Int. Fin. Markets, Inst.

and Money 31 (2014) 1429

Contents lists available at ScienceDirect

Journal of International Financial

Markets, Institutions & Money
j ou rn al ho me pa ge : w w w . e l s e v i e r . c o m / l o c a t e / i n t f i n

Determinants of stock returns: Factors or

systematic co-moments? Crisis versus
non-crisis periods
Chi-Hsiou D. Hung a,,1, A.S.M. Sohel Azad b, Victor Fang b
a
Department of Accounting and Finance, Adam Smith Business School, University of Glasgow, Main
Building, University Avenue, Glasgow G12 8QQ, United Kingdom
b
School of Accounting, Economics and Finance, Faculty of Business and Law, Deakin University, 221
Burwood Highway, Burwood Vic-3125, Australia

a r t i c l e

i n f o

Article history:
Received 28 January 2013
Accepted 14 March 2014
Available online 26 March 2014
JEL classication:
G11
G12
Keywords:
Systematic co-moment
Size
Value
Momentum
Liquidity

a b s t r a c t
In this paper we evaluate the intertemporal pricing performance
of stock return determinants over the periods surrounding, and
outside of, nancial crises. The analysis focuses on the variables of
size, book-to-market ratio, momentum, liquidity, and higher-order
systematic co-moments. The evidence reveals that over non-crisis
periods the market beta plays an important role in determining the
cross-section of stock returns. Size, value, momentum, and liquidity also exhibit associations with the cross-section of stock returns.
However, over crisis periods most of the variables we examined lose
their explanatory power, suggesting that their usefulness is limited
for investment purposes when nancial markets experience crises.
There is some evidence of coskewness pricing surrounding market
crashes. Practitioners may consider coskewness over crisis periods.
2014 Elsevier B.V. All rights reserved.

Corresponding author. Tel.: +44 (0) 1413305666; fax: +44 1413304939.

E-mail addresses: Chi-Hsiou.Hung@glasgow.ac.uk (C.-H.D. Hung), s.azad@deakin.edu.au (A.S.M.S. Azad),
v.fang@deakin.edu.au (V. Fang).
1
This paper is an extension of the rst authors project entitled Higher-Order Capital Asset Pricing Models and Stock Market
Investment Strategies. Chi-Hsiou Daniel Hung gratefully acknowledges the research grant from INQUIRE UK.
http://dx.doi.org/10.1016/j.intn.2014.03.005
1042-4431/ 2014 Elsevier B.V. All rights reserved.

C.-H.D. Hung et al. / Int. Fin. Markets, Inst. and Money 31 (2014) 1429

15

1. Introduction
Many empirical studies document strong evidence that the cross-section of expected stock returns
is associated with rm-specic characteristics such as market equity capitalization (size) (Banz,
1981), book-to-market ratio (value) (Fama and French, 1993), return momentum (Jegadeesh and
Titman, 1993), and liquidity (Pastor and Stambaugh, 2003). In addition to the market beta of the capital
asset pricing model (CAPM), the SMB and HML portfolios that mimic the risk dimensions of size and
value become pricing factors in Fama and Frenchs (FF) (1993) three-factor model. In the setting of
multifactor models, recent research further evaluates the momentum factor WML, the winner minus
loser hedge portfolio, and liquidity (see, e.g., Avramov and Chordia, 2006).
As opposed to the approach of incorporating the empirically documented pricing factors, Harvey
and Siddique (2000) and Dittmar (2002) examine the role of coskewness and cokurtosis, respectively,
in the U.S. stock markets.2 Furthermore, Chung, Johnson and Schill (2006) provide empirical evidence
to suggest that both the SMB and HML factors proxy for higher-order systematic co-moments. Harvey
and Siddique (2000) compare portfolio characteristics and suggest an association between skewness
and momentum returns.
There has been extensive theoretical and empirical research on the higher-order co-moment models (Kraus and Litzenberger, 1976; Friend and Westereld, 1980; Lim, 1989; Hung et al., 2004; Smith,
2007; Hung, 2008; Doan et al., 2010; among others). However, the literature has not examined the
pricing roles of coskewness and cokurtosis as well as the empirically documented pricing factors in
different phases of the nancial market as dictated by the prominent phenomena of stock market
crashes, bubbles, and the recent credit crunch.3 This study aims to ll this gap and performs litmus
tests to assess the performance of coskewness, cokurtosis, and the SMB, HML, WML, and liquidity (LIQ)
factors in explaining the cross-section of stock returns.
The paper contributes to the literature by revealing important insights into the asset pricing roles
of systematic co-moments and each of the empirical factors over different phases of the nancial
market. We explicitly analyze the WML- and LIQ-augmented FF models, and the four-moment CAPM
that restricts investors preferences to depend on the rst four moments of returns. In addition, we
estimate a model that includes the systematic co-moments together with the size, book-to-market
ratio, momentum, and liquidity factors.
The analyses involve generalized least squares (GLS) estimation, which is crucial for asset pricing
tests (see, Lewellen, Nagel, & Shanken, 2010). We apply the method of Shanken (1992) to correct for
the errors-in-variables bias in the standard errors of coefcient estimates. Furthermore, in order to
avoid the survivorship bias, the sample includes all stocks listed on the NYSE, AMEX, and NASDAQ for
the period from 1926 to 2012.4
We examine two prominent crisis periods: (i) the stock market crash in October 1929 and the
crash in October 1987; and (ii) the dot-com bubble and the credit crunch. These crises have had
signicant impacts on the stock market and the economy. The 1929 crash was followed by the Great
Recession. The 1987 crash set the largest one-day percentage loss of 22.61% ever recorded for the Dow
Jones Industrial Average. The dot-com bubble was built up in the late 1990s and then was followed by
the burst in the early 2000s. The credit crunch evolved from 2006 to 2008 and led to the recent global
nancial crisis.
The paper presents evidence that the return distributions of stock portfolios are signicantly different from normal and exhibits signicant market coskewness and cokurtosis (the third and fourth
systematic co-moments, respectively). Over non-crisis periods the market beta plays an important role
in determining the cross-section of stock returns, while in crisis periods the signicance of coskewness
emerges.

2
Other approaches in identifying systematic risk without the normality restriction include Rosss arbitrage pricing theory
(1976) and the co-lower partial moments of Bawa and Lindenberg (1977).
3
We are grateful for the referees valuable suggestions.
4
Some prior articles only use stocks that were continuously listed on the NYSE. For example, Kraus and Litzenberger (1976)
use stocks that survived through 19261935. Fang and Lai (1997) use stocks that survived through 19691988.

16

C.-H.D. Hung et al. / Int. Fin. Markets, Inst. and Money 31 (2014) 1429

Further, we nd that the loadings on the SMB, HML, WML, and LIQ factors often have signicant
explanatory power for the cross-section of returns on the test portfolios over the non-crisis periods.
However, over crisis periods almost all of these factors entirely lose their explanatory power. Interestingly, there is some evidence of a statistically signicant coskewness premium over the crisis
periods.
Moreover, in the presence of market beta and coskewness the signicance of the empirical factors (size, book-to-market ratio, momentum, and liquidity) is mostly diminished in both crisis and
non-crisis periods. The SMB factor loading and coskewness are the exceptions that show statistical signicance only for size-momentum sorted portfolios over non-crisis periods. Liquidity shows marginal
signicance only for the industry portfolios over the recent credit crunch. Likewise, in the presence
of the empirical factors the market beta and coskewness cease to exhibit statistical signicance in
explaining the cross-section of stock returns.
Our ndings are important for a number of reasons. First, the evidence adds subtle insights to the
literature that the market beta and the empirical factors we have examined only work over noncrisis periods. However, over crisis periods, nancial economists face both theoretical and empirical
challenges in discovering the factors, if any, that determine the cross-section of stock returns.
Second, from a practical viewpoint our nding that the relations between stock returns and the
empirical factors disappear in crisis periods has important implications. Asset managers who build
these empirical factors into investment strategies would nd that when it comes to nancial crises,
their strategies may not deliver the desired investment outcome. Moreover, over crisis periods performance evaluation of managed funds will be biased due to omitted factors in an asset pricing model
(see, Banerjee and Hung, 2011). Banerjee and Hung (2011) thus develop a score function that is robust
to market booms and crashes, and omitted factors in asset pricing models.
On the other hand, employing the coskewness risk in investment strategies over crisis periods
might be benecial as it may help capture the so-called tail risk in the joint return distribution of
risky assets and the market. Given the academic attention to the practical usefulness of higher-order
moments in optimal portfolio allocation (see e.g., Jondeau and Rockinger, 2006), our ndings provide
some support for the importance of coskewness over nancial crisis periods.
The rest of the paper is structured as follows. In Section 2, we outline the higher co-moment models.
Section 3 describes the data and provides summary statistics for the sample. Section 4 contains the
test methodologies. Section 5 presents empirical results, and Section 6 concludes the paper.

