25
[57.6]
[3.87]
22.84
[2.16]
Specification III
1.08
7:82
[34.5]
[5.21]
Specification IV
1.60
26:97
[7.70]
[3.42]
35.73
[2.56]
0:95
[0.50]
10:24
[4.41]
3:63
[1.52]
11:09
[4.11]
H0 : Spec I
vs H1 : Spec II
3.85
(0.05)
H0 : Spec III
vs H1 : Spec IV
9.49
(0.00)
This table reports GMM estimates and GMM specification tests using the 25 Fama and French (1993)
size and book-to-market sorted portfolios as base assets. Each linear factor model specification takes the
form of Equation (13), with specifications IIII being special cases of Equation (13). In each case, we
estimate the coefficients using the moment conditions in Equation (14). The column labeled J
provides a 2 difference test of a restricted specification H0 versus an alternative specification H1 .
We report the 2 statistic with p-values below in parentheses. The sample period is from July 1963 to
December 2001. In the table, robust t-statistics are reported in square brackets.
where r is a vector of the 25 Fama-French portfolios. The coefficient bm

reveals the exposure of the test portfolios to downside risk.11 We conduct
a 2 specification test to examine the fit with and fit without the downside risk exposure of various specifications of Equation (13). Specifically,
we compute a J 2 difference test of Newey and West (1987b) using an
optimal weighting matrix of the moment conditions under the unrestricted model of the alternative hypothesis. In particular, if we reject the
null hypothesis that bm 0, then we conclude that the restrictions
imposed by the null hypothesis model that there is no downside risk
exposure is too restrictive.
We consider two null models in Table 6, the null of the CAPM (Specification I) and the null of the Fama-French model (Specification III). In
both alternatives (Specifications II and IV), Table 6 shows that the
coefficient bm is statistically significant at the 5% level. This indicates
that, for pricing the size and book-to-market portfolios, the downside
portion of market return plays a significant role, even in the presence of
11
Note that the factor r

m is not traded, and thus the alpha of a time-series regression using the factors in
Equation (13) does not represent the return of an investable strategy. Therefore, the alpha cannot be
tested against the value of zero to examine possible mispricing. Similarly, we cannot compute a premium
for a traded downside risk factor from Equation (13).
1219
Specification II
1.35
17:73
[8.70]
[3.03]
3. Predicting Future Downside Risk

The previous section demonstrates a strong positive relationship between
stocks that exhibit high downside risk and returns for holding such stocks
over the same period. While this is the essence of the relationship implied
by a risk-to-reward explanation, knowing this relationship may not be of
practical value if we cannot predict downside risk prior to the holding
period. Therefore, we now examine if we can predict downside risk in a
future period using past information. If todays information can predict
future downside risk, then we can form an investable trading strategy that
has exposure to downside risk. In particular, one potential valuable
predictor of future downside risk exposure may be the covariation of a
stock with market down movements in the past if downside risk is a
persistent risk characteristic of stocks.
In Section 3.1, we explore the determinants of downside risk using past
information. Section 3.2 examines returns of portfolios of stocks sorted
by past downside risk, as well as sorts by past coskewness. We find that
forecasting future downside risk is difficult for stocks with high volatility
and we explore why in Section 3.3. Nevertheless, we can forecast future
downside risk, and predict high future returns for stocks with high past
downside risk for a large portion of the market. We conduct additional
robustness checks in Section 3.4.
3.1 Determinants of Future Downside Risk Exposure
In this section, we begin by exploring the cross-sectional determinants of
downside risk. Since we have very little theoretical guidance as to what
firm characteristics determine riskiness of a stock, our investigation is
merely exploratory in nature. Our analysis complements Harvey and
Siddique (1999), who characterize coskewness loadings, and Ang and
Chen (2002), who examine how some stock characteristics are related
1220
the standard market factor. This is true even when we allow for SMB and
HML to be present in the model. This is a strong result because the SMB
and HML factors are constructed specifically to explain the size and
value premia of the 25 Fama-French portfolios. For both the CAPM
and the Fama-French model, the J test strongly rejects both specifications in favor of allowing for downside market risk. For the FamaFrench model, the p-value of the rejection is almost zero.
Thus, not only do individual stocks sorted directly on loadings
reveal a large reward for stocks with high downside risk exposure, but
other portfolios commonly used in asset pricing also exhibit exposure to
downside risk. In particular, linear factor model tests using the Fama and
French (1993) size and book-to-market portfolios reject the hypothesis
that these portfolios do not have exposure to downside market risk.
Downside Risk
12
Industry classifications are based on groupings of two-digit Standard Industrial Classification (SIC)
codes following Ferson and Harvey (1991). They are Miscellaneous, Petroleum, Finance, Durables, Basic
Industry, Food & Tobacco, Construction, Capital Goods, Transportation, Utilities, Textile & Trade,
Service, and Leisure. In unreported results, we find very little pattern of downside risk exposure across
industries, except that utilities generally exhibit lower exposure to downside risk than other industries.
This is consistent with the notion that utilities are traditionally defensive stocks that tend to hold their
value relative to other industries during market downturns.
1221
with upside and downside correlations. Neither one of these studies

examine how downside beta is related to firm characteristics at the
individual stock level. In Table 7, we explore how relative downside
betas are correlated with other cross-sectional variables. We run FamaMacBeth (1973) regressions of realized relative downside beta on various
firm characteristics that are known ex ante and on other future risk
characteristics measured ex post. In particular, finding some firm characteristics that are cross-sectionally correlated with future downside beta
can help us develop some possible investable strategy that generates a
future spread in downside beta, and hence, future returns.
In Table 7, we first consider regressions of future realized relative downside beta over the next 12 months on variables that are estimated over the
past 12 months in the columns labeled Past variables. All the independent variables in these regressions are measured in a period prior to the
realization of relative downside beta. The regressions are run at a monthly
frequency, so we use 12 Newey-West (1987a) lags. Regressions IX use one
independent variable at a time, in addition to industry dummy variables,
while regression XI uses all the past variables simultaneously.12
Regression I shows that past relative beta does predict future relative
beta over the next 12 months. However, while highly statistically significant, the coefficient is only 0.077 and far from one. Hence, future relative
downside beta is difficult to predict simply by using past relative downside risk. In fact, the average 12-month autocorrelation of relative downside risk across all NYSE stocks is only 0.082, which is significantly lower
than the average autocorrelation of downside beta at 0.435 and regular
beta at 0.675. Thus, although past downside risk can also be used to
predict future downside risk, we should expect that using only past downside risk to predict future downside risk may be difficult.
In regressions IIX, we examine how past firm characteristics are
related to future relative downside beta. Past standard volatility increases
relative downside beta, while smaller stocks tend to exhibit greater relative downside risk. While regression IV suggests that value stocks exhibit
greater relative downside risk, controlling for additional characteristics in
regression XI indicates that it is growth stocks that tend to have more
relative downside risk. This is consistent with the results in Table 6 that
downside risk is present in the Fama-French size and book-to-market
portfolios. Stocks that are past winners also exhibit greater relative downside risk, consistent with the interpretation in Rouwenhorst (1998).
Past variables
I
II
III
IV
V
VI
VII
VIII
IX
X
Past
rel. beta
Past
std dev
Log
size
Bk-mkt
Past
ret
ROE
Asset
growth
Sales
growth
Leverage
Ind
[Div]
Std
dev
Coskew
0.007
[10.7]
Cokurt
Liquidity
beta
Indstry
dummy?
Yes
0.251
[6.10]
Yes
0:038
[10.8]
Yes
0.007
[2.04]
Yes
0.052
[4.85]
Yes
0:020
[3.04]
Yes
0.003
[0.43]
Yes
0.006
[1.04]
Yes
0.000
[0.13]
Yes
0:066
[6.82]
Regression
Realized risk measures
Yes
1222
Table 7
Determinants of Relative Downside Beta
Past variables
Regression
XI
XII
XIII
Past
rel. beta
0.040
[10.7]
0.022
[9.19]
0.017
[5.72]
Past
std dev
0.094
[2.58]
0:150
[6.39]
0:145
[5.88]
Log
size
Bk-mkt
Past
ret
ROE
0:035
[12.8]
0:042
[18.8]
0:041
[18.5]
0:010
[3.56]
0:006
[3.08]
0:005
[2.64]
0.063
[8.12]
0.003
[0.78]
0.001
[0.33]
0:011
[1.10]
0.007
[1.21]
0.008
[1.33]
Realized risk measures

