Crash Sensitivity and the Cross-Section of Expected

Stock Returns
Stefan Ruenzi and Florian Weigert

This Version: December 2012
Abstract
We examine whether investors receive a compensation for holding crash-sensitive
stocks. We capture the crash sensitivity of stocks by their lower tail dependence with
the market based on copulas. Stocks with strong contemporaneous crash sensitivity
clearly outperform stocks with weak crash sensitivity and a trading strategy based on
past crash sensitivity delivers positive abnormal returns of about 4% p.a. This eect
cannot be explained by traditional risk factors and is dierent from the impact of beta,
downside beta, and coskewness. Our ndings are consistent with results from the em-
pirical option pricing literature and support the notion that stock market investors are
crash-averse.
Keywords: Asset Pricing, Downside Risk, Tail Risk, Crash Aversion, Asymmetric depen-
dence, Copulas.
JEL Classication Numbers: C12, G01, G11, G12, G17.

Stefan Ruenzi is from the Chair of International Finance at the University of Mannheim, Address:
L9, 1-2, 68131 Mannheim, Germany, Telephone: ++49-621-181-1640, e-mail: ruenzi@bwl.uni-mannheim.de.
Florian Weigert is from the Chair of International Finance and CDSB at the University of Mannheim,
Address: L9, 1-2, 68131 Mannheim, Germany, Telephone: ++49-621-181-1625, e-mail: weigert@bwl.uni-
mannheim.de. The authors thank Andres Almazan, Tobias Berg, Joseph Chen, Fousseni Chabi-Yo, Knut
Griese, Allaudeen Hameed, Maria Kasch, Alexandra Niessen-Ruenzi, Alexander Puetz, Sheridan Titman,
and seminar participants at the 2011 European Economics Association Meeting, 2011 European Finance
Association Meeting, 2011 German Finance Association Meeting, 2011 Inquire Europe Autumn Seminar,
2012 EFM Asset Management Symposium, 2012 Swiss Financial Market Research (SGF) Conference, 2012
FMA European Conference, University of Mannheim and University of Texas at Austin for their helpful
comments. This paper was previously circulated under the title Extreme Dependence Structures and the
Cross-Section of Expected Stock Returns. All errors are our own.
Crash Sensitivity and the Cross-Section of Expected
Stock Returns
Abstract
We examine whether investors receive a compensation for holding crash-sensitive
stocks. We capture the crash sensitivity of stocks by their lower tail dependence with
the market based on copulas. Stocks with strong contemporaneous crash sensitivity
clearly outperform stocks with weak crash sensitivity and a trading strategy based on
past crash sensitivity delivers positive abnormal returns of about 4% p.a. This eect
cannot be explained by traditional risk factors and is dierent from the impact of beta,
downside beta, and coskewness. Our ndings are consistent with results from the em-
pirical option pricing literature and support the notion that stock market investors are
crash-averse.
Keywords: Asset Pricing, Asymmetric dependence, Copulas, Coskewness, Downside Risk,
Tail Risk, Crash Aversion
JEL Classication Numbers: C12, G01, G11, G12, G17.
1 Introduction
Rare-disaster risk has caught a lot of attention in the nance literature in recent years (e.g.,
Barro (2006) and (2009)). Bollerslev and Todorov (2011) nd that most of the aggregate
equity risk premium is a compensation for the risk of extreme events and Gabaix (2012)
shows that time-varying rare-disaster risk can explain the equity premium puzzle (as well as
several other puzzles in macro-nance). Similarly, there is now a small number of recent
papers that examine the time-series relationship between tail risk and aggregate stock market
returns (e.g., Bali, Demirtas, and Levy (2009), Kelly (2012), Bollerslev and Todorov (2011)).
They nd that proxies for tail risk can predict aggregate market returns.
Surprisingly, the potential impact of rare disasters and crash aversion has not caught
much attention in the empirical literature on the pricing of the cross-section of individual
stocks.
1
If investors are crash-averse, they derive disproportionately large disutility from
large losses. Then, stocks that do particularly badly when the market performs badly, i.e.,
crash-sensitive stocks, are unattractive assets to hold. In this case, crash-sensitive stocks
should bear a premium. In this paper, we document that crash-sensitive stocks indeed
deliver higher returns than crash-insensitive stocks.
While crash aversion is not a feature of standard expected utility theory, there is sub-
stantial evidence that individuals are particularly averse to large losses. The accumulated
evidence from experiments designed to verify the cumulative prospect theory by Kahne-
man and Tversky (1979) shows that individuals distort the probabilities of low-probability
outcomes like market crashes heavily upwards (e.g., Tversky and Fox (1995), Bleichrodt
and Pinto (2000) and Abdellaoui (2002)).
2
More recently, Polkovnichenko and Zhao (2012)
1
Notable exceptions are the contemporaneous papers by Kelly (2012) and Cholette and Lu (2011) that
we will discuss in more detail below.
2
He and Zhou (2012) show that this can have important implications for investors optimal portfolio
1
conrm this pattern using market data from traded nancial options to derive empirical
probability weighting functions. These results are also consistent with earlier ndings from
the empirical option pricing literature that deep out-of-the-money puts, i.e., instruments
that oer protection against extreme market downturns, have a very high implied volatility
and are relatively expensive,
3
which is typically interpreted as investors being crash-averse
or showing signs of crash-o-phobia (Rubinstein (1994), Bates (2008)).
To measure crash sensitivity at the individual asset level, we need a dependence concept
that allows us to focus on joint extreme events. Standard asset pricing models since Sharpe
(1964) and Lintner (1965) argue that the joint distribution of individual stock returns and the
market portfolio return determines the cross-section of expected stock returns. According to
the empirical interpretation of the traditional CAPM, a stocks expected return only depends
on its beta, i.e., its scaled linear correlation with the market, without any focus on tail events.
However, the correlation only describes the full dependence structure in the case of bivariate
normally distributed variables. It can not characterize the full dependence structure of non-
normally distributed random variables such as realized stock returns (Embrechts, McNeil,
and Straumann (2002)). Particularly, it can not capture clustering in the lower tail of
the bivariate return distribution between individual securities and the market, which is
of foremost importance if investors are crash-averse. Thus, we develop a novel proxy for
stock-individual crash sensitivity using copula methods based on extreme value theory.
4
choice and leads to demand for portfolio insurance.
3
See, e.g., Rubinstein (1994), Jackwerth and Rubinstein (1996), At-Sahalia and Lo (2000), Bates (2000),
Jackwerth (2000), Rosenberg and Engle (2002), Broadie, Chernov, and Johannes (2009). Garleanu, Pedersen,
and Poteshman (2009) show that this eect is driven by high demand for out-of-the-money puts. Bollen
and Whaley (2004) also show that buying pressure for index puts increases their price. Grossman and Zhou
(1996) suggest an equilibrium model where a group of investors buys portfolio insurance, which can explain
the high prices of out-of-the-money put options, too.
4
An intuitive alternative approach would be to look at the correlation between individual stock and
market returns conditional on the market return being below a certain threshold. This approach is similar
to the downside beta used in Ang, Chen, and Xing (2006). However, to really capture crash sensitivity, one
2
We capture stock individual crash sensitivity based on the extreme dependence between
individual stock returns and market returns in the lower left tail of their joint distribution
(also called lower tail dependence) and then investigate its inuence on the cross-section
of individual stock returns.
5
All else being equal, securities that exhibit strong lower tail
dependence (LTD) are unattractive assets for crash-averse investors to hold, because they
realize their lowest payos when investors wealth is already very low (given that one accepts
the market as a proxy for investor wealth). Thus, investing in stocks with only weak or no
LTD serves as insurance against extreme negative portfolio returns (similar to buying out-
of-the-money index puts) as these stocks are unlikely to realize their worst returns when
the market realizes its worst return. Consequently, investors who are sensitive to extreme
downside losses will require a return premium for holding stocks with strong LTD.
Based on daily return data for all US stocks from 1963 to 2009, we calculate LTD coef-
cients for each stock and year.
6
Aggregate LTD peaks around the 1987 crash as well as
during the recent nancial crisis starting in 2008. We then relate individual LTD to con-
temporaneous returns on an annual basis. In doing so, we follow Lewellen and Nagel (2006)
who suggest using short, non-overlapping periods and daily data in asset pricing exercises
when risk exposures might be time varying. Our empirical results using portfolio sorts and
multivariate regression analysis on the individual rm level show a strong positive impact
of LTD on contemporaneous returns. Top quintile LTD stocks outperform bottom quintile
LTD stocks by more than 10% p.a. Our results from Fama-MacBeth (1973) regressions show
has to focus on the really bad market outcomes (not just below average market outcomes as in their paper)
and it is not possible to estimate such conditional betas reliably (see Section 3.4).
5
A positive inuence of LTD on returns is expected (but not empirically shown) in Poon, Rockinger, and
Tawn (2004): If tail events are systematic as well, one might expect the extremal dependence between the
asset returns and the market factor returns to also command a risk premium. (p. 586).
6
To do so, we rst determine the convex combination of basic parametric copulas that best explains the
empirical bivariate distribution of an individual stocks return and the market return. Then we compute the
respective tail dependence coecients.
3
that an increase of one standard deviation in LTD is associated with an average return pre-
mium of about 5.28% p.a.
7
In contrast, weak LTD stocks have signicantly higher returns
than strong LTD stocks during extreme market downturns, i.e., weak LTD stocks indeed
oer some protection against extreme market downturns. Finally, we nd that an investible
trading strategy consisting of buying a portfolio of stocks with the strongest past lower tail
dependence and selling a portfolio of stocks with the weakest past lower tail dependence over
the previous twelve months delivers a signicantly positive return of 4.4% p.a. The Carhart
(1997) four-factor alpha of this strategy amounts to 4.6% p.a.
The impact of LTD has to be distinguished from the impact of downside beta documented
in Ang, Chen, and Xing (2006) as well as from the impact of other higher co-moments.
Downside beta focuses on individual securities exposure to market returns conditional on
below-average market returns. Thus, it is dierent from LTD, as it places no particular em-
phasis on tail events.
8
Consequently, downside beta captures general downside risk aversion
rather than crash aversion. Lower tail dependence is a fundamentally dierent concept and
captures the dependence in the extreme left tail of return distributions, i.e., it focuses on how
individual securities behave during the worst market return realizations within in a given pe-
riod. We nd a strong impact of LTD even after controlling for the impact of the Ang, Chen,
and Xing (2006) downside beta. Adding lower tail dependence as an explanatory variable
drives out most of the impact of downside beta in our multivariate analysis. Thus, crash
aversion has an additional impact on the cross-section of returns that can be distinguished
7
Additionally, we conduct several robustness tests including various risk- and industry-adjustments of
returns, changes in the weighting scheme and analysis of the temporal stability of our results. We also show
that our results are very similar if we do not calibrate the optimal copula structure for each stock and year
but instead just estimate tail dependence coecients based on some xed ad-hoc copula combinations. This
result might be useful for future research, because selecting the right copula combination is computationally
costly.
8
Downside betas conditional on very low market returns (instead of just below mean market returns)
are intuitively more related to LTD. However, they can not be estimated reliably, as we show in Section 3.4.
4
from the impact of downside risk aversion. We can also show that the risk premium associ-
ated with LTD cannot be captured by coskewness (Harvey and Siddique (2000)), cokurtosis
(Fang and Lai (1997)), idiosyncratic volatility (Ang, Hodrick, Xing, and Zhang (2006)), or a
stocks lottery characteristics (Bali, Cakici, and Whitelaw (2011a)), and holds after control-
ling for a host of systematic risk factors suggested in the literature. Furthermore, we nd
that most of the abnormal returns of the Jegadeesh and Titman (1993) momentum strategy
are due to its exposure to a contemporaneous LTD factor which suggests that momentum
prots are highly crash-sensitive.
Taken together, the main contribution of our study is twofold: (1) We document that the
lower tail dependence between individual stock and market returns has an impact on the
cross-section of stock returns. (2) We apply copula methods to capture LTD and use it in
an asset pricing context for the rst time.
Besides its pertinence to the literature on rare-disaster risk cited above, our study is
related to the literature on downside risk and loss aversion. Downside risk aversion is already
discussed in Roy (1952), who argues that investors display safety-rst preferences, and in
Markowitz (1959), who suggests using the semi-variance as a measure of risk.
9
Kahneman
and Tversky (1979) argue that individuals evaluate outcomes relative to reference points and
show that individuals are loss-averse. Although aversion to losses and downside risk aversion
is discussed extensively in the literature, only a few papers investigate the eect of loss or
disappointment aversion on expected asset returns.
10
However, these papers (as well as the
9
Many subsequent contributions analyze the impact of higher co-moments on expected returns. Exten-
sions of the basic CAPM that allow for preferences for skewness and lower partial moments of security and
market returns are developed by Kraus and Litzenberger (1976) and Bawa and Lindenberg (1977). Kraus
and Litzenberger (1976), Friend and Westereld (1980), and Harvey and Siddique (2000) document that
investors dislike negative coskewness of stock returns with the market return. Fang and Lai (1997) and
Dittmar (2002) show that stocks with high cokurtosis earn high average returns.
10
Barberis and Huang (2001), in one of their model variants, study equilibrium rm-level stock returns
when investors are loss averse over the uctuations of the individual stocks in their portfolio. They predict
5
study by Ang, Chen, and Xing (2006) discussed above) are concerned with general downside
risk aversion rather than crash aversion.
Crash aversion has still caught relatively little attention in the cross-sectional asset pricing
literature. One exception is Berkman, Jacobsen, and Lee (2011) who examine whether in-
dustries that are sensitive to a real crises index deliver higher returns and nd some evidence
for this to be the case. The only other papers we are aware of that investigate whether crash
sensitivity (or tail risk exposure) has an impact on the cross-section of individual stock re-
turns are two contemporaneous papers by Kelly (2012) and Cholette and Lu (2011).
11
These
papers predict aggregate tail risk based on individual extreme events by applying the tail
risk estimator of Hill (1975) to the cross-section of all daily stock returns in a given month.
Consistent with our results, Kelly (2012) also documents that a long-short portfolio that
is based on individual stocks exposure to an aggregate tail risk factor that hedges against
tail events delivers signicantly negative returns. Similar results are obtained by Cholette
and Lu (2011). Our paper diers from these two papers conceptually: We capture crash
sensitivity using lower tail dependence between a stock and the market. Thus, our proxy for
crash sensitivity of an individual stock is directly based on the joint distribution of its return
and the market return. This approach oers the advantage that we only need a time series
of an individual stocks return and the market return (while Kellys implementation of the
Hill (1975) estimator requires the whole cross-section of all individual daily stock returns at
a large value premium in the cross-section. Other models with loss-averse investors include Benartzi and
Thaler (1995) and Barberis, Huang, and Santos (2001), while Ang, Bekaert, and Liu (2005) model investors
with disappointment aversion.
11
A related approach is followed in Bali, Cakici, and Whitelaw (2011b). They extend the Bawa and
Lindenberg (1977) mean-lower partial moment CAPM and look at co-lower partial moments of stocks with
the market conditional on the individual stock being below a specied threshold. Conditioning on the
individual stock (rather than the market) being below a specic threshold is motivated by the argument
that many investors are not diversied. They nd that stocks with such high co-lower partial moments
outperform stocks with low co-lower partial moments.
6
each point in time).
Our paper is also related to the literature on the application of extreme value theory in
nance and contributes to the nance literature methodologically. Despite its long history
in statistics, multivariate extreme value theory has been applied to the analysis of nancial
markets only recently.
12
It is used to describe dependence patterns across dierent markets
and assets (Longin and Solnik (2001)). However, to the best of our knowledge, ours is the
rst paper to investigate extreme dependence structures in the bivariate distribution of the
returns of individual stocks and the market based on copulas. Our application details how
to t exible combinations of basic parametric copulas to this bivariate distribution and
how to derive the corresponding tail dependence coecients. The copula approach has the
advantage that extreme dependence is not estimated based on a small number of observations
in the tail exclusively, but that information from the whole joint distribution can be used.
The rest of this paper is organized as follows. Section 2 gives a short overview of copula
theory and presents the estimation procedure for tail dependence coecients. Section 3
demonstrates that stocks with strong lower tail dependence have, at the same time, high
average returns. In Section 4, we examine the persistence of tail dependence and evaluate a
trading strategy based on lower tail dependence. Section 5 concludes and briey discusses
the implications of our results.
12
Longin and Solnik (2001) and Rodriguez (2007) use extreme value theory to model the bivariate return
distributions between dierent international equity markets. Focusing on risk management applications, Ane
and Kharoubi (2003) propose modeling the dependence structure of international stock index returns via
parametric copulas and derive their tail dependence behavior. Poon, Rockinger, and Tawn (2004) present a
general framework for identifying joint-tail distributions based on multivariate extreme value theory. They
argue that the use of traditional dependence measures could lead to inaccurate portfolio risk assessment.
Patton (2004) uses copula theory to model time-varying dependence structures of stock returns and assesses
the importance of skewness and asymmetric depependence for asset allocation. Patton (2009) applies copula
functions to assess dierent denitions of market neutrality for hedge funds.
7
2 Copula Methodology and Data
As copula concepts are not yet regularly used in standard asset pricing applications, we
will rst give a short intuitive introduction into the concept (Section 2.1) and explain how
we compute measures of tail dependence based on copulas (Section 2.2). We then describe
our data and the development of aggregate tail dependence over time (Section 2.3).
13
2.1 Copulas
Most of the standard empirical asset pricing literature focuses on risk factors based on
linear correlation coecients. However, this measure of stochastic dependence is not typically
able to completely characterize the dependence structure of non-normally distributed random
variables (Embrechts, McNeil, and Straumann (2002)). It is now widely recognized that
many nancial time series, including stock returns, are non-normally distributed.
14
For
example, they are often characterized by leptokurtosis. This is problematic because when
we are dealing with a fat-tailed bivariate distribution F(x
1
, x
2
) of two random variables
X
1
and X
2
, the linear correlation fails to capture the dependence structure in the extreme
lower left and upper right tail. As an example, consider the following illustrations of 2,000
simulated bivariate realizations based on dierent dependence structures between (X
1
, X
2
)
shown in Figure 1.
15
In all models, X
1
and X
2
have standard normal marginal distributions and a linear cor-
relation of 0.8, but other aspects of the dependence structure are clearly dierent. For
comparison, in Panel A we rst show an example where we did not allow for clustering in
13
A more precise technical, but still accessible, treatment of copula concepts is contained in Nelsen (2006).
14
For some early evidence, see Mandelbrot (1963) and Fama (1965).
15
The realizations plotted in Figure 1 are based on simulations of dierent popular copula functions. The
statistical models used to simulate these realizations are the Gauss-copula (Panel A), the Gumbel-copula
(Panel B), the Clayton-copula (Panel C), and the Student t-copula (Panel D), all as dened in Table A.1.
8
either tail of the distribution. Panels B to D show examples of increased dependence in the
upper right tail, in the lower left tail, and symmetric increased dependence in both tails.
Still, all of these bivariate distributions are characterized by a linear correlation coecient of
0.8. These examples show that it is not always possible to describe the dependence structure
by the linear correlation alone.
Copulas oer an elegant way to describe the complete dependence structure between ran-
dom variables. Every distribution function of two random variables X
1
and X
2
(e.g., an
individual asset return and the market return) implicitly contains both, a description of
the marginal distribution functions F
1
(x
1
) and F
2
(x
2
) and their dependence structure. The
copula approach allows us to isolate the description of the dependence structure from the
univariate marginal distributions of the bivariate distribution. Sklar (1959) shows that all
bivariate distribution functions F(x
1
, x
2
) can be completely described based on the univari-
ate marginal distributions and a copula C: [0, 1]
2
[0, 1]. Sklars Theorem explicitly states
that all bivariate distributions can be decomposed into copulas and that the marginal distri-
butions and bivariate distributions can be constructed by combining the univariate marginal
distributions using copulas (McNeil, Frey, and Embrechts (2005)). Formally, Sklars Theo-
rem states:
Theorem 1 (Sklar 1959). Let F be a bivariate distribution function with margins F
1
and
F
2
. Then there exists a copula C : [0, 1]
2
[0, 1] such that, for all x
1
, x
2
in R = [, ],
F(x
1
, x
2
) = C(F
1
(x
1
), F
2
(x
2
)). (1)
If the margins are continuous, then C is unique. Conversely, if C is a copula and F
1
and
F
2
are univariate distribution functions, then the function F dened in (1) is a bivariate
9
distribution function with margins F
1
and F
2
.
There are many dierent parametric copulas C that can model dierent tail dependence
structures. In this study, we use combinations of simple basic parametric copulas that either
show no tail dependence (the Gauss-, the Frank-, the FGM-, and the Plackett-copula), lower
tail dependence (the Clayton-, the Rotated Gumbel-, the Rotated Joe-, and the Rotated
Galambos-copula), or upper tail dependence (the Gumbel-, the Joe-, the Galambos-, and
the Rotated Clayton-copula). By using a convex combination of one copula from each class,
we allow for maximum exibility in modeling dependence structures (see Section 2.2).
We will later use copulas to estimate coecients of upper and lower tail dependence (Sibuya
(1960)), i.e., measures of the strength of dependence in the tails of a bivariate distribution.
2.2 Computation of Tail Dependence Coecients
Intuitively, the lower (upper) tail dependence coecient between two variables reects the
probability that a realization of one random variable is in the extreme lower (upper) tail
of its distribution conditional on the realization of the other random variable also being in
the extreme lower (upper) tail of its distribution. Formally, lower tail dependence (LTD) is
dened as
LTD := LTD(X
1
, X
2
) := lim
u0+
P(X
1
F
1
1
(u)|X
2
F
1
2
(u)), (2)
where u (0, 1) is the argument of the distribution function, i.e., lim
u0+
indicates the limit
if we approach the left tail of the distribution from above. Analogously, we can dene the
upper tail dependence coecient (UTD) as:
10
UTD := UTD(X
1
, X
2
) := lim
u1
P(X
1
> F
1
1
(u)|X
2
> F
1
2
(u)). (3)
If LTD (UTD) is equal to zero, the two variables are asymptotically independent in the
lower (upper) tail. Simple expressions for LTD and UTD in terms of the copula C of the
bivariate distribution can be derived based on
LTD = lim
u0+
C(u, u)
u
(4)
and
UTD = lim
u1
1 2u + C(u, u)
1 u
, (5)
if F
1
and F
2
are continuous (McNeil, Frey, and Embrechts (2005)). The coecients of tail
dependence have closed form solutions for the basic parametric copulas used in this study
(see Table A.1 in Appendix A).
However, basic copulas do not allow us to model upper and lower tail dependence simul-
taneously. Thus, we obtain exible copula structures by allowing for convex combinations of
these basic copulas. To allow for the maximum possible exibility, we consider all 64 possible
convex combinations of the afore mentioned basic copulas from Table A.1 that each consist of
one copula that allows for asymptotic dependence in the lower tail, C
LTD
, one copula that is
asymptotically independent, C
NTD
, and one copula that allows for asymptotic dependence
in the upper tail, C
UTD
:
C(u
1
, u
2
, ) = w
1
C
LTD
(u
1
, u
2
;
1
)
+w
2
C
NTD
(u
1
, u
2
;
2
) + (1 w
1
w
2
) C
UTD
(u
1
, u
2
;
3
), (6)
11
where denotes the set of the basic copula parameters
i
, i = 1, 2, 3 and the weights w
1
and
w
2
. We pick the copula C