2. Higher-order co-moment models

In their pioneering studies, Jean (1971) and Scott and Horvath (1980) show that if returns are
not normally distributed, moments of returns higher than variance matter in maximizing investors
expected utility. Rubinstein (1973) derives higher-moment CAPM that links expected returns to all
moments of returns. Consider an investor who constructs a portfolio by investing his current wealth
in risk-free and risky assets with an objective of maximizing the expected utility of end-of-period
wealth. An investors expected utility can be expressed as follows by using a Taylor series expansion
and ignoring terms of order higher than n for approximation5 :
around mean wealth W
)] =
E[U(W


)
U (n) (W
n=0

n!

)]
E(W
E[W

(1)

) is the nth derivative of the utility U(W

) evaluated at the mean of the investors termiwhere U (n) (W
nal wealth, and E[ ] is the expectation operator. Thus, the expected return of asset i in excess of the

5
The assumptions are that the investors utility function is continuously differentiable and measurable and that the rst n
moments of terminal wealth exist and are nite.

C.-H.D. Hung et al. / Int. Fin. Markets, Inst. and Money 31 (2014) 1429

17

risk-free rate Rf is equal to the weighted sum of co-moments with the weights reecting the measures
of the investors risk aversion:
E(Ri ) Rf =

U n
)]
(n 1)!E[U  (W

)
in (Ri , W

(2)

where Ri is the return on risky asset i. The nth co-moment of risky asset i with the investors wealth
portfolio is the contribution of a marginal increase in the holdings of the security to the corresponding
central moments of the investors future wealth,
E(W
)]n1
in E[Ri E(Ri )][W

for n 2

(3)

At the aggregate market level, assuming homogeneous subjective probability beliefs and separable cubic utility, Kraus and Litzenberger (1976) develop a three-moment CAPM that incorporates
coskewness. Fang and Lai (1997) and Christie-David and Chaudhry (2001) consider a four-moment
CAPM that further includes cokurtosis. The expected return of an asset is linearly associated with the
contributions of an asset to the variance, skewness, and kurtosis of the market portfolio as:
E[Ri ] Rf =  i +  i +  i

(4)

where Ri , and Rf are the return on risky asset i and the risk-free rate, respectively.  ,  , and 
are the market prices of beta (covariance scaled by the market return variance), gamma (coskewness
scaled by the market return skewness) and delta (cokurtosis scaled by the market return kurtosis),
respectively. The beta, gamma, and delta of asset i, respectively, with the market portfolio are6 :
i =
i =
i =

E[(Ri R i )(RM R M )]
2
M

E[(Ri R i )(RM R M ) ]
3
sM

(5)

E[(Ri R i )(RM R M ) ]
4
kM

where Ri , and RM are returns on risky asset i and the market, respectively. The notations of R M and
R i are mean returns on the market and the asset, respectively, and  M , sM , and kM are the standard
deviation, skewness, and kurtosis of the market portfolio, respectively.
Hung (2008) demonstrates a cubic market model for the time-series of security returns, as shown
in Eq. (6). This equation gives the denitions of the beta, gamma, and delta of either a single security
or a portfolio that are consistent with the four-moment CAPM in the cross-section:
2
3
Rpt Rft = C0p + C1p (Rmt Rft ) + C2p (Rmt R mt ) + C3p (Rmt R mt ) + t

(6)

where Rpt and Rft are the returns on portfolio p and the risk-free asset at time t, respectively. Rmt is
the return at time t on the value-weighted global market portfolio. R mt is the population mean value
of RM .
Chung et al. (2006) examine the signicance of SMB and HML factors in the presence of systematic
co-moments using an extended model to the case of n co-moments in which the expected return of
an asset is:
E[Ri ] Rf =

N


n bni

(7)

n=2

6
i ,  i , and i are additive and equal to unity for the market portfolio. Note that  in in Eq. (3) is the co-moment measure at
the level of an individual investors wealth portfolio, and thus is different from Eq. (5).

18

C.-H.D. Hung et al. / Int. Fin. Markets, Inst. and Money 31 (2014) 1429

where Ri , and Rf are the return on risky asset i and the risk-free rate, respectively. bni is the nth-order
co-moment of asset i with the market portfolio, and n is the market measure of risk aversion for the
nth-order co-moment.
3. Data and summary statistics of portfolio returns
The empirical analyses focus on the monthly data of common equities of all NYSE, AMEX, and
NASDAQ rms in the Center for Research in Security Price (CRSP) database from January 1926 to
December 2012. Book value data are collected from Compustat from 1962 to 2012 and from Kenneth
Frenchs data library from 1926 to 1961. The market portfolio is the CRSP value-weighted index, and
the risk-free rate is the one-month Treasury bill rate. The factor returns for SMB, HML, WML, and LIQ
(the monthly return differences between the small and large size portfolios, the high and low book-tomarket portfolios, the high and low prior return portfolios, and the high and low liquidity portfolios,
respectively) are obtained from Kenneth Frenchs data library.7
We set two subsamples corresponding to the two crisis periods. The rst subsample includes the
1929 and 1987 stock market crashes. For the period surrounding the 1929 crash we consider the 162
months from January 1927 to June 1940. For the period surrounding the 1987 crash we consider the
87 months from October 1984 to February 1991. These two crash periods give a total of 249 months.
The second subsample includes the dot-com bubble and the credit crunch from July 1996 to December
2012, a total of 198 months. We then classify 595 months in our sample to examine the non-crisis
periods that lie outside of the above two subsample periods.
For consistency with the extant research (see e.g., Fama and French, 1993; Chung et al., 2006), we
use three sets of test portfolios. First, the 25 two-way momentum-size sorted portfolios are constructed
every month by prior return over the past 212 months and size (market equity). Second, the 25 twoway book-to-market value (B/M)-size sorted portfolios are constructed at the end of each June. Third,
the 30 industry portfolios are formed on the basis of the four-digit SIC code at the end of June each
year. Equally weighted portfolio returns are computed from July each year to June in the next year.
All the portfolio returns are obtained from Kenneth Frenchs data library. The time-series of portfolio
returns has 1032 monthly observations from January 1927 to December 2012. For tests that require
the liquidity factor (only available from August 1962), the time-series of portfolio returns is computed
during the 605 months from August 1962 to December 2012.
Portfolio beta (pt ), gamma ( pt ) and delta (pt ) are computed in each month t = , by using portfolio
returns from t =  60 to  1 as:

t=1

(r r pt )(rMt r Mt )
t=60 pt
t=1
(r r Mt )2
t=60 Mt
t=1
(r r pt )(rMt r Mt )2
t=60 pt

pt =


pt =

t=1

(r r Mt )3
t=60 Mt
t=1
(r r pt )(rMt r Mt )3
t=60 pt


pt =

t=1

t=60

(8)

(rMt r Mt )4

where rp and rM are the returns on portfolio p and the market portfolio (the CRSP value-weighted
portfolio of NYSE, AMEX, and NASDAQ stocks) in excess of the one-month Treasury bill rate; and r pt
and r Mt are the average excess returns in the preceding 60 months for the portfolio and the market,
respectively.
Table 1 presents summary statistics for the return distributions of the CRSP value-weighted market portfolio, the momentum-size, the value-size, and industry portfolios. The mean and volatility
of the monthly market return are 0.91% and 5.43%, respectively. The distribution of the market
returns is leptokurtic, with a maximum of 37.84% and a minimum of 28.97% per month. The

http://mba.tuck.dartmouth.edu/pages/faculty/ken.french/data library.html.