Asset
growth
0.032
[2.62]
0.016
[2.02]
0.011
[1.40]
Sales
growth
0:001
[0.09]
0:009
[1.34]
0:001
[1.46]
Leverage
0.001
[1.47]
0.000
[0.63]
0:001
[1.46]
Ind
[Div]
0:005
[1.08]
0.004
[1.16]
0.004
[1.10]
Std
dev
Coskew
Cokurt
Liquidity
beta
Indstry
dummy?
Yes
0.273
[7.08]
0.275
[6.79]
2:441
[23.5]
2:484
[22.5]
0.092
[3.21]
0.106
[3.48]
Yes
0.006
[1.54]
Yes
This table reports the results of Fama-MacBeth (1973) regressions of realized relative downside risk over a 12-month period on various firm characteristics and risk
measures. The independent variables include both past variables in the information set observable at time t (Past variables) and also risk measures with contemporaneous
realization as the dependent variables (Realized risk measures). For all specifications except regression XIII, the sample period is from July 1963 to December 2001, with the
last 12-month period from January 2001 to December 2001, and observations are at a monthly frequency (451 months) for all stocks listed on the NYSE. For regression XIII,
the sample period is from January 1967 to December 2001 (397 months). The t-statistics in square brackets are computed using Newey-West (1987a) heteroskedastic-robust
standard errors with 12 lags. All regressions include industry dummy variables using the industry classification codes of Ferson and Harvey (1991). The past variables include
realized relative downside risk beta (Past relative ) and realized volatility (Past std dev) over the previous 12-month period. The firm characteristics are log of market
capitalization (Log-size), the firm book-to-market ratio (Bk-mkt), and past 12-month excess returns (Past ret), all computed at the beginning of each period. We also
include firm growth measured over the most recently available four-quarter periodreturn on equity (ROE), the growth rate of assets (Asset growth), and the growth rate
of sales (Sales growth)as well as book leverage (Leverage), and an indicator which equals one if the firm pays dividends (Ind[Div]). The realized risk characteristics
measured contemporaneously as the realized relative downside beta dependent variable are the standard deviation of excess returns (Std dev), coskewness, and cokurtosis. All
of the realized characteristics are computed over the same 12-month period as the relative downside beta using daily continuously compounded returns. We also include the
realized Pastor-Stambaugh (2003) liquidity beta, L , for January 1967 to January 2001. All independent variables are Winsorized at the 1% level and at the 99% within each
month.
Downside Risk
Table 7
(continued)
1223
3.2 Past Downside Risk and Future Returns

The results of Section 2 suggest that downside betas are contemporaneously correlated with high average returns. However, a trading strategy
to exploit this pattern of expected returns must successfully predict future
downside beta loadings from past information. We now examine if stocks
13
In an unreported table, we do find that firms with low ROE, high growth, and high leverage have high
downside betabut, these characteristics do not add more than what is already captured by regular beta.
We also find that dividend-paying firms exhibit less downside risk, without accounting for the regular
beta.
1224
Regressions VIVIII relate accounting measures of performance

(return on equity, ROE; asset growth; and sales growth) with downside
risk. There is strong evidence that stocks with high past ROE tend to have
high future exposure to downside risk. In contrast, the evidence for asset
or sales growth as predictors of future downside relative beta is weak, as
either one of these variables are statistically insignificant on their own, or
statistically insignificant in the full regression XI. Regression IX shows
that there is little leverage effect in relative downside risk. In regression X,
we use a dummy variable that equals one if the firm has paid a dividend in
the past 12 months. Dividend-paying firms exhibit less relative downside
risk, though this relation disappears when we include all of these variables
simultaneously in regression XI.13
In regressions XII and XIII of Table 7, we include other contemporaneous risk measures on the right-hand side. The motivation behind these
regressions is that we want to remove any possible confounding effects of
other risk factors that are correlated with downside risk. We include as
independent variables the stocks standard deviation, coskewness, cokurtosis, and Pastor-Stambaugh (2003) liquidity betas, all of which are
measured over the same 12-month interval as the realized relative downside beta. We find that high volatility, negative coskewness, and high
cokurtosis are all related to greater future relative downside risk. Note
that realized high relative downside beta is contemporaneously correlated
with negative coskewness and high cokurtosis, both of which are associated with average high returns [see Harvey and Siddique (2000) and
Dittmar (2002)]. The earlier Tables 2 and Table 3 show that the downside
beta is different from coskewness and cokurtosis risk. However, the effect
of standard deviation is just the oppositehigh downside beta is contemporaneously associated with high average returns, but high standard
deviation is associated with lower average returns. This is consistent with
Ang et al. (2006). Therefore, regression XII suggests that, in order to find
an investable strategy that provides a spread in downside beta, controlling for volatility may be necessary. Finally, regression XIII shows that
the Pastor-Stambaugh liquidity effect is not related to relative downside
risk.
Downside Risk
14
We find that past has no predictive ability for future returns, even excluding the most volatile stocks
as in Section 3.3. This is consistent with the results in Section 2, which fails to find a consistent
contemporaneous pattern in expected returns for realized risk. These results are available on request.
1225
simply sorted on past downside risk provide enough variation in future

