Measuring and Pricing Risk in A Quadrant-Dependent

Model ∗
René Garcia Octave Jokung Jean-Christophe Meyfredi
Edhec Business School Edhec Business School Edhec Business School

September 23, 2008

Abstract

Recent empirical evidence has revealed that betas of equity portfolios or individual stocks
are state-dependent or time-varying. In this paper, we explore a four-state model where
betas may differ not only in the positive and negative quadrants but also when the market
return is positive while the portfolio return is negative, and inversely when the market re-
turn is negative while the portfolio return is positive. We find strong support for the model
both in market capitalization and book-to-market decile portfolios and in individual secu-
rities. We also estimate the conditional exposure to coskewness risk and the corresponding
prices. We further estimate quadrant-dependent prices of risk and show that estimated risk
premia are close to the average observed excess returns.

Keywords: Asymmetric Beta Models, State-Dependent Models, Time-varying Betas JEL

Classification: C1,C5,G1

∗
Address for correspondence: René Garcia, Edhec Business School, Finance, Law and Accounting De-
partment, 393, Promenade des Anglais, BP 1116, 06 E-mail : [email protected]. The first and
third coauthors are members of the Edhec-Risk Research Center in Risk and Asset Management. The first
author is a research fellow of CIRANO and an associate fellow of CIREQ.
1 Introduction

Recent empirical evidence has revealed that betas of equity portfolios or individual stocks
are state-dependent or time-varying. Ang and Chen (2002) show that betas of equity
portfolios computed conditionally on the positive or negative orthants, so-called upside and
downside beta coefficients, exhibit some asymmetry. This is analogous to the asymmetry in
the upside and downside exceedance correlation put forward by Longin and Solnik (2001)
for international equity markets. Cho and Engle (1999) use a double-beta model with
E-GARCH specifications to show that news affect asymmetrically the betas of individual
stocks. These asymmetries are important to account properly for risk when explaining
expected returns.
In this paper, we explore a four-state model where betas may differ not only in the
positive and negative quadrants but also when the market return is positive while the
portfolio return is negative (plus-minus), and inversely when the market return is negative
while the portfolio return is positive (minus-plus). We first estimate this model with daily
returns on the market capitalization and book-to-market decile portfolios. We find that
the upside and downside betas are positive and significantly different from each other for
almost all portfolios. An important finding is that the usual single-beta estimated value is
found to be determined mainly by joint downturn periods between market and portfolios,
especially for portfolios ranked by size. We confirm these results by estimating the same
model for all stocks of the S&P 500 index over a ten-year period at a daily frequency.
An important reason for separating betas according to the four quadrants is to increase
the stability of the risk measure over a long period of time. According to several measures of
stability, we find that the conditional betas exhibit much more stability than the single beta
especially for individual stocks. This result is particularly important for event studies where
a risk model is essential to determine abnormal returns. Maintaining the market model,
up to differentiating the states of the market, will provide a good and stable correction for
risk.
A key issue related to quadrant-dependent betas is to establish whether or not the
market prices these risks differently. There is some evidence that preferences are state-
dependent (see Campbell and Cochrane (1999), and Chabi-Yo, Garcia, and Renault (2008)),
which will rationalize a state-dependent risk premium. However, the state-dependence in
these papers are more related to the business cycle. The model developed in this paper is
more in line with loss aversion preferences, where investors do not have the same marginal
utility for gains and losses (see Barberis, Huang, and Santos (2001)). To estimate the
prices of risk, we run a cross-sectional regression for all individual stocks over the full