C.-H.D. Hung et al. / Int. Fin. Markets, Inst. and Money 31 (2014) 1429

19

Table 1
Summary statistics of portfolio returns.
Portfolio sorts

Market

Momentum-size

BM-size

Industry

Mean
Maximum
Minimum
Volatility
Skewness
Kurtosis
JarqueBera statistic
KolmogorovSmirnov statistic
% of portfolio beta signicance
% of portfolio gamma signicance
% of portfolio delta signicance

0.0091
0.3784
0.2897
0.0543
0.13
10.24
2259**
0.4165**

0.0144
0.8098
0.4106
0.0699
0.87
14.85
221,585**
0.4265**
100%
98%
100%

0.0133
0.8120
0.4074
0.0672
0.71
12.82
156,023**
0.4216**
100%
92.5%
100%

0.0128
1.7313
0.4148
0.0823
1.89
26.06
708,560**
0.4152**
100%
100%
100%

The sample uses monthly data of all CRSP NYSE/AMEX and NASDAQ common equities from 1926 to 2012 and book value
data from Kenneth French from 1926 to 1961 and Compustat from 1962 to 2012. MARKET is the CRSP value-weighted market
return. Time-series of equally weighted returns for the 25 momentum-size portfolios and for the 25 size and book-to-market
portfolios are computed from January 1927 to December 2012. The 30 equally weighted industry portfolios are from January
1927 to June 2012. JarqueBera statistic tests the normality hypothesis based on skewness and kurtosis. KolmogorovSmirnov
statistic tests the hypothesis of cumulative standard normal distribution. * and ** denote statistical signicance at the 5% and
1% level, respectively. The bottom three rows show the percentages of the portfolios that exhibit signicant beta, gamma, and
delta (dened in Section 3) at the 5% level, respectively.

non-normality in the market return distribution is further evidenced by the JarqueBera statistic and
the KolmogorovSmirnov statistic.
The portfolios sorted by momentum and size have an average return of 1.44% per month with a
standard deviation of 6.99%, a positive skewness of 0.87, and a kurtosis of 14.85. Both the JarqueBera
statistic and KolmogorovSmirnov statistic show that the return distributions of the momentum-size
portfolios are different from normal at the 1% signicance level. All of the 25 momentum-size portfolios
exhibit signicant beta and delta at the 5% level, and 98% of these portfolios exhibit signicant gamma
at the 5% level.
The value (B/M) size two-way sorted portfolios have an average return of 1.33% per month, a
standard deviation of 6.72%, a positive skewness of 0.71, and a kurtosis of 12.82. The return distributions of these portfolios are signicantly different from normal at the 1% level. All 25 value-size
portfolios exhibit signicant beta and delta at the 5% level, and 92.5% of these portfolios exhibit
signicant gamma at the 5% level.
The industry portfolios have an average return of 1.28% per month, a standard deviation of 8.23%,
a positive skewness of 1.89, and a kurtosis of 26.06. All of the industry portfolios are signicantly
different from normal distribution at the 1% level and exhibit signicant beta, gamma, and delta at the
5% level. These statistics suggest that coskewness and cokurtosis might be important risk factors.

4. Tests of models
4.1. Time-series return explanatory power of the cubic market model
We rst examine whether any non-linear association exists in the time-series variation in the
returns between the portfolios and the market using the following regression (as re-printed for
convenience):
2
3
Rpt Rft = C0p + C1p (Rmt Rft ) + C2p (Rmt R mt ) + C3p (Rmt R mt ) + t

(9)

where Rpt and Rft are the returns on portfolio p and the risk-free asset at time t, respectively, and Rmt
is the return at time t on the CRSP value-weighted market portfolio. R mt is the mean value of Rmt for
the entire sample period.

20

C.-H.D. Hung et al. / Int. Fin. Markets, Inst. and Money 31 (2014) 1429

4.2. Cross-sectional return explanatory power of the models

We use the two-pass regression tests of Fama and MacBeth (1973) to evaluate the pricing power
for the (two, three, and four-moment) CAPM, the FF, the WML- and LIQ-augmented FF models. The
rst-stage regression is based on a ve-year estimation window with one month rolling forward,
and the second stage consists of monthly cross-sectional regressions. The rolling regressions give
different coefcients in different months, and thus allow for the factor loadings to evolve over time,
as in Brennan, Chordia and Subrahmanyam (1998) and Chung et al. (2006).
In the cross-sectional tests, we perform both OLS (results available upon request) and GLS regressions, and report the GLS results for brevity. For each of the coefcient estimates we present the
p-value that corresponds to the t-statistic, which is calculated as the mean of the coefcient divided
by its standard error, as in the FamaMacBeth procedure. As noted earlier, the p-values using the
adjustment of Shanken (1992) to correct for the errors-in-variables bias are also presented. The GLS
method performs transformed regressions that use scaled versions of the dependent and independent
variables, and thus reduces the noise caused by the variability of observations. In addition, the GLS R2
is determined by the factors proximity to the minimum-variance frontier of the test portfolios (see
Lewellen et al., 2010). Since all the adjusted R2 values from the GLS tests are very high and close to
one, we do not report these values.
4.2.1. The four-moment CAPM
As presented in Table 1, the test portfolios exhibit non-normally distributed returns with signicant
beta, coskewness and cokurtosis, implying that risk-averse investors may be concerned about extreme
outcomes and that coskewness and cokurtosis may be priced. We test the prediction of the fourmoment CAPM concerning the intercept and risk premia associated with market beta, gamma and
delta. Having estimated portfolio beta, gamma and delta as in (7), we run a cross-sectional regression
of the four-moment CAPM across portfolios in each month to estimate the risk premia t , t , and
t associated with portfolio pt ,  pt , and pt :
rpt = 0t + t pt + t pt + t pt + pt .

(10)

The model predicts that the intercept in the regressions is insignicantly different from zero and
that the coefcients of beta, gamma, and delta are signicant. Since investors prefer positive skewness,
which is implied by non-increasing absolute risk aversion (Kraus and Litzenberger, 1976), the gamma
premium is expected to have the opposite sign of the market skewness. Thus, investors require higher
expected return from assets that have positive coskewness when the market skewness is negative.
Investors aversion to kurtosis is implied by decreasing absolute prudence. Thus, the delta premium
is expected to have a positive sign and investors must be compensated by a higher expected return
for holding a risky asset that positively contributes to market kurtosis.
4.2.2. The FamaFrench model with momentum and liquidity factors
We proceed to analyze whether the WML- and LIQ-augmented FF models explain the cross-section
of stock returns. Specically, in each month t = , for each portfolio, the factor loadings sp , hp , mp , and
lp for SMB, HML, WML, and LIQ, respectively, are estimated on a rolling basis from regressing portfolio
returns on the SMB, HML, WML, and LIQ factor returns from t =  60 to t =  1.8 Next, for each month
t = , we estimate across portfolios a cross-sectional regression of portfolio returns on portfolio beta
p (estimated as in Eq. (8)) and the factor loadings sp , hp , mp , and lp as:
rpt = 0t + t pt + st spt + ht hpt + mt mpt + lt lpt + pt ,

(11)

where  , s , h , m , and l are the market prices of beta and the slope coefcients of sp , hp , mp , and lp ,
respectively. From Eq. (11) we obtain the time-series of monthly cross-sectional regression estimates
and then calculate the test statistics.