(, ;

) that best ts the data (as described in the appendix) for

each stock and year and compute the tail dependence coecients LTD and UTD implied by
the estimated parameters

of this copula. As this procedure is repeated for each stock

and year, we end up with a panel of tail dependence coecients LTD

i,t
and UTD

i,t
at the
year-rm level. For a more detailed description of the estimation and selection method, we
refer the reader to the Appendix.
2.3 Data and the Evolution of Aggregate Tail Dependence
Our sample consists of all common stocks (CRSP share codes 10 and 11) from CRSP
trading on the NYSE and AMEX between January 1, 1963 through December 31, 2009.
Copulas and tail dependence coecients are estimated for each rm and each year separately.
To estimate the yearly tail dependence coecients, we use daily data for all days on which
the stocks price at the end of the previous trading day was at least $2. We retain all stocks
that have at least 100 valid daily return observations per year. Overall, there are 96, 676
rm-year observations. The number of rms in each year over our sample period ranges from
1, 489 to 2, 440. Summary statistics are provided in Table 1.
The rst four columns show the mean as well as the 25%, the 50%, and the 75% quantile
for key variables. The mean (median) yearly excess return over the riskfree rate of all stocks
in our sample is nearly 10.98% (3.73%) and the mean (median) LTD coecient is 0.13 (0.11).
We also observe considerable variation in LTD, with an interquartile range of nearly 0.18.
The mean (median) of UTD is 0.08 (0.05) and is signicantly lower than the mean (median)
of LTD. The general tendency for stronger asymptotic dependence in the left tail than in
the right tail of the distributions is consistent with the well-documented nding that return
12
correlations generally increase in down markets.
16
The rest of the table provides information
on the summary statistics regarding other rm characteristics and return patterns that we
later use in our empirical analysis. All variable denitions are contained in Appendix B.
The last three columns of Table 1 show the average characteristics of stocks with an above
and below, respectively, value of LTD in the respective year as well as the dierence between
the two. Contemporaneous excess returns for high LTD stocks are 13.28% p.a., while they
are only 8.68% p.a. for low LTD stocks. The dierence amounts to 4.6% and is statistically
signicant on the 1%-level. At the same time, high LTD stocks also have signicantly higher
regular betas () and particularly downside betas (

), tend to be somewhat larger and have

lower book-to-market ratios. Cross-correlations between the independent variables used in
this study are shown in Table 2 and conrm these patterns.
The correlation between LTD and UTD is relatively moderate at 0.12. The low correlation
shows that rms with strong tail dependence in one tail of the distribution do not necessarily
exhibit strong tail dependence in the other tail. This nding also justies our exible model-
ing approach for tail dependence which allows for asymmetric tail dependence in the upper
and lower tail. LTD is correlated with downside beta with a correlation coecient of 0.49
and with regular beta with a correlation coecient of 0.38. There is also a positive (negative)
correlation between LTD and size as well as cokurtosis (illiquidity as well as coskewness).
We carefully take into account the impact of these correlations in our later analysis.
To get some idea about the temporal variation of tail dependence, we investigate the time
series of aggregate LTD. We dene aggregate LTD of the market, LTD
m,t
, as the yearly
cross-sectional, equal-weighted, average of LTD
i,t
over all stocks i in our sample. In Figure
16
In unreported tests, we also look at 5-year subperiods and nd that UTD is signicantly weaker than
LTD in each period. Increased extreme dependence among international markets during bear markets is also
documented in Longin and Solnik (2001) and Poon, Rockinger, and Tawn (2004).
13
2, we plot the time series of LTD
m,t
.
There is no particular time trend in LTD
m,t
.
17
However, the graph does exhibit occasional
spikes in LTD
m,t
that roughly correspond to worldwide nancial crises. The highest value in
LTD
m,t
corresponds to 1987, the year of Black Monday, with the largest one-day percentage
decline in US stock market history. Another spike in market LTD occurs during the years
2007 through 2009, the years of the recent worldwide nancial crisis. This pattern suggests
that LTD
m,t
similar to return correlations increases in times of nancial crises.
18
Figure
2 also plots aggregate UTD, UTD
m,t
, dened as the yearly cross-sectional, equal-weighted,
average of UTD
i,t
. The time series of LTD
m,t
and UTD
m,t
are positively correlated with a
linear correlation coecient of 0.29. We nd that in most of the years of our sample LTD
m,t
clearly exceeds UTD
m,t
.
3 Crash Sensitivity and Realized Returns
We start our empirical investigation by looking at the contemporaneous relationship be-
tween extreme dependence and returns in univariate sorts and double-sorts (Section 3.1). In
Section 3.2, we conduct Fama-MacBeth (1973) regressions to control for various other po-
tential determinants of returns. In Section 3.3 we examine the impact of a stocks riskiness
on our results, and in Section 3.4 we examine the relationship between variants of the Ang,
Chen, and Xing (2006) downside beta and LTD in more detail. In Section 3.5 we analyze
17
Performing an augmented Dickey-Fuller test rejects the null hypothesis that LTD
m,t
contains a unit
root with a p-value smaller that 2%.
18
The time-series correlation between LTD and the market return is -0.21 and the time-series
correlation between LTD and market volatility is 0.59. The time-series correlation between the
Yale Crash Condence Index (based on data from October 1989 till December 2009 obtained from
http://icf.som.yale.edu/stock-market-confidence-indices-united-states-crash-index-data)
based on institutional and individual survey participants is roughly -0.5, suggesting that if condence that
no crash will happen is low, aggregate LTD is high.
14
how LTD and momentum are related. Furthermore, we investigate whether we get similar
results if we use a simplication of the estimation procedure for tail dependence coecients
in Section 3.6 and nally, results from a battery of robustness tests are presented in Section
3.7.
In the empirical analysis in this section we relate realized tail dependence coecients to
portfolio and individual security returns over the same period (while we relate returns to
lagged LTD in Section 4). Looking at the contemporaneous relationship closely follows papers
like Ang, Chen, and Xing (2006) and Lewellen and Nagel (2006), and implicitly assumes that
realized returns are on average a good proxy for expected returns. This procedure is mainly
motivated by the fact that several studies document that risk exposures (like regular beta)
are time-varying (see, e.g., Fama and French (1992), and Ang and Chen (2007)). It is likely
that a stocks LTD is also time-varying and past LTD might not necessarily be a good
predictor of future LTD.
19
Thus, as advocated by Lewellen and Nagel (2006), we use daily
return data for non-overlapping intervals of one year and over each annual period t calculate
a stock is LTD-coecient.
20
Using an annual horizon trades o two concerns: First, we need
a suciently large number of observations to get reliable estimates for our tail dependence
coecients. Second, by investigating contemporaneous relationships over relatively short
horizons we are able to account for time-varying extreme dependence.
19
In unreported tests we nd evidence for limited predictive power of past LTD on current LTD (see also
Section 4.1).
20
The estimation procedure of the LTD- and UTD-coecients for each stock is performed according to
Section 2.2 and Appendix A.
15
3.1 Portfolio Sorts
3.1.1 Average Returns and Alphas of LTD-sorted Portfolios
To examine whether stocks with strong lower tail dependence earn a premium, we rst
look at simple univariate portfolio sorts. For each year t we sort stocks into ve quintile
portfolios based on their realized LTD in the same year. Results are reported in Table 3.
The rst column shows average LTD coecients of the stocks in the quintile portfolios.
There is considerable cross-sectional variation in LTD: average LTD ranges from 0.01 in the
weakest LTD quintile up to 0.29 in the strongest LTD quintile.
In the second column, we report the annual equal-weighted average excess return over
the risk free rate of these portfolios as well as dierences in average excess returns between
quintile portfolio 5 (strong LTD) and quintile portfolio 1 (weak LTD).
21
We nd that stocks with strong LTD have signicantly higher average returns than stocks
with weak LTD. Stocks in the quintile with the weakest (strongest) LTD earn an annual
average excess return of 3.99% p.a. (19.70% p.a.). The return spread between quintile port-
folio 1 and 5 is 15.71% p.a., which is statistically signicant at the 1% level. These results
are consistent with investors being crash averse and requiring a premium for holding stocks
with strong LTD. However, these ndings are only univariate and LTD is correlated with
several other variables that are contemporaneously related to returns such as regular beta
or size (see Table 2). Thus, we also compute the alphas generated by the quintile as well as
the dierence portfolios based on the one-factor CAPM, the three-factor Fama and French
(1993), as well as the four-factor Carhart (1997) model.
22
Results presented in the last three
columns show that alphas always increase monotonically from the weakest to the strongest
21
Results are stable if we use value-weighting instead of equal-weighting (see Section 3.7).
22
Alphas are estimated based on monthly portfolio and factor returns over the whole sample period.
16
LTD quintile portfolios. The CAPM-alpha (3-factor, 4-factor alpha) of the dierence portfo-
lio is economically large and amounts to 13.53% (16.19%, 14.06%) p.a., which is statistically
signicant at the 1% level.
In summary, results from Table 3 suggest that LTD determines the cross-section of stock
returns. Stocks with strong LTD earn high average contemporaneous raw- and risk-adjusted
returns. This nding is consistent with the view that stocks with weak LTD oer protection
against extremely low returns in crisis periods and thus on average earn lower returns.
3.1.2 Returns of LTD-sorted Portfolios During Crises
As a consistency check and to check whether weak LTD stocks really oer relatively good
returns during crises, we now also compute average returns of stocks in dierent LTD quintiles
on the most relevant nancial crisis days in our sample period. We examine Black Monday
(October 19, 1987), the Asian Crisis (October 27, 1997), the burst of the dot-com bubble
(April 14, 2000), and the recent Lehman crises (October 15, 2008). If LTD really captures
crash sensitivity, we should see an underperformance of strong LTD stocks as compared to
weak LTD stocks on these days. Results are presented in Table 4.
As expected (and opposite to what we nd in the overall sample; see Table 3), strong
LTD stocks strongly underperform weak LTD stocks on each of these individual days. The
dierences are economically large: the daily return of the weak LTD portfolio is from 4.4%
to 9.2% higher than that of the strong LTD portfolio. To assess the statistical signicance,
we also jointly analyze all days on which the excess return of the market over the riskfree
rate is less than 5%, which was the case on 16 days in our sample. Results are presented
in the last column of Table 4 and show that the weak LTD portfolio outperforms the strong
LTD portfolio by nearly 5% per day on those days. The eect is statistically signicant at
17
the 1% level. These ndings show that weak LTD stocks can indeed serve as an insurance
for loss-averse investors and might explain why overall they earn lower returns than strong
LTD stocks, as documented above.
3.1.3 Double-Sorts
Our univariate result of higher average returns of strong LTD stocks from Section 3.1.1
(Table 3) could be driven by dierences in beta, downside beta or dierences with respect to
other return characteristics like coskewness across the dierent LTD quintiles. This possibil-
ity is again suggested by the results from the correlation Table 2, where downside beta and
coskewness are the variables most strongly correlated with LTD. Thus, as a next step, we
conduct double sorts based on LTD as well as regular beta, downside beta, and coskewness.
23
We rst form quintile portfolios sorted on .
24
Then, within each quintile, we sort stocks
into ve portfolios based on LTD.
Panel A of Table 5 reports equal-weighted contemporaneous portfolio excess returns over
the risk-free rate of the 25 LTD portfolios. In all LTD quintiles, high stocks outperform
low stocks. More importantly, in all quintiles, we document a nearly monotonic increase
in returns from the weak LTD to the strong LTD portfolio. The return dierence between the
weakest LTD quintile and the strongest LTD quintile within all quintiles is economically
large and statistically signicant at the one percent level. It ranges from 5.23% p.a. in the
lowest quintile to 18.29% p.a. in the highest quintile. The average spread in excess
returns amounts to 10.40% p.a. Hence, regular risk cannot account for the reward earned
by holding stocks with strong LTD. This nding is consistent with our earlier ndings of a
23
Although we already control for linear beta exposure of our portfolios by looking at CAPM alphas
above, we now additionally analyze dependent portfolio double-sorts on LTD and regular , which allows us
to also control for a possible non-linear impact of .
24
Our results (not reported) are stable if we reverse the sorting order or conduct independent sorts.
18
highly positive CAPM-alpha of the dierence portfolio from Table 3.
LTD is also related to downside beta (

) as dened in Ang, Chen, and Xing (2006) as a

stocks conditional on the market return being below its mean (see Appendix B). Thus,
as above, in a rst step we form quintile portfolios sorted on

. Then, within each

quintile, we sort stocks into ve portfolios based on LTD. Panel B of Table 5 reports equal-
weighted average excess returns of the 25

LTD portfolios. We nd that the returns

of the low

portfolios tend to be smaller than those of the high

portfolios. Overall,
these ndings generally conrm the results of Ang, Chen, and Xing (2006). Turning to the
impact of LTD, we nd that stocks in the weak LTD portfolios have an average (across all

quintiles) excess return of 7.09% p.a., while stocks in the strong LTD portfolios have an
average excess return of 15.57% p.a. The spread is signicant at the 1% level. Amounting
to 8.48% p.a., it is also economically large. We nd the strongest eect of LTD on returns in
the highest

quintile, but it is highly statistically signicant and economically large within

each individual

quintile (ranging from 5.16% p.a. to 12.98% p.a.). Hence the impact of
LTD on returns does not seem to be driven by

, either.
25
Harvey and Siddique (2000) show that lower coskewness, coskew, is associated with higher
expected returns. Thus, to explicitly control for the impact of coskewness, we rst form
quintile portfolios sorted on coskew. Then, within each coskew quintile, we sort stocks into
ve portfolios based on LTD. Panel C of Table 5 shows equal-weighted average excess returns
of the 25 coskew LTD portfolios. In most coskew quintiles, we document a monotonic
increase of returns with LTD. The return dierence between the weakest LTD quintile and
the strongest LTD quintile ranges from 9.74% p.a. up to 13.73% p.a., with an average spread
of 12.34% p.a. This spread is again highly signicant within each individual coskew quintile,
25
We show later that our results can also not be explained by alternative denitions of downside beta
that condition on more extreme bad market outcomes (see Section 3.4).
19
which indicates that coskewness risk also cannot account for the reward earned by holding
stocks with strong LTD. At the same time, we can conrm the negative impact of coskewness
on returns documented in Harvey and Siddique (2000).
To summarize, based on double sorts we provide strong evidence that the risk associated
with LTD is dierent from risks associated with regular market beta, downside beta, and
coskewness. Double sorts oer the advantage that they allow us to control for any potential
non-linear impact. However, in double sorts we can only control for one return characteristic
at a time. Thus, we now turn to a multivariate approach that allows us to examine the joint
impact of dierent return and other characteristics of the rm that might have an impact
on the cross-section of stock returns.
3.2 Multivariate Evidence
We run Fama-MacBeth (1973) regressions on the individual rm level in the period from
1963 - 2009 using non-overlapping data.
26
Panel A of Table 6 presents the regression results
of yearly excess returns on realized LTD and various combinations of control variables in the
rst six columns.
In regression (1), we only include LTD as explanatory variable. It has a highly statistically
as well economically signicantly positive impact. A one standard deviation increase in
LTD leads to additional returns of 6.26% p.a. In regression (2), we add the stocks UTD
coecient. It has a signicantly negative impact on returns, but the economic magnitude
is much smaller than the impact of LTD. To check whether we can conrm the results
26
This econometric procedure has the disadvantage that risk factors are estimated less precisely in com-
parison to using portfolios as test assets. However, Ang, Liu and Schwarz (2010) show analytically and
demonstrate empirically that the smaller standard errors of risk factor estimates from creating portfolios
does not necessarily lead to smaller standard errors of cross-sectional coecient estimates. Creating portfo-
lios destroys information by shrinking the dispersion of risk factors and leads to larger standard errors.
20
from Ang, Chen, and Xing (2006) in our sample, in regression (3), we replace the two tail
dependence coecients by downside- and upside beta,

and
+
. We nd that the impact
of

is highly signicant, while there is no signicant impact of

+
. This nding is broadly
consistent with the analysis and results in Ang, Chen, and Xing (2006) who also nd that

is an important determinant of the cross-section of stock returns while

+
has much
less of an impact. In the following regressions, we expand regression model (2) and add
other rm characteristics like size, book-to-market and several other return characteristics
that might have an impact on returns.
27
Specically, in regression (4) we add coskewness,
coskew, and the Amihud (2002) illiquidity ratio, illiq, as a liquidity proxy, while regression
(5) additionally includes previous year returns, idiosyncratic return volatility, cokurtosis of
individual returns with the market return, and a stocks lottery features captured by the
maximum daily return over the past year, max, similar as in Bali, Cakici, and Whitelaw
(2011a).
28
Results show that the impact of LTD is very similar to before and still highly
signicant in economic as well as statistical terms after the inclusion of the control variables.
LTD exhibits the strongest inuence of all variables in terms of statistical power (t-statistic
of 11.59 and 9.89, respectively).
Several of the control variables have a signicant impact on returns, too, that conrm
ndings from the existing literature: Firm size (book-to-market ratio) has a negative (posi-
tive) impact (e.g., Fama and French (1993)), coskewness (Harvey and Siddique (2000)) and
illiquidity (Amihud (2002)) have a positive impact, while idiosyncratic volatility (Ang, Ho-
drick, Xing, and Zhang (2006) and (2009)) and max (Bali, Cakici, and Whitelaw (2011a))
27
We winsorize all realizations of our independent variables at the 1% and 99% levels in order to avoid
outliers driving our results. Our results do not hinge on this winsorization (see Section 3.7).
28
Bali, Cakici, and Whitelaw (2011a) dene max based on a monthly horizon. Our later examination
of a trading strategy is conducted on a monthly basis. Thus, we will later include a lottery-factor in that
analysis, which matches the time-horizon of Bali, Cakici, and Whitelaw (2011a) more closely (see Section
4.2).
21
have a negative impact. Consistent with Fang and Lai (1997) and Dittmar (2002), we also
nd a positive coecient for the impact of cokurtosis. However, the eect is not statistically
signicant in our setting.
In regression (6) we replace by

and
+
. Regressions (7) to (9) are identical to
regression (6), but we use rm-individual yearly alphas estimated based on daily data from
the CAPM, the Fama and French (1993), and the Carhart (1997) models, respectively, as
dependent variable. In all cases, our earlier ndings are conrmed: there is a very strong
positive impact of LTD and the t-statistic for the impact of LTD is always the highest t-
statistic of all explanatory variables. Interestingly, including downside beta does not reduce
the impact of LTD. Rather, in regressions (7) to (9) downside beta has no signicant impact
on returns anymore when LTD is also included.
29
The last column presents the economic
signicance based on a one-standard deviation change of each explanatory variable based on
the results from regression (9): a one-standard deviation increase of LTD leads to a increase
in the Carhart (1997) four factor alpha of 5.01% p.a. This is the largest eect in terms of
economic magnitude of all variables included. In contrast, the economic impact of

is
economically small, amounting to 0.21% p.a. only.
3.3 Impact of LTD in Risk-, VaR-, and Conditional VaR-sorted
Samples
So far, we have focused on the impact of LTD, dened in (2) as the limit of the probability
of a joint left tail realization of the market and an individual rms return. We use this
29
In unreported tests, we also run variants of regressions (6) to (9) where we include

and
+
and
the other controls, but not LTD and UTD. In this case, we nd an economically meaningful and typically
statistically signicant coecient for the impact of

. This nding shows that

does not loose its

signicant impact due to the inclusion of the other control variables, but due the inclusion of LTD.
22
measure to capture a stocks crash sensitivity, because it tells us how likely it is that a
stock realizes its worst return exactly at the time when the market realizes its worst return.
However, it does not take into account how bad the worst return realization of the stock
really is (similar to the dierence between the correlation of a stocks return with the market
and its market beta). Thus, we now want to check whether the impact of LTD is stronger if
we can expect the worst return realization of a stock to be particularly low, i.e., if we weight
the probability of the joint outcome with a proxy for the potential severity of the outcome.
We use three ad-hoc proxies to capture how bad a bad outcome would be: a stocks annual
return standard deviation based on daily returns,
30
its Value-at-Risk (VaR, dened as the
5% percentile of the daily returns of the stock), and its conditional Value-at-Risk (CoVaR,
dened as the conditional mean of all daily returns below the 5% percentile). Stocks with a
high return standard deviation (a low VaR and a low CoVaR) tend to have particularly low
worst returns (i.e., they have particularly bad return realizations in an absolute sense if they
realize their relatively worst returns). To examine whether there is a stronger impact of LTD
on returns for stocks with a high return standard deviation (a low VaR and CoVaR), we sort
stocks into two categories according to the median of the three proxies in the respective year
and repeat regressions (6) to (9) from Panel A for these subsamples. Results for the impact
of LTD are shown in Panel B of Table 6. All other explanatory variables are included in
the regressions, but suppressed in the Panel. Our ndings clearly indicate that the impact
of LTD is much stronger for rms where bad outcomes are particularly severe, i.e., for rms
with a high return standard deviation (low VaR and low CoVaR). Irrespective of which
proxy we look at, we always nd that the coecient for the impact of LTD on returns and
alphas is about three times as large among high standard deviation (low VaR, low CoVaR)
30
We also use a stocks semi-variance (variance conditional on the return being below its mean) instead
of its variance. Findings (not reported) are very similar.
23
rms than among low standard deviation (high VaR, high CoVaR) rms. These ndings
conrm our conjecture, that tail dependence matters more for returns among more risky (in
a traditional, i.e., univariate, sense) rms.
3.4 Downside Beta vs. Lower Tail Dependence
Results in Panel B of Table 5 show that our results are not driven by downside beta (

).
Although the concepts of LTD and

seem related, this is not surprising, because the latter

focuses on all market returns below the mean, while the former explicitly focuses on extreme
events. However, one could argue that alternative denitions of

that focus more on the

left tail of the market return distribution capture eects similar to LTD. To analyze this
idea more closely, we repeat our

LTD double sorts from Table 5 for alternative

denitions in Table 7. Specically, we calculate downside betas as betas conditional on the

market return being below its 20%, 10%, 5%, 2%, and 1% quantile (rather than being below
the mean, as before and as in Ang, Chen, and Xing (2006)).
We report results on the returns of the strong minus weak LTD portfolios within each
downside beta quintile in the rst ve columns as well as the average of this dierence
portfolio return across all downside beta quintiles in the last column (like in the last row of
Panel B in Table 5) for all alternative

denitions. The average dierence returns range

from 9.46% (for the

denition based on the 20% quantile) up to 15.23% (for the

denition based on the 1% quantile) and is signicant at the 1% level in each case. Thus,
our results not only hold after adjusting for various

alternatives, they even get stronger if

we look at more restrictive

denitions.
31
At rst glance, this result might seem surprising,
as more restrictive

s (that focus more on extremely bad market returns) should be more

31
We also replace the standard downside beta by alternative downside beta denitions in our multivariate
regressions from Table 6 (Panel A) and obtain similar results. Results are available upon request.
24
related to our LTD measure. The reason we nd even stronger results for the alternative

s is that they are actually not able to reliably capture dependence in the tails because
they are estimated based on a very small number of observations (e.g., only about 12 daily
return observations per year for the 5% quintile

) and are thus very noisy.