1
sample period to explain mean returns in each quadrant. The R2 are very high in both the
upside and downside quadrants, rehabilitating the market beta as the main risk factor.
Another way of capturing the asymmetric relationship between portfolios or individ-
ual stock returns and market returns is to introduce a measure of coskewness. This is
the approach proposed by Kraus and Litzenberger (1976) for unconditional skewness and
more recently Harvey and Siddique (2000) for unconditional skewness. It is natural to
ask whether there remains some asymmetry after accounting for the up and down states.
Therefore, we enrich the model by adding the market return squared to the market model.
Without distinguishing for the upside and downside states, the coefficient γ of this new
skewness factor is mostly unsignificantly different from zero for size portfolios and signifi-
cantly negative for book-to-market portfolios. Separating the upside and downside obser-
vations, we now find that γ is positive and strongly significant for almost all size portfolios
in both up and down states. For the book-to-market portfolios, we observe that it is mainly
the downside quadrant that explains the overall result. The downside γ is negative and
strongly significant, while the upside γ is not significantly different from zero for almost all
portfolios. This confirms the presence of a bear market asymmetry put forward in Longin
and Solnik (2001) on top of the negative and positive return asymmetry.
We also test whether this coskewness is priced by the market. We first run a cross-
sectional regression for all stocks in the S&P 500 index. The coskewness price is positive
and significantly different form zero. It adds marginally but significantly to the R2 of the
regressions.
When we put together the estimated risk measures and risk prices to compute the
risk premia, we can explain a large part of the portfolios ranked according to market
capitalization and book-to-market value. The beta premium is the most important, but
the premium for coskewness, although small, is not negligible for some portfolios.
Several strands of literature are related to state-dependent measures of risk and market
risk premia. Pettengill, Sundaram, and Mathur (1995) examine the conditional relation
between beta and returns, controlling for the sign of realized market premia, while Hung,
Shackleton, and Xu (2004) extend their model with higher-moment factors. Another class is
the regime-switching modeling of beta risk. Turner, Startz, and Nelson (1989) account for a
time-varying risk premium in stock returns, while Maheu and McCurdy (2000) use a Markov
switching model which incorporates duration dependence to capture nonlinear structure in
both the conditional mean and the conditional variance of stock returns. Markov switching
models have also been used in consumption-based asset pricing models (see for example
Cecchetti, Lam, and Mark (1990) and Bonomo and Garcia (1994) ).
Breakdowns of the market beta along cash flow news and discount rate news have also

2
been proposed. Campbell and Vuolteenaho (2004) use a model with two betas, where the
required return on a stock is determined by its bad beta (covariation with market cash-
flow shocks), that earns a high premium, and by its good beta (covariation with market
discount rates), that earns a low premium. In an earlier paper, Campbell and Mei (1993)
proposed another beta decomposition. The market beta of a portfolio is the sum of a
cash-flow beta (covariation with market return of portfolio cash-flow news) and a discount
rate beta (covariation with market return of portfolio discount rate news). Campbell, Polk,
and Vuolteenaho (2005) have a four-beta encompassing model, which determines whether
stocks bad betas and good betas are determined by the characteristics of their cash flows,
or whether instead they arise from the discount rates, possibly driven by sentiment, that
investors apply to those cash flows.
Of course, beta and risk premia instability can be captured by time-variation through
conditioning on firm characteristics or macroeconomic or global financial factors (see in
particular Shanken (1992), Harvey (1991), and Ferson and Harvey (1991) and Ferson and
Harvey (1999) for a relationship between betas and proxies for the state of the economy).
Wang (2003) notes that conditional beta models are empirically challenging since a mod-
eling assumption has to be made to relate the betas to conditioning variables. Ghysels
(1998) shows that several well-known time-varying betas are seriously misspecified.
The rest of the paper is organized as follows. Section 2 describes the four-quadrant
beta model and its estimation, as well as asymmetry and stability tests conducted on the
model. In section 3, we describe the data used to estimate and test the model. Section
4 presents and discusses the results. We conclude and outline some implications of our
results in section 5.

2 The Four-Quadrant Beta Model

Most tests of the CAPM start from the following benchmark model:

rAt = αA + βA rM t + εAt , (1)

where rAt is the return on an asset or portfolio A at time t, rM t is the return on the market at
time t, and εAt denotes a mean-zero residual term. The coefficients α and β are the intercept
and the beta of asset i respectively. This standard benchmark model postulates a linear
relationship between the assets returns and the markets returns for any individual asset
or portfolio. To measure the comovements between the individual assets and the market
portfolio, it rests on the concept of correlation. However, several studies have suggested
that this relationship could change with the state of the market. Recently, Ang, Chen, and

3
Xing (2006) have separated the betas into the positive and negative comovements in their
study of an symmetric correlation between portfolios and the market. This asymmetric
correlation had been put forward by Longin and Solnik (2001), Ang and Chen (2002) and
Harvey and Siddique (2000), among others, with various techniques. The first two papers
use exceedance correlation to measure correlation beyond some return levels, while the last
one uses coskewness.
Norsworthy, Schuler, Morgan, and Li (2001) and Norsworthy, Schuler, Morgan, and Li
(2003) propose a so-called four-state model to capture potential asymmetric and skewed
relationships between asset returns and market returns. The idea is to partition the ob-
servations into four sets according to the signs of the asset and the market returns as
follows:
- State 1 corresponds to observations when both the sign of the asset return and the
sign of the market return are non negative;
- State 2 is related to a positive sign of the market return and a negative for the asset
return;
- State 3 corresponds to observations where both the sign of the asset return and the
sign of the market return are negative;
- State 4 groups all the other observations.
Norsworthy, Schuler, Morgan, and Li (2003) rationalize such a model by prospect theory
and provide some axiomatic foundations.