8
The estimation of factor loadings using 120 months produces similar results and does not change the conclusion of the
paper.

C.-H.D. Hung et al. / Int. Fin. Markets, Inst. and Money 31 (2014) 1429

21

Table 2
Four-moment-consistent cubic model of industry portfolio returns.
Industry

C0

C1

C2

C3

Adj. R2

Non-durable

0.2801
(3.32***)
0.0311
(0.24)
0.0654
(1.04)
0.2142
(1.66)
0.0134
(0.12)
0.1838
(1.82*)
0.1926
(2.08**)
0.3207
(2.67***)
0.2140
(1.71*)
0.1836
(2.04**)

0.7428
(28.17***)
1.1567
(28.22***)
1.0856
(60.23***)
0.8829
(26.57***)
1.2351
(36.57***)
0.6821
(21.60***)
0.9358
(31.43***)
0.7945
(22.25***)
0.7177
(19.02***)
1.1030
(39.16***)

0.0021
(1.78*)
0.0027
(1.03)
0.0035
(3.70***)
0.0003
(0.16)
0.0011
(0.93)
0.0018
(1.44)
0.0027
(1.77*)
0.0014
(0.67)
0.0028
(1.06)
0.0031
(1.23)

0.0001
(0.90)
0.0003
(1.82*)
0.0001
(2.44**)
0.0001
(1.10)
0.0000
(0.17)
0.0001
(1.07)
0.0001
(1.18)
0.0002
(2.69***)
0.0002
(2.00**)
0.0001
(0.40)

0.7786

Durable
Manufacturing
Energy
High Tech
Telecom
Shops
Health
Utilities
Others

0.7567
0.9293
0.5941
0.8253
0.5938
0.7813
0.6528
0.5830
0.8781

Time-series regressions are performed for evaluating models in explaining returns of the industry portfolios from January
1927 to December 2012. The t-statistics of the slope coefcients in parentheses are calculated by applying the NeweyWest
heteroskedasticity-and-autocorrelation-consistent standard errors. The asterisk of *** denotes the signicance at the 1% level.
2
3
Rpt Rft = C0p + C1p (Rmt Rft ) + C2p (Rmt R mt ) + C3p (Rmt R mt ) + t .

4.2.3. The FamaFrench model with momentum, liquidity and coskewness

Finally, we test the statistical signicance of the empirical factors when coskewness is added into
the model. In each month t, we run a cross-sectional regression as in Eq. (12) of portfolio returns on
factor loadings of size, value, momentum, and liquidity, together with systematic coskewness, i.e.,
portfolio gamma  p as:
rpt = 0t + t pt + st spt + ht hpt + mt mpt + lt lpt + t pt + pt .

(12)

5. Empirical results
5.1. Time-series test of the cubic market model
Table 2 presents the regression results for the cubic market model over the entire sample period for
ten generally classied industry portfolios. The NeweyWest standard errors are applied to correct for
heteroskedasticity and autocorrelation of residuals. The results show that the loading on the excess
market return, C1 (i.e., the market beta), is highly signicant for all the industry portfolios, demonstrating that the excess market return explains the return variations in these industry portfolios. The High
Tech industry portfolio has the highest market beta of 1.24, while the Telecom industry portfolio
has the lowest market beta of 0.68.
The slope coefcient C2 on the squared market return deviation is positive and statistically significant at the 1% level for the Manufacturing industry portfolio. This indicates that the effect of the
market return deviation exhibits a positive and quadratic impact on the Manufacturing industry
portfolio. Both the portfolios of Non-Durable and Shops industries have negative exposures to the
squared market term, albeit marginally signicant at the 10% level. The other industry portfolios do
not appear to have a statistically signicant relationship with the squared market term.
The slope coefcient C3 is positive and statistically signicant for the portfolios of Manufacturing,
Health, Utilities, and Durable industries. The results show that the time-series return variation
in these industry portfolios is not only associated with the variation in the excess market return, but

22

C.-H.D. Hung et al. / Int. Fin. Markets, Inst. and Money 31 (2014) 1429

also exposed to the cubed term of the market return deviation. Overall, the squared and the cubed
terms of the market return deviation in the cubic market model explain the time-variation in returns
on some of the industry portfolios.
5.2. The four-moment CAPM
Table 3 presents the results for the two-, three- and four-moment CAPMs. Panel A, for the size-B/M
portfolios, shows that over the non-crisis periods the estimate of the beta premium is 0.8% per month
and highly statistically signicant in the two-moment CAPM. In the three-moment CAPM the beta
premium is marginally signicant according to the Shanken-adjusted p-value; it becomes insignicant
in the four-moment CAPM. Coskewness and cokurtosis do not exhibit statistically signicant risk
premia.
The results for the two crisis periods show that the beta premium no longer exhibits statistical
signicance in all the three models. Importantly, coskewness is the only variable that shows a negative
and statistically signicant coefcient in the four-moment CAPM for the periods over the 1929 and
1987 crashes and those periods surrounding the dot-com bubble and the credit crunch.
Panel B, for the size-momentum portfolios, documents that in the two-moment CAPM the market
beta commands a highly signicant market premium of 0.86 percent per month during the non-crisis
periods. However, in both the crisis and non-crisis periods, all the coefcients of the market beta,
coskewness, and cokurtosis are insignicant in the corresponding three- and four-moment CAPMs.
As shown in Panel C, for the industry portfolios, the beta premium over the non-crisis periods
remains statistically signicant in both the two- and three-moment CAPMs, but loses its statistical
signicance in the four-moment CAPM. For the periods over the 1929 and 1987 crashes, the coskewness premium is negative and statistically signicant in the three-moment CAPM. However, for the
recent dot-com bubble and credit crunch none of the co-moment models are capable of describing
the cross-section of returns on the industry portfolios.
Overall, the market beta alone appears to have a signicant role in explaining the cross-section of
returns on the three sets of portfolios over the non-crisis periods. In the crisis periods the coskewness
premium shows its statistical signicance sporadically. Importantly, the coskewness premium has a
negative sign when it does show its statistical signicance for all of the test assets.
5.3. WML-augmented FamaFrench model
Table 4 presents the estimation results for the FF and WML-augmented FF models. Over the noncrisis periods for portfolios sorted by size and B/M, Panel A reports that, apart from the WML factor,
the coefcients on all the three-factor loadings on the excess market return (i.e., the market beta), SMB
and HML exert signicant inuences on the cross-section of portfolio returns. The results conrm the
ndings of Fama and French (1993) and Chung et al. (2006).
Importantly, the results for the periods surrounding the 1929 and 1987 crashes show a clear contrast to those of the non-crisis periods. The FF three-factors cease to exert signicant inuences on the
cross-section of portfolio returns, while only the WML factor loading shows statistical signicance.
Over the periods surrounding the dot-com bubble and credit crunch all the factor loadings fail to
explain the cross-section of returns on portfolios sorted by size and B/M.
Panel B reports the results for portfolios sorted by size and momentum. Remarkably, over the noncrisis periods in the WML-augmented FF model, the market price for portfolio beta and all the loadings
on the SMB, HML, and WML factors are signicantly positive. The evidence suggests the importance of
these factors in explaining the cross-section of portfolio returns.
In a stark contrast, the results for the periods surrounding the 1929 and 1987 crashes as well
as for the periods surrounding the dot-com bubble and credit crunch show very little evidence for
the ability of these factors to explain the cross-section of returns on portfolios sorted by size and
momentum. Only the size factor commands a statistically signicant and positive premium over the
periods surrounding the 1929 and 1987 crashes.
Panel C of Table 4 summarizes the results for the industry portfolios. In the FF model the portfolio
beta, size, and value factors all have positive and statistically signicant coefcients over the non-crisis