32
These results also illustrate the advantage of the copula approach in estimating extreme
dependence: in estimating the whole dependence structure between individual and market
returns using our semi-parametric approach, we make use of all available daily return ob-
servations within a year, which allows for a relatively more precise estimation. We then
calculate LTD coecients based on the procedure described in Section 2.2. Thus, the com-
putation of LTD ultimately does not rely on a very small number of extreme observations
only (like the more restrictive

s discussed above) and consequently is much less noisy and

more informative about the true crash sensitivity of a stock.
3.5 Lower Tail Dependence and Momentum
Chen, Hong, and Stein (2001) and Boyer, Mitton, and Vorkink (2010) show that (con-
ditional) skewness is related to past returns. In particular, they document that negative
skewness (which is also a proxy for the downside exposure of stocks) is most pronounced in
stocks that have expericenced high positive returns over the last 12 to 36 months. In unre-
ported tests, we also nd that past returns are a strong positive predictor of a stocks LTD.
33
32
Correlations (not reported in tables) between the

alternatives and LTD actually decrease from 0.49

for the standard Ang, Chen, and Xing (2006)

, to 0.41 for the correlation between LTD and

based
on the 20% quintile and even further to 0.17, 0.07, and 0.03 for the correlation of LTD with

based on
the 5%, 2%, and 1% quintile. Additionally, if we compute the returns of a portfolio going long in high

stocks and short in low

stocks, we typically nd no positive abnormal returns using the more restrictive

denitions, irrespective of whether we look at univariate sorts or double sorts based on

and LTD (not

reported in tables).
33
Regressing a stocks LTD in year t on its past return in year t 1 in a univariate model delivers a
positive coecient estimate of 0.029 with a high degree of statistical signicance (t-stat: 6.00).
25
Thus, it is not unlikely that the LTD-premium we document is related to the Jegadeesh and
Titman (1993) momentum eect.
To examine whether such a relationship exists, we run a simple time series regression based
on yearly data where the momentum strategy return is regressed on the excess market return,
the Fama and French (1993) size and book-to-market factor as well as the LTD premium
(computed as the yearly strong LTD - weak LTD portfolio return). Results are presented in
Table 8.
In the rst column we do not include the LTD premium factor and nd evidence for
the well-documented momentum eect. The yearly alpha after controlling for the exposure
to the size- and book-to-market-factor amounts to 12.8%. In column (2) we include the
equal-weighted return of the strong LTD - weak LTD portfolio and in column (3) the value-
weighted return of the strong LTD - weak LTD portfolio as additional factors. We nd that
in both cases the momentum strategy return positively loads on LTD. The coecient is
around 0.7 and statistically signicant at the 5%-level, showing that a momentum strategy
indeed is crash-sensitive. Interestingly, the alpha of the momentum strategy is reduced to
1.88% and 3.91%, respectively, if we include the equal- or value-weighted LTD factor, and
not statistically signicant at conventional levels anymore. Thus, our results suggest that
part of the momentum eect can be explained by the crash sensitivity of this strategy.
34
34
It should be noted that our LTD factor is not based on an implementable trading strategy, because rms
are sorted based on contemporaneous LTD. In unreported tests, we also construct a traded factor based on
past LTD. The return of the momentum factor still has a signicantly positive loading on this LTD factor,
but the alpha of the momentum strategy is only reduced by about 1.5% p.a. and remains signicant in this
case.
26
3.6 Simplifying Tail Dependence Coecient Estimation
The copula selection procedure described in Section 2.2 and Appendix A is computation-
ally costly.
35
We now investigate whether this estimation procedure could be simplied.
Instead of selecting the appropriate parametric copula by minimizing the distance between
64 dierent convex copula combinations and the empirical copula, we ex ante choose var-
ious xed convex copula combinations. As our ad-hoc xed copula combinations, we con-
sider the Rotated Joe/F-G-M/Joe (3-D-I)-, the Rotated Galambos/F-G-M/Joe (4-D-I)-, the
Rotated Gumbel/F-G-M/Joe (2-D-I)-, the Rotated Gumbel/Frank/Gumbel (2-B-II)-, the
Rotated Galambos/Frank/Gumbel (4-B-II)-, and the Rotated Galambos/Frank/Galambos
(4-B-III)-copula out of Table A.1. We choose the copula combinations (3-D-I), (4-D-I), and
(2-D-I) because they are the copulas most often selected in the estimation procedure in Table
A.2. In contrast, the remaining three copula combinations are the copulas least often se-
lected in the estimation procedure. We perform Fama-Macbeth (1973) regressions of excess
returns on LTD (estimated based on the xed copula combinations) as well as the full set
of control variables (as in Regression (6) of Table 6, Panel A). Results on the coecient
estimates for the inuence of LTD are displayed in the left part of Table 9.
We nd that LTD is a highly signicant explanatory factor for the cross-section of expected
stocks returns independent of the specied convex copula combination. The magnitude
of the coecients and the signicance levels of the control variables (not shown in the
table) also remain stable across the dierent specications. These results document that our
main ndings are not driven by the tail dependence coecient estimation procedure. The
estimation procedure, which is computationally intensive, can be dramatically simplied by
35
The selection of the optimal copula combination and the estimation of the tail dependence coecients
on a grid cluster takes about 20 seconds per stock and year. This amounts to an entire estimation time for
the LTD coecients used in this study of well above 500 hours.
27
just picking a reasonable convex copula combination (based on the copulas in Table A.1).
This might be a helpful result for researchers working on the impact of tail dependence in
similar settings.
3.7 Additional Analysis and Robustness Checks
In this section we conduct a battery of additional robustness tests to analyze whether
our main results from above are stable. We examine the inuence of return adjustments
(Section 3.7.1) and the weighting scheme in the portfolio sorts (Section 3.7.2), the temporal
stability of our results (Section 3.7.3), and variations of the regression setup employed in the
multivariate tests (Section 3.7.4).
36
3.7.1 Industry- and DGTW(1997)-Adjusted Returns
Some extreme market downturns can mainly be driven by a few industries. Consequently,
some of our ndings might be driven by industry eects. Thus, we repeat our multivariate
regressions with the full set of controls (i.e., Regression (6) from Table 6, Panel A) but
use industry-adjusted returns instead of raw returns as dependent variable. Results on the
estimates for the impact of LTD are presented in the right part of Table 9 .
In the rst two lines, we use the Fama-French 12 (FF12) and 48 (FF48) industry classi-
cations. In both cases, the coecient for the impact of LTD is statistically signicant at
the one percent level and similar in magnitude to the ndings above. As Johnson, Moore,
and Sorescu (2009) point out, tests of long-term abnormal returns based on FF48 industry
adjustments might not be well-specied if there is industry clustering (because rms in the
36
Besides the robustness tests described here, we also use weekly data instead of daily data and a longer
estimation horizon to determine our tail dependence coecients. Our results (not reported) are similar if
we use a longer estimation horizon of 2 years, 3 years, or 5 years instead of 12 months.
28
same FF48 industry often belong to very dierent SIC-3 industries). Thus, we also follow
their suggestion and use the SIC-3 digit industry classication to industry-adjust returns.
Additionally, we use the SIC 2- and 4-digit industry classication. In all cases, our prior
results are conrmed. These ndings show that our results are not just driven by industry
eects.
Furthermore, instead of controlling for the impact of stock characteristics by including
them as explanatory variables, we also adjust the return of each stock by subtracting the
return of its corresponding Daniel, Grinblatt, Titman, and Wermers (1997) characteristic-
based benchmark (DGTW).
37
Results are presented in the last line of the same table. Again,
our result of a strongly positive impact of LTD on expected returns remains unaected.
3.7.2 Weighting Scheme
Our sorts and double sorts in Section 3.1 focus on equal-weighted portfolios. Thus, results
could be inuenced by over-weighting the importance of small or tiny stocks.
38
Therefore,
we now examine value-weighted average excess returns of the same univariate as well as
bivariate sorts as in Tables 3 and 5. Results are presented in Table 10.
Value-weighted returns for our LTD quintile portfolios are presented in Panel A. We nd
that stocks in the quintile with the weakest (strongest) LTD earn an annual average excess
return of 1.03% p.a. (9.45% p.a.). The return spread between quintile portfolios 1 and 5
is 10.48% p.a. and statistically signicant at the 1% level.
In Panels B to D of Table 10 we perform value-weighted double-sorts on LTD and ,

,
37
The DGTW benchmarks are available via http://www.smith.umd.edu/faculty/rwermers/ftpsite/
Dgtw/coverpage.htm.
38
Similar concerns can also be raised with respect to the multivariate regression results because the
regression evidence presented in Table 6 is essentially also based on equal weighting as each observations
enters the cross-sectional Fama-MacBeth (1973) regressions with the same weight.
29
and coskew, respectively. As with equal-weighted portfolios, we document a clear outperfor-
mance of strong LTD stocks over weak LTD stocks in each quintile (Panel B). The return
spread ranges from 2.75% p.a. in the lowest quintile up to 16.22% p.a. in the highest
quintile. The return spread is signicant at the 1% level within all quintiles except for the
very lowest quintile, where it is signicant at the 5% level. We nd similar patterns for
the double sorts based on

and LTD (Panel C) as well as double sorts based on coskew

and LTD (Panel D).
Overall, value-weighting portfolios rather than equal-weighting does not change our main
nding of signicantly higher returns of stocks with strong LTD. Thus, our results are not
driven by the returns of very small rms.
3.7.3 Temporal Stability
As a rst test of the temporal stability of our main nding, we reproduce the results of the
univariate sorts from Table 3 for 5 subperiods of roughly equal length: 1963-1972, 1973-1981,
1982-1990, 1991-1999, and 2000-2009. Results are presented in the rst ve columns in Panel
A of Table 11.
In each subperiod we nd returns that are monotonically increasing with LTD. The return
spread between the strong and the weak LTD portfolios is highest (21.21%) in the 1991-1999
period and lowest (7.86%) in the most recent period, 2000-2009. It is always statistically
signicant at least at the 5% level. In the last two columns we split the sample in 1987
in order to check whether the results dier prior to the crash of 1987 and after. In the
option pricing literature it is sometimes argued that investors became crash-o-phobic after
the experience of the 1987 crash (Rubinstein (1994)). If investors were less crash-averse prior
to 1987, we should see a weaker compensation for LTD in the earlier subperiod.
30
We document a return spread between stocks with strong LTD and stocks with weak LTD
of 16.24% p.a. in the period 1963 to 1986. Results for the later subperiod from 1987 to
2009 show that the return spread is of similar magnitude at 15.17% p.a. In both cases, the
magnitude of the spread is similar to that from the complete sample and signicant at the
1% level. This shows that investors on the stock market were compensated for holding stocks
with strong LTD prior to 1987 and suggests that they were crash-averse even prior to the
crash of 1987.
39
In Panel B, we report results from the same double sorts as in Table 5 separately for the
pre- and post-1987 period. We only report the average return across the ,

, and coskew,
respectively, portfolios for the 5 LTD portfolios as well as the dierence portfolio. While the
dependent sorts based on and LTD as well as those based on

and LTD show a more

pronounced return spread in the later period (8.45% vs. 12.45% and 6.77% vs. 10.26%),
return spreads from dependent sorts based on coskew and LTD are very similar in the two
subperiods (12.71% and 11.95%). Irrespective of these patterns, results are economically
meaningful and statistically signicant in either case.
Finally, we look at the results from our multivariate regressions with the full set of ex-
planatory variables for the same ve subperiods as in Panel A. Results of the respective
Fama-MacBeth (1973) regressions are presented in the rst ve columns in Panel C of Ta-
ble 11. Even within these relatively short subsamples, we always nd a coecient for the
impact of LTD that is of similar magnitude to the full sample. It ranges from 0.385 in the
last subperiod of our sample up to 0.605 in the fourth subperiod. The coecient estimate
for LTD is signicant at the 5% level in the rst subperiod and signicant at the 1% level in
39
While seemingly contradictory to some of the empirical option pricing literature it should be noted that
studies that nd no strong crash-fear eect prior to 1987 typically rely on very short pre-87 samples due
to the lack of option data availability for earlier years.
31
all other subperiods. LTD is the most consistently priced factor of all explanatory variables
across the subperiods.
We also look at subsamples based on up and down market returns. Results are presented
in the last two columns. As expected, the eect of LTD on the cross-section of stock returns
is stronger in times of up markets (market return > 0) than in down markets (market return
< 0). However, the eect is signicant at the 1% level even in the latter case. Overall, the
strong positive cross-sectional impact of LTD on stock returns is remarkably stable over time
and holds in up and down markets.
3.7.4 Alternative Regression Methods
Our previous multivariate regression evidence in Section 3.2 relies on standard Fama-
MacBeth (1973) regressions with winsorized independent variables. We now perform several
variations of this basic regression approach with the full set of independent variables. Results
are presented in Table 12.
Regression (1) repeats the baseline regression (6) from Panel A in Table 6, but we now use
Newey-West standard errors in the second stage of the Fama-MacBeth (1973) regressions
to determine statistical signicance. Regression (2) repeats the standard Fama-MacBeth
(1973) regression, but we do not winsorize the independent variables. In regression (3)
we perform a pooled OLS regression with time-xed eects and standard errors clustered
by stock. Regression (4) is identical, but we cluster standard errors by industry (based
on the Fama-French 12 industry classication).
40
Regressions (5) and (6) perform panel
data regressions with rm xed eects. In regression (6) standard errors are additionally
clustered by rm. Finally, in regression (7) we regress excess returns on the independent
40
Results are virtually unchanged whether we cluster by Fama-French 48 or SIC industries.
32
variables via a random-eect panel data regression. We document that LTD is a highly
signicant explanatory factor for the cross-section of expected stocks returns independent of
the specic regression setup. The point estimate for the inuence of LTD is roughly 0.45
and very similar across regressions.
Overall, the results from our robustness checks show that our main ndings are conrmed
and do not depend on the weighting scheme or return adjustments, are stable over time, and
are not driven by the specic regression technique or by a specic dependence structure of
the error terms.
41
4 Crash Sensitivity Persistence and Trading Strategy
The previous section documents that LTD has a strong impact on contemporaneous re-
turns. We now examine whether tail dependence is persistent over time (Section 4.1) and
whether it also predicts cross-sectional return dierences. Although not the focus of our
paper, we thereby want to check whether a protable trading strategy could be implemented
solely based on past information on tail dependence (Section 4.2).
4.1 Persistence of Crash Sensitivity
To examine whether LTD is time-varying, we compute the average LTD of the stocks in
the LTD quintile portfolios over time. Firms are sorted into quintiles based on their realized
LTD in year t = 1. Panel A of Figure 3 displays the evolution of the average equal weighted
LTD of these portfolios over the following four years t = 2 to t = 5.
41
In unreported tests, we also examine the impact of LTD within subsamples consisting of rms from one
Fama-French 12 industry at a time. The eect is signicant within most individual industry subsamples and
is strongest in Business Equipment, Whole Sales, Retail, Services, and Manufacturing.
33
The Figure shows that the stocks in the strong LTD portfolio also have stronger LTD than
the stocks from the weak LTD portfolio in the following years. However, the dierence quickly
shrinks considerably. In unreported results from a regression analysis, we nd evidence for
limited predictive power of past LTD on current LTD. Regressing current LTD on past LTD
in a univariate model delivers a coecient estimate of 0.248 with a high degree of statistical
signicance (t-stat: 12.56), but a relatively low R
2
of about 7%. The patterns documented
indicate that there is some predictability in extreme dependence, but the sharply decreasing
dierences shown in Figure 3 suggest that many stocks are exposed to time-varying tail
risk.
42
4.2 Past LTD and Future Returns
4.2.1 Trading Strategy
We now examine whether it is possible to generate abnormal returns based on prior infor-
mation about past extreme dependence. Our strategy consists of going long in stocks with
strong past LTD and going short in stocks with weak past LTD with monthly rebalancing.
Specically, we sort stocks into ve quintile portfolios at the beginning of each month based
on past LTD estimated over the previous twelve months. Then, we examine equal-weighted
returns of these portfolios over the next month. We use data from January to December
1963 to determine the LTD estimates for the rst portfolio formation in January 1964. Thus,
our trading strategy spans the period January 1964 to December 2009.
Our empirical setup requires the estimation of LTD-coecients for each stock for 552
overlapping 12 month periods. To reduce the computational eort, we rely on the results
42
This latter nding justies our previous approach of relating returns to contemporaneous tail dependence
realizations over relatively short horizons (as suggested for time-varying risk exposures in Lewellen and Nagel
(2006)).
34
from Section 3.6 and simplify the estimation procedure for LTD in selecting the Rotated
Joe/F-G-M/Joe (3-D-I)-copula as our xed copula convex combination for all stocks and
periods.
43
Table 13 reports the monthly average excess return over the risk free rate for all quintile
portfolios based on past LTD and the return and alpha dierences between quintile portfolio
5 (strong LTD) and quintile portfolio 1 (weak LTD).
Column 1 of Table 13 shows that stocks in the strongest past LTD quintile earn an average
equal-weighted excess return over the risk free rate of 0.862% per month, while the stocks
in the weakest past LTD quintile earn 0.499% per month. Thus, our trading strategy of
investing in strong LTD stocks and shorting weak LTD stocks delivers an economically sig-
nicant future return of 0.363% per month, which translates into an average return premium
of 4.36% p.a. The spread in average excess returns between quintile portfolios 5 and 1 is
statistically signicant at the 1% level.
44
To check whether exposures to systematic risk factors drives this nding, we regress the
monthly return time series of the past LTD quintile portfolios as well as the dierence
portfolio returns on the monthly excess market return and other systematic risk factors.
The alphas from these regressions are shown in columns (2)-(4) of Table 13. Results from
the CAPM one-factor regressions show that a part of the premium from our trading strategy
is indeed due to its sensitivity to current market beta. However, the market factor exposure
cannot explain the full LTD return premium. The trading strategy still delivers a monthly
alpha of 0.253% (signicant at the 5% signicance level) after controlling for market beta.
43
Results based on other ad-hoc chosen copula combinations are very similar.
44
In unreported tests, we also investigate a trading strategy based on UTD. We nd no signicant return
spreads based on such a strategy. Furthermore, we also analyze trading strategies that are double sorted on
past LTD as well as past beta, past downside beta, and past coskewness, respectively. In each case we nd
a signicant outperformance of the strong past LTD over the weak past LTD portfolio.
35
When taking into account the size factor (SMB) and the book-to-market factor (HML), the
alpha of our strategy increases to 0.454% per month. The returns of our trading strategy
load signicantly negatively on both factors. In the fourth column, we also control for the
momentum factor. We still obtain a highly signicant monthly alpha of 0.383% per month.
Results from the four-factor Carhart model also show, that the long and the short side of the
trading strategy contribute roughly equally to the overall performance. While the portfolio
with the weakest past LTD stocks delivers an alpha of 0.20% per month, the quintile
portfolio with the strongest past LTD stocks delivers an alpha of 0.19% per month (both
dierent from zero with a signicance level of 5%). In the columns (5) and (6) we repeat
the same analysis, but estimate the LTD of the stocks not over the past 12 months, but over
the past 24 and 36 months, respectively. Results are generally conrmed. Finally, in the
last three columns, we present the Carhart (1997) four factor alpha based on value-weighted
portfolio rather than equal-weighted portfolios for the three estimation horizons. Results are
slightly weaker than for equal-weighted portfolios, but still positive and signicant.
As downside beta and LTD are dierent but related concepts (see Section 3.4), we also
want to check whether a trading strategy based on

delivers similar results. In the last line

of the table, we report dierence returns and alphas of a trading strategy going long in stocks
in the top quintile and going short in stocks in the bottom quintile according to their lagged

. The return of the trading strategy based on

estimated over the previous 12 months

is 0.09% and not statistically signicant. The CAPM (Fama and French (1993), Carhart
(1997)) alpha of the strategy is also negative at 0.319% (0.333%, 0.260%). Similar
results are obtained for longer estimation horizons for

and for value-weighted portfolios.