downside risk to provide large spreads in future returns. However, given
the results of Table 7, we expect that using only past downside beta to
predict future downside beta might be difficult, since past downside beta
is a poor predictor of future downside beta. We confirm that this is indeed
the case using a portfolio trading rule. Fortunately, the Fama-MacBeth
regressions in Table 7 also suggest a way of refining the simple trading
strategies, which we implement in Section 3.3.
To construct these investable portfolios, we sort stocks into portfolios
at time t based on preformation characteristics and then examine
monthly holding period returns from t to t 1. Our main preformation
criterion is downside beta computed using daily returns over the past 12
months. At the beginning of each month t, we sort stocks into five
quintiles based on their past and coskewness. In the column labeled
Return in Table 8, we report the average realized excess return over the
next month from t to t 1. The sample period is from July 1962 to
January 2001, with our first 12-month risk measurement period ending
in June 1963 for the portfolio formation in July 1963.
Panel A of Table 8 reports the differences in the excess returns between
the highest and the lowest past quintile portfolios in the row labeled
High-low. We also report the cross-sectional realized , , and of
each quintile portfolio. These realized factor loadings are computed over
the following 12 months. The last column of Table 8 shows the crosssectional realized coskewness of each quintile portfolios, also computed
over the next 12 months.14
Panel A reports a monotonically increasing pattern in the realized
downside betas of stocks sorted on past downside betas. That is, stocks
with low (high) in the past continue to have low (high) going
forward. The difference in the realized between portfolios 5 and 1
sorted on past is 0.80. Hence, past seems to predict future , but
this range is far less than the range of loadings for portfolios sorted on
realized downside beta. In contrast, Table 1 shows that the range of
realized loadings is much wider at 1.72. Although past downside
beta is persistent, it is far from a perfect predictor of future downside
risk exposure. This is also illustrated by the 12-month autocorrelation of
among stocks listed on the NYSE, which is only 0.43. The difference
in realized , at 0.77, is similar to the difference in realized . Panel A
also shows that the past variation of produces little variation in
realized coskewness. Hence, the past sorts are not reflecting coskewness risk.
Table 8
Returns of Stocks Sorted by Past Asymmetry Measures
Realized statistics
Portfolio
Average next
month return
Panel A: Stocks sorted by past b
Coskew
0.59%
0.71%
0.77%
0.84%
0.54
0.70
0.85
1.02
0.61
0.77
0.93
1.11
0.48
0.65
0.80
0.96
0:13
0:15
0:17
0:18
5 High
0.70%
1.31
1.41
1.25
0:18
High-low
t-stat
0.11%
[0.60]
0.77
0.80
0.77
0:05
1 Low coskew
2
3
4
5 High coskew
0.84%
0.82%
0.76%
0.60%
0.57%
0.91
0.90
0.89
0.87
0.85
1.01
0.99
0.97
0.95
0.91
0.85
0.84
0.83
0.82
0.81
0:19
0:17
0:16
0:15
0:13
0:28%
[2.76]
0:06
0:10
0:04
0.05
Panel B: Stocks sorted by past coskewness
High-low
t-stat
This table reports the equal-weighted average returns and risk characteristics of stocks sorted by past
(panel A) and past coskewness (panel B). For each month, we compute and coskewness with respect
to the market of all stocks listed on the NYSE using daily continuously compounded returns over the
previous 12 months. For each risk characteristic, we rank stocks into quintiles (15) and form equalweighted portfolios at the beginning of each month. The sample period is from July 1962 to January
2001. The number of stocks in each portfolio varies across time from 221 to 346 stocks. The column
labeled Average next month return reports the average return in excess of the one-month Treasury bill
rate over the next month. The row labeled High-low reports the difference between the returns of
portfolio 5 and portfolio 1. The entry labeled t-stat is the simple OLS t-statistic in square brackets. The
columns labeled , , and report the time-series averages of equal-weighted cross-sectional
averages of individual stock betas over the next 12-month period. The column labeled Coskew reports
the time-series averages of equal-weighted cross-sectional averages of individual stock coskewness over
the next 12-month period of the High-low difference.
While the ex post loadings maintain their monotonic pattern, panel

A disappointingly shows a very weak relationship between past and
future returns. We do observe a strictly increasing pattern moving from
the first-quintile portfolio to the fourth-quintile portfolio from
0.59% to 0.84% per month. This difference is statistically significant with
a t-statistic of 2.06. However, the highest quintile portfolio has a low
average excess return of 0.70% per month. Thus, the predictive relation is
not as strong among stocks with the highest past .
In contrast, panel B of Table 8 shows the relation between past coskewness, future returns, and future risk attributes. The portfolios of stocks
sorted by past coskewness do not exhibit large ex post variation in
coskewness. The average preranking spread in coskewness between
1226
1 Low
2
3
4
Downside Risk
3.3 Investable Portfolios with a Spread in Downside Risk

The Fama-MacBeth regressions in Table 7 suggest that a simple trading
strategy of forming portfolios on past to predict future exposure to
downside risk can be refined by considering subsets of portfolios focusing
on volatility, book-to-market, and momentum interactions with downside risk. In this section, we consider one possible refinement of predicting future using past , which the Fama-MacBeth regressions of the
determinants of in Table 7 suggest is an important confounding
influence. Specifically, we examine the role of stock return volatility and
and show that only for stocks with the highest volatility levels does
past high fail to predict future high average returns.
There are two reasons why high stock volatility causes past to be a
poor predictor of future . First, in order for the strong contemporaneous pattern between loadings and average returns to be exploited in
an investable portfolio strategy, we must accurately forecast future
exposure. When return volatility is very high, the past estimates
contain substantial measurement error. Thus, high volatility causes the
persistence of to be lower, and thus future loadings to be less
predictable, because of sampling error.
Second, Ang et al. (2006) identify a puzzling anomaly that stocks with
very high total or idiosyncratic volatility have extremely low returns. We
suspect that this effect confounds the relationship between high and
high return. High-volatility stocks also tend to be high-beta stocks
because, holding correlation between the market and the stock return
constant, a high individual stock volatility implies a high . To clearly see
the confounding interaction between downside beta and volatility, we
rewrite Equation (5) as:

i
;

m
15
1227
portfolios 5 and 1 is 0.50, while the average postranking spread in realized

coskewness is only 0.05. Nevertheless, we observe a strong, strictly
decreasing pattern in the returns of stocks sorted by past coskewness.
The portfolio of the most negative coskewness stocks (quintile 1) has an
average excess return of 0.84% per month, whereas the portfolio of the
least negative coskewness stocks (quintile 5) has an average excess return
of 0.57% per month. This pattern is consistent with Harvey and Siddique
(2000). The strong predictive pattern of past coskewness and future
returns does not have a relationship with downside beta. The future
loadings from past coskewness are almost flat. Hence, the predictive
pattern for cross-sectional returns from past coskewness is not picking
up downside risk.
15

In contrast, when
i is held constant, increasing downside correlation can only increase . Hence, we
tend to see high average future returns for stocks with high past downside correlation. Ang, Chen, and
Xing (2002) report that the difference in average future returns over the next month between the tenth
and first decile portfolios sorted on past is approximately 5% per annum.
16
Since we use all stocks listed on the NYSE, AMEX, and NASDAQ in constructing size and book-tomarket benchmark portfolios, the average adjusted returns of all stocks listed on the NYSE does not sum
up to 0.0%.
1228
wherep
corrri ; rm jrm < m p
is
downside correlation,
and

i varri jrm < m and
varrm jrm < m represent stock
m
and market volatilities conditional on down markets, respectively. High
downside beta can be produced by high downside correlation, , or by