Table 3
FamaMacBeth regressions for the four-moment CAPM.
Model

Non-crisis periods, 595 months



0.2477
{0.8133)
{0.8413}

0.0895
(0.3545)
{0.3557}
0.0654
(0.3775)
{0.4135}
0.0337
(0.5501)
{0.5800}

0.5977
(0.4196)
{0.4207}
0.7907
(0.4504)
{0.4844}
1.1905
(0.7383)
{0.7571}

0.0959
(0.9288)
{0.9397}

0.1958
(0.0863)
{0.0870}
0.0200
(0.7974)
{0.8120}
0.0135
(0.7789)
{0.7949}

0.6078
(0.3943)
{0.3955}
0.2896
(0.8299)
{0.8423}
0.0519
(0.9877)
{0.9886}

0.0860
(0.8971)
{0.9127}

0.0160
(0.8536)
{0.8540}
0.0004
(0.9933)
{0.9938}
0.0381
(0.3929)
{0.4289}

0.6068
(0.3931)
{0.3943}
1.6804
(0.0951)
{0.1219}
1.3699
(0.6448)
{0.6695}

Dot-com bubble and credit

crunch, 198 months


0.0927
(0.8964)
{0.9040}
3.8863
(0.0428)
{0.0605}

0.7372
(0.5230)
{0.5539}
3.1509
(0.1385)
{0.1700}

1.0276
(0.0649)
{0.0870}
10.1471
(0.4161)
{0.4514}



5.7423
(0.2394)
{0.2759}

-0.1335
(0.3511)
{0.3531}
0.0506
(0.6204)
{0.6734}
0.0967
(0.1025)
{0.1665}

0.4601
(0.2813)
{0.2833}
0.1254
(0.8804)
{0.8981}
1.5569
(0.4382)
{0.5117}

0.4012
(0.5267)
{0.5900}
5.4918
(0.0424)
{0.0853}

6.9923
(0.0707)
{0.1256}

3.6516
(0.4581)
{0.4921}

0.1065
(0.4559)
{0.4577}
0.0482
(0.5259)
{0.5893}
0.0377
(0.5761)
{0.6361}

0.6522
(0.1417)
{0.1434}
0.4782
(0.6382)
{0.6891}
1.1855
(0.4583)
{0.5301}

0.8451
(0.3232)
{0.4004}
0.5975
(0.8695)
{0.8894}

2.0330
(0.6378)
{0.6903}

4.1860
(0.2738)
{0.3108}

0.0943
(0.3055)
{0.3075}
0.0029
(0.9693)
{0.9739}
0.0276
(0.6289)
{0.6825}

0.5841
(0.2129)
{0.2148}
0.3120
(0.7401)
{0.7777}
2.3886
(0.1618)
{0.2360}

0.5275
(0.3567)
{0.4327}
0.9647
(0.5826)
{0.6418}

3.9865
(0.1157)
{0.1826}

C.-H.D. Hung et al. / Int. Fin. Markets, Inst. and Money 31 (2014) 1429

Panel A 25 size-B/M two-way sorted portfolios

0.8032
0.1440
2-M
(0.1816)
(0.0000)
{0.1882}
{0.0000}
0.5322
0.1821
0.0476
3-M
(0.0461)
(0.3572)
(0.2699)
{0.3494}
{0.0906}
{0.4349}
0.1919
0.2136
0.0158
4-M
(0.8235)
(0.5254)
(0.6688)
{0.7169}
{0.8501}
{0.5904}
Panel B 25 size-momentum two-way sorted portfolios
0.8572
0.1278
2-M
(0.0000)
(0.0903)
{0.0952}
{0.0000}
0.4951
0.3203
0.0824
3-M
(0.0968)
(0.1287)
(0.0775)
{0.1589}
{0.1976}
{0.1342}
0.4340
31.6478
0.0251
4-M
(0.6406)
(0.3211)
(0.3991)
{0.6923}
{0.4003}
{0.4747}
Panel C 30 industry portfolios
0.9732
0.0606
2-M
(0.0000)
(0.2298)
{0.0000}
{0.2369}
0.8781
0.0908
0.0253
3-M
(0.0003)
(0.5194)
(0.3397)
{0.0020}
{0.5848}
{0.4181}
0.6899
0.0197
0.0332
4-M
(0.1640)
(0.2476)
(0.9264)
{0.2380}
{0.3270}
{0.9376}

1929 and 1987 crashes,

249 months

The table presents the GLS results of cross-sectional regressions of excess portfolio returns on portfolio beta, gamma, and delta for two-way sorted portfolios based on size and book-tomarket (Panel A), size and momentum (Panel B), and industries (Panel C). Portfolio beta, gamma and delta are computed on a rolling basis every month according to Eq. (8). The slope
coefcients are mean values across all test months. The p-values in parentheses use raw standard errors, and the p-values in brackets use standard errors as in Shanken (1992).
rpt = t + t pt + t pt + t pt + pt .
23

24

Table 4
FamaMacBeth regressions for the WML-augmented FamaFrench model.
Model

Non-crisis periods, 595 months



s

h

Panel A 25 size-B/M two-way sorted portfolios

F-F
0.5906
0.3393
0.4860
0.0141
(0.5320) (0.0006) (0.0029) (0.0000)
{0.5491} {0.0010} {0.0043} {0.0000}
WML-FF
0.0336
0.5800
0.3371
0.4661
(0.1083) (0.0007) (0.0032) (0.0000)
{0.1362} {0.0016} {0.0062} {0.0001}
Panel B 25 size-momentum two-way sorted portfolios
F-F
0.5812
0.4281
0.1600
0.0161
(0.5676) (0.0007) (0.0003) (0.3498)
{0.5837} {0.0012} {0.0006} {0.3700}
WML-FF
0.6119
0.3745
0.3941
0.0115
(0.5645) (0.0003) (0.0009) (0.0034)
{0.5929} {0.0009} {0.0021} {0.0066}
Panel C 30 industry portfolios
F-F
0.7070
0.2870
0.3367
0.0043
(0.8610) (0.0001) (0.0256) (0.0060)
{0.8667} {0.0003} {0.0324} {0.0084}
WML-FF
0.0023
0.6336
0.3654
0.1777
(0.9106) (0.0003) (0.0039) (0.1460)
{0.9170} {0.0008} {0.0074} {0.1773}