Thus, we can conclude that a trading strategy based on information about lagged

does
not deliver a positive outperformance, while a trading strategy based on information about
36
lagged LTD seems protable.
45
4.2.2 Alternative Factor Models
In the following, we test the robustness of our ndings from Table 13, by evaluating our
trading strategy using alternative factor models. Results are presented in Table 14.
First, we include the Pastor and Stambaugh (2003) traded liquidity risk factor in regression
(1). The monthly alpha of our LTD trading strategy from above is now 0.38% and remains
signicant at the 1% level. In regression (2), we replace the Pastor and Stambaugh (2003)
liquidity factor with the Sadka (2006) liquidity factor that is based on the permanent (vari-
able) component of the price impact function. In regression (3) we include the Bali, Cakici,
and Whitelaw (2011a) factor to control for exposure of our strategy to lottery-type stocks
and in regression (4) we include the Baker and Wurgler (2006) sentiment index orthogonoal-
ized with respect to a set of macroeconomic conditions.
46
In regression (5) we replace the
momentum factor with the Fama-French short- and long-term reversal factors. Finally, as
alternative factor models, in regressions (6) and (7) we use the models suggested in Cremers,
Petajisto, Zitzewitz (2010) that contain various combinations of return dierences between
the S&P 500 and Russell 2000 and 3000 subindexes as well as the momentum factor.
47
In
each case, we document a signicant positive alpha of our trading strategy ranging from
0.30% up to 0.55% per month, showing that the results from our basic trading strategy are
45
Ang, Chen, and Xing (2006) also nd no positive outperformance of a simple trading strategy based on
information about past

only.
46
The lottery factor is provided by Nusret Cakici (http://www.bnet.fordham.edu/cakici/) and the
time series of the sentiment factor is taken from http://people.stern.nyu.edu/jwurgler/.
47
Cremers, Petajisto, Zitzewitz (2010) propose using either four or seven factors out of the ten calculated
by them. The data on the factor returns are taken from http://www.petajisto.net/data.html. The four
factor model contains the factors s5rf, r2s5, r3vr3g, and the momentum factor (umd), while the seven factor
model contains the factors s5rf, rms5, r2rm, s5vs5g, rmvrmg, r2vs2g, and the momentum factor (umd) (as
explained in their data library).
37
robust with respect to alternative factor model specications.
4.2.3 Further Robustness Checks: Multivariate Evidence
To check for the impact of rm characteristics we also repeat the multivariate regression
specications from Table 6, but use lagged values for LTD as well as all other independent
variables. Results are presented in Table 15.
We only show the coecient estimate for the impact of lagged LTD. The rst four lines in
the left part of the table show results from regressions equivalent to regressions (1), (2), (5),
and (6) from Table 6 (Panel A). The impact of past LTD is always positive and typically
signicant at the 1% level (5% level if we use LTD as the only independent variable in
regression (1)). In the last line, we again use regression (6) with the full set of controls,
but use DGTW-adjusted returns as dependent variable. Again, the impact of past LTD is
highly signicant and of similar magnitude. In the right part of the table, we present results
from regression (6) where we use industry adjusted returns as dependent variables. To dene
industries we use the Fama-French 12 and 48 industry classication as well as SIC 2-, 3-, and
4-digit codes. Results are again very similar in all cases and the coecient estimate for the
impact of past LTD is always signicant at the 1% level. These results conrm our previous
multivariate results based on contemporaneous regressions also in a predictive regression
setup.
Overall, the ndings from Section 4 suggest that it is possible to create a protable and
implementable trading strategy based on information about past LTD. However, these results
are only indicative, as we do not take into account any trading costs and other limits of
arbitrage. Limits of arbitrage are likely to be relevant here, because we short stocks with
weak LTD (which tend to be small and low stocks). Furthermore, this strategy would of
38
course only be protable on a risk-adjusted basis if we assume that investors do not require
a risk premium for holding crash-sensitive assets (or at least only one that is lower than the
documented LTD premium).
5 Conclusion
The cross-section of expected stock returns reects a premium for crash sensitivity as
measured by a stock returns lower tail dependence, LTD, with the market return. Stocks
that are characterized by strong LTD earn signicantly higher average returns than stocks
with weak LTD. We nd that the high average returns earned by stocks with strong LTD
are not explained by alternative cross-sectional eects, including market beta, size, book-to-
market, momentum, liquidity, coskewness, cokurtosis, idiosyncratic volatility, and downside
beta. Controlling for these and other cross-sectional eects, we nd that an increase of one
standard deviation in LTD is associated with an expected return premium of about 5% p.a.
Our ndings also suggest that most of the impact of the downside beta of Ang, Chen, and
Xing (2006) seems to be driven by extreme dependence in the lower tail of the bivariate
distribution of individual security and market returns. Furthermore, we document some
predictability of extreme dependence based on past extreme dependence. We can form an
investable trading strategy based on past extreme dependence structures that is protable
at least before trading costs
In contrast, if we focus exclusively on periods of heavy market downturns, we nd that
stocks with weak LTD outperform stocks with strong LTD. As stocks with weak LTD thus
essentially oer an insurance against extreme negative portfolio returns, our results are
consistent with the view that investors are willing to pay higher prices and eventually accept
39
lower returns for stocks with weak LTD. The conjecture that the higher returns of stocks
with strong LTD is a reection of higher equilibrium returns in the presence of crash-averse
investors is consistent with ndings from the empirical literature on option prices (e.g.,
Rubinstein (1994) and Polkovnichenko and Zhao (2012)). Our results are also consistent
with the literature on rare disaster risk that nds that much of the aggregate equity risk
premium seems to be a compensation for rare disaster or tail risk (e.g., Bollerslev and
Todorov (2011) or Gabaix (2012)). Furthermore, we nd that momentum prots can be
explained by the exposure of the momentum strategy returns to a contemporaneous LTD
factor. This nding suggests that momentum prots are to a large part a compensation for
the crash-sensitivity of the momentum strategy.
On a broader level, the fact that investors can earn a premium for bearing LTD risk has
serious implications for nancial stability: If nancial institutions do not have to bear the
expected costs of a severe market downturn (e.g., because regulatory capital requirements do
not take into account LTD or because they expect to be bailed out in a severe crisis), they
have incentives to invest in exactly those securities that are characterized by strong lower
tail dependence with the market in order to earn the associated premium.
48
Such incentives
would make those institutions heavily exposed to market crises and could lead to systemic
instability. Whether nancial institutions are heavily invested in strong LTD assets is an
interesting open question for future research.
48
Some suggestive evidence along these lines is again provided in the empirical option market literature.
Garleanu, Pedersen, and Poteshman (2009) document that dealers on aggregate hold short positions in out-
of-the-money puts - that also oer protection against downturns - while end-users (dened as customers of
brokers), seem to hold long positions, i.e., they insure against extreme downside risk. Furthermore, Kelly
and Jiang (2012) document that hedge funds often have a large exposure to downside tail risk.
40
A Appendix: Estimating Tail Dependence Coecients
This appendix provides the technical details of the copula estimation and selection pro-
cedure and the calculation of the respective tail dependence coecients. The estimation
of LTD- and UTD-coecients can either be based on the entire set of observations or on
data on extreme events only. In the univariate setting, the extreme value distributions can
be expressed in parametric form (see Fisher and Tippett (1928)) and parametric extreme
value theory (EVT) is the natural choice for inferences on extreme values. On the contrary,
bivariate extreme value distributions (such as in this paper) cannot be characterized by a
fully parametric model in general, which leads to more complicated estimation techniques.
Our estimation approach for the estimation of the lower tail dependence of a stock with the
market relies on the entire set of return observations of a rm and the market in a given
year.
Coecients of tail dependence have closed form solutions for several basic parametric
copulas (see Table A.1). Unfortunately, these basic copulas do not allow us to model upper
and lower tail dependence simultaneously. However, Tawn (1988) shows that every convex
combination of existing copula functions is again a copula. Thus, if C
1
(u
1
, u
2
), C
2
(u
1
, u
2
),
. . ., C
n
(u
1
, u
2
) are bivariate copula functions, then
C(u
1
, u
2
) = w
1
C
1
(u
1
, u
2
) + w
2
C
2
(u
1
, u
2
) + . . . + w
n
C
n
(u
1
, u
2
)
is again a copula for w
i
0 and

n
i=1
w
i
= 1.
To allow for the maximum possible exibility, we consider all 64 possible convex combi-
nations of the afore mentioned basic copulas from Table A.1 that each consist of one copula
that allows for asymptotic dependence in the lower tail, C
LTD
, one copula that is asymp-
41
totically independent, C
NTD
, and one copula that allows for asymptotic dependence in the
upper tail, C
UTD
:
C(u
1
, u
2
, ) = w
1
C
LTD
(u
1
, u
2
;
1
)
+w
2
C
NTD
(u
1
, u
2
;
2
) + (1 w
1
w
2
) C
UTD
(u
1
, u
2
;
3
), (7)
where denotes the set of the basic copula parameters
i
, i = 1, 2, 3 and the weights w
1
and
w
2
.
49
Our estimation approach for the upper and lower tail dependence coecients then follows
a three-step procedure. First, based on daily return data for the market and each stock, we
estimate a set of copula parameters
j
for j = 1, . . . , 64 dierent copulas C
j
(, ;
j
) between
an individual stock return r
i
and the market return r
m
for each year.
50
Each of these convex
combinations requires the estimation of ve parameters: one parameter
i
(i = 1, 2, 3) for
each of the three basic copulas and two parameters for the weights w
1
and w
2
. The copula
parameters
j
are estimated via the canonical maximum likelihood procedure of Genest,
Ghoudi, and Rivest (1995). The details of this step are described in Section A.1.
Second, we follow Ane and Kharoubi (2003) and select the appropriate parametric copula
C

(, ;

) by minimizing the distance between the dierent estimated parametric copulas

C
j
(, ;

j
) and the empirical copula

C based on the Integrated Anderson-Darling distance.
51
49
These convex combinations are similar to other copulas such as the BB1 to BB7 copulas suggested in
Joe (1997), but oer more exibility. Particularly, as our convex combinations also contain one copula that is
asymptotically independent, ours is an extremely exible and ecient way to model dependence structures.
50
In computing the market return r
m
we exclude stock i, so the market return r
m
is slightly dierent for
each stocks time series regression. This removes potential endogeneity problems when calculating LTD- and
UTD-coecients for each stock.
51
Results are very similar if we select the copula based on the Kolmogorov-Smirnov distance or the
estimated log-likelihood value.
42
The details of the selection procedure are described in more detail in Section A.2. The result
of this step is summarized in Table A.2, in which we present the absolute and percentage
frequency by which each of the possible 64 combinations is chosen.
All combinations are chosen regularly and no specic copula clearly dominates. The three
copula combinations that are most often selected are the Rotated-Joe/F-G-M/Joe-copula
(3.29%), the Rotated-Galambos/F-G-M/Joe-copula (3.08%), and the Rotated-Gumbel/F-
G-M/Joe-copula (2.65%).
Third, we compute the tail dependence coecients LTD and UTD implied by the estimated
parameters

of the selected copula C

(, ;

). The computation of LTD and UTD is

straightforward if the copula in question has a closed form, as all the basic copulas used in
this study do. The lower and upper tail dependence coecient of the convex combination
are calculated as the weighted sum of the LTD and UTD coecients from the individual
copulas, respectively, where the weights from (6) are used, i.e., LTD

= w

1
LTD(

1
) and
UTD

= (1 w

1
w

2
) UTD(

3
). As this procedure is repeated for each stock and year, we
end up with a panel of tail dependence coecients LTD

i,t
and UTD

i,t
at the year-rm level.
A.1 Estimation of the Copula Parameters
The estimation of the set of copula parameters
j
for the dierent copula combinations
C
j
(, ;
j
) is performed as follows:
Let {r
i,k
, r
m,k
}
n
k=1
be a random sample from the bivariate distribution
F(r
i
, r
m
) = C(F
i
(r
i
), F
m
(r
m
))
between an individual stock return r
i
and the market return r
m
, where n denotes the number
43
of daily return observations in a period. We estimate the marginal distributions F
i
and F
m
of an individual stock return r
i
and the market return r
m
non-parametrically by their scaled
empirical distribution functions

F
i
(x) =
1
n + 1
n

k=1
1
r
i,k
x
and

F
m
(x) =
1
n + 1
n

k=1
1
r
m,k
x
. (8)
The estimation of F
i
and F
m
by their empirical counterparts avoids an incorrect speci-
cation of the marginal distributions. We then have to estimate the set of copula parameters

j
. Since we assume a parametric form of the copula functions, the parameters
j
can be
estimated via the maximum likelihood estimator

j
= argmax

j
L
j
(
j
) with L
j
(
j
) =
n

k=1
log(c
j
(

F
i,r
i,k
,

F
m,r
m,k
;
j
)), (9)
where L
j
(
j
) denotes the log-likelihood function and c
j
(, ;
j
) the copula densitiy. Assum-
ing that {r
i,k
, r
m,k
}
n
k=1
is an i.i.d. random sample,

is a consistent and asymptotic normal
estimate of the set of copula parameters under standard regularity conditions (e.g., Genest,
Ghoudi, and Rivest (1995)).
52
A.2 How to Select the Right Copula
So far we have pointed out an estimation procedure under the assumption that the copula
C
j
(, ;
j
) is known up to a set of parameters
j
. The choice of the copula C

(, ;

) obvi-
ously aects the resulting bivariate distribution and the resulting tail dependence coecients
52
Obviously, daily return data often violate the assumption of an i.i.d. random sample. An alternative
approach to the problem of non-i.i.d. data due to serial correlation in the rst and the second moment of
the time series would be to specify an ARMA-GARCH model for the univariate processes and analyze the
dependence structure of the residuals. We decide not to lter our data, since ltering will also change the
datas dependence structure.
44
LTD and UTD. However, most applications presented in the literature do not discuss this
issue and rely on an arbitrary choice of the copula. To avoid this problem, we follow Ane
and Kharoubi (2003) and use the empirical copula function introduced by Deheuvels (1981)
to evaluate the t of dierent parametric copula families. We proceed as follows:
Let {R
i,k
, R
m,k
}
n
k=1
denote the rank statistic of {r
i,k
, r
m,k
}
n
k=1
, i.e., the smallest individual
(market) return observation of r
i,k
(r
m,k
) has rank R
i,k
= 1 (R
m,k
= 1), the second smallest
individual (market) return observation of r
i,k
(r
m,k
) has rank R
i,k
= 2 (R
m,k
= 2),..., and the
largest individual (market) return observation of r
i,k
(r
m,k
) has rank R
i,k
= n (R
m,k
= n).
Deheuvels (1981) introduces the empirical copula

C
(n)
on the lattice
L =
__
t
i
n
,
t
m
n
_
, t
i
= 0, 1, . . . , n, t
m
= 0, 1, . . . , n
_
by the following equation:

C
(n)
_
t
i
n
,
t
m
n
_
=
1
n
n

k=1
1
R
i,k
t
i
1
R
m,k
tm
. (10)
We compute Integrated Anderson-Darling distances D
j,IAD
between the parametric copulas
C
j
(, ;

j
) and the empirical copula

C
(n)
via
D
j,IAD
=
n

t
i
=1
n

tm=1
_

C
(n)
_
t
i
n
,
tm
n
_
C
j
_
t
i
n
,
tm
n
;

j
__
2
C
j
_
t
i
n
,
tm
n
;

j
_

_
1 C
j
_
t
i
n
,
tm
n
;

j
__. (11)
Hence, we calculate the distance between the predicted value of the parametric copulas
C
j
(, ;

j
) and the empirical copula

C
(n)
for every grid point on the lattice L. The estimation
of the tail dependence coecients LTD and UTD is based on the estimated parameters

of the copula C

(, ;

) that minimizes D
j,IAD
. In unreported robustness checks, we also
45
apply the Kolmogorov-Smirnov distances D
j,KS
between the parametric copulas C
j
(, ;

j
)
and the empirical copula

C
(n)
, i.e.,
D
j,KS
= max
1t
i
,tmn
|

C
(n)
_
t
i
n
,
t
m
n
_
C
j
_
t
i
n
,
t
m
n
;

j
_
|, (12)
as well as log-likelihood values to select the appropriate dependence structure. Independent
of the selected evaluation measure (D
j,IAD
, D
j,KS
, or log-likelihood values), we obtain very
similar results for the selected parametric copula.
46
T
a
b
l
e
A
.
1
:
B
i
v
a
r
i
a
t
e
C
o
p
u
l
a
F
u
n
c
t
i
o
n
s
w
i
t
h
T
a
i
l
D
e
p
e
n
d
e
n
c
e
C
o
e

c
i
e
n
t
s
C
o
p
u
l
a
P
a
r
a
m
e
t
r
i
c
F
o
r
m
L
T
D
U
T
D
C
l
a
y
t
o
n
(
1
)
C
C
l
a
(
u
1
,
u
2
;

)
=
(
u

1
+
u

1
)

1
/

1
/

R
o
t
a
t
e
d
-
G
u
m
b
e
l
(
2
)
C
R
G
u
m
(
u
1
,
u
2
)
=
u
1
+
u
2

1
+
e
x
p

(
(

l
o
g
(
u
1
)
)

+
(

l
o
g
(
u
2
)
)

)
1
/

2
1
/

R
o
t
a
t
e
d
-
J
o
e
(
3
)
C
R
J
o
e
(
u
1
,
u
2
)
=
u
1
+
u
2

(
u
1
+
u
2

u
1

u
2
)
1
/

2
1
/

R
o
t
a
t
e
d
-
G
a
l
a
m
b
o
s
(
4
)
C
R
G
a
l
(
u
1
,
u
2
)
=
u
1
+
u
2

1
+
(
u
1
)

(
u
2
)

e
x
p

(
(

l
o
g
(
u
1
)
)

+
(

l
o
g
(
u
2
)
)

1
/

1
/

G
a
u
s
s
(
A
)
C
G
a
u
(
u
1
,
u
2
;

)
=

1
(
u
1
)
,

1
(
u
2
)
)

F
r
a
n
k
(
B
)
C
F
r
a
(
u
1
,
u
2
;

)
=

1
l
o
g

e
x
p
(

(
1

e
x
p
(

u
1
)
)
(
1

e
x
p
(

u
2
)
)
1

e
x
p
(

P
l
a
c
k
e
t
t
(
C
)
C
P
l
a
(
u
1
,
u
2
;

)
=
12
(

1
)

1
+
(

1
)
(
u
1
+
u
2
)

[
(
1
+
(

1
)
(
u
1
+
u
2
)
)
2

u
1
u
2
]
1
/
2

F
-
G
-
M
(
D
)
C
F
g
m
(
u
1
,
u
2
;

)
=
u
1
u
2
(
1
+

(
1

u
1
)
(
u
2
)
)

J
o
e
(
I
)
C
J
o
e
(
u
1
,
u
2
;

)
=
1

(
(
u
1
)

+
(
u
2
)

(
u
1
)

(
u
2
)

)
1
/

2
1
/

G
u
m
b
e
l
(
I
I
)
C
G
u
m
(
u
1
,
u
2
;

)
=
e
x
p

(
(

l
o
g
(
u
1
)
)

+
(

l
o
g
(
u
2
)
)

)
1
/

2
1
/

G
a
l
a
m
b
o
s
(
I
I
I
)
C
G
a
l
(
u
1
,
u
2
;

)
=
u
1

u
2

e
x
p

(
(

l
o
g
(
u
1
)
)

+
(

l
o
g
(
u
2
)
)

1
/

1
/

R
o
t
a
t
e
d
-
C
l
a
y
t
o
n
(
I
V
)
C
R
C
l
a
(
u
1
,
u
2
)
=
u
1
+
u
2

1
+
(
(
u
1
)

+
(
u
2
)

1
)

1
/

1
/

T
h
i
s
t
a
b
l
e
r
e
p
o
r
t
s
t
h
e
p
a
r
a
m
e
t
r
i
c
f
o
r
m
s
o
f
t
h
e
b
i
v
a
r
i
a
t
e
c
o
p
u
l
a
f
u
n
c
t
i
o
n
s
c
o
n
s
i
d
e
r
e
d
i
n
t
h
i
s
s
t
u
d
y
i
n
t
h
e
s
e
c
o
n
d
c
o
l
u
m
n
a
n
d
t
h
e
c
o
r
r
e
s
p
o
n
d
i
n
g
l
o
w
e
r
a
n
d
u
p
p
e
r
t
a
i
l
d
e
p
e
n
d
e
n
c
e
c
o
e

c
i
e
n
t
s
,
L
T
D
a
n
d
U
T
D
,
i
n
t
h
e
l
a
s
t
t
w
o
c
o
l
u
m
n
s
.
T
h
e
C
l
a
y
t
o
n
-
,
t
h
e
R
o
t
a
t
e
d
J
o
e
-
,
t
h
e
R
o
t
a
t
e
d
G
u
m
b
e
l
-
,
a
n
d
t
h
e
R
o
t
a
t
e
d
G
a
l
a
m
b
o
s
-
c
o
p
u
l
a
e
x
h
i
b
i
t
l
o
w
e
r
t
a
i
l
d
e
p
e
n
d
e
n
c
e
.
T
h
e
G
a
u
s
s
-
,
t
h
e
F
r
a
n
k
-
,
t
h
e
P
l
a
c
k
e
t
t
-
,
a
n
d
t
h
e
F
G
M
-
c
o
p
u
l
a
a
r
e
a
s
y
m
p
t
o
t
i
c
a
l
l
y
i
n
d
e
p
e
n
d
e
n
t
i
n
b
o
t
h
t
a
i
l
s
.
T
h
e
J
o
e
-
,
t
h
e
G
u
m
b
e
l
-
,
t
h
e
G
a
l
a
m
b
o
s
-
,
a
n
d
t
h
e
R
o
t
a
t
e
d
C
l
a
y
t
o
n
-
c
o
p
u
l
a
e
x
h
i
b
i
t
u
p
p
e
r
t
a
i
l
d
e
p
e
n
d
e
n
c
e
.
I
n
b
r
a
c
k
e
t
s
w
e
a
s
s
i
g
n
a
l
a
b
e
l
t
o
e
a
c
h
b
a
s
i
c
c
o
p
u
l
a
.
W
e
d
e

n
e
u
1
=
1

u
1
a
n
d
u
2
=
1

u
2
.

d
e
n
o
t
e
s
t
h
e
s
t
a
n
d
a
r
d
n
o
r
m
a
l
N
(
0
,
1
)
d
i
s
t
r
i
b
u
t
i
o
n
f
u
n
c
t
i
o
n
,

1
t
h
e
f
u
n
c
t
i
o
n
a
l
i
n
v
e
r
s
e
o
f

a
n
d

i
s
t
h
e
b
i
v
a
r
i
a
t
e
s
t
a
n
d
a
r
d
n
o
r
m
a
l
d
i
s
t
r
i
b
u
t
i
o
n
f
u
n
c
t
i
o
n
w
i
t
h
c
o
r
r
e
l
a
t
i
o
n