2.1 Model Estimation

We summarize the partition as:

State Sign of Asset Return Sign of Market Return

1 rA ≥ 0 rM ≥ 0
2 rA < 0 rM ≥ 0
3 rA < 0 rM < 0
4 rA ≥ 0 rM < 0
Given an asset, for each state, we have a single market model described as follows:

rAit = αAi + βAi rM t + εAit i = 1, ...4 (2)

The four-state or four-quadrant model1 is obtained by taking into account the last four
1
Norsworthy, Schuler, Morgan, and Li (2001, 2003))also investigate a so-called rotated model when the
reference point is not zero. They consider the expected asset return and the expected return of the market,
and rotate the horizontal axis in order to pass through the expected point given by the two expected
values. Jokung and Meyfredi (2004) introduce also a so-called translated version of the model obtained
by changing the origin, the coordinates of the new origin being exactly the expected values of the returns.

4
equations:

X X
rAt = αi I i + βi rM t Ii + eAt i = 1, ...4 (3)
i i

1 if rA and rM are in state i
where Ii = .
0 otherwise

This type of partitioning is more in line with the concept of quadrant-dependence used in
insurance economics (see Denuit and Scaillet (2004) and references therein). To generalize
asymmetric correlation, Garcia and Tsafack (2006) model and test asymmetric dependence
in international equity and bond markets.
Harvey and Siddique (2000) estimate another type of asymmetry or skewness by intro-
ducing the square of the market return along with the market return:

rAt = αA + βA rmt + γA rm2t + εAt (4)

Given our four-quadrant model, an important issue is to verify whether the squared-
return remains a significant risk factor in each quadrant. Indeed, the asymmetry captured
by the previous partition may not account for the coskewness relationship put forward by
Harvey and Siddique (2000). Therefore we estimate the following model:

X X X
2
rAit = αAi Ii + βAi rM t Ii + γAi rM t Ii + εAt i = 1, ...4 (5)
i i i

1 if rA and rM are in state i
where Ii = .
0 otherwise

2.2 Asymmetry and Stability Tests

Estimating this model will allow us to test several hypotheses of interest. The most inter-
esting one, already looked at by Ang, Chen, and Xing (2006) for a number of portfolios,
is the equality of β1 and β3 . This asymmetry between up and down concomitant moves
would be consistent with the exceedance correlation evidence. Another important test is
symmetry in up and down markets. This will be tested by β1 + β2 = β3 + β4 .
One main reason for partitioning the data is to hope that the quadrant regressions will
exhibit more stability of the beta coefficients, especially when looking at long spans of data.
Indeed, there is by now a lot of evidence in favor of time-varying betas when considering
These two models rest on expected values, which are themselves determined by the model. Taking simply
the average past returns is not consistent with the postulated model. These concepts could be interesting
to explore with respect to expectations formulated by analysts outside the model itself.

5
a single beta model (see in particular Bollerslev, Engle, and Woolridge (1988), and Chou,
Engle, and Kane (1992)). This time-variation may come from not distinguishing between
upside and downside betas.
First, we conduct a classic Chow test over two sub-samples by splitting the sample
into two equal parts. Second, we analyze the stability of the beta parameters over moving
windows of 252 and 1260 trading days (one and five years of data respectively). We also
look at the evolution of the betas over non-overlapping time periods of 252 trading days.
We run tests for the presence of coskewness per quadrant. In other words, we conduct
the same type of tests as Harvey and Siddique (2000) but on a quadrant-per-quadrant
basis. Last, we test if these asymmetric risks are priced by the market, and to what extent
the prices of risk are different in the different states. We will do it in the usual Fama and
MacBeth (1973) framework, by running cross-sectional regressions on individual securities
and on size and book-to-market portfolios.