m

s

h

0.0294
(0.8890)
{0.8970}

0.0729
(0.0925)
{0.0950}
0.0829
(0.0388)
{0.0410}

0.2746
(0.6830)
{0.6853}
0.2017
(0.7518)
{0.7545}

0.5647
(0.1270)
{0.1300}
0.4860
(0.1722)
{0.1770}

0.2648
(0.6346)
{0.6373}
0.0698
(0.8902)
{0.8914}

0.8103
(0.0000)
{0.0000}

0.0031
(0.9589)
{0.9592}
0.0789
(0.0545)
{0.0572}

0.4907
(0.4392)
{0.4428}
0.3492
(0.5880)
{0.5922}

0.5956
(0.1470)
{0.1502}
0.7769
(0.0587)
{0.0615}

0.6927
(0.1851)
{0.1886}
0.1406
(0.8018)
{0.8040}

0.1879
(0.3309)
{0.3670}

0.0094
(0.8456)
{0.8468}
0.0011
(0.9774)
{0.9777}

0.5275
(0.4251)
{0.4288}
0.5374
(0.4127)
{0.4179}

0.2360
(0.4548)
{0.4583}
0.0669
(0.8288)
{0.8307}

0.0759
(0.8678)
{0.8688}
0.3948
(0.3939)
{0.3991}

m

s

h

m

1.5228
(0.0534)
{0.0560}

0.0088
(0.8766)
{0.8785}
0.0152
(0.7274)
{0.7323}

0.2291
(0.5892)
{0.5952}
0.3392
(0.4051)
{0.4139}

0.1896
(0.8412)
{0.8437}
0.4557
(0.3695)
{0.3785}

0.2575
(0.7346)
{0.7387}
0.1937
(0.6180)
{0.6246}

0.2288
(0.7652)
{0.7695}

0.1827
(0.7337)
{0.7365}

0.0900
(0.1728)
{0.1798}
0.0277
(0.5986)
{0.6055}

0.2589
(0.5399)
{0.5465}
0.2153
(0.5764)
{0.5836}

0.4360
(0.6559)
{0.6611}
0.9436
(0.2169)
{0.2256}

0.2734
(0.7757)
{0.7792}
0.4345
(0.5027)
{0.5108}

0.1793
(0.7127)
{0.7179}

0.4586
(0.4756)
{0.4805}

0.0036
(0.9337)
{0.9348}
0.0265
(0.5116)
{0.5196}

0.0179
(0.9716)
{0.9721}
0.1467
(0.7700)
{0.7742}

0.3515
(0.4669)
{0.4741}
0.5576
(0.2341)
{0.2430}

0.6034
(0.1854)
{0.1925}
0.2384
(0.5674)
{0.5748}

0.0369
(0.9489)
{0.9498}

The table presents the GLS results of cross-sectional regressions of excess portfolio returns on portfolio betas and factor loadings, spt , hpt , and mpt for two-way sorted portfolios based on
size and book-to-market (Panel A), size and momentum (Panel B), and industries (Panel C). Portfolio beta and factor loadings are computed on a rolling basis every month by regressing
portfolio returns for the preceding 60 months on the market, SMB, HML, and WML factors, respectively. The slope coefcients are mean values across all test months. The p-values in
parentheses use raw standard errors, and the p-values in brackets use standard errors as in Shanken (1992).
rpt = t + t pt + st spt + ht hpt + mt mpt + pt .

C.-H.D. Hung et al. / Int. Fin. Markets, Inst. and Money 31 (2014) 1429

Dot-com bubble and credit

crunch, 198 months

1929 and 1987 crashes,

249 months

C.-H.D. Hung et al. / Int. Fin. Markets, Inst. and Money 31 (2014) 1429

25

periods. The inclusion of the WML factor loading, although with its insignicant coefcient, causes the
value factor loading to become statistically insignicant. The results for the periods over the 1929
and 1987 crashes, the dot-com bubble and the credit crunch show that none of these factors are able
to explain the cross-section of returns on the industry portfolios. Overall, the FF three factors often
display statistical signicance in explaining the cross-section of portfolio returns over the non-crisis
periods. During the crisis periods, however, the statistical signicance of almost all of the empirical
factors is diminished.

5.4. LIQ-WML-augmented FamaFrench model

Table 5 presents the results for further including the liquidity factor LIQ into the FF and the WMLaugmented FF models. Since the data for the liquidity factor are only available from August 1962, the
sample period for all the tests is restricted to start from this point onward. Hence, the 1929 crash is
not included for the tests in this sub-section (results reported in Table 5) and the next sub-section
(results reported in Table 6).
Panel A of Table 5 reports that over the non-crisis periods the loadings on the size and value factors
appear to be important determinants for the cross-section of returns on the portfolios sorted by size
and B/M. Over the periods surrounding the 1987 crash the liquidity factor is statistically signicant.
However, for the periods surrounding the dot-com bubble and the credit crunch none of these factors
exhibit statistical signicance.
The results for portfolios sorted by size and momentum presented in Panel B show that the loadings
on the SMB, WML, and LIQ factors have positive and signicant coefcients over the non-crisis periods.
Over the 1987 crash the size factor loses its statistical signicance once the momentum factor is added.
When both the WML and LIQ factors are included, none of the factors exhibit statistical signicance.
Similar to the results reported in Panel A, none of these factors are able to show their associations
with the cross-section of returns on the size-momentum portfolios over the periods surrounding the
dot-com bubble and the credit crunch.
Panel C of Table 5 presents the results for the industry portfolios. Over the non-crisis periods the
coefcient of the portfolio beta is signicant in all the models. The statistical signicance of the value
factor disappears once the momentum factor is added. The results for the periods surrounding the 1987
crash, the dot-com bubble and the credit crunch show that the HML factor loading exhibits evidence
for explaining the cross-section of returns on the industry portfolios. The coefcient on the LIQ factor
loading is only marginally signicant at the 10% level for the period surrounding the 1987 crash, and
becomes insignicant for the periods surrounding the dot-com bubble and the credit crunch.

5.5. Size, value, and momentum factors with beta and coskewness
Table 6 documents the results from estimating Eq. (12) that includes the market beta, coskewness,
and the loadings on the SMB, HML, WML, and LIQ factors. The results reported in Panel A for the sizeB/M portfolios show that only the loading on the HML factor displays statistical signicance over the
non-crisis periods. This suggests that the HML factor loading has explanatory power for the crosssection of returns on the size-B/M portfolios. In both of the crisis periods none of the variables show
statistical signicance.
Panel B presents the results for the size-momentum portfolios. It is interesting to see that over the
non-crisis periods, the size and liquidity factors work very well and both command a positive market
price in which the size premium is economically signicant at 1.12 percent per month. Again, in the
two subsamples for the crisis periods, none of the variables are able to explain the cross-section of
stock portfolio returns. The overall pattern of the results presented in Panel C for the industry portfolios
shows that when all the variables are included in the model none of the variables are able to explain
the cross-section of portfolio returns either over the non-crisis or the crisis periods, except for the
liquidity factor, which shows marginal signicance over the periods of the dot-com bubble and the
credit crunch.

26

Table 5
FamaMacBeth regressions for the WML- and LIQ-augmented FamaFrench model.
Model

Non-crisis periods, 320 months

h
0.6450
(0.0001)
{0.0003}
0.6000
(0.0005)
{0.0010}
0.5512
(0.0011)
{0.0021}
0.2160
(0.4233)
{0.4404}
0.2800
(0.1937)
{0.2228}
0.2302
(0.2872)
{0.3182}
0.3900
(0.0458)
{0.0543}
0.3013
(0.1018)
{0.1248}
0.2510
(0.1677)
{0.1956}