.
47
T
a
b
l
e
A
.
2
:
F
r
e
q
u
e
n
c
y
a
n
d
R
e
l
a
t
i
v
e
P
e
r
c
e
n
t
a
g
e
o
f
C
o
p
u
l
a
S
e
l
e
c
t
i
o
n
C
o
p
u
l
a
F
r
e
q
P
e
r
c
C
o
p
u
l
a
F
r
e
q
P
e
r
c
C
o
p
u
l
a
F
r
e
q
P
e
r
c
C
o
p
u
l
a
F
r
e
q
P
e
r
c
(
1
-
A
-
I
)
1
,
4
4
8
1
.
5
0
(
2
-
A
-
I
)
1
,
5
5
7
1
.
6
1
(
3
-
A
-
I
)
1
,
7
5
6
1
.
8
1
(
4
-
A
-
I
)
1
,
3
3
7
1
.
3
8
(
1
-
A
-
I
I
)
1
,
3
2
3
1
.
3
7
(
2
-
A
-
I
I
)
9
4
9
0
.
9
8
(
3
-
A
-
I
I
)
1
,
2
9
9
1
.
3
4
(
4
-
A
-
I
I
)
8
8
8
0
.
9
2
(
1
-
A
-
I
I
I
)
1
,
5
4
4
1
.
6
0
(
2
-
A
-
I
I
I
)
1
,
1
5
0
1
.
1
9
(
3
-
A
-
I
I
I
)
1
,
5
6
6
1
.
6
2
(
4
-
A
-
I
I
I
)
8
9
0
0
.
9
2
(
1
-
A
-
I
V
)
1
,
5
9
6
1
.
6
5
(
2
-
A
-
I
V
)
1
,
8
4
3
1
.
9
0
(
3
-
A
-
I
V
)
1
,
6
7
9
1
.
7
4
(
4
-
A
-
I
V
)
1
,
2
9
4
1
.
3
4
(
1
-
B
-
I
)
1
,
5
8
6
1
.
6
4
(
2
-
B
-
I
)
1
,
0
4
6
1
.
0
8
(
3
-
B
-
I
)
1
,
5
0
9
1
.
5
6
(
4
-
B
-
I
)
1
,
2
6
5
1
.
3
1
(
1
-
B
-
I
I
)
1
,
1
6
3
1
.
2
0
(
2
-
B
-
I
I
)
5
2
4
0
.
5
4
(
3
-
B
-
I
I
)
9
4
0
0
.
9
7
(
4
-
B
-
I
I
)
6
6
1
0
.
6
8
(
1
-
B
-
I
I
I
)
1
,
6
3
1
1
.
6
9
(
2
-
B
-
I
I
I
)
1
,
0
1
9
1
.
0
5
(
3
-
B
-
I
I
I
)
1
,
3
6
0
1
.
4
1
(
4
-
B
-
I
I
I
)
8
0
8
0
.
8
3
(
1
-
B
-
I
V
)
1
,
4
8
6
1
.
5
4
(
2
-
B
-
I
V
)
1
,
4
0
7
1
.
4
5
(
3
-
B
-
I
V
)
1
,
2
8
7
1
.
3
3
(
4
-
B
-
I
V
)
1
,
2
0
6
1
.
2
5
(
1
-
C
-
I
)
1
,
8
0
0
1
.
8
6
(
2
-
C
-
I
)
1
,
5
5
5
1
.
6
1
(
3
-
C
-
I
)
2
,
3
0
9
2
.
3
9
(
4
-
C
-
I
)
1
,
8
0
2
1
.
8
6
(
1
-
C
-
I
I
)
1
,
4
0
0
1
.
4
5
(
2
-
C
-
I
I
)
8
0
3
0
.
8
3
(
3
-
C
-
I
I
)
1
,
1
8
4
1
.
2
2
(
4
-
C
-
I
I
)
1
,
0
1
9
1
.
0
5
(
1
-
C
-
I
I
I
)
1
,
6
1
8
1
.
6
7
(
2
-
C
-
I
I
I
)
1
,
2
2
1
1
.
2
6
(
3
-
C
-
I
I
I
)
1
,
5
5
8
1
.
6
1
(
4
-
C
-
I
I
I
)
8
3
1
0
.
8
6
(
1
-
C
-
I
V
)
1
,
3
5
9
1
.
4
0
(
2
-
C
-
I
V
)
1
,
5
6
5
1
.
6
2
(
3
-
C
-
I
V
)
1
,
3
6
1
1
.
4
1
(
4
-
C
-
I
V
)
1
,
3
0
0
1
.
3
4
(
1
-
D
-
I
)
2
,
4
5
2
2
.
5
3
(
2
-
D
-
I
)
2
,
5
6
9
2
.
6
5
(
3
-
D
-
I
)
3
,
1
8
0
3
.
2
9
(
4
-
D
-
I
)
2
,
9
8
5
3
.
0
8
(
1
-
D
-
I
I
)
1
,
6
6
6
1
.
7
2
(
2
-
D
-
I
I
)
1
,
1
4
4
1
.
1
8
(
3
-
D
-
I
I
)
1
,
4
9
8
1
.
5
5
(
4
-
D
-
I
I
)
1
,
2
8
1
1
.
3
2
(
1
-
D
-
I
I
I
)
2
,
0
9
2
2
.
1
6
(
2
-
D
-
I
I
I
)
1
,
4
6
8
1
.
5
2
(
3
-
D
-
I
I
I
)
2
,
1
7
3
2
.
2
5
(
4
-
D
-
I
I
I
)
1
,
6
3
3
1
.
6
9
(
1
-
D
-
I
V
)
2
,
1
3
0
2
.
2
0
(
2
-
D
-
I
V
)
2
,
4
7
7
2
.
5
6
(
3
-
D
-
I
V
)
2
,
0
2
0
2
.
0
9
(
4
-
D
-
I
V
)
2
,
2
9
7
2
.
3
7
T
h
i
s
t
a
b
l
e
r
e
p
o
r
t
s
t
h
e
a
b
s
o
l
u
t
e
a
n
d
p
e
r
c
e
n
t
a
g
e
f
r
e
q
u
e
n
c
y
o
f
t
h
e
s
e
l
e
c
t
e
d
p
a
r
a
m
e
t
r
i
c
c
o
p
u
l
a
c
o
m
b
i
n
a
t
i
o
n
s
.
T
h
e
a
p
p
r
o
p
r
i
a
t
e
d
e
p
e
n
d
e
n
c
e
s
t
r
u
c
t
u
r
e
i
s
s
e
l
e
c
t
e
d
b
y
m
i
n
i
m
i
z
i
n
g
t
h
e
d
i
s
t
a
n
c
e
b
e
t
w
e
e
n
t
h
e
p
a
r
a
m
e
t
r
i
c
c
o
p
u
l
a
s
a
n
d
t
h
e
e
m
p
i
r
i
c
a
l
c
o
p
u
l
a
v
i
a
t
h
e
I
n
t
e
g
r
a
t
e
d
A
n
d
e
r
s
o
n
-
D
a
r
l
i
n
g
d
i
s
t
a
n
c
e
.
I
n
c
o
l
u
m
n
s
1
,
4
,
7
,
a
n
d
1
0
w
e
i
n
d
i
c
a
t
e
t
h
e
l
a
b
e
l
o
f
t
h
e
r
e
s
p
e
c
t
i
v
e
c
o
p
u
l
a
c
o
m
b
i
n
a
t
i
o
n
b
a
s
e
d
o
n
t
h
e
b
a
s
i
c
c
o
p
u
l
a
l
a
b
e
l
s
f
r
o
m
T
a
b
l
e
A
.
1
.
T
h
e
t
h
r
e
e
c
o
p
u
l
a
s
t
h
a
t
a
r
e
m
o
s
t
o
f
t
e
n
s
e
l
e
c
t
e
d
a
r
e
:
(
3
-
D
-
I
)
-
R
o
t
a
t
e
d
-
J
o
e
/
F
-
G
-
M
/
J
o
e
-
c
o
p
u
l
a
,
(
4
-
D
-
I
)
-
R
o
t
a
t
e
d
-
G
a
l
a
m
b
o
s
/
F
-
G
-
M
/
J
o
e
-
c
o
p
u
l
a
,
a
n
d
(
2
-
D
-
I
)
-
R
o
t
a
t
e
d
-
G
u
m
b
e
l
/
F
-
G
-
M
/
J
o
e
-
c
o
p
u
l
a
(
t
h
e
s
e
c
o
p
u
l
a
c
o
m
b
i
n
a
t
i
o
n
s
a
r
e
m
a
r
k
e
d
i
n
b
o
l
d
)
.
48
B Appendix: Brief Denitions and Data Sources of
Main Variables
This table briey denes the main variables used in the empirical analysis. The data sources
are: (i) CRSP: CRSP Stocks Database, (ii) KF: Kenneth French Data Library, (iii) CS:
Compustat. EST indicates that the variable is estimated or computed based on original
variables from the respective data sources.
Panel A: Return-Based Variables
Variable Name Description Source
Return (return) Raw excess return of a portfolio (stock) over the riskfree rate. As
riskfree rate the 1-month T-Bill rate is used.
CRSP,
KF, EST
CAPM-Alpha,
FF-Alpha,
CAR-Alpha
CAPM, Fama and French (1993) three-factor, and Carhart (1997)
performance Alpha of a portfolio over the sample period. We use
monthly portfolio returns to estimate the alphas.
CRSP,
KF, EST
capm-alpha,
-alpha,
car-alpha
CAPM, Fama and French (1993) three-factor, and Carhart (1997)
performance Alpha of an individual stock per year. We use daily
return data to estimate alphas for each stock and year. We require
at least 100 daily return observations.
CRSP,
KF, EST
LTD Lower tail dependence coecient of a stock. Estimated based on
daily data from one year as detailed in Appendix A.
CRSP,
EST
UTD Upper tail dependence coecient of a stock. Estimated based on
daily data from one year as detailed in Appendix A.
CRSP,
EST
Factor loading on the market factor from a CAPM one-factor regres-
sion estimated based on daily data from one year: =
COV(r
i
,rm)
VAR(rm)
.
CRSP,
EST

Downside beta estimated based on daily return data from one year
as dened in Ang, Chen, and Xing (2006):
CRSP,
EST

=
COV(r
i
,rm|rm<m)
VAR(rm|rm<m)
, where
m
is the mean of the daily market
return.

+
Upside beta estimated based on daily return data from one year as
dened in Ang, Chen, and Xing (2006):
+
=
COV(r
i
,rm|rm>m)
VAR(rm|rm>m)
.
CRSP,
EST
49
Variable Name Description Source
illiq The Amihud (2002) illiquidity ratio dened as: illiq
i,t
=
1
Days
i
t

Days
d=1
|r
i,d
|
V ol
i,d
, where V ol
i,d
is security is trading volume in dollars on
day d and Days
i
t
is the number of trading days in year t.
CRSP,
EST
idiovola A stocks idiosyncratic volatility, dened as the standard deviation
of the CAPM-residuals of its daily returns.
CRSP,
EST
coskew The co-skewness of a stocks daily returns with the market:
coskew =
E[(r
i

i
)(rmm)
2
]

VAR(r
i
)VAR(rm)
.
CRSP,
EST
cokurt The co-kurtosis of a stocks daily returns with the market:
cokurt =
E[(r
i

i
)(rmm)
3
]

VAR(r
i
)VAR(rm)
3/2
.
CRSP,
EST
max The maximum daily return over the last year or month, respectively. CRSP
Panel B: Other Firm Characteristics
Variable Name Description Source
size The natural logarithm of a rms equity market capitalization in
million USD.
CS
bookmarket A rms book-to-market ratio computed as the ratio of CS book value
of equity per share (i.e., book value of common equity less liquidation
value (CEQL) divided by common share outstanding (CSHO)) to
share price (i.e., market value of equity per share).
CS
50
References
Abdellaoui, M., 2000. Paramater-Free Elicitation of Utility and Probability Weighting Func-
tions. Management Science 46, 1497-1512.
Acharya, V., Pedersen, L., 2005. Asset pricing with liquidity risk. Journal of Financial Eco-
nomics 77, 375-410.
At-Sahalia, Y., Lo, A.W., 2000. Nonparametric risk management and implied risk aversion.
Journal of Econometrics 94, 9-51.
Amihud, Y., 2002. Illiquidity and stock returns: cross-section and time-series eects. Journal
of Financial Markets 5, 31-56.
Ane, T., Kharoubi, C., 2003. Dependence structure and risk measure. Journal of Business
76, 411-438.
Ang, A., Bekaert, G., Liu, J., 2005. Why stocks may disappoint. Journal of Financial Eco-
nomics 76, 471-508.
Ang, A., Chen, J., 2007. CAPM over the long-run: 1926-2001. Journal of Empirical Finance
14, 1-40.
Ang, A., Chen, J., Xing, Y., 2006. Downside risk. Review of Financial Studies 19, 1191-1239.
Ang, A., Hodrick, R.J., Xing, Y., Zhang, X., 2006. The cross-section of volatility and ex-
pected returns. Journal of Finance 61, 259-299.
Ang, A., Hodrick, B., Xing, Y., Zhang, X., 2009. High idiosyncratic volatility and low returns:
international and further U.S. evidence, Journal of Financial Economics 91, 1-23.
51
Ang, A., Liu, J., Schwarz, K., 2010. Using stocks or portfolios in tests of factor models.
Unpublished Working Paper, Columbia University and University of Pennsylvania.
Arzac, E.R., Bawa, V.S., 1977. Portfolio choice and equilibrium in capital markets with
safety-rst investors. Journal of Financial Economics 4, 277-288.
Baker, M., Wurgler, J., 2006. Investor sentiment and the cross-section of stock returns.
Journal of Finance 61, 1645-1680.
Bali, T.G., Cakici, N., Whitelaw, R.F., 2011a. Maxing out: stocks as lotteries and the
cross-section of expected returns. Journal of Financial Economics 99, 427-446.
Bali, T.G., Cakici, N., Whitelaw, R.F., 2011b. Hybrid tail risk and expected stock returns:
when does the tail wag the dog? Unpublished Working Paper, Georgetown University,
Fordham University, and New York University.
Bali, T.G., Demirtas, K.O., Levy, H., 2009. Is there an intertemporal relation between
downside risk and expected returns? Journal of Financial and Quantitative Analysis 44,
883-909.
Banz, R., 1981. The relation between return and market value of common stocks. Journal of
Financial Economics 9, 3-18.
Barro, R., 2006. Rare disasters and asset markets in the twentieth century. Quarterly Journal
of Economics 121, 823-866.
Barro, R., 2009. Rare disasters, asset prices, and welfare costs. American Economic Review
99, 243-264.
52
Basu, S., 1983. The relationship between earnings yield, market value, and return for NYSE
common stocks: further evidence. Journal of Financial Economics 12, 129-156.
Barberis, N., Huang, M., 2001. Mental accounting, loss aversion, and individual stock re-
turns. Journal of Finance 56, 1247-1292.
Barberis, N., Huang, M., Santos, T., 2001. Prospect theory and asset prices. Quarterly
Journal of Economics 116, 1-53.
Bates, D.S., 2000. Post 87 crash fears in the S&P 500 futures options market. Journal of
Econometrics 94, 181-238.
Bates, D.S., 2008. The market for crash risk. Journal of Economic Dynamics and Control
32, 2291-2321.
Baumol, W.J., 1963. An expected gain-condence limit criterion for portfolio selection. Man-
agement Science 10, 174-182.
Bawa, V.S., Lindenberg, E.B., 1977. Capital market equilibrium in a mean-lower partial
moment framework. Journal of Financial Economics 5, 189-200.
Benartzi, S., Thaler, R.H., 1995. Myopic loss aversion and the equity premium puzzle. Quar-
terly Journal of Economics 110, 73-92.
Berkman, H., Jacobsen, B., Lee, J.B., 2011. Time-varying rare disaster risk and stock returns.
Journal of Financial Economics 101, 313-332.
Black, F., Jensen, M., Scholes, M., 1973. The capital asset pricing model: some empirical
tests. In: Jensen, M. (Ed.), Studies in the Theory of Capital Markets. Praeger, New York.
53
Bleichrodt, H., Pinto, J.L., 2000. A Parameter-Free Elicitation of the Probability Weighting
Function in Medical Decision Analysis. Management Science 46, 1485-1496.
Bollen, N.P.B., Whaley, R.E., 2004. Does net buying pressure aect the shape of implied
volatility functions? Journal of Finance 59, 711-753.
Bollerslev, T., Todorov, V., 2011. Tails, fears, and risk premia. Journal of Finance 66, 2165-
2211.
Boyer, B., Mitton, T., Vorkink, K., 2010. Expected Idiosyncratic Skewness. Review of Fi-
nancial Studies 23, 169-202.
Broadie, M., Chernov, M., Johannes, M., 2009. Understanding index option returns. Review
of Financial Studies 22, 4493-4529.
Carhart, M., 1997. On persistence in mutual fund performance. Journal of Finance 52, 57-82.
Chen, J., Hong, H., Stein, J., 2001. Forecasting crashes: trading volume, past returns and
conditional skewness in stock prices. Journal of Financial Economics 61, 345-381.
Chen, L., Novy-Marx, R., Zhang, L., 2011. An alternative three-factor model. Unpublished
Working Paper, Cheung Kong Graduate School of Business, University of Rochester, and
Ohio State University.
Cholette, L., Lu, C.-C., 2011. The market premium for dynamic tail risk. Unpublished
Working Paper, University of Stavanger and National Chengchi University.
Cremers, M., Petajisto, A., Zitzewitz, E., 2010. Should benchmark indices have alpha? Re-
visiting performance evaluation. Unpublished Working Paper, University of Notre Dame,
New York University, and Dartmouth College.
54
Daniel, K., Grinblatt, M., Titman, S., Wermers, R., 1997. Measuring mutual fund perfor-
mance with characteristic-based benchmarks. Journal of Finance 52, 1035-1058.
Deheuvels, P., 1981. A non parametric test for independence. Publications de lInstitut de
Statistique de lUniversite de Paris 26, 29-50.
Dittmar, R., 2002. Nonlinear pricing kernels, kurtosis preference, and the cross-section of
equity returns. Journal of Finance 57, 369-403.
Embrechts, P:, McNeil, A., Straumann, D., 2002. Correlation and dependence in risk man-
agement: properties and pitfalls. In: M.A.H. Dempster (Ed.), Risk Management: Value
at Risk and Beyond. Cambridge University Press, Cambridge, 176-223.
Fama, E.F., 1965. The Behavior of stock market prices. Journal of Business 38, 34-105.
Fama, E.F., French, K.R., 1992. The cross-section of expected stock returns. Journal of
Finance 47, 427-465.
Fama, E.F., French, K.R., 1993. Common risk factors in the returns on stocks and bonds.
Journal of Financial Economics 33, 3-56.
Fama, E.F., French, K.R., 1997. Industry costs of equity. Journal of Financial Economics
43, 153-193.
Fama, E.F., French, K.R., 2006. The value premium and the CAPM. Journal of Finance 61,
2163-2185.
Fama, E.F., MacBeth, J.D., 1973. Risk, return, and equilibrium: empirical tests. Journal of
Political Economy 81, 607-636.
55
Fang, H., Lai, T.-Y., 1997. Co-kurtosis and capital asset pricing. Financial Review 32, 293-
307.
Ferson, W.E., Harvey, C.R., 1991. The variation of economic risk premiums. Journal of
Political Economy 99, 385-415.
Fisher, R.A., Tippett, L.H.C., 1928. Limiting forms of the frequency distributions of the
largest or smallest member of a sample. Proceedings of the Cambridge Philosophical So-
ciety 24, 180-190.
Friend, I., Westereld, R., 1980. Co-skewness and capital asset pricing. Journal of Finance
35, 897-913.
Gabaix, X., 2012. Variable rare disasters: an exactly solved framework for ten puzzles in
macro-nance. Quarterly Journal of Economics 127, 645-700.
Garleanu, N.G., Pedersen, L.H., Poteshman, A., 2009. Demand-based option pricing. Review
of Financial Studies 22, 4259-4299.
Genest, C., Ghoudi, K., Rivest, L.P., 1995. A semiparametric estimation procedure of de-
pendence parameters in multivariate families of distributions. Biometrika 82, 543-552.
Gibbons, M.R., 1982. Multivariate tests of nancial models. Journal of Financial Economics
10, 3-27.
Grossman, S.J., Zhou, Z., 1996. Equilibrium analysis of portfolio insurance. Journal of Fi-
nance 51, 1379-1403.
Gul, F., 1992. A theory of disappointment aversion. Econometrica 59, 667-686.
56
Harvey, C.R., Siddique, A., 2000. Conditional skewness in asset pricing tests. Journal of
Finance 55, 1263-1295.
He, X.D., Zhou, X.Y., 2012. Hope, Fear and Aspirations. Unpublished Working Paper,
Columbia University and University of Oxford.
Hill, B., 1975. A simple general approach to inference about the tail of a distribution. Annals
of Statistics 3, 1163-1164.
Jackwerth, J.C., 2000. Recovering risk aversion from option prices and realized returns.
Review of Financial Studies 13, 433-451.
Jackwerth, J.C., Rubinstein, M., 1996. Recovering probability distributions from option
prices. Journal of Finance 51, 1611-1631.
Jagannathan, R., Wang, Z., 1996. The conditional CAPM and the cross-section of expected
returns. Journal of Finance 51, 3-53.
Jegadeesh, N., Titman, S., 1993. Returns to buying winners and selling losers: implications
for stock market eciency. Journal of Finance 48, 65-91.
Joe, H., 1997. Multivariate models and dependence concepts. Chapman & Hall / CRC Mono-
graphs on Statistics & Applied Probability.
Johnson, S. A., Moorman, T., Sorescu, S., 2009. A reexamination of corporate governance
and equity prices. Review of Financial Studies 22, 4753-4786.
Kahneman, D., Tversky, A., 1979. Prospect theory: an analysis of decision under risk.
Econometrica 47, 263-291.
57
Kelly, B., 2012. Tail risk and asset prices. Unpublished Working Paper, University of Chicago.
Kelly, B., Jiang, H., 2012. Tail risk and hedge fund returns. Unpublished Working Paper,
University of Chicago and Erasmus University Rotterdam School of Management.
Kraus, A., Litzenberger, R.H., 1976. Skewness preference and the valuation of risk assets.
Journal of Finance 31, 1085-1100.
Kumar, A., 2009. Who gambles in the stock market? Journal of Finance 64, 1889-1933.
Levy, H., Sarnat, M., 1972. Safety rst: an expected utility principle. Journal of Financial
and Quantitative Analysis 7, 1829-1834.
Lewellen, J., Nagel, S., 2006. The conditional CAPM does not explain asset-pricing anoma-
lies. Journal of Financial Economics 82, 289-314.
Lintner, J., 1965. The valuation of risk assets and the selection of risky investments in stock
portfolios and capital budgets. Review of Economics and Statistics 47, 13-37.
Longin, F., Solnik, B., 2001. Extreme correlation of international equity markets. Journal of
Finance 56, 649-676.
Markowitz, H., 1959. Portfolio selection. Yale University Press, New Haven, CT.
Mandelbrot, B., 1963. The variation in certain speculative prices. Journal of Business 36,
394-419.
McNeil, A.J., Frey, R., Embrechts, P., 2005. Quantitative risk management. Princeton Uni-
versity Press.
Nelsen, R.B., 2006. An introduction to copulas. Springer Series in Statistics, Second Edition.
58
Newey, W.K., West, K.D., 1987. A Simple Positive Semi-Denite, Heteroskedasticity and
Autocorrelation Consistent Covariance Matrix. Econometrica 55, 703-708.
Pastor, L., Stambaugh, R.F., 2003. Liquidity risk and expected returns. Journal of Political
Economy 111, 642-685.
Patton, A.J., 2004. On the Out-of-Sample Importance of Skewness and Asymmetric Depen-
dence for Asset Allocation. Journal of Financial Econometrics, 2, 130-168.
Patton, A.J., 2004. Are Market Neutral Hedge Funds Really Market Neutral? Review of
Financial Studies, 22, 2495-2530.
Polkovnichenko, V., Zhao, F., 2012. Probability weighting functions implied by options
prices. Journal of Financial Economics, forthcoming.
Poon, S.H., Rockinger, M., Tawn, J., 2004. Extreme value dependence in nancial markets:
diagnostics, models, and nancial implications. Review of Financial Studies 17, 581-610.
Rodriguez, J.C., 2007. Measuring Financial Contagion: A Copula Approach. Journal of
Empirical Finance 14, 401-423.
Rosenberg, J.V., Engle, R.F., 2002. Empirical pricing kernels. Journal of Financial Economics
64, 341-372.
Roy, A.D., 1952. Safety rst and the holdings of assets. Econometrica 20, 431-449.
Rubinstein, M., 1994. Implied binomial trees. Journal of Finance 49, 771-813.
Sadka, R., 2006. Momentum and post-earnings-announcement drift anomalies: the role of
liquidity risk. Journal of Financial Economics 80, 309-349.
59
Sibuya, M., 1960. Bivariate extreme statistics. Annals of the Institute of Statistical Mathe-
matics 11, 195-210.
Sharpe, W.F., 1964. Capital asset prices: a theory of market equilibrium under conditions
of risk. Journal of Finance 19, 425-442.
Sklar, A., 1959. Fonctions de repartition `a n dimensions et leurs marges. Publications de
lInstitut de Statistique de lUniversite de Paris 14, 229-231.
Tawn, J., 1988. Bivariate extreme value theory: models and estimation. Biometrika 75,
397-415.
Tversky, A., Fox, C.R, 1995. Weighting Risk and Uncertainty. Psychological Review 102,
269-283.
60
Figure 1: Dierent Copula Dependence Structures
(a) Panel A: Gauss-copula structure (b) Panel B: Gumbel-copula structure
(c) Panel C: Clayton-copula structure (d) Panel D: Student t-copula structure
This gure displays 2, 000 random variates from four bivariate distributions with standard normal marginal
distributions and the Gauss-copula (Panel A), the Gumbel-copula (Panel B), the Clayton-copula (Panel
C), and the Student t-copula (Panel D) determining the dependence structure. In each case, the linear
correlation is set to 0.8.
61
Figure 2: Aggregate Tail Dependence over Time
This gure displays the evolution of aggregate lower tail dependence, LTD, and aggregate upper tail depen-
dence, UTD, over time. Aggregate LTD (UTD) is dened as the yearly cross-sectional, equal-weighted, aver-
age of the individual lower tail dependence coecients, LTD
i,t
(upper tail dependence coecients, UTD
i,t
)
between stock returns and market returns over all stocks i in year t in our sample. The sample covers all
U.S. common stocks traded on the NYSE / AMEX and the sample period is from January 1963 to December
2009.
62
Figure 3: Persistence of LTD
This gure displays the evolution of the average equal-weighted lower tail dependence, LTD, of ve quintile
portfolios. Firms are sorted into quintiles based on their realized LTD between their stock return and the
market return in year t = 1. Then the equal-weighted average of LTD of these portfolios is computed again
for each of the following four years. The sample covers all U.S. common stocks traded on the NYSE / AMEX
and the sample period is from January 1963 to December 2009.
63
T
a
b
l
e
1
:
S
u
m
m
a
r
y
S
t
a
t
i
s
t
i
c
s
M
e
a
n
2
5
%
-
Q
u
a
n
t
i
l
e
M
e
d
i
a
n
7
5
%
-
Q
u
a
n
t
i
l
e
S
t
d
.
D
e
v
.
A
b
o
v
e
L
T
D
M
e
d
i
a
n
B
e
l
o
w
L
T
D
M
e
d
i
a
n
A
b
o
v
e
-
B
e
l
o
w
r
e
t
u
r
n
1
0
.
9
8
%