high downside volatility,
i . But, holding constant , stocks with high

volatility, or i , tend to have low returns, which is exactly opposite to
the high , high average return effect that we wish to observe.15 Therefore the Ang et al. (2006) volatility effect works in the opposite way as the
expected return pattern for , making it hard to predict downside risk
for stocks with very high volatility.
Table 9 shows that past poorly predicts future only for high
volatility stocks. In panel A, we report selected summary statistics of
portfolios sorted by volatility, measured using daily continuously compounded returns over the previous 12 months. For the overall sample and
for each of the highest volatility groups, we report the average market
capitalization, past , and average returns adjusted for size and bookto-market using a characteristic control similar to Daniel, Grinblatt,
Titman, and Wermers (1997). While the average stock has an annualized
volatility of 36%, stocks in the highest volatility quintile have a considerably higher average volatility of 61% per annum. Stocks in the highest
demi-decile (5%-tile) have an average volatility of over 85% per annum.
The second row of panel A reports that the quintile of stocks with the
highest volatility constitutes, on average, only 3.9% of the overall market
capitalization. Hence, by excluding the highest quintile of volatile stocks,
we exclude stocks that represent only a small fraction of the market. In
fact, the highest demi-decile (5%-tile) of the most volatile stocks constitutes only 0.4% of the total market capitalization. Hence, not surprisingly, the highest volatility stocks tend to be small. In the third row, we
report the size and book-to-market adjusted returns.16 Stocks in the
highest volatility quintile (demi-decile) underperform their benchmark
portfolios by, on average, a large 0.38% (0.67%) per month. This is the
puzzling Ang et al. (2006) effectstocks with very high total or idiosyncratic volatility have low average returns.
The last two rows of panel A explore the interaction between volatility
and . While an average stock has a past loading of almost one
(0.99), high-volatility stocks tend to have high past . In particular, the
average past is 1.44 for the stocks in the highest volatility quintile. We
Downside Risk
Table 9

Stocks Sorted by Past b Excluding the Most Volatile Stocks
Panel A: Selected characteristics of volatility portfolios
Stocks of the highest volatility
Quintile
Octile
Decile
5%-tile
100.0%
35.9%
0.08%
0.99
43.5%
3.9%
61.0%
0.38%
1.44
25.8%
1.9%
68.4%
0.50%
1.47
21.8%
1.3%
72.5%
0.56%
1.48
20.5%
0.4%
85.6%
0.67%
1.44
17.3%
Panel B: Average excess returns

Quintiles excluding most volatile stocks
1 Low
0.58%
0.69%
0.82%
4
0.82%
5 High
0.92%
High-low
Q4-low
0.34%
[2.31]
0.25%
[2.28]
In panel A, each month, we calculate individual stock volatility of all stocks listed on the NYSE using
daily continuously compounded returns over the previous 12 months. We sort stocks according to
volatility into quintiles, octiles, deciles, and demi-deciles (5%-tiles). Panel A reports selected average
characteristics of stocks in each volatility group. The first column reports the characteristics over the
entire sample. The other columns report the characteristics within the highest volatility groups. The row
labeled Annualized volatility reports the average stock volatility over the past 12 months, while the row
labeled Market cap reports the time-series averages of cumulative market capitalization represented by
the stocks in each group. The other rows report the returns adjusted for size and book-to-market using a
characteristic control similar to Daniel et al. (1997), the past over the previous 12 months, and the
autocorrelation of between the past 12 months and the following 12 months. For each characteristic,
we report the time-series averages of equal-weighted cross-sectional averages. In panel B, we report the
average return in excess of the 1-month Treasury bill rate over the next one month of portfolios sorted
on past that exclude the highest volatility quintile of stocks. Each month, we first sort stocks into
quintiles according to volatility measured using daily continuously compounded returns over the
previous 12 months. Then, we exclude the stocks that fall into the highest volatility quintile and rank
the remaining stocks into equal-weighted quintiles (15) according to past measured using continuously compounded returns over the previous 12 months. We report the average excess return over the
next month. The row labeled High-low (Q4-low) reports the difference between the returns of
portfolio 5 (portfolio 4) and portfolio 1. We report simple t-statistics in square brackets. The number
of stocks in each portfolio varies across time and groupings from 177 to 346 stocks. The sample period is
from July 1962 to January 2001.
compare this with stocks in the highest realized quintile of Table 1,

which have an average of 1.92. Thus, on the one hand, stocks with
high volatility tend to have high . On the other hand, high stocks
are not completely stocks with high volatilities.
Moreover, the fact that high-volatility stocks have high past loadings does not imply that they continue to exhibit high loadings in the
following period. For stocks in the highest volatility quintile, the average
loading over the next 12 months is only 1.25, compared to the over
the past 12 months of 1.44. This accounts for the fairly low persistence of
the 12-month downside betas of 26%, compared to the average 12-month
autocorrelation of 43.5%. If we narrow our focus to stocks on the
1229
Proportion of market capitalization

Annualized volatility
Size/book-to-market adjusted returns
Past
12-month autocorrelation of
All stocks
3.4 Robustness Checks

In Table 10, we check that the predictive relation between past and
future returns, excluding the most volatile stocks, is not due to size,
book-to-market, momentum, coskewness, or liquidity effects. The first
column of Table 10 shows the average size and book-to-market adjusted
returns of portfolios sorted by past . We observe a monotonically
increasing relation between past and next month returns. Controlling
for size and book-to-market increases the average difference in returns
between the lowest and highest quintile to 0.44% per month from
0.34% per month in panel B of Table 9. This difference is highly statistically significant, with a t-statistic of 3.36.
17
If we exclude fewer stocks and only exclude stocks in the highest volatility octile or decile, the return
difference between highest portfolio and the lowest portfolio is about the same order of magnitude
(roughly 0.31% per month), but the statistical significance is somewhat weaker, with p-values of 0.051
and 0.063, respectively. However, the return difference between quintiles 1 and 4 is unaffected for all the
volatility exclusions at 0.25% per month and is always statistically significant at the 5% level. This is due
to the fact that the volatility effect is strongest among high-volatility stocks which tend to have the
highest past .
1230
highest volatility octile, decile or demi-decile, we find that these stocks

exhibit even lower persistence of across the formation period and
future holding periods. We conjecture that this lower correspondence of
past to future arises, in part, from the fact that high volatility
increases the sampling error of the past as a predictor of future
downside risk. Large measurement error makes the preformation
loadings a more unreliable predictor of actual, postformation, downside
betas. If we focus on stocks with lower volatility, our past loadings
are more accurate ex ante predictors of future ex post downside risk
loadings.
In panel B of Table 9, we examine the average return patterns to
ranking on past when we exclude the most volatile stocks. Fortunately, panel A shows that the segment of stocks where past provides
a poor predictor of future constitutes only a small fraction of total
market capitalization (less than 4%). We first sort stocks based on their
past total volatility over the past year. Then, we exclude stocks in the
highest volatility quintile and sort the remaining stocks into quintiles
according to past . We report the difference in average (unadjusted)
returns between the highest quintile and the lowest quintile portfolios,
as well as the difference in returns between the second highest and the
lowest quintile portfolios. Panel B shows that, by excluding stocks in the
highest volatility quintile, the average difference between the highest and
the lowest quintile portfolios of 0.34% per month is statistically
significant with a t-statistic of 2.31. The difference between the second
highest and the lowest quintile portfolios is slightly smaller at 0.25%
per month, but has approximately the same statistical significance.17
Downside Risk
Table 10