m

0.0603
(0.8618)
{0.8703}
0.0171
(0.9604)
{0.9629}

0.7694
(0.0004)
{0.0008}
0.8536
(0.0001)
{0.0002}

0.2299
(0.4679)
{0.4960}
0.2029
(0.5371)
{0.5627}

Dot-com bubble and credit crunch,

198 months

l

s

h

0.0030
(0.5583)
{0.5830}

0.0283
(0.7011)
{0.7079}
0.0324
(0.6487)
{0.6674}
0.0023
(0.9674)
{0.9693}

0.4299
(0.5721)
{0.5813}
0.4003
(0.6225)
{0.6424}
0.3495
(0.6726)
{0.6907}

0.2530
(0.5801)
{0.5892}
0.2742
(0.5051)
{0.5295}
0.0702
(0.8400)
{0.8492}

0.0431
(0.8867)
{0.8894}
0.1070
(0.7378)
{0.7522}
0.1843
(0.4860)
{0.5117}

0.0153
(0.0258)
{0.0364}

0.0282
(0.8295)
{0.8335}
0.1572
(0.0615)
{0.0772}
0.0437
(0.4016)
{0.4295}

0.4800
(0.5630)
{0.5724}
0.4631
(0.5762)
{0.5980}
0.5705
(0.4907)
{0.5162}

0.3069
(0.4973)
{0.5076}
0.4611
(0.2843)
{0.3122}
0.3404
(0.4455)
{0.4724}

1.5390
(0.0396)
{0.0444}
0.1181
(0.8670)
{0.8745}
0.8366
(0.1279)
{0.1511}

0.0007
(0.8763)
{0.8839}

0.0054
(0.9026)
{0.9049}
0.0223
(0.8064)
{0.8172}
0.0055
(0.9152)
{0.9201}

0.1081
(0.8928)
{0.8953}
0.4566
(0.5294)
{0.5529}
0.5877
(0.4695)
{0.4957}

0.0460
(0.9280)
{0.9297}
0.2694
(0.5974)
{0.6184}
0.3019
(0.5362)
{0.5601}

0.7752
(0.1081)
{0.1167}
0.9181
(0.0973)
{0.1173}
1.0044
(0.0442)
{0.0576}

m

0.3086
(0.7441)
{0.7581}
0.0342
(0.9683)
{0.9702}

0.9902
(0.0409)
{0.0532}
0.7987
(0.1477)
{0.1723}

0.4502
(0.4546)
{0.4805}
0.4890
(0.4369)
{0.4640}

l

s

h

m

l

0.0373
(0.0201)
{0.0281}

0.0088
(0.8766)
{0.8785}
0.0152
(0.7274)
{0.7323}
0.0430
(0.3388)
{0.3479}

0.2291
(0.5892)
{0.5952}
0.3392
(0.4051)
{0.4139}
0.4163
(0.3038)
{0.3130}

0.1896
(0.8412)
{0.8437}
0.4557
(0.3695)
{0.3785}
0.4013
(0.4822)
{0.4905}

0.2575
(0.7346)
{0.7387}
0.1937
(0.6180)
{0.6246}
0.2626
(0.5711)
{0.5784}

0.2288
(0.7652)
{0.7695}
0.6572
(0.3787)
{0.3877}

0.0087
(0.4014)
{0.4103}

0.0188
(0.1434)
{0.1677}

0.0900
(0.1728)
{0.1798}
0.0277
(0.5986)
{0.6055}
0.0690
(0.1760)
{0.1842}

0.2589
(0.5399)
{0.5465}
0.2153
(0.5764)
{0.5836}
0.4836
(0.2296)
{0.2384}

0.4360
(0.6559)
{0.6611}
0.9436
(0.2169)
{0.2256}
0.8377
(0.2282)
{0.2370}

0.2734
(0.7757)
{0.7792}
0.4345
(0.5027)
{0.5108}
0.4389
(0.4711)
{0.4795}

0.1793
(0.7127)
{0.7179}
0.2530
(0.5873)
{0.5944}

0.0028
(0.8462)
{0.8490}

0.0186
(0.0799)
{0.0985}

0.0036
(0.9337)
{0.9348}
0.0265
(0.5116)
{0.5196}
0.0114
(0.7233)
{0.7283}

0.0179
(0.9716)
{0.9721}
0.1467
(0.7700)
{0.7742}
0.1914
(0.6954)
{0.7008}

0.3515
(0.4669)
{0.4741}
0.5576
(0.2341)
{0.2430}
0.0397
(0.9340)
{0.9353}

0.6034
(0.1854)
{0.1925}
0.2384
(0.5674)
{0.5748}
0.8487
(0.0508)
{0.0553}

0.0369
(0.9489)
{0.9498}
0.1356
(0.8056)
{0.8092}

0.0183
(0.0977)
{0.1042}

The table presents the GLS results of cross-sectional regressions of excess portfolio returns on portfolio betas and factor loadings, spt , hpt , mpt , and lpt for two-way sorted portfolios based
on size and book-to-market (Panel A), size and momentum (Panel B) and industries (Panel C). Portfolio beta and factor loadings are computed on a rolling basis every month by regressing
portfolio returns for the preceding 60 months on the market, SMB, HML, WML, and LIQ factors, respectively. The slope coefcients are mean values across all test months. The p-values in
parentheses use raw standard errors, and the p-values in brackets use standard errors as in Shanken (1992).
rpt = 0t + t pt + st spt + ht hpt + mt mpt + lt lpt + pt .

C.-H.D. Hung et al. / Int. Fin. Markets, Inst. and Money 31 (2014) 1429


s
Panel A 25 size-B/M two-way sorted portfolios
0.0324
0.2318
0.5575
F-F
(0.3673)
(0.0042)
(0.3285)
{0.3465}
{0.3851}
{0.0058}
0.2388
0.5839
0.0246
WML-FF
(0.3862)
(0.3490)
(0.0024)
{0.3797}
{0.0044}
{0.4164}
0.2235
0.6017
0.0337
LIQ-WML-FF
(0.1897)
(0.0019)
(0.3829)
{0.2186}
{0.4132}
{0.0035}
Panel B 25 size-momentum two-way sorted portfolios
0.0777
0.1607
0.6763
F-F
(0.0593)
(0.5317)
(0.0006)
{0.5468}
{0.0009}
{0.0692}
0.2575
0.6487
0.0472
WML-FF
(0.0962)
(0.3071)
(0.0008)
{0.3381}
{0.0016}
{0.1186}
0.6233
0.2714
0.0557
LIQ-WML-FF
(0.0668)
(0.2921)
(0.0013)
{0.0855}
{0.3231}
{0.0025}
Panel C 30 industry portfolios
0.0188
0.5747
0.3190
F-F
(0.4940)
(0.0426)
(0.1427)
{0.1579}
{0.0507}
{0.5100}
0.0042
0.5133
0.3605
WML-FF
(0.0522)
(0.0988)
(0.8572)
{0.8660}
{0.0685}
{0.1215}
0.0323
0.7005
0.2711
LIQ-WML-FF
(0.0132)
(0.1968)
(0.2678)
{0.0201}
{0.2260}
{0.2986}

1987 crashes, 87 months

Non-crisis periods, 320 months

s

1987 crashes, 87 months

h

Dot-com bubble and credit crunch, 198 months

l



s

h

m

l



s

h

m

l



Panel A 25 size-b/m two-way sorted portfolios

0.1320 0.2574
0.6399
0.2267
0.0448
(0.1001) (0.7110) (0.3489) (0.0007) (0.5160)
{0.1697} {0.7572} {0.4343} {0.0046} {0.5878}

0.0004
(0.9522)
{0.9601}

0.3574
(0.2294)
{0.3157}

0.0711
(0.1233)
{0.1698}

1.4986
(0.3253)
{0.3816}

0.6918
(0.3178)
{0.3743}

0.2541
(0.4611)
{0.5124}

0.7216
(0.4429)
{0.4951}

0.0050
(0.8385)
{0.8563}

1.9952
(0.1605)
{0.2114}

0.0375
(0.4379)
{0.5137}

0.1610
(0.8044)
{0.8349}

5.2053
(0.3206)
{0.4030}

0.5535
(0.2483)
{0.3310}

0.5367
(0.5753)
{0.6372}

0.2503
(0.3142)
{0.3968}

0.2384
(0.6100)
{0.6677}

Panel B 25 size-momentum two-way sorted portfolios

0.1438
1.1233
-2.1271
0.7409
0.0549
(0.0319) (0.7072) (0.0184) (0.2854) (0.1696)
{0.0730} {0.7540} {0.0488} {0.3726} {0.2515}