2
0
.
0
0
%
3
.
7
3
%
3
0
.
1
4
%
0
.
6
0
3
1
3
.
2
8
%
8
.
6
8
%
+
4
.
6
0
%

L
T
D
0
.
1
3
1
0
.
0
2
8
0
.
1
1
2
0
.
2
0
3
0
.
1
1
4
0
.
2
2
3
0
.
0
3
7
+
0
.
1
8
6

U
T
D
0
.
0
8
3
0
.
0
0
0
0
.
0
5
4
0
.
1
3
9
0
.
0
9
4
0
.
0
9
4
0
.
0
7
1
+
0
.
0
2
3

0
.
8
4
4
0
.
4
3
3
0
.
7
7
0
1
.
1
6
8
0
.
5
5
9
1
.
0
2
4
0
.
6
6
3
+
0
.
3
6
1

0
.
9
7
5
0
.
4
7
6
0
.
8
8
4
1
.
3
7
0
0
.
7
3
2
1
.
2
9
0
0
.
6
6
1
+
0
.
6
2
9

+
0
.
7
5
7
0
.
2
4
9
0
.
6
8
5
1
.
1
8
4
0
.
7
7
2
0
.
8
7
9
0
.
6
3
4
+
0
.
2
4
5

s
i
z
e
1
2
.
0
4
1
0
.
3
9
1
1
.
9
2
1
3
.
5
6
2
.
1
1
9
1
2
.
5
1
1
1
.
5
5
+
0
.
9
6

b
o
o
k
m
a
r
k
e
t
0
.
9
0
4
0
.
4
1
9
0
.
7
1
6
1
.
1
6
3
0
.
7
2
8
0
.
8
4
3
0
.
9
6
6

0
.
1
2
3

i
l
l
i
q
0
.
1
6
4
0
.
0
1
5
0
.
1
7
8
0
.
2
8
4
0
.
1
3
3
0
.
1
4
5
0
.
1
8
9

0
.
0
4
4

i
d
i
o
v
o
l
a
0
.
0
2
4
0
.
0
1
5
0
.
0
2
2
0
.
0
3
1
0
.
0
1
2
0
.
0
2
4
0
.
0
2
5

0
.
0
0
1

c
o
s
k
e
w

0
.
0
7
8

0
.
1
7
0

0
.
0
4
7
0
.
0
7
2
0
.
3
0
7

0
.
1
6
5
0
.
0
0
9

0
.
1
7
4

c
o
k
u
r
t
1
.
8
1
1
0
.
6
1
6
1
.
2
0
6
2
.
0
4
3
2
.
6
7
3
2
.
6
2
5
0
.
9
9
6
+
1
.
6
2
9

m
a
x
1
1
.
6
0
%
6
.
4
4
%
9
.
6
9
%
1
4
.
6
2
%
0
.
0
7
3
1
1
.
4
3
%
1
1
.
8
0
%

0
.
3
7
%

T
h
i
s
t
a
b
l
e
d
i
s
p
l
a
y
s
s
u
m
m
a
r
y
s
t
a
t
i
s
t
i
c
s
f
o
r
t
h
e
m
a
i
n
v
a
r
i
a
b
l
e
s
u
s
e
d
i
n
t
h
i
s
s
t
u
d
y
.
T
h
e

r
s
t

v
e
c
o
l
u
m
n
s
s
h
o
w
t
h
e
m
e
a
n
,
t
h
e
2
5
%
-
q
u
a
n
t
i
l
e
,
t
h
e
5
0
%
-
q
u
a
n
t
i
l
e
(
m
e
d
i
a
n
)
,
t
h
e
7
5
%
-
q
u
a
n
t
i
l
e
,
a
n
d
t
h
e
s
t
a
n
d
a
r
d
d
e
v
i
a
t
i
o
n
o
f
e
a
c
h
v
a
r
i
a
b
l
e
.
T
h
e
l
a
s
t
t
h
r
e
e
c
o
l
u
m
n
s
d
i
s
p
l
a
y
m
e
a
n
v
a
l
u
e
s
o
f
t
h
e
v
a
r
i
a
b
l
e
s
c
o
n
d
i
t
i
o
n
a
l
o
n
L
T
D
b
e
i
n
g
a
b
o
v
e
(
b
e
l
o
w
)
i
t
s
5
0
%
-
q
u
a
n
t
i
l
e
a
s
w
e
l
l
a
s
t
h
e
d
i

e
r
e
n
c
e
w
i
t
h
c
o
r
r
e
s
p
o
n
d
i
n
g
s
t
a
t
i
s
t
i
c
a
l
s
i
g
n
i

c
a
n
c
e
l
e
v
e
l
.
W
e
p
r
e
s
e
n
t
r
e
s
u
l
t
s
f
o
r
y
e
a
r
l
y
e
x
c
e
s
s
s
t
o
c
k
r
e
t
u
r
n
s
o
v
e
r
t
h
e
1
-
m
o
n
t
h
T
-
B
i
l
l
r
a
t
e
(
r
e
t
u
r
n
)
,
l
o
w
e
r
t
a
i
l
d
e
p
e
n
d
e
n
c
e
(
L
T
D
)
,
u
p
p
e
r
t
a
i
l
d
e
p
e
n
d
e
n
c
e
(
U
T
D
)
,
b
e
t
a
(

)
,
d
o
w
n
s
i
d
e
b
e
t
a
(

)
,
u
p
s
i
d
e
b
e
t
a
(

+
)
,
t
h
e
l
o
g
o
f
m
a
r
k
e
t
c
a
p
i
t
a
l
i
z
a
t
i
o
n
(
s
i
z
e
)
,
b
o
o
k
-
t
o
-
m
a
r
k
e
t
v
a
l
u
e
(
b
o
o
k
m
a
r
k
e
t
)
,
i
l
l
i
q
u
i
d
i
t
y
(
i
l
l
i
q
)
,
i
d
i
o
s
y
n
c
r
a
t
i
c
v
o
l
a
t
i
l
i
t
y
(
i
d
i
o
v
o
l
a
)
,
c
o
s
k
e
w
n
e
s
s
(
c
o
s
k
e
w
)
,
c
o
k
u
r
t
o
s
i
s
(
c
o
k
u
r
t
)
,
a
n
d
t
h
e
m
a
x
i
m
u
m
d
a
i
l
y
r
e
t
u
r
n
o
v
e
r
t
h
e
p
a
s
t
o
n
e
y
e
a
r
(
m
a
x
)
.
A
d
e
t
a
i
l
e
d
d
e
s
c
r
i
p
t
i
o
n
o
f
t
h
e
c
o
m
p
u
t
a
t
i
o
n
o
f
t
h
e
s
e
v
a
r
i
a
b
l
e
s
i
s
g
i
v
e
n
i
n
t
h
e
m
a
i
n
t
e
x
t
a
n
d
i
n
A
p
p
e
n
d
i
x
B
.
T
h
e
s
a
m
p
l
e
c
o
v
e
r
s
a
l
l
U
.
S
.
c
o
m
m
o
n
s
t
o
c
k
s
t
r
a
d
e
d
o
n
t
h
e
N
Y
S
E
/
A
M
E
X
a
n
d
t
h
e
s
a
m
p
l
e
p
e
r
i
o
d
i
s
f
r
o
m
J
a
n
u
a
r
y
1
9
6
3
t
o
D
e
c
e
m
b
e
r
2
0
0
9
.
t
-
s
t
a
t
i
s
t
i
c
s
a
r
e
i
n
p
a
r
e
n
t
h
e
s
e
s
.

,
a
n
d

i
n
d
i
c
a
t
e
s
i
g
n
i

c
a
n
c
e
a
t
t
h
e
o
n
e
,

v
e
,
a
n
d
t
e
n
p
e
r
c
e
n
t
l
e
v
e
l
,
r
e
s
p
e
c
t
i
v
e
l
y
.
64
T
a
b
l
e
2
:
C
o
r
r
e
l
a
t
i
o
n
s
L
T
D
U
T
D

+
s
i
z
e
b
o
o
k
m
a
r
k
e
t
i
l
l
i
q
p
a
s
t
r
e
t
u
r
n
i
d
i
o
v
o
l
a
c
o
s
k
e
w
c
o
k
u
r
t
m
a
x
L
T
D
1
.
0
0
-
-
-
-
-
-
-
-
-
-
-
-
U
T
D
0
.
1
2
1
.
0
0
-
-
-
-
-
-
-
-
-
-
-

0
.
3
8
0
.
3
1
1
.
0
0
-
-
-
-
-
-
-
-
-
-

0
.
4
9
0
.
0
7
0
.
7
7
1
.
0
0
-
-
-
-
-
-
-
-
-

+
0
.
1
9
0
.
4
8
0
.
7
8
0
.
4
7
1
.
0
0
-
-
-
-
-
-
-
-
s
i
z
e
0
.
2
9
0
.
3
0
0
.
0
5
0
.
0
3
0
.
1
6
1
.
0
0
-
-
-
-
-
-
-
b
o
o
k
m
a
r
k
e
t

0
.
0
9

0
.
1
1

0
.
1
1

0
.
0
7

0
.
0
8

0
.
3
4
1
.
0
0
-
-
-
-
-
-
i
l
l
i
q

0
.
2
8

0
.
2
8

0
.
2
2

0
.
0
8

0
.
1
9

0
.
8
4
0
.
3
0
1
.
0
0
-
-
-
-
-
p
a
s
t
r
e
t
u
r
n
0
.
0
8

0
.
0
5
0
.
1
0
0
.
1
2
0
.
0
5
0
.
0
7
0
.
1
8

0
.
0
1
1
.
0
0
-
-
-
-
i
d
i
o
v
o
l
a

0
.
0
9

0
.
0
9
0
.
2
3
0
.
2
7
0
.
1
2

0
.
4
2
0
.
0
5
0
.
3
1

0
.
1
3
1
.
0
0
-
-
-
c
o
s
k
e
w

0
.
3
7
0
.
2
3
0
.
0
7

0
.
1
2
0
.
2
4

0
.
0
5
0
.
0
1
0
.
0
7

0
.
0
9
0
.
0
3
1
.
0
0
-
-
c
o
k
u
r
t
0
.
3
8
0
.
2
3
0
.
2
1
0
.
1
6
0
.
2
0
0
.
2
1

0
.
0
9

0
.
2
1

0
.
0
0

0
.
0
6

0
.
6
9
1
.
0
0
-
m
a
x

0
.
0
8

0
.
1
0
0
.
1
4
0
.
1
8
0
.
0
7

0
.
3
9
0
.
1
4
0
.
3
1
0
.
0
4
0
.
5
9
0
.
0
3

0
.
0
8
1
.
0
0
T
h
i
s
t
a
b
l
e
d
i
s
p
l
a
y
s
l
i
n
e
a
r
c
o
r
r
e
l
a
t
i
o
n
s
b
e
t
w
e
e
n
t
h
e
i
n
d
e
p
e
n
d
e
n
t
v
a
r
i
a
b
l
e
s
u
s
e
d
i
n
t
h
i
s
s
t
u
d
y
:
l
o
w
e
r
t
a
i
l
d
e
p
e
n
d
e
n
c
e
(
L
T
D
)
,
u
p
p
e
r
t
a
i
l
d
e
p
e
n
d
e
n
c
e
(
U
T
D
)
,
b
e
t
a
(

)
,
d
o
w
n
s
i
d
e
b
e
t
a
(

)
,
u
p
s
i
d
e
b
e
t
a
(

+
)
,
t
h
e
l
o
g
o
f
m
a
r
k
e
t
c
a
p
i
t
a
l
i
z
a
t
i
o
n
(
s
i
z
e
)
,
b
o
o
k
-
t
o
-
m
a
r
k
e
t
v
a
l
u
e
(
b
o
o
k
m
a
r
k
e
t
)
,
i
l
l
i
q
u
i
d
i
t
y
(
i
l
l
i
q
)
,
p
a
s
t
r
e
t
u
r
n
,
i
d
i
o
s
y
n
c
r
a
t
i
c
v
o
l
a
t
i
l
i
t
y
(
i
d
i
o
v
o
l
a
)
,
c
o
s
k
e
w
n
e
s
s
(
c
o
s
k
e
w
)
,
c
o
k
u
r
t
o
s
i
s
(
c
o
k
u
r
t
)
,
a
n
d
t
h
e
m
a
x
i
m
u
m
d
a
i
l
y
r
e
t
u
r
n
o
v
e
r
t
h
e
p
a
s
t
o
n
e
y
e
a
r
(
m
a
x
)
.
A
d
e
t
a
i
l
e
d
d
e
s
c
r
i
p
t
i
o
n
o
f
t
h
e
c
o
m
p
u
t
a
t
i
o
n
o
f
t
h
e
s
e
v
a
r
i
a
b
l
e
s
i
s
g
i
v
e
n
i
n
t
h
e
m
a
i
n
t
e
x
t
a
n
d
i
n
A
p
p
e
n
d
i
x
B
.
T
h
e
s
a
m
p
l
e
c
o
v
e
r
s
a
l
l
U
.
S
.
c
o
m
m
o
n
s
t
o
c
k
s
t
r
a
d
e
d
o
n
t
h
e
N
Y
S
E
/
A
M
E
X
a
n
d
t
h
e
s
a
m
p
l
e
p
e
r
i
o
d
i
s
f
r
o
m
J
a
n
u
a
r
y
1
9
6
3
t
o
D
e
c
e
m
b
e
r
2
0
0
9
.
65
Table 3: Univariate Equal-weighted Portfolio Sorts: Tail Dependence, Returns and Alphas
Portfolio LTD Return CAPM-Alpha FF-Alpha CAR-Alpha
1 Weak LTD 0.01 +3.99% 1.28% 6.27%

3.30%

2 0.06 +8.84% +2.82% 1.86%

0.04%
3 0.12 +10.39% +4.06%

0.07% +1.15%
4 0.18 +14.07% +7.12%

+3.24%

+4.85%

5 Strong LTD 0.29 +19.70% +12.25%

+9.92%

+10.76%

Strong - Weak 0.28

15.71%

13.53%

16.19%

14.06%

(8.70) (7.09) (5.83) (4.77)

This table reports results from univariate portfolio sorts based on realized LTD. In each year, we rank stocks
into quintiles (1-5) and form equal-weighted portfolios at the beginning of each annual period. The column
labeled Return reports the average annual return in excess of the one-month T-bill rate of the portfolios.
The column labeled CAPM-Alpha (FF-Alpha, CAR-Alpha) reports the yearly alpha with regard to
Sharpe (1964)s capital asset pricing model (Fama and French (1993)s three factor model, Carhart (1997)s
four factor model). The row labeled Strong - Weak reports the dierence between the returns of portfolio 5
and portfolio 1 with corresponding statistical signicance levels. The sample covers all U.S. common stocks
traded on the NYSE / AMEX and the sample period is from January 1963 to December 2009. t-statistics
are in parentheses.

,

, and

indicate signicance at the one, ve, and ten percent level, respectively.
66
T
a
b
l
e
4
:
E
x
c
e
s
s
R
e
t
u
r
n
s
o
f
L
T
D
-
P
o
r
t
f
o
l
i
o
s
d
u
r
i
n
g
F
i
n
a
n
c
i
a
l
C
r
i
s
e
s
P
o
r
t
f
o
l
i
o
B
l
a
c
k
M
o
n
d
a
y
A
s
i
a
C
r
i
s
i
s
D
o
t
-
C
o
m
B
u
b
b
l
e
B
u
r
s
t
L
e
h
m
a
n
C
r
i
s
i
s
r
m
,
d
<

5
%
1
W
e
a
k
L
T
D

9
.
5
%

2
.
4
%

1
.
7
%

5
.
9
%

4
.
4
4
%
2

1
3
.
3
%

4
.
4
%

3
.
1
%

6
.
9
%

5
.
9
7
%
3

1
5
.
7
%

5
.
7
%

4
.
3
%

9
.
4
%

7
.
2
8
%
4

1
6
.
3
%

6
.
3
%

5
.
9
%

1
1
.
2
%

8
.
3
9
%
5
S
t
r
o
n
g
L
T
D

1
8
.
7
%

6
.
8
%

7
.
3
%

1
1
.
8
%

9
.
1
8
%
S
t
r
o
n
g
-
W
e
a
k

9
.
2
%

4
.
4
%

5
.
6
%

5
.
9
%

4
.
7
4
%

1
2
.
1
9
)
T
h
i
s
t
a
b
l
e
r
e
p
o
r
t
s
e
q
u
a
l
-
w
e
i
g
h
t
e
d
d
a
i
l
y
e
x
c
e
s
s
r
e
t
u
r
n
s
r
e
t
u
r
n
s
o
f
s
t
o
c
k
s
s
o
r
t
e
d
b
y
r
e
a
l
i
z
e
d
L
T
D
.
E
a
c
h
y
e
a
r
w
e
r
a
n
k
s
t
o
c
k
s
i
n
t
o
q
u
i
n
t
i
l
e
s
(
1
-
5
)
a
n
d
f
o
r
m
e
q
u
a
l
-
w
e
i
g
h
t
e
d
p
o
r
t
f
o
l
i
o
s
a
t
t
h
e
b
e
g
i
n
n
i
n
g
o
f
e
a
c
h
a
n
n
u
a
l
p
e
r
i
o
d
.
W
e
i
n
v
e
s
t
i
g
a
t
e
d
a
i
l
y
e
x
c
e
s
s
r
e
t
u
r
n
s
o
f
t
h
e
s
e
p
o
r
t
f
o
l
i
o
s
d
u
r
i
n
g

B
l
a
c
k
M
o
n
d
a
y

(
O
c
t
o
b
e
r
1
9
,
1
9
8
7
)
,
t
h
e
A
s
i
a
n
C
r
i
s
i
s
(
O
c
t
o
b
e
r
2
7
,
1
9
9
7
)
,
t
h
e
b
u
r
s
t
o
f
t
h
e
d
o
t
-
c
o
m
b
u
b
b
l
e
(
A
p
r
i
l
1
4
,
2
0
0
0
)
,
a
n
d
t
h
e
r
e
c
e
n
t
L
e
h
m
a
n
c
r
i
s
e
s
(
O
c
t
o
b
e
r
1
5
,
2
0
0
8
)
.
T
h
e
l
a
s
t
c
o
l
u
m
n
r
e
p
o
r
t
s
a
v
e
r
a
g
e
d
a
i
l
y
r
e
t
u
r
n
s
i
n
e
x
c
e
s
s
o
f
t
h
e
o
n
e
-
m
o
n
t
h
T
-
b
i
l
l
r
a
t
e
o
f
t
h
e
p
o
r
t
f
o
l
i
o
s
w
h
e
n
t
h
e
d
a
i
l
y
e
x
c
e
s
s
m
a
r
k
e
t
r
e
t
u
r
n
,
r
m
,
d
,
o
n
d
a
y
d
w
a
s
b
e
l
o
w

5
%
(
1
6
d
a
y
s
)
.
T
h
e
r
o
w
l
a
b
e
l
e
d

S
t
r
o
n
g
-
W
e
a
k

r
e
p
o
r
t
s
t
h
e
d
i

e
r
e
n
c
e
b
e
t
w
e
e
n
t
h
e
r
e
t
u
r
n
s
o
f
p
o
r
t
f
o
l
i
o
5
a
n
d
p
o
r
t
f
o
l
i
o
1
w
i
t
h
c
o
r
r
e
s
p
o
n
d
i
n
g
s
t
a
t
i
s
t
i
c
a
l
s
i
g
n
i

c
a
n
c
e
l
e
v
e
l
s
(
o
n
l
y
l
a
s
t
c
o
l
u
m
n
)
.
T
h
e
s
a
m
p
l
e
c
o
v
e
r
s
a
l
l
U
.
S
.
c
o
m
m
o
n
s
t
o
c
k
s
t
r
a
d
e
d
o
n
t
h
e
N
Y
S
E
/
A
M
E
X
a
n
d
t
h
e
s
a
m
p
l
e
p
e
r
i
o
d
i
s
f
r
o
m
J
a
n
u
a
r
y
1
9
6
3
t
o
D
e
c
e
m
b
e
r
2
0
0
9
.
t
-
s
t
a
t
i
s
t
i
c
s
a
r
e
i
n
p
a
r
e
n
t
h
e
s
e
s
.