Characteristic Controls on Stocks Sorted by Past b
Including additional controls for
Momentum
Coskewness
Liquidity
1 Low
2
3
4
5 High
0:25%
0:09%
0.05%
0.07%
0.20%
0:21%
0:07%
0.04%
0.10%
0.12%
0:21%
0:07%
0.07%
0.04%
0.15%
0:17%
0:02%
0.05%
0.10%
0.13%
High-low
t-stat
0.44%
[3.36]
0.32%
[2.71]
0.36%
[2.69]
0.30%
[2.15]
The table reports robustness checks of the results in Table 9. For each month, we compute individual stock
volatility and with respect to the market of all NYSE stocks using daily continuously compounded
returns over the previous 12 months. We first sort stocks according to volatility into quintiles and exclude
stocks that fall within the highest volatility quintile. We rank the remaining stocks into quintiles (15)
according to past and form equal-weighted portfolios at the beginning of each month. The table reports
characteristic-adjusted holding period returns over the next month of the quintiles that exclude stocks
in the highest quintile. In column labeled Size/bk-mkt adjusted, we report the average returns in excess
of size and book-to-market matched benchmark portfolios similar to Daniel et al. (1997). In the next three
columns, we include additional controls for momentum (as measured by past 12-month returns), past
coskewness, and past liquidity betas, computed following Pastor and Stambaugh (2003). For each additional control, we first perform a quintile sort based on the characteristic and then on past excluding the
highest quintile of stocks. Then, we average the quintiles across the characteristic quintiles and report
size and book-to-market matched returns within each quintile. The number of stocks in each portfolio
varies across time from 177 to 277 stocks. The row labeled High-low reports the difference between the
returns of quintile portfolios 1 and 5. The entry labeled t-stat in square brackets is the simple t-statistic
for the High-low difference. The sample period is from July 1962 to January 2001, except in the last
column, where the sample period is from January 1967 to January 2001.
In the next three columns, we control for additional past return characteristics: momentum, coskewness, and liquidity. For each additional
control, we first perform a quintile sort based on the characteristic and
then on past excluding the highest quintile of stocks. Then, we
average the quintiles across the characteristic quintiles and report size
and book-to-market matched returns within each quintile. To control
for momentum, we use past 12-month returns. Liquidity is measured
using the historical liquidity betas of Pastor and Stambaugh (2003).
Table 10 clearly shows that the spreads in size and book-to-market
adjusted returns between the quintiles 5 and 1 remain significant
at the 5% level after controlling for momentum, coskewness, and liquidity. In each case, the point estimates of the differences average over 0.30%
per month. Therefore, our predictive pattern of returns for past are
not due to size, book-to-market, past return, coskewness, or liquidity
effects.
Finally, as an alternative control for volatility, we consider orthogonalizing the portfolios with respect to volatility in Table 11. We first
rank stocks according to past 12-month volatility into quintiles. Then, we
1231
Size/bk-mkt adjusted
Table 11
Returns of Stocks Sorted by Past Volatility and Past Realized Downside Beta
Past volatility quintiles
Portfolio
5 High
Average
1 Low
2
3
4
5 High
0:26%
0:23%
0:15%
0.01%
0.08%
0:07%
0.01%
0.07%
0.11%
0.17%
0:21%
0:01%
0.03%
0.23%
0.22%
0:27%
0.08%
0.00%
0.17%
0.14%
0:43%
0:37%
0:31%
0:47%
0:31%
0:25%
0:13%
0:07%
0.01%
0.06%
High-low
t-stat
0.35%
[3.25]
0.25%
[2.35]
0.43%
[3.59]
0.41%
[2.84]
0.12%
[0.55]
0.31%
[3.14]
The table examines the relation between past realized volatility and past realized . For each month, we
compute past volatility and past with respect to the market of all stocks listed on NYSE using daily
continuously compounded returns over the past 12 months. We first rank stocks into quintiles (15) at
the beginning of each month based on volatility calculated over the previous 12 months. Then, we rank
stocks within each first-sort quintile into additional quintiles according to realized , computed over
the previous 12 months. For each 5 5 grouping, we form an equal-weighted portfolio. In each cell, we
report the average returns in excess of size and book-to-market matched benchmark portfolios similar to
Daniel et al. (1997). In the column labeled Average, we report the average size and book-to-market
adjusted return of stocks in each second sort quintile, which controls for realized volatility. The row
labeled High-low reports the difference between the returns of quintile portfolios 5 and portfolio 1.
The entry labeled t-stat in square brackets is the simple t-statistic. The sample period is from July 1963
to December 2001.
perform a second ranking within each volatility quintile into quintiles

according to over the past 12 months. We then form equal-weighted
portfolios within each doubly sorted group, hold the portfolio over the
next one month, and report average size and book-to-market adjusted
returns.
Table 11 shows a strong positive relation between past downside beta
and the next one-month size and book-to-market adjusted return within
each volatility quintile. In fact, the differences between the highest and the
lowest downside beta quintile returns are significant in all quintiles,
except in the highest volatility quintile. When we average the returns of
portfolios across volatility quintiles (in the last column), we find that
there is a statistically significant difference in returns of 0.31% per month
across the highest quintile downside beta stocks and the lowest quintile
downside beta stocks. Hence, controlling for volatility, there is a strong
predictive relation between past and future returns.
In summary, although there is a strong contemporaneous relation
between downside risk and average returns, this relation does not hold
predictively for stocks with very high volatility. For the vast majority of
stocks, we can use past downside beta to predict future returns. For
stocks with high volatility, factor loadings contain large measurement
error, making past downside beta a less reliable predictor of future downside risk, and Ang et al. (2006) report that stocks with high volatility have
1232
1 Low
Downside Risk
low returns. Fortunately, these very volatile stocks constitute a small

fraction of the total market capitalizations of only 4%. If we focus on
stocks not in the highest volatility quintile, we find a strong pattern
between past and future holding period returns. However, by not
examining the predictive ability of downside risk among stocks with very
high volatility, we completely abstract from the puzzling Ang et al. (2006)
volatility effect, which we cannot resolve in this article.
4. Conclusion
1233
The cross section of stock returns reflects a premium for downside risk.
Stocks that covary strongly with the market, conditional on market
declines, have high average returns. This riskreturn relation is consistent
with an economy where agents place greater weight on downside risk than
they place on upside gains. Agents with aversion to downside risk require
a premium to hold assets that have high sensitivities to market downturns. Hence, stocks with high downside risk exposure, or high downside
betas, have high average returns.
We find that the contemporaneous high average returns earned by
stocks with high downside betas are not explained by a list of crosssectional effects, including size and book-to-market effects, coskewness
risk, liquidity risk, and momentum effect. The effect is also different from
regular market beta. Controlling for these and other cross-sectional
effects, we estimate that the cross-sectional premium for bearing downside beta risk is approximately 6% per annum. The downside premium is
robust across a battery of robustness tests. In particular, we find that the
premium captured by downside beta is quite different from the coskewness effect of Harvey and Siddique (2000). Downside beta measures risk
conditional only on market declines, whereas coskewness captures the
unconditional covariation of a stock with extreme downside movements
of the market. In contrast, we find that the premium for upside risk is
weak in the data and often changes signs depending on the set of crosssectional risk controls.
Past downside beta is a good predictor of future covariation with
down market movements, except for stocks that are extremely volatile.
For the vast majority of stocks, past downside beta cross-sectionally
predicts future returns. However, for stocks with very high volatility,
consisting of less than 4% of market capitalization, past downside beta
provides a poor predictor of future downside risk. While high-volatility
stocks constitute only a small fraction of the total market, so a predictive downside beta relationship holds for the vast majority of stocks,
it remains to be explored why the cross-sectional predictive relation
for downside risk does not hold for stocks with very high levels of
volatility.
Appendix
A Solution of the Disappointment Aversion Asset Allocation Problem

@UW
@UW
x1fW W g A E
x1fW >W g 0
@W
@W

@UW
@UW
y1fW W g A E
y1fW >W g 0;
E
@W
@W
E
and
A-1
where 1fg is an indicator function.