0.0245
(0.0031)
{0.0135}

0.0922
(0.7242)
{0.7684}

0.0096
(0.8690)
{0.8835}

0.7563
(0.7320)
{0.7609}

0.1779
(0.8230)
{0.8425}

0.8142
(0.1924)
{0.2459}

0.5211
(0.6443)
{0.6816}

0.0119
(0.6768)
{0.7111}

0.2591
(0.8927)
{0.9046}

0.0196
(0.6321)
{0.6870}

0.5072
(0.3739)
{0.4541}

0.6659
(0.3382)
{0.4200}

0.3027
(0.5945)
{0.6541}

0.1883
(0.7033)
{0.7485}

0.0085
(0.5234)
{0.5912}

0.7941
(0.0707)
{0.1276}

Panel C 30 industry portfolios

0.3057
0.0015 0.3999
(0.9431) (0.3278) (0.3397)
{0.9525} {0.4140} {0.4255}

0.0006
(0.9115)
{0.9261}

0.1882
(0.5076)
{0.5803}

0.0413
(0.4461)
{0.4982}

0.9666
(0.4743)
{0.5248}

0.2294
(0.7677)
{0.7930}

0.9497
(0.0878)
{0.1280}

0.7542
(0.3272)
{0.3835}

0.0159
(0.4836)
{0.5335}

0.2388
(0.8504)
{0.8670}

0.0453
(0.1595)
{0.2360}

0.4662
(0.4640)
{0.5376}

0.4836
(0.3031)
{0.3858}

0.5209
(0.2176)
{0.2991}

0.0541
(0.9219)
{0.9342}

0.0200
(0.0461)
{0.0927}

0.5764
(0.1919)
{0.2717}

0.2118
(0.2896)
{0.3768}

m

0.0320
(0.9350)
{0.9457}

The table presents the GLS results of cross-sectional regressions of excess portfolio returns on portfolio betas and gamma, and factor loadings, spt , hpt , mpt , and lpt for two-way sorted
portfolios based on size and book-to-market (Panel A), size and momentum (Panel B), and industries (Panel C). Portfolio beta and factor loadings are computed on a rolling basis every
month by regressing portfolio returns for the preceding 60 months on the market, SMB, HML, WML, and LIQ factors, respectively. The slope coefcients are mean values across all test
months. The p-values in parentheses use raw standard errors, and the p-values in brackets use standard errors as in Shanken (1992).
rpt = 0t + t pt + st spt + ht hpt + mt mpt + lt lpt + t pt + pt .

C.-H.D. Hung et al. / Int. Fin. Markets, Inst. and Money 31 (2014) 1429

Table 6
FamaMacBeth Regressions for Empirical Factors and Coskewness.

27

28

C.-H.D. Hung et al. / Int. Fin. Markets, Inst. and Money 31 (2014) 1429

6. Conclusions
A number of empirical studies show that the cross-section of stock returns is strongly associated
with the factors of market capitalization, book-to-market equity ratio, return momentum, and liquidity. Prior literature also suggests that the higher-order co-moments of returns are associated with
the cross-section of stock returns. This article conducts a close scrutiny of the pricing ability of these
competing variables in explaining the cross-section of stock returns. The analysis explicitly focuses on
the different phases of the nancial market, as dictated by prominent phenomena of the stock market.
In particular, we examine three sample periods: (i) the periods surrounding the 1929 and 1987 stock
market crashes; (ii) the periods surrounding the dot-com bubble from the late 1990s and the recent
credit crunch; and (iii) the periods outside of these two crisis periods between 1926 and 2012.
The main ndings can be summarized as follow. First, over the non-crisis periods the market beta
exhibits a strong ability to explain the cross-section of stock returns in the two-moment CAPM, the
FamaFrench (FF) model and the WML-augmented FF model. The size factor consistently shows strong
explanatory power in the FF models with or without the WML factor, and the value factor generally
works well for the cross-section of stock returns. Both the WML and the LIQ factors only show statistical
signicance for portfolios sorted by momentum and size.
Second, over the crisis periods almost all of the variables examined lose their explanatory power for
the cross-section of stock returns, and do not show consistent statistical signicance across different
sets of the test portfolios. There is some evidence of coskewness pricing surrounding the 1929 and
1987 stock market crashes.
Our ndings contribute subtle insights to the literature, and show that over crisis periods the
market beta and the empirical factors we examined fail to explain the cross-section of stock returns.
Our results suggest a new frontier for future theoretical and empirical research in this area. As Pastor
and Stambaugh (1999) pointed out, the differences between pricing models are eliminated when
investors beliefs in models contain certain mispricing uncertainty. This mispricing uncertainty may be
more severe during nancial crises. From a practitioners viewpoint, our results suggest that building
these empirical factors into investment strategies will only deliver the desired outcome or achieve the
purposes of risk management and performance evaluation in the absence of nancial crises. There is
some evidence that suggests the need for consideration of the coskewness risk over crisis periods.

References
Avramov, D., Chordia, T., 2006. Asset pricing models and nancial market anomalies. Review of Financial Studies 19, 10011040.
Banerjee, A., Hung, C.-H.D., 2011. Informed momentum trading versus uninformed naive investors strategies. Journal of
Banking and Finance 35, 30773089.
Banz, R., 1981. The relationship between return and market value of common stocks. Journal of Financial Economics 9, 318.
Bawa, V., Lindenberg, E., 1977. Capital market equilibrium in a mean-lower partial moment framework. Journal of Financial
Economics 5, 189200.
Brennan, M.J., Chordia, T., Subrahmanyam, A., 1998. Alternative factor specications, security characteristics, and the crosssection of expected stock returns. Journal of Financial Economics 49, 345373.
Christie-David, R., Chaudhry, M., 2001. Coskewness and cokurtosis in futures markets. Journal of Empirical Finance 8, 5581.
Chung, Y., Johnson, H., Schill, M., 2006. Asset pricing when returns are nonnormal: FamaFrench factors versus higher-order
systematic co-moments. Journal of Business 79, 923940.
Dittmar, R., 2002. Nonlinear pricing kernels, kurtosis preference, and evidence from the cross section of equity returns. Journal
of Finance 57, 369403.
Doan, M., Lin, C.-T., Zurbruegg, R., 2010. Pricing assets with higher moments: evidence from the Australian and US stock markets.
Journal of International Financial Markets, Institutions and Money 20, 5167.
Fama, E., French, K., 1993. Common risk factors in the returns on stocks and bonds. Journal of Financial Economics 33, 356.
Fama, E., MacBeth, J., 1973. Risk, return and equilibrium: empirical tests. Journal of Political Economy 81, 607636.
Fang, H., Lai, T., 1997. Co-kurtosis and capital asset pricing. Financial Review 32, 293307.
Friend, I., Westereld, R., 1980. Coskewness and capital asset pricing. Journal of Finance 35, 897913.
Harvey, C., Siddique, A., 2000. Conditional skewness in asset pricing tests. Journal of Finance 55, 12631295.
Hung, C.-H. D., 2008. Return predictability of higher-moment CAPM market models. Journal of Business Finance and Accounting
35, 9981022.
Hung, C.-H.D., Shackleton, M., Xu, Z., 2004. CAPM, higher co-moment and factor models of UK stock returns. Journal of Business
Finance and Accounting 31, 97122.
Jean, W., 1971. The extension of portfolio analysis to three or more parameters. Journal of Financial and Quantitative Analysis
6, 505515.

C.-H.D. Hung et al. / Int. Fin. Markets, Inst. and Money 31 (2014) 1429

29

Jegadeesh, N., Titman, S., 1993. Returns to buying winners and selling losers: implications for stock market efciency. Journal
of Finance 48, 6591.
Jondeau, E., Rockinger, M., 2006. Optimal portfolio allocation under higher moments. European Financial Management 12,
2955, 2006.
Kraus, A., Litzenberger, R., 1976. Skewness preference and the valuation of risk assets. Journal of Finance 31, 10851100.
Lewellen, J., Nagel, S., Shanken, J., 2010. A skeptical appraisal of asset-pricing tests. Journal of Financial Economics 96, 175194.
Lim, K.-G., 1989. A new test of the three-moment capital asset pricing model. Journal of Financial and Quantitative Analysis 24,
205216.
Pastor, L., Stambaugh, R., 1999. Comparing asset pricing models: an investment perspective. Journal of Financial Economics 56,
335381.
Pastor, L., Stambaugh, R., 2003. Liquidity risk and expected stock returns. Journal of Financial Economics 111, 643685.
Ross, S., 1976. The arbitrage theory of capital asset pricing. Journal of Economic Theory 13, 341360.
Rubinstein, M., 1973. The fundamental theorem of parameter-preference security valuation. Journal of Finance and Quantitative
Analysis 8, 6169.
Scott, R., Horvath, P., 1980. On the direction of preference for moments of higher order than the variance. Journal of Finance 35,
915919.
Shanken, J., 1992. On the estimation of beta-pricing models. Review of Financial Studies 5, 133.
Smith, D., 2007. Conditional coskewness and asset pricing. Journal of Empirical Finance 14, 91119.