,
a
n
d

i
n
d
i
c
a
t
e
s
i
g
n
i

c
a
n
c
e
a
t
t
h
e
o
n
e
,

v
e
,
a
n
d
t
e
n
p
e
r
c
e
n
t
l
e
v
e
l
,
r
e
s
p
e
c
t
i
v
e
l
y
.
67
Table 5: Dependent Portfolio Sorts
Panel A: Beta () and Lower Tail Dependence (LTD)
Portfolio 1 Low 2 3 4 5 High Average
1 Weak LTD 4.48% 3.57% 4.53% 5.97% 10.40% 5.79%
2 6.60% 7.67% 8.81% 12.64% 14.28% 10.00%
3 6.48% 9.47% 9.29% 13.53% 20.32% 11.82%
4 6.95% 9.05% 12.78% 14.29% 23.92% 13.40%
5 Strong LTD 9.71% 12.04% 13.76% 16.76% 28.69% 16.19%
Strong - Weak 5.23%

8.48%

9.23%

10.79%

18.29%

10.40%

(3.28) (5.67) (5.12) (5.03) (6.88) (5.20)

Panel B: Downside Beta (

) and Lower Tail Dependence (LTD)

Portfolio 1 Low

2 3 4 5 High

Average
1 Weak LTD 3.12% 4.86% 4.58% 7.88% 15.01% 7.09%
2 3.07% 6.79% 8.30% 11.71% 17.76% 9.52%
3 6.18% 7.95% 9.48% 12.26% 21.97% 11.57%
4 7.60% 8.65% 12.38% 13.99% 23.79% 13.28%
5 Strong LTD 10.01% 10.02% 12.56% 17.27% 27.99% 15.57%
Strong - Weak 6.89%

5.16%

7.98%

9.40%

12.98%

8.48%

(5.02) (3.57) (4.24) (4.32) (3.95) (4.02)

Panel C: Coskewness (coskew) and Lower Tail Dependence (LTD)
Portfolio 1 Low coskew 2 3 4 5 High coskew Average
1 Weak LTD 10.68% 7.49% 5.18% 2.60% 3.13% 5.81%
2 12.58% 11.15% 10.03% 6.65% 4.46% 8.97%
3 15.35% 12.08% 9.91% 7.97% 6.88% 10.44%
4 18.94% 15.41% 13.10% 12.42% 8.70% 13.71%
5 Strong LTD 23.58% 19.94% 18.06% 16.32% 12.87% 18.16%
Strong - Weak 12.90%

12.46%

12.88%

13.73%

9.74%

12.34%

(5.12) (6.44) (6.28) (7.29) (4.84) (5.99)

This table reports equal-weighted average annual excess returns over the one-month T-Bill rate of 25 portfolios double-sorted
on realized LTD and realized beta (Panel A), realized downside beta (Panel B), and realized coskewness (Panel C), respectively.
First, we form quintile portfolios sorted on beta, downside beta, and coskewness, respectively. Then, within each of those
quintiles, we sort stocks into quintile portfolios based on LTD. The row labeled Strong - Weak reports the dierence between
the returns of portfolio 5 and portfolio 1 in each beta, downside beta, or coskewness quintile with corresponding statistical
signicance levels. The sample covers all U.S. common stocks traded on the NYSE / AMEX and the sample period is from
January 1963 to December 2009. t-statistics are in parentheses.

,

, and

indicate signicance at the one, ve, and ten
percent level, respectively.
68
T
a
b
l
e
6
:
M
u
l
t
i
v
a
r
i
a
t
e
R
e
g
r
e
s
s
i
o
n
s
P
a
n
e
l
A
:
F
a
m
a
-
M
a
c
B
e
t
h
(
1
9
7
3
)
R
e
g
r
e
s
s
i
o
n
s
(
1
)
(
2
)
(
3
)
(
4
)
(
5
)
(
6
)
(
7
)
(
8
)
(
9
)
E
c
o
n
o
m
i
c
S
i
g
n
.
r
e
t
u
r
n
r
e
t
u
r
n
r
e
t
u
r
n
r
e
t
u
r
n
r
e
t
u
r
n
r
e
t
u
r
n
c
a
p
m
-
a
l
p
h
a

-
a
l
p
h
a
c
a
r
-
a
l
p
h
a
B
a
s
e
d
o
n
(
9
)
L
T
D
0
.
5
5
1

0
.
5
8
4

0
.
5
5
5

0
.
4
4
8

0
.
4
5
2

0
.
4
5
7

0
.
4
5
4

0
.
4
4
1

5
.
0
1
%
(
8
.
4
4
)
(
9
.
1
1
)
(
1
1
.
5
9
)
(
9
.
8
9
)
(
1
0
.
1
6
)
(
1
0
.
0
9
)
(
1
0
.
1
5
)
(
9
.
8
4
)
U
T
D
-
0
.
3
2
6

-
0
.
2
5
4

-
0
.
2
9
6

-
0
.
2
9
2

-
0
.
2
7
4

-
0
.
2
5
4

-
0
.
2
4
1

-
2
.
2
5
%
(
-
5
.
0
0
)
(
-
4
.
4
5
)
(
-
7
.
0
9
)
(
-
6
.
5
3
)
(
-
6
.
0
7
)
(
-
5
.
8
2
)
(
-
5
.
4
5
)

0
.
0
8
1
5

0
.
0
3
7
5

0
.
0
2
0
3
0
.
0
1
3
3
0
.
0
0
2
9
2
+
0
.
2
1
%
(
3
.
7
4
)
(
2
.
8
0
)
(
1
.
5
8
)
(
0
.
8
9
)
(
0
.
1
9
)

+
-
0
.
0
1
0
7
0
.
0
0
9
6
0
-
0
.
0
0
2
6
2
0
.
0
0
4
9
3
0
.
0
0
8
4
6
+
0
.
6
5
%
(
-
1
.
3
6
)
(
0
.
8
7
)
(
-
0
.
2
3
)
(
0
.
4
0
)
(
0
.
7
0
)

0
.
0
7
4
8

0
.
1
4
0

(
2
.
6
4
)
(
5
.
6
1
)
s
i
z
e
-
0
.
0
1
2
1

-
0
.
0
3
0
2

-
0
.
0
2
7
9

-
0
.
0
2
8
7

-
0
.
0
2
5
0

-
0
.
0
2
2
0

-
4
.
6
6
%
(
-
1
.
7
0
)
(
-
5
.
8
7
)
(
-
5
.
7
0
)
(
-
5
.
6
9
)
(
-
6
.
0
7
)
(
-
5
.
6
3
)
b
o
o
k
m
a
r
k
e
t
0
.
0
3
8
3

0
.
0
3
0
0

0
.
0
2
8
6

0
.
0
2
9
3

0
.
0
1
6
5

0
.
0
1
6
7

+
1
.
2
2
%
(
4
.
5
0
)
(
3
.
9
1
)
(
3
.
7
3
)
(
3
.
8
4
)
(
3
.
0
4
)
(
3
.
1
1
)
c
o
s
k
e
w
0
.
1
2
7

0
.
0
8
6
3

0
.
1
1
4

0
.
1
1
3

0
.
1
3
2

0
.
1
2
7

+
3
.
9
0
%
(
2
.
6
3
)
(
2
.
1
4
)
(
2
.
3
4
)
(
2
.
2
9
)
(
2
.
1
9
)
(
2
.
0
4
)
i
l
l
i
q
0
.
2
2
8

0
.
1
9
8

0
.
1
7
2

0
.
1
7
7

0
.
1
2
2

0
.
1
0
8

+
1
.
4
3
%
(
4
.
2
0
)
(
3
.
6
3
)
(
3
.
1
4
)
(
3
.
1
3
)
(
2
.
2
1
)
(
1
.
7
7
)
p
a
s
t
r
e
t
u
r
n
-
0
.
0
1
7
3
-
0
.
0
1
3
6
-
0
.
0
1
6
5
-
0
.
0
1
1
1
-
0
.
0
1
9
2
-
0
.
8
8
%
(
-
0
.
8
1
)
(
-
0
.
6
3
)
(
-
0
.
7
6
)
(
-
0
.
5
1
)
(
-
1
.
1
0
)
i
d
i
o
v
o
l
a
-
3
.
7
5
8

-
1
.
6
8
2
-
2
.
4
3
4
-
3
.
6
0
3

-
2
.
9
9
4

-
3
.
5
4
%
(
-
2
.
1
9
)
(
-
0
.
9
0
)
(
-
1
.
3
9
)
(
-
2
.
5
8
)
(
-
2
.
1
8
)
c
o
k
u
r
t
0
.
0
0
6
3
0
.
0
0
5
7
0
.
0
0
3
5
0
.
0
0
2
4
0
.
0
0
1
9
+
0
.
5
1
%
(
0
.
5
8
)
(
0
.
2
8
)
(
0
.
8
4
)
(
0
.
7
6
)
(
0
.
3
8
)
m
a
x
-
0
.
2
5
0

-
0
.
2
2
2

-
0
.
2
3
1

-
0
.
2
2
7

-
0
.
2
1
4

-
1
.
5
7
%
(
-
2
.
5
3
)
(
-
2
.
2
9
)
(
-
2
.
3
6
)
(
-
2
.
4
5
)
(
-
2
.
3
1
)
c
o
n
s
t
a
n
t
0
.
0
4
6
0
0
.
0
6
7
8

0
.
0
3
9
1

0
.
0
5
5
8
0
.
3
8
5

0
.
3
2
5

0
.
3
4
4

0
.
3
2
3

0
.
2
9
6

(
1
.
3
8
)
(
1
.
9
2
)
(
1
.
7
6
)
(
0
.
5
6
)
(
4
.
9
7
)
(
4
.
5
1
)
(
4
.
5
8
)
(
4
.
9
1
)
(
4
.
5
6
)
R
2
0
.
0
1
9
0
.
0
2
4
0
.
0
4
0
0
.
1
1
0
0
.
1
5
0
0
.
1
4
6
0
.
1
3
7
0
.
1
0
0
0
.
0
8
6
69
Table 6: Continued
Panel B: Conditional Fama-MacBeth (1973) Regressions
low Std.Dev. high Std.Dev. low VaR high VaR low CoVaR high CoVaR
LTD 0.241

0.811

0.673

0.269

0.730

0.238

(based on (6)) (9.88) (9.31) (9.07) (9.64) (9.08) (7.86)

LTD 0.242

0.812

0.675

0.270

0.732

0.239

(based on (7)) (9.86) (9.27) (9.01) (9.64) (9.02) (7.85)

LTD 0.239

0.814

0.675

0.269

0.729

0.244

(based on (8)) (9.42) (9.28) (9.08) (9.41) (9.00) (7.93)

LTD 0.236

0.785

0.654

0.264

0.705

0.240

(based on (9)) (9.17) (9.00) (8.78) (9.41) (8.55) (7.98)

This table displays the results of multivariate Fama-MacBeth (1973) regressions. In Panel A we report the
results of regressions of yearly stock-level excess returns over the risk free rate on LTD, UTD, downside
beta (

), upside beta (
+
), beta (), the log of market capitalization (size), the book-to-market ratio
(bookmarket), coskewness (coskew), the Amihud Illiquidity Ratio (illiq), the past 12-month excess returns
(past return), idiosyncratic volatility (idiovola), cokurtosis (cokurt), and the maximum daily return over the
past one year (max). All risk characteristics (LTD, UTD,

,
+
, , coskew, idiovola, cokurt) are calculated
contemporaneously to the yearly excess return. Size, bookmarket, and illiq for year t are calculated using
data from (the end of) year t 1. The last column displays the change in annualized excess returns for
a one standard deviation increase in the respective independent variable based on regression (9). Panel B
repeats the Fama-MacBeth (1973) regression specications (6) to (9) from Panel A conditional on the stocks
standard deviation, Value-at-Risk, (VaR, dened as the 5% percentile of the daily returns of the stock), and
its conditional Value-at-Risk (CoVaR, dened as the conditional mean of all daily returns below the 5%
percentile), respectively, being below (above) the median in the respective year. Panel B only reports the
coecient estimate for the impact of LTD. All other explanatory variables are included in the regressions,
but their coecient estimates suppressed. The rst two columns report the coecient estimates of LTD
if a stocks standard deviation is below (above) its median. The next two columns report the coecient
estimates of LTD if a stocks Value-at-Risk (VaR) is below (above) its median. Finally, the last two columns
report the coecient estimates of LTD if a stocks conditional Value-at-Risk (CoVaR) is below (above))
its median. The independent variables are winsorized at the 1% level and at the 99% level. The sample
covers all U.S. common stocks traded on the NYSE / AMEX and the sample period is from January 1963
to December 2009. t-statistics are in parentheses.

,

, and

indicate signicance at the one, ve, and
ten percent level, respectively.
70
T
a
b
l
e
7
:
D
e
p
e
n
d
e
n
t
P
o
r
t
f
o
l
i
o
S
o
r
t
s
:
L
T
D
a
n
d
A
l
t
e
r
n
a
t
i
v
e
D
o
w
n
s
i
d
e
B
e
t
a
D
e

n
i
t
i
o
n
s
C
u
t
-
O

L
T
D
P
o
r
t
f
o
l
i
o
1
L
o
w

2
3
4
5
H
i
g
h

A
v
e
r
a
g
e
2
0
%
Q
u
a
n
t
i
l
e
S
t
r
o
n
g
-
W
e
a
k
4
.
9
9
%

6
.
8
7
%

9
.
0
6
%

9
.
1
0
%

1
7
.
3
0
%

9
.
4
6
%

(
3
.
9
8
)
(
6
.
3
6
)
(
4
.
8
9
)
(
4
.
0
6
)
(
6
.
2
1
)
(
5
.
1
0
)
1
0
%
Q
u
a
n
t
i
l
e
S
t
r
o
n
g
-
W
e
a
k
6
.
9
6
%

8
.
8
2
%

1
2
.
5
7
%

1
3
.
6
9
%

1
7
.
8
4
%

1
1
.
9
7
%

(
4
.
5
6
)
(
7
.
1
0
)
(
7
.
5
3
)
(
7
.
2
6
)
(
6
.
8
9
)
(
6
.
6
7
)
5
%
Q
u
a
n
t
i
l
e
S
t
r
o
n
g
-
W
e
a
k
1
0
.
7
3
%

1
0
.
2
1
%

1
2
.
1
5
%

1
6
.
3
4
%

1
7
.
5
6
%

1
3
.
4
0
%

(
5
.
4
4
)
(
6
.
1
5
)
(
7
.
7
2
)
(
8
.
2
1
)
(
5
.
8
4
)
(
6
.
6
7
)
2
%
Q
u
a
n
t
i
l
e
S
t
r
o
n
g
-
W
e
a
k
1
2
.
5
2
%

1
1
.
3
9
%

1
2
.
5
5
%

1
5
.
7
2
%

1
9
.
8
2
%

1
4
.
4
0
%

(
6
.
1
4
)
(
6
.
1
5
)
(
7
.
9
2
)
(
9
.
8
5
)
(
6
.
4
5
)
(
7
.
3
0
)
1
%
Q
u
a
n
t
i
l
e
S
t
r
o
n
g
-
W
e
a
k
1
5
.
9
3
%

1
3
.
3
1
%

1
3
.
1
5
%

1
5
.
8
1
%

1
7
.
9
7
%

1
5
.
2
3
%

(
7
.
4
6
)
(
6
.
2
5
)
(
7
.
3
1
)
(
8
.
1
3
)
(
8
.
5
3
)
(
7
.
5
3
)
T
h
i
s
t
a
b
l
e
r
e
p
o
r
t
s
r
e
s
u
l
t
s
o
f
p
o
r
t
f
o
l
i
o
s
d
o
u
b
l
e
-
s
o
r
t
e
d
o
n
c
o
n
t
e
m
p
o
r
a
n
e
o
u
s
L
T
D
a
n
d
d
i

e
r
e
n
t
a
l
t
e
r
n
a
t
i
v
e
d
e

n
i
t
i
o
n
s
o
f
c
o
n
t
e
m
p
o
r
a
n
e
o
u
s
d
o
w
n
s
i
d
e
b
e
t
a
s
.
F
i
r
s
t
,
w
e
f
o
r
m
q
u
i
n
t
i
l
e
p
o
r
t
f
o
l
i
o
s
s
o
r
t
e
d
o
n
d
o
w
n
s
i
d
e
b
e
t
a
.
T
h
e
n
,
w
i
t
h
i
n
e
a
c
h
o
f
t
h
o
s
e
q
u
i
n
t
i
l
e
s
,
w
e
s
o
r
t
s
t
o
c
k
s
i
n
t
o
q
u
i
n
t
i
l
e
p
o
r
t
f
o
l
i
o
s
b
a
s
e
d
o
n
L
T
D
.
A
s
a
l
t
e
r
n
a
t
i
v
e
d
o
w
n
s
i
d
e
b
e
t
a
d
e

n
i
t
i
o
n
s
,
w
e
c
a
l
c
u
l
a
t
e
d
o
w
n
s
i
d
e
b
e
t
a
s
a
s
b
e
t
a
s
c
o
n
d
i
t
i
o
n
a
l
o
n
t
h
e
m
a
r
k
e
t
r
e
t
u
r
n
b
e
i
n
g
b
e
l
o
w
i
t
s
2
0
%
q
u
a
n
t
i
l
e
,
i
t
s
1
0
%
q
u
a
n
t
i
l
e
,
i
t
s
5
%
q
u
a
n
t
i
l
e
,
i
t
s
2
%
q
u
a
n
t
i
l
e
,
a
n
d
i
t
s
1
%
q
u
a
n
t
i
l
e
,
r
e
s
p
e
c
t
i
v
e
l
y
.
W
e
o
n
l
y
r
e
p
o
r
t
r
e
s
u
l
t
s
o
n
t
h
e
r
e
t
u
r
n
s
o
f
t
h
e
s
t
r
o
n
g
m
i
n
u
s
w
e
a
k
L
T
D
p
o
r
t
f
o
l
i
o
s
w
i
t
h
i
n
e
a
c
h
d
o
w
n
s
i
d
e
b
e
t
a
q
u
i
n
t
i
l
e
i
n
t
h
e

r
s
t

v
e
c
o
l
u
m
n
s
a
s
w
e
l
l
a
s
t
h
e
a
v
e
r
a
g
e
o
f
t
h
i
s
d
i

e
r
e
n
c
e
p
o
r
t
f
o
l
i
o
r
e
t
u
r
n
a
c
r
o
s
s
a
l
l
d
o
w
n
s
i
d
e
b
e
t
a
q
u
i
n
t
i
l
e
s
i
n
t
h
e
l
a
s
t
c
o
l
u
m
n
(
l
i
k
e
i
n
t
h
e
l
a
s
t
r
o
w
o
f
P
a
n
e
l
B
i
n
T
a
b
l
e
5
)
f
o
r
a
l
l
a
l
t
e
r
n
a
t
i
v
e
d
o
w
n
s
i
d
e
b
e
t
a
d
e

n
i
t
i
o
n
s
.
T
h
e
s
a
m
p
l
e
c
o
v
e
r
s
a
l
l
U
.
S
.
c
o
m
m
o
n
s
t
o
c
k
s
t
r
a
d
e
d
o
n
t
h
e
N
Y
S
E
/
A
M
E
X
a
n
d
t
h
e
s
a
m
p
l
e
p
e
r
i
o
d
i
s
f
r
o
m
J
a
n
u
a
r
y
1
9
6
3
t
o
D
e
c
e
m
b
e
r
2
0
0
9
.
t
-
s
t
a
t
i
s
t
i
c
s
a
r
e
i
n
p
a
r
e
n
t
h
e
s
e
s
.

,
a
n
d

i
n
d
i
c
a
t
e
s
i
g
n
i

c
a
n
c
e
a
t
t
h
e
o
n
e
,

v
e
,
a
n
d
t
e
n
p
e
r
c
e
n
t
l
e
v
e
l
,
r
e
s
p
e
c
t
i
v
e
l
y
.
71
Table 8: Factor Models: Momentum and LTD
(1) (2) (3)
mom mom mom
market -0.255 -0.456

-0.409

(-1.54) (-2.53) (-2.43)

smb -0.0880 -0.0766 -0.0416
(-0.44) (-0.40) (-0.22)
hml -0.361

-0.104 -0.0580
(-1.76) (-0.46) (-0.25)
LTD
ew
Strong-Weak 0.676

(2.31)
LTD
vw
Strong-Weak 0.757

(2.46)
alpha 0.128

0.0188 0.0391
(3.97) (0.33) (0.82)
N 47 47 47
R
2
0.106 0.206 0.218
This table lists OLS-regression results of the yearly Jegadeesh and Titman (1993) momentum factor (mom)
on the yearly CRSP value-weighted market excess return (market), the Fama and French (1993) small-minus-
big (smb) factor, the Fama and French (1993) high-minus-low (hml) factor, as well as equal (value)-weighted
returns from a portfolio going long in stocks with strong contemporaneous LTD and going short in stocks
with weak contemporaneous LTD. The LTD portfolios are rebalanced on a yearly basis. The sample covers
all U.S. common stocks traded on the NYSE / AMEX and the sample period is from January 1963 to
December 2009. t-statistics are in parentheses.

,

, and

indicate signicance at the one, ve, and ten
percent level, respectively.
72
Table 9: FMB with Fixed Copula Combinations & Industry- and DGTW-adjusted Returns
Fixed LTD R
2
return LTD R
2
Copula (t-stat) adjustment (t-stat)
(3-D-I) 0.519

0.147 FF-12 0.430

0.124
(9.95) (9.43)
(4-D-I) 0.514

0.146 FF-48 0.399

0.108
(9.93) (9.66)
(2-D-I) 0.497

0.145 SIC-2 0.398

0.106
(7.44) (10.19)
(2-B-II) 0.491

0.145 SIC-3 0.344

0.087
(7.06) (9.60)
(4-B-II) 0.587

0.147 SIC-4 0.281

0.071
(10.15) (8.91)
(4-B-III) 0.624

0.147 DGTW 0.453

0.082
(10.17) (9.31)
This table shows results for the estimate of the inuence of LTD from Fama-MacBeth (1973) regressions
of 1-year excess returns over the riskfree rate on LTD and the full set of controls as in Regression (6)
from Table 6 (included in the regression but coecient estimates suppressed in the table) in the rst three
columns. LTD coecients are calculated based on the Rotated Joe/F-G-M/Joe copula (3-D-I), the Ro-
tated Galambos/F-G-M/Joe copula (4-D-I), the Rotated Gumbel/F-G-M/Joe copula (2-D-I), the Rotated
Joe/Frank/Gumbel copula (2-B-II), the Rotated Galambos/Frank/Gumbel copula (4-B-II), and the Rotated
Galambos/Frank/Galambos copula (4-B-III). The last three columns repeat the same Fama-MacBeth (1973)
regressions with the full set of controls as in Regression (6) from Table 6 (Panel A) with alternative dependent
variables. We adjust returns by industry (based on Kenneth Frenchs 12 (FF-12) and 48 (FF-48) industry
portfolios as well as SIC 2-, 3-, and 4-digit codes. In the last line, we report the regression results when we
adjust returns for DGTW characteristic-based benchmarks (as in Daniel, Grinblatt, Titman, and Wermers
(1997)). The sample covers all U.S. common stocks traded on the NYSE / AMEX and the sample period is
from January 1963 to December 2009. t-statistics are in parentheses.