Over a discrete-state space over states xs ; ys indexed by s, the definition of the certainty
equivalent W in Equation (1) can be written as:
1
W
1
K
ps Ws1
s:Ws W
!
Aps Ws1
A-2
s:Ws >W
where wealth in state s is given by:

W s R f wx x s wy y s ;
and the FOC (A-1) take the form:

"
s:Ws W
"
and
"
ps Ws xs A
s:Ws >W
#
ps Ws ys
#
ps Ws xs 0
"
s:Ws W
A-3
ps Ws ys 0:
s:Ws >W
Ang, Bekaert, and Liu (2005) note that Equation (A-3) is a standard CRRA maximization
problem with a changed probability measure, where the probabilities for wealth above the
certainty equivalent are down-weighted. That is, defining the probabilities as:
i
p1 ; ; pi ; Api1 ; ; ApN
;
p1 pi Api1 pN
A-4
allows Equation (A-3) to be re-written as:

X
s Ws xs 0
and
s Ws ys 0:
A-5
The algorithm starts with a state i, solves the standard CRRA problem with probability
distribution f i g for the optimal portfolio weights w*xi and w*yi , and then computes the
certainty equivalent *Wi , given by:
*Wi
N
X
s1
1234
1
!1
Ws* 1 is
Ang, Bekaert, and Liu (2005) develop an algorithm for solving the portfolio allocation
problem for DA utility that transforms the DA asset allocation problem in Equation (3)
into a series of standard CRRA problems under a transformed measure that involves the
degree of disappointment aversion A. The simplicity of their algorithm relies crucially on the
assumption of a discretized state space that is ordered by wealth. However, their setup is only
for a single risky asset. We extend their algorithm to a multiple-asset case, by considering all
possible combinations of the six states. This appendix outlines this numerical solution.
Epstein and Zin (1990, 2001) show that the First-Order Conditions (FOC) for Equation
(3) are given by:
Downside Risk
We must find the state i where:

*Wi 2 Rf w*xi xi w*yi yi ; Rf w*xi xi1 w*yi yi1 :
If this condition holds, then the optimal portfolio weights for x and y, w*x and w*y ,
found, so w*x w*xi and w*y w*yi and the optimal utility is given by *W *Wi .
A-6
have been
x
y
mkt
mean
stdev
0.1168
0.1250
0.1200
0.1863
0.2750
0.1375
0.6944
1.3800
1.0000
To obtain the relations between the , downside beta and coskewness and alphas, we
alter xd from 0:16 to 0:30. Figure 1 shows the risk-return relations for asset x.
With an alternative set of parameters, CAPM alphas increase with increasing but also
increasing coskewness. For this case, 6, A 0:7, Rf 1:05, xu 0:50 ,
xm 0:20 , xd 0:30 , yu 0:35, and yd 0:7. The probabilities are given by
p1 0:10, p2 0:20, p3 0:20, p4 0:20, and p5 0:20. If p1 is altered between 0.08 and
0.10 and solved for each case to obtain equilibrium, then we have:
CAPMa
0.0273
0.0317
0.0348
b
co-skew
0.9567
0.8473
0.7326
2.9883
3.1752
3.3873
0.1417
0.1087
0.0598
However, this case is unrealistic because the values of the s are extremely high.
B Data and Portfolio Construction

We use data from the Center for Research in Security Prices (CRSP) to construct portfolios
of stocks sorted by various characteristics of returns. We confine our attention to ordinary
common stocks listed on NYSE, AMEX, and NASDAQ, omitting ADRs, REITs, closedend funds, foreign firms, and other securities which do not have a CRSP share type code of
10 or 11. We use daily and monthly returns from CRSP for the period covering July 3, 1962
to December 31, 2001, including NASDAQ data, which is only available after 1972. We use
the one-month Treasury bill rate from Ibbotson Associates as the risk-free rate and take
CRSPs value-weighted returns of all stocks as the market portfolio. All our returns are
expressed as continuously compounded returns. We also use book value information found
on COMPUSTAT.
1235
The Condition (A-6) relies on ordering the states in increasing wealth. To modify this
algorithm, we take all possible M orderings of the states. Then, we find state i of ordering j
where Condition (A-6) is satisfied. This provides the solution to the DA asset allocation
problem over the two assets x and y.
For our calibrations, we set 6, A 0:8 and set the gross-risk free rate to be
Rf 1:05. For a baseline case, we take ux 0:25 , mx 0:16 , dx 0:25 ,
uy 0:40, and dy 0:15. The six states have probabilities given by p1 0:15, p2 0:20,
p3 0:15, p4 0:25, and p5 0:20. In equilibrium, the value of 0:0021. This gives us
equilibrium weights of w*x 0:5543 and w*y 0:4457, which sum to 1 and which represents
the market. In this specification, the mean excess returns, standard deviations, and betas of
the two assets and the market are given by:
For every 12-month period, we construct portfolios based on measures of risk between
asset is excess return, rit , and the markets excess return, rmt . We exclude stocks with more
than five missing observations from our analysis. We first demean returns within each
period and denote ~rit as the demeaned excess return of asset i and ~rmt as the demeaned
^ and the
market excess return. We obtain estimates of the regular market , denoted ,
individual stock volatility , denoted ^, in the usual manner as:
r
P
~rit~rmt
1X 2
~rit ;
^ P 2 ; and
~rmt
T
B-7
P
P
r
r
r
r
it ~
it ~
mt
mt
fr <^ g ~
fr >^ g ~
^ P mt m 2 ; and ^ P mt m 2 :
rmt
rmt
^m g ~
frmt <
frmt >^m g ~
B-8
We calculate higher-order moments of stock returns using continuously compounded

daily returns over each 12-month period. Coskewness and cokurtosis are estimated as:
P 2
P 3
1
1
~ri ~rm
~ri ~rm
T
T
d q
d q
coskew
and
cokurt
P 2 1 P 2
P 2 1 P 2 3=2 :
1
1
~
~
~
~rmt
r
r
r
it
it
mt
T
T
T
T
B-9
We also collect market capitalizations, book-to-market ratio, and past 12-month returns at
the beginning of each 12-month period for each stock.
To calculate the liquidity betas for individual stocks, at the end of each month, we identify
stocks listed on NYSE, AMEX, and NASDAQ with at least five years of monthly returns.
For each stock, we estimate a liquidity beta, iL , by running the following regression using
the most recent five years of monthly data:
ri;t i0 iL Lt iM rm;t iS SMBt iH HMLt
i;t ;
B-10
where Lt is the innovation in aggregate liquidity, and SMBt and HMLt are size and bookto-market factors of Fama and French (1993). The details of the construction of Lt is in
Pastor and Stambaugh (2003).
Once portfolios are formed, we calculate the returns to holding these portfolios. Over
each 12-month period, we collect the cumulative returns of each stock in excess of the onemonth Treasury bill rate over the period. We also collect the excess stock return over the next
month, as well as stock returns in excess of size and book-to-market matched benchmark
portfolios. These size and book-to-market adjusted returns are calculated in a manner
similar to Daniel et al. (1997). Each month, stocks listed on NYSE, AMEX, and NASDAQ
are sorted into quintiles according to their beginning-of-period market capitalizations based
on NYSE breakpoints. Then within each of these quintiles, stocks are further sorted into
quintiles according to their book-to-market ratios based on NYSE breakpoints. For each
5 5 grouping, we calculate the return on an equal-weighted portfolio consisting of all
stocks that fall into that grouping. For each stock, size and book-to-market adjusted returns
are defined as the return in excess of the portfolio return of the 5 5 grouping to which the
stock belongs. All of these returns are calculated with an adjustment for delisting by taking
1236
where T is the number of trading days in a period. We estimate downside beta and upside
beta by conditioning the observations for which the market realization is below or above the
P
sample mean, ^m T1 rmt . We calculate demeaned excess return of asset i and demeaned
market excess returns conditional on market excess return being below the sample means,
denoted ~r
r
it and ~
mt , respectively. We also calculate demeaned excess return of asset i and
demeaned market excess returns conditional on market excess return being above the
^
^
sample means, denoted ~r
r
it and ~
mt , respectively. We then calculate and as:
Downside Risk
the delisting return at the time the stock is delisted. If a return is missing, we take the next
available return.
References
Ang, A., G. Bekaert, and J. Liu, 2005, Why Stocks May Disappoint, Journal of Financial Economics,
76, 471508.
Ang, A., and J. Chen, 2002, Asymmetric Correlations of Equity Returns, Journal of Financial Economics, 63, 443494.
Ang, A., and J. Chen, 2005, CAPM Over the Long-Run: 19262001, forthcoming in Journal of
Empirical Finance.
Ang, A., R. J. Hodrick, Y. Xing, and X. Zhang, 2006, The Cross-Section of Volatility and Expected
Returns, Journal of Finance, 51, 259299.
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. A. Hsieh, and S. Viswanathan, 1993, A New Approach to International Arbitrage
Pricing, Journal of Finance, 48, 17191747.
Banz, R., 1981. The Relation between Return and Market Value of Common Stocks, Journal of
Financial Economics, 9, 318.
Barberis, N., and M. Huang, 2001, Mental Accounting, Loss Aversion, and Individual Stock Returns,
Journal of Finance, 56, 12471292.
Basu, S., 1983, The Relationship between Earnings Yield, Market Value, and Return for NYSE
Common Stocks: Further Evidence, Journal of Financial Economics, 12, 129156.
Bawa, V. S., and E. B. Lindenberg, 1977, Capital Market Equilibrium in a Mean-Lower Partial Moment
Framework, Journal of Financial Economics, 5, 189200.
Bekaert, G., R. J. Hodrick, and D. A. Marshall, 1997, The Implications of First-Order Risk Aversion
for Asset Market Risk Premiums, Journal of Monetary Economics, 40, 339.
Black, F., M. Jensen, and M. Scholes, 1972, The Capital Asset Pricing Model: Some Empirical Tests, in
M. Jensen (ed.), Studies in the Theory of Capital Markets, Praeger, New York.
Chen, J., Hong, H., and J. Stein, 2001, Forecasting Crashes: Trading Volume, Past Returns and
Conditional Skewness in Stock Prices, Journal of Financial Economics, 61, 345381.
Daniel, K., M. Grinblatt, S. Titman, and R. Wermers, 1997, Measuring Mutual Fund Performance with
Characteristic Based Benchmarks, Journal of Finance, 52, 10351058.
Daniel, K., and S. Titman, 1997, Evidence on the Characteristics of Cross-Sectional Variation in Stock
Dittmar, R., 2002, Nonlinear Pricing Kernels, Kurtosis Preference, and Evidence from the CrossSection of Equity Returns, Journal of Finance, 57, 369403.
Epstein, L. G., and S. E. Zin, 1990, First-Order Risk Aversion and the Equity Premium Puzzle, Journal
of Monetary Economics, 26, 387407.
Epstein, L. G., and S. E. Zin, 2001, The Independence Axiom and Asset Returns, Journal of Empirical
Finance, 8, 537572.
Fama, E. F., and K. R. French, 1992, The Cross-Section of Expected Stock Returns, Journal of
Finance, 47, 427465.
1237
Ang, A., J. Chen, and Y. Xing, 2002, Downside Correlation and Expected Stock Returns, NBER
Working Paper 8643.
Fama, E. F., and K. R. French, 1993, Common Risk Factors in the Returns on Stocks and Bonds,
Journal of Financial Economics, 33, 356.
Fama, E. F., and K. R. French, 1997, Industry Costs of Equity, Journal of Financial Economics, 43,
153193.
Fama, E. F., and K. R. French, 2005, The Value Premium and the CAPM, forhtcoming Journal of
Finance.
Fama, E. F., and J. D. MacBeth, 1973, Risk, Return, and Equilibrium: Empirical Tests, Journal of
Political Economy, 71, 607636.
Ferson, W. E., and C. R. Harvey, 1991, The Variation of Economic Risk Premiums, Journal of Political
Economy, 99, 385415.
Gibbons, M. R., 1982, Multivariate Tests of Financial Models, Journal of Financial Economics, 10, 3
27.
Guidolin, M., and A. Timmermann, 2002, Optimal Portfolio Choice Under Regime Switching, Skew
and Kurtosis Preferences, working paper, University of California, San Diego.
Gul, F., 1991, A Theory of Disappointment Aversion, Econometrica, 59, 667686.
Harlow, W., and R. Rao, 1989, Asset Pricing in a Generalized Mean-Lower Partial Moment Framework: Theory and Evidence, Journal of Financial and Quantitative Analysis, 24, 285311.
Harvey, C. R., J. C. Liechty, M. W. Liechty, and P. Muller, 2003, Portfolio Selection with Higher
Moments, working paper, Duke University.
Harvey, C. R., and A. Siddique, 1999, Autoregressive Conditional Skewness, Journal of Financial and
Quantitative Analysis, 34, 465477.
Harvey, C. R., and A. Siddique, 2000, Conditional Skewness in Asset Pricing Tests, Journal of Finance,
55, 12631295.
Hong, H., and J. Stein, 2003, Differences of Opinion, Short-Sales Constraints and Market Crashes,
Review of Financial Studies, 16, 487525.
Isakov, D., 1999, Is Beta Still Alive? Conclusive Evidence from the Swiss Stock Market, European
Journal of Finance, 5, 202212.
Jagannathan, R., and Z. Wang, 1996, The Conditional CAPM and the Cross-Section of Expected
Jahankhani, A., 1976, E-V and E-S Capital Asset Pricing Models: Some Empirical Tests, Journal of
Financial and Quantitative Analysis, 11, 513528.
Jegadeesh, N., and S. Titman, 1993, Returns to Buying Winners and Selling Losers: Implications for
Stock Market Efficiency, Journal of Finance, 48, 6591.
Jondeau, E., and M. Rockinger, 2006, How Higher Moments Affect the Allocation of Assets, forthcoming in European Financial Management.
Kahneman, D., and A. Tversky, 1979, Prospect Theory: An Analysis of Decision Under Risk,
Econometrica, 47, 263291.
Knez, P. J., and M. J. Ready, 1997, On the Robustness of Size and Book-to-Market in Cross-Sectional
Regressions, Journal of Finance, 52, 13551382.
Kothari, S. P., J. Shanken, and R. G. Sloan, 1995, Another Look at the Cross Section of Expected Stock
1238
Friend, I., and R. Westerfield, 1980, Co-Skewness and Capital Asset Pricing, Journal of Finance, 35,
897913.
Downside Risk
Kraus, A., and R. Litzenberger, 1976, Skewness Preference and the Valuation of Risk Assets, Journal of
Finance, 31, 10851100.
Kraus, A., and R. Litzenberger, 1983, On the Distributional Conditions for a Consumption-Oriented
Three Moment CAPM, Journal of Finance, 38, 13811391.
Kyle, A. W., and W. Xiong, 2001, Contagion as a Wealth Effect of Financial Intermediaries, Journal of
Finance, 56, 14011440.
Lettau, M., and S. Ludvigson, 2001, Resurrecting the (C)CAPM: A Cross-Sectional Test When Risk
Premia Are Time-Varying, Journal of Political Economy, 109, 12381287
Lewellen, J., and S. Nagel, 2005, The Conditional CAPM Does Not Explain Asset-Pricing Anomalies,
forthcoming in Journal of Financial Economics.
Newey, W. K., and K. D. West, 1987a, A Simple Positive Semi-Definite, Heteroskedasticity and
Autocorrelation Consistent Covariance Matrix, Econometrica, 55, 7038.
Newey, W. K., and K. D. West, 1987b, Hypothesis Testing with Efficient Method of Moments,
International Economic Review, 28, 777787.
Pastor, L., and R. F. Stambaugh, 2003, Liquidity Risk and Expected Stock Returns, Journal of
Political Economy, 111, 642685.
Pettengill, G. N., S. Sundaram, and I. Mathur, 1995, The Conditional Relation between Beta and
Returns, Journal of Financial and Quantitative Analysis, 30, 101116.
Price, K., B. Price, and T. J. Nantell, 1982, Variance and Lower Partial Moment Measures of Systematic
Risk: Some Analytical and Empirical Results, Journal of Finance, 37, 843855.
Rouwenhorst, K. G., 1998, International Momentum Strategies, Journal of Finance 53, 267284.
Routledge, B. R., and S. E. Zin, 2003, Generalized Disappointment Aversion and Asset Prices, NBER
Working Paper 10107.
Roy, A. D., 1952, Safety First and the Holding of Assets, Econometrica, 20, 431449.
Rubinstein, M., 1973, The Fundamental Theory of Parameter-Preference Security Valuation, Journal
of Financial and Quantitative Analysis, 8, 6169.
Scholes, M., and J. T. Williams, 1977, Estimating Betas from Nonsynchronous Data, Journal of
Financial Economics, 5, 30927.
Scott, R. C., and P. A. Horvath, 1980, On the Direction of Preference for Moments of Higher Order
than the Variance, Journal of Finance, 35, 915919.
Shanken, J., 1992, On the Estimation of Beta-Pricing Models, Review of Financial Studies, 5, 133.
Shumway, T., 1997, Explaining Returns with Loss Aversion, working paper, University of Michigan.
1239
Markowitz, H., 1959, Portfolio Selection. Yale University Press, New Haven, CT.