,

, and

indicate signicance at
the one, ve, and ten percent level, respectively.
73
Table 10: Value-weighted Portfolio Sorts
Panel A: Univariate Sorts on Lower Tail Dependence (LTD)
Portfolio 1 Weak LTD 2 3 4 5 Strong LTD Strong - Weak
Return 1.03% 2.23% 3.73%

5.53%

9.45%

10.48%

(0.46) (0.96) (1.64) (2.30) (3.56) (6.36)

Panel B: Dependent Double-sorts on Beta () and Lower Tail Dependence (LTD)
Portfolio 1 Low 2 3 4 5 High Average
Strong - Weak 2.75%

5.04%

7.95%

9.87%

16.22%

8.37%

(2.01) (2.85) (5.22) (5.52) (7.12) (4.54)

Panel C: Dependent Double-sorts on Downside Beta (

) and Lower Tail Dependence (LTD)

Portfolio 1 Low

2 3 4 5 High

Average
Strong - Weak 2.52% 7.07%

5.62%

11.05%

13.99%

8.05%

(0.79) (4.51) (3.78) (6.10) (5.64) (4.16)

Panel D: Dependent Double-sorts on Coskewness (coskew) and Lower Tail Dependence (LTD)
Portfolio 1 Low coskew 2 3 4 5 High coskew Average
Strong - Weak 11.02%

9.42%

8.58%

10.51%

6.89%

9.29%

(4.25) (3.90) (4.59) (5.35) (2.98) (4.21)

This table reports the results of value-weighted portfolio sorts. Panel A reports the results of value-weighted
univariate sorts. Each year we rank stocks into quintiles (1-5) and form value-weighted portfolios at the
beginning of each annual period. The row labeled Return reports the average return in excess of the
one-month T-bill rate over the next year in the respective quintiles and the dierence between the returns
of portfolio 5 and portfolio 1 with corresponding statistical signicance levels. In Panels B to D we show
the results of value-weighted double-sorts on LTD and beta (Panel B), LTD and downside beta (Panel C),
and LTD and coskewness (Panel D). First, we form quintile portfolios sorted on beta, downside beta, and
coskewness, respectively. Then, within each beta, downside beta, and coskewness quintile, we sort stocks into
ve portfolios based on LTD. We only report results on the returns of the strong minus weak LTD portfolios
within each beta, downside beta, and coskewness, respectively, quintile in the rst ve columns as well as
the average of this dierence portfolio return across all beta, downside beta, and coskewness quintiles in
the last column with corresponding statistical signicance levels. The sample covers all U.S. common stocks
traded on the NYSE / AMEX and the sample period is from January 1963 to December 2009. t-statistics
are in parentheses.

,

, and

indicate signicance at the one, ve, and ten percent level, respectively.
74
Table 11: Temporal Stability: Portfolio Sorts and Fama-MacBeth (1973) Regressions
Panel A: Univariate Sorts
Jan1963 Jan1973 Jan1982 Jan1991 Jan2000 Jan1963 Jan1987
Portfolio - - - - - - -
Dec1972 Dec1981 Dec1990 Dec1999 Dec2009 Dec1986 Dec2009
1 Weak LTD 5.12% 4.21% 2.25% 4.31% 7.98% 4.67% 3.27%
2 9.59% 7.10% 3.25% 11.54% 12.23% 8.98% 8.69%
3 11.29% 7.77% 4.93% 14.94% 12.64% 9.80% 11.00%
4 16.56% 12.41% 9.26% 18.87% 13.06% 14.54% 13.57%
5 Strong LTD 23.21% 19.47% 14.48% 25.52% 15.84% 20.91% 18.44%
5 Strong - Weak 18.09%

15.26%

16.73%

21.21%

7.86%

16.24%

15.17%

(4.85) (3.02) (6.22) (5.65) (2.29) (6.62) (5.60)

Panel B: Bivariate Sorts
Beta Downside Beta Coskewness
Portfolio 1963-1986 1987-2009 1963-1986 1987-2009 1963-1986 1987-2009
1 Weak LTD 7.64% 3.86% 8.97% 5.13% 6.73% 4.86%
2 10.10% 9.89% 10.25% 8.77% 8.20% 9.78%
3 11.89% 11.74% 10.80% 12.36% 10.47% 10.40%
4 13.29% 13.50% 13.09% 13.48% 14.19% 13.22%
5 Strong LTD 16.08% 16.31% 15.74% 15.39% 19.44% 16.82%
5 Strong - Weak 8.45%

12.45%

6.77%

10.26%

12.71%

11.95%

(3.21) (4.06) (3.01) (3.65) (4.29) (3.62)

Panel C: Fama-MacBeth (1973) Regressions
Jan1963 - Jan1973 - Jan1982 - Jan1991 - Jan2000 - Up Down
Dec1972 Dec1981 Dec1990 Dec1999 Dec2009 Market Market
LTD 0.385

0.475

0.463

0.605

0.342

0.535

0.261

(3.16) (4.26) (4.50) (8.08) (4.26) (10.54) (3.93)

75
Table 11: Continued
This tables reports the results of temporal stability checks. Panel A shows equal-weighted average returns of
stocks sorted by realized LTD for dierent 9- and 10 year subsamples as well as the time period from January
1963 to December 1986, and from January 1987 to December 2009. Each year we rank stocks into quintiles
(1-5) and form equal-weighted portfolios at the beginning of each annual period. We report the average
return in excess of the one-month T-bill rate over the next year. The row labeled Strong - Weak reports
the dierence between the returns of portfolio 5 and portfolio 1 with corresponding statistical signicance
levels. In Panel B, we report the results from the same double sorts as in Table 5 seperately for the pre-
and post-1987 period. We only report the average return across the beta, downside beta, and coskewness
portfolios for the 5 LTD portfolios as well as the dierence portfolio. Panel C displays the results of Fama-
MacBeth (1973) regressions of 1-year excess returns on rm characteristics and realized risk characteristics
for, respectively, 9- and 10 year subsamples. We include the full set of independent variables from Regression
(6) in Table 6 (included in the regression but coecient estimates suppressed in the table). All independent
variables are winsorized at the 1% level and at the 99%. The sample covers all U.S. common stocks traded
on the NYSE / AMEX and the sample period is from January 1963 to December 2009. t-statistics are in
parentheses.

,

, and

indicate signicance at the one, ve, and ten percent level, respectively.
76
Table 12: Dierent Regression Methods
Regression (1) (2) (3) (4) (5) (6) (7)
LTD 0.452

0.455

0.480

0.480

0.455

0.455

0.484

(11.35) (9.87) (23.48) (18.68) (22.03) (21.29) (23.76)

Controls yes yes yes yes yes yes yes
Method fmb fmb ols ols panel panel panel
Windsorized yes no yes yes yes yes yes
Year Eects yes yes yes yes yes
Firm Eects no no xed xed random
Clustered SE rm industry no rm no
Newey-West SE yes no no no no no no
R
2
0.146 0.142 0.221 0.221 0.275 0.275
This table reports the results from regressions of excess returns on rm- and risk characteristics using various
regression techniques. The independent variables are the same as in regression (6) of Table 6 (Panel A). We
only display the results for the coecient estimate for the impact of LTD. All other variables are included
in the regression but coecient estimates are suppressed in the table. In regression (1) we perform a Fama-
MacBeth (1973) regression with Newey-West (1987) adjusted standard errors. Regression (2) displays the
results of a Fama-MacBeth (1973) regression with unwinsorized independent variables. In regression (3)
we perform a pooled OLS-regression with time-xed eects and standard errors clustered by single stocks.
Regression (4) is identical, but we cluster standard errors by Fama-French 12 industry. Regression (5)
performs a panel data regression with rm xed eects. Regression (6) performs a panel data regression
with rm xed eects and standard errors clustered by single stocks. In regression (7) we regress excess
returns on the independent variables via a random-eect panel data regression. The sample covers all U.S.
common stocks traded on the NYSE / AMEX and the sample period is from January 1963 to December
2009. t-statistics are in parentheses.

,

, and

indicate signicance at the one, ve, and ten percent
level, respectively.
77
T
a
b
l
e
1
3
:
T
r
a
d
i
n
g
S
t
r
a
t
e
g
y
B
a
s
e
d
o
n
P
a
s
t
L
o
w
e
r
T
a
i
l
D
e
p
e
n
d
e
n
c
e
(
1
)
(
2
)
(
3
)
(
4
)
(
5
)
(
6
)
(
7
)
(
8
)
(
9
)
W
e
i
g
h
t
i
n
g
S
c
h
e
m
e
E
W
E
W
E
W
E
W
E
W
E
W
V
W
V
W
V
W
E
s
t
i
m
a
t
i
o
n
H
o
r
i
z
o
n
1
y
1
y
1
y
1
y
2
y
3
y
1
y
2
y
3
y
C
A
P
M
-
F
F
-
C
A
R
-
C
A
R
-
C
A
R
-
C
A
R
-
C
A
R
-
C
A
R
-
P
o
r
t
f
o
l
i
o
R
e
t
u
r
n
A
l
p
h
a
A
l
p
h
a
A
l
p
h
a
A
l
p
h
a
A
l
p
h
a
A
l
p
h
a
A
l
p
h
a
A
l
p
h
a
1
W
e
a
k
L
T
D
0
.
4
9
9
%
0
.
0
9
7
%

0
.
3
3
3
%

0
.
1
9
5
%

0
.
1
7
7
%

0
.
1
0
3
%

0
.
1
0
4
%

0
.
1
2
2
%

0
.
1
0
9
%
2
0
.
6
7
1
%
0
.
2
3
6
%

0
.
1
9
2
%

0
.
0
2
9
%

0
.
0
0
9
%
0
.
0
2
1
%
0
.
0
2
4
%

0
.
0
9
0
%

0
.
0
6
6
%
3
0
.
7
1
3
%
0
.
2
5
6
%

0
.
1
2
9
%

+
0
.
0
1
9
%
0
.
0
4
0
%
0
.
1
0
1
%

0
.
0
3
0
%

0
.
1
0
0
%
0
.
0
2
9
%
4
0
.
7
7
5
%
0
.
2
9
5
%

0
.
0
3
9
%
+
0
.
1
0
9
%

0
.
1
2
3
%

0
.
0
9
7
%

0
.
0
0
2
%
0
.
0
4
2
%

0
.
0
0
1
%
5
S
t
r
o
n
g
L
T
D
0
.
8
6
2
%
0
.
3
5
0
%
+
0
.
1
2
2
%
+
0
.
1
8
7
%

0
.
2
2
0
%

0
.
2
0
3
%

0
.
0
8
4
%
0
.
1
1
0
%

0
.
0
9
4
%

S
t
r
o
n
g
-
W
e
a
k
0
.
3
6
3
%

0
.
2
5
3
%

0
.
4
5
4
%

0
.
3
8
3
%

0
.
3
9
6
%

0
.
3
0
6
%

0
.
1
8
8
%

0
.
2
3
2
%

0
.
2
0
3
%

(
2
.
9
9
)
(
2
.
2
9
)
(
4
.
6
5
)
(
3
.
8
6
)
(
3
.
2
6
)
(
2
.
3
0
)
(
1
.
7
0
)
(
2
.
0
8
)
(
1
.
7
6
)

:
S
t
r
o
n
g
-
W
e
a
k

0
.
0
9
1
%

0
.
3
1
9
%

0
.
3
3
3
%

0
.
2
6
0
%

0
.
1
3
7
%

0
.
1
5
3
%

0
.
1
2
3
%

0
.
0
3
5
%

0
.
0
2
4
%
(

0
.
4
7
)
(

2
.
2
8
)
(

2
.
2
1
)
(

1
.
8
8
)
(

1
.
4
7
)
(

1
.
5
6
)
(

0
.
6
7
)
(

0
.
7
1
)
(

0
.
3
5
)
T
h
i
s
t
a
b
l
e
r
e
p
o
r
t
s
e
q
u
a
l
-
w
e
i
g
h
t
e
d
a
n
d
v
a
l
u
e
-
w
e
i
g
h
t
e
d
a
v
e
r
a
g
e
r
e
t
u
r
n
s
a
n
d
a
l
p
h
a
s
o
f
s
t
o
c
k
p
o
r
t
f
o
l
i
o
s
s
o
r
t
e
d
b
y
p
a
s
t
L
T
D
.
I
n
e
a
c
h
m
o
n
t
h
,
w
e
r
a
n
k
s
t
o
c
k
s
i
n
t
o
q
u
i
n
t
i
l
e
s
(
1
-
5
)
a
n
d
f
o
r
m
p
o
r
t
f
o
l
i
o
s
b
a
s
e
d
o
n
p
a
s
t
L
T
D
.
I
n
c
o
l
u
m
n
s
(
1
)
-
(
3
)
,
w
e
f
o
r
m
e
q
u
a
l
-
w
e
i
g
h
t
e
d
p
o
r
t
f
o
l
i
o
s
s
o
r
t
e
d
b
y
p
a
s
t
1
-
y
e
a
r
L
T
D
a
n
d
r
e
p
o
r
t
t
h
e
a
v
e
r
a
g
e
e
x
c
e
s
s
r
e
t
u
r
n
(

R
e
t
u
r
n

)
,
S
h
a
r
p
e
(
1
9
6
4
)

C
A
P
M
-
A
l
p
h
a

,
a
n
d
F
a
m
a
a
n
d
F
r
e
n
c
h
(
1
9
9
3
)

F
F
-
A
l
p
h
a

o
v
e
r
t
h
e
f
o
l
l
o
w
i
n
g
m
o
n
t
h
.
C
o
l
u
m
n
s
(
4
)
-
(
6
)
r
e
p
o
r
t
C
a
r
h
a
r
t
(
1
9
9
7
)

s
f
o
u
r
f
a
c
t
o
r
m
o
d
e
l
a
l
p
h
a
f
o
r
e
q
u
a
l
-
w
e
i
g
h
t
e
d
p
o
r
t
f
o
l
i
o
s
s
o
r
t
e
d
b
y
p
a
s
t
1
-
y
e
a
r
,
2
-
y
e
a
r
,
a
n
d
3
-
y
e
a
r
L
T
D
e
x
p
o
s
u
r
e
.
I
n
c
o
l
u
m
n
s
(
7
)
-
(
9
)
w
e
r
e
p
o
r
t
C
a
r
h
a
r
t
(
1
9
9
7
)

s
f
o
u
r
f
a
c
t
o
r
m
o
d
e
l
a
l
p
h
a
f
o
r
v
a
l
u
e
-
w
e
i
g
h
t
e
d
p
o
r
t
f
o
l
i
o
s
s
o
r
t
e
d
b
y
p
a
s
t
1
-
y
e
a
r
,
2
-
y
e
a
r
,
a
n
d
3
-
y
e
a
r
L
T
D
e
x
p
o
s
u
r
e
.
T
h
e
r
o
w
l
a
b
e
l
e
d

S
t
r
o
n
g
-
W
e
a
k

r
e
p
o
r
t
s
t
h
e
d
i

e
r
e
n
c
e
b
e
t
w
e
e
n
t
h
e
r
e
t
u
r
n
s
a
n
d
a
l
p
h
a
s
o
f
p
o
r
t
f
o
l
i
o
5
a
n
d
p
o
r
t
f
o
l
i
o
1
w
i
t
h
c
o
r
r
e
s
p
o
n
d
i
n
g
s
t
a
t
i
s
t
i
c
a
l
s
i
g
n
i

c
a
n
c
e
l
e
v
e
l
s
.
T
h
e
l
a
s
t
r
o
w
(
l
a
b
e
l
e
d

:
S
t
r
o
n
g
-
W
e
a
k

)
r
e
p
o
r
t
s
t
h
e
d
i

e
r
e
n
c
e
b
e
t
w
e
e
n
t
h
e
r
e
t
u
r
n
s
a
n
d
a
l
p
h
a
s
o
f
t
h
e
t
o
p
q
u
i
n
t
i
l
e
p
o
r
t
f
o
l
i
o
w
i
t
h
r
e
s
p
e
c
t
t
o
l
a
g
g
e
d
d
o
w
n
s
i
d
e
b
e
t
a
a
n
d
o
f
t
h
e
b
o
t
t
o
m
q
u
i
n
t
i
l
e
p
o
r
t
f
o
l
i
o
w
i
t
h
r
e
s
p
e
c
t
t
o
l
a
g
g
e
d
d
o
w
n
s
i
d
e
b
e
t
a
w
i
t
h
c
o
r
r
e
s
p
o
n
d
i
n
g
s
t
a
t
i
s
t
i
c
a
l
s
i
g
n
i

c
a
n
c
e
l
e
v
e
l
s
.
T
h
e
s
a
m
p
l
e
c
o
v
e
r
s
a
l
l
U
.
S
.
c
o
m
m
o
n
s
t
o
c
k
s
t
r
a
d
e
d
o
n
t
h
e
N
Y
S
E
/
A
M
E
X
a
n
d
t
h
e
s
a
m
p
l
e
p
e
r
i
o
d
i
s
f
r
o
m
J
a
n
u
a
r
y
1
9
6
3
t
o
D
e
c
e
m
b
e
r
2
0
0
9
.
t
-
s
t
a
t
i
s
t
i
c
s
a
r
e
i
n
p
a
r
e
n
t
h
e
s
e
s
.

,
a
n
d

i
n
d
i
c
a
t
e
s
i
g
n
i

c
a
n
c
e
a
t
t
h
e
o
n
e
,

v
e
,
a
n
d
t
e
n
p
e
r
c
e
n
t
l
e
v
e
l
,
r
e
s
p
e
c
t
i
v
e
l
y
.
78
Table 14: Trading Strategy - Alternative Factor Models
(1) (2) (3) (4) (5) (6) (7)
Return Return Return Return Return trad strat trad strat
marketrf 0.281

0.233

0.309

0.308

0.302

(12.04) (7.49) (12.34) (12.98) (13.19)

smb -0.413

-0.453

-0.397

-0.397

-0.308

(-13.37) (-11.00) (-10.45) (-12.39) (-9.19)

hml -0.295

-0.220

-0.235

-0.240

-0.149

(-8.26) (-4.56) (-6.09) (-6.68) (-3.91)

mom 0.100

0.0498

0.0767

0.0741

0.0695

0.0126
(4.29) (1.72) (3.34) (3.26) (2.57) (0.41)
ps liqui -0.0000332
(-1.00)
sadka tf -1.398

(-2.47)
lot max ew 0.00451
(0.16)
sent orth -0.000969
(-0.99)
rev short -0.0767

(-2.54)
rev long -0.259

(-5.95)
s5rf 0.331

0.275

(11.08) (7.85)
r2s5 -0.279

(-7.35)
r3vr3g -0.104

(-2.18)
rms5 -0.0177
(-0.19)
r2rm -0.522

(-6.70)
s5vs5g -0.0563
(-0.68)
rmvrmg -0.0624
(-0.63)
r2vr2g 0.00986
(0.09)
alpha 0.383%

0.551%

0.385%

0.397%

0.509%

0.296%

0.379%

(3.76) (4.32) (3.84) (3.89) (5.34) (2.35) (2.62)

R
2
0.427 0.412 0.398 0.401 0.430 0.356 0.354
79
Table 14: Continued
This table shows regression results of the return from the same LTD trading strategy as in Table 13 on various combinations
of systematic risk factors and various factor models. Regression (1) includes factors of the Carhart (1997) model (smb, hml,
mom) plus the Pastor and Stambaugh (2003) traded liquidity risk factor (ps liqui). In regression (2), the Pastor and Stambaugh
(2003) liquidity factor is replaced by the Sadka (2006) liquidity factor (sadka tf). In regressions (3) and (4) we include the Bali,
Cakici, and Whitelaw (2011a) lottery factor (lot max ew) and the Baker and Wurgler (2006) sentiment index orthogonalized
with respect to a set of macroeconomic conditions (sent orth). In regression (5), we replace the Carhart (1997) momentum
factor by the Fama-French short- and long-term reversal factors (rev short and rev long). In regressions (6) and (7), the models
suggested in Cremers, Petajisto, Zitzewitz (2010) that contain various combinations of return dierences between the S&P 500
and Russell 2000 and 3000 subindexes as well as the momentum factor are used (the denitions of their factors are given in
their data library). Portfolios are re-balanced on a monthly basis. The alpha of the strategies is shown in the second to last
row. The sample covers all U.S. common stocks traded on the NYSE / AMEX and the sample period is from January 1963
to December 2009. t-statistics are in parentheses.

,

, and

indicate signicance at the one, ve, and ten percent level,
respectively.
80
Table 15: Predictive FMB & Industry- and DGTW-adjusted Returns
LTD R
2
return LTD R
2
(t-stat) adjustment (t-stat)
(1) 0.00984

0.007 FF-12 0.0138

0.063
(2.34) (6.29)
(2) 0.0113

0.011 FF-48 0.0139

0.055
(2.99) (6.17)
(5) 0.0132

0.077 SIC-2 0.0138

0.054
(6.12) (6.71)
(6) 0.0144

0.078 SIC-3 0.0132

0.044
(5.74) (6.67)
(DGTW) 0.0122

0.043 SIC-4 0.0106

0.037
(3.95) (5.86)
This table displays the results of multivariate Fama-MacBeth (1973) regressions to assess the impact of past
LTD on future monthly returns. We repeat variants of the multivariate specications from Panel A in Table
6, but use lagged values for LTD as well as all other independent variables. We only show the coecient
estimate for the impact of lagged LTD. The rst four lines in the left part of the table show results from
regressions equivalent to regressions (1), (2), (5), and (6) from Table 6 (Panel A). In the last line, we again
use regression (6) with the full set of controls, but use DGTW-adjusted returns as dependent variable. In the
right part of the table, we present results from regression (6) with industry-adjusted returns as dependent
variables. To dene industries we use the Fama-French 12 and 48 industry classication as well as SIC 2-,
3-, and 4-digit codes. The sample covers all U.S. common stocks traded on the NYSE / AMEX and the
sample period is from January 1963 to December 2009. t-statistics are in parentheses.

,

, and

indicate
signicance at the one, ve, and ten percent level, respectively.
81