Ang 2006

Uploaded by

Copyright:

Available Formats

Ang 2006

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Ang 2006

Uploaded by

Copyright:

Available Formats

Downside Risk

If an asset tends to move downward in a declining market more than it

Downloaded from http://rfs.oxfordjournals.org/ at Oakland University on September 30, 2014

The Review of Financial Studies / v 19 n 4 2006

Downloaded from http://rfs.oxfordjournals.org/ at Oakland University on September 30, 2014

average returns. In this article, we show that the cross-section of stock

Downloaded from http://rfs.oxfordjournals.org/ at Oakland University on September 30, 2014

The Review of Financial Studies / v 19 n 4 2006

1. A Simple Model of Downside Risk

Downloaded from http://rfs.oxfordjournals.org/ at Oakland University on September 30, 2014

average one-year autocorrelation of one-year betas for very volatile stocks is

where UW is the felicity function over end-of-period wealth W , which

Downloaded from http://rfs.oxfordjournals.org/ at Oakland University on September 30, 2014

Rather than considering models with one-sided constraints or

The Review of Financial Studies / v 19 n 4 2006

The optimal portfolio weight for a DA investor is given by the solution

where the certainty equivalent is defined in equation (1), wx (wy ) is the

and Rf is the gross risk-free rate. An equilibrium is characterized by a set

Downloaded from http://rfs.oxfordjournals.org/ at Oakland University on September 30, 2014

covri ; rm jrm < m

where i is the average excess return of asset i. Harvey and Siddique

where W is the total wealth of the representative agent, and U can be

Downloaded from http://rfs.oxfordjournals.org/ at Oakland University on September 30, 2014

where ri (rm ) is security is (the markets) excess return, and m is the

The Review of Financial Studies / v 19 n 4 2006

Downloaded from http://rfs.oxfordjournals.org/ at Oakland University on September 30, 2014

covri ; rm jrm > m

Downloaded from http://rfs.oxfordjournals.org/ at Oakland University on September 30, 2014

The Review of Financial Studies / v 19 n 4 2006

2. Downside Risk and Realized Returns

Downloaded from http://rfs.oxfordjournals.org/ at Oakland University on September 30, 2014

downside beta, we compute a relative upside beta (denoted by ).

Downloaded from http://rfs.oxfordjournals.org/ at Oakland University on September 30, 2014

The Review of Financial Studies / v 19 n 4 2006

Downloaded from http://rfs.oxfordjournals.org/ at Oakland University on September 30, 2014

Downloaded from http://rfs.oxfordjournals.org/ at Oakland University on September 30, 2014

example, the coskewness effect of Harvey and Siddique (2000) is strongest

The Review of Financial Studies / v 19 n 4 2006

Panel E: Stocks sorted by realized relative b

Panel F: Stocks sorted by realized b b

Panel D: Stocks sorted by realized b

Panel C: Stocks sorted by realized relative b

high , , and regular . Note that, for these portfolios sorted by

Downloaded from http://rfs.oxfordjournals.org/ at Oakland University on September 30, 2014

Panel B: Stocks sorted by realized b

Downloaded from http://rfs.oxfordjournals.org/ at Oakland University on September 30, 2014

The Review of Financial Studies / v 19 n 4 2006

Downloaded from http://rfs.oxfordjournals.org/ at Oakland University on September 30, 2014

the effects of . Instead, the increasing pattern of returns in panel D

discounts. However, these relations do not control for other known

Downloaded from http://rfs.oxfordjournals.org/ at Oakland University on September 30, 2014

2.2 Fama-MacBeth Regressions

The Review of Financial Studies / v 19 n 4 2006

In order to avoid putting too much weight on extreme observations,

Downloaded from http://rfs.oxfordjournals.org/ at Oakland University on September 30, 2014

Downloaded from http://rfs.oxfordjournals.org/ at Oakland University on September 30, 2014

We begin with regression I in Table 2 to show the familiar, standard set

The Review of Financial Studies / v 19 n 4 2006