3 Data Description

Three sets of daily data are used for empirical validation. Two of them consist in Fama
and French (FF) portfolios (Market Equity and Book-to-Market decile portfolios) over
the period 1 January 1963 to 29 December 2006, resulting in 10951 observations for each
portfolio, collected on the Kenneth French’s website. The last set corresponds to the shares
constituting the Standard and Poor’s equity index between 25 June 1998 and 23 August
20072 , resulting in 2610 observations for each share, obtained from Datastream.
Descriptive statistics for the two first sets and for the market index are presented in
Table 1. As we can observe, all data series exhibit means near zero. All t-tests confirm that
the means are not statistically different from zero, except for portfolio 1 in the book-to-
market panel. All series are negatively skewed and exhibit strong excess kurtosis whatever
the criterion chosen to form the portfolios. This presumption of non-normality is confirmed
by Kolmogorov-Smirnov tests (not reported here).
We also report the number of observations in each quadrant to determine whether
specific patterns exist. As we can see in Table 2, the first and third quadrants contain
the largest part of the total observations, with about 20% more observations in the first
than in the third quadrant. In general, the higher the market capitalization, the higher
the number of times the asset returns follow the market move (both a rise or a drop in
the market). This is in accordance with the view that large firms behave roughly as the
market. Of course, we obtain the opposite conclusion for adverse moves. Only 7% of the
2
Composition of the SP500 is that on 23 August 2007.

6
total number of opposite move observations fall in the second and fourth quadrants for the
highest size decile portfolio, while the percentage increases to close to 23% for the lowest
size decile portfolio.
When we look at the book-to-market portfolios, we find a similar pattern but not as
monotone as for the size criterion. Growth companies behave more like the market than
value companies. About 10% of the observations fall in the opposite-move quadrants in the
lowest decile portfolio, while about 20% belong to these quadrants for the highest decile.
This table already gives us an intuition about the potential failures of the standard
market model. Given that the numbers of observations in concomitant up and down moves
are of the same order of magnitude, the market model will fail if the beta loadings are
different in the positive and negative quadrants. A second source of failure may come from
a non negligible number of opposite moves for some portfolios. This is most notably the
case for small and value portfolios, where the strongest departures of the single beta model
have been documented.
In the next section, we will explore the empirical support for these two potential sources
of failure of the single beta model.

4 Empirical Results

In this section we start by estimating the four-quadrant model for the size and book-
to-market portfolios. We then test in various ways the stability of these risk exposures
compared to a model with a single beta exposure. We also estimate the exposures of the
size and book-to-market portfolios to the coskewness factor. In the last subsection, we
estimate the prices of the beta and coskewness risks.

4.1 Estimates of betas per quadrant

We start our validation process with the estimation of both the market model and the
4-quadrant model parameters for equally-weighted size and book-to-market portfolios. Es-
timation results are reported in table 3.3
Slope coefficients of the market model, hereafter called traditional betas or simply betas
and denoted by β, are all positive and strongly significant. For the book-to-market portfo-
lios, betas decrease as the book-to-equity (B/E) ratios increase except for the two highest
portfolios. We can also note that offensive (higher than one) betas correspond only to the
3
We chose to use equally-weighted portfolios rather than value-weighted portfolios in order to limit
the influence of the market value on the book-to-market portfolios. For homogeneity reasons, we also
selected equally-weighted portfolios for the size criterion. All parameters have been estimated using a
White-consistent correlation matrix.

7
two lowest book-to-equity ratios. Considering the set of size portfolios, β is generally an
increasing function of market capitalization. For this criterion, offensive betas are obtained
only for the highest market capitalization portfolio. This means that small capitalization
portfolios are more defensive assets than large capitalization portfolios, which tend to over-
react to market moves. Of course, it is the opposite for the book-to-market criterion.
For the four-quadrant model alternative, results obtained for β1 and β3 are quite similar
to the results of the single-beta model. These two slope coefficients behave like the betas
in the traditional market model. They both increase as the market equity increases and
decrease with the book-to-equity ratio, except for the two last portfolios. However, the
estimated values display interesting differences with respect to the single-beta estimated
values. For the portfolios sorted on size, the values for β1 are consistently smaller than the
ones for β, except for the last two portfolios. The difference is also decreasing with size.
For the smallest portfolio, the difference is quite sizable at more than 30%, but it is still
in the order of 10% for portfolio 6 and 7. The estimated values for β3 are much closer to
the values of the single-beta model. This is quite an important result since it says that the
single-beta value, used extensively in many applications from event studies to investment
evaluation, is determined by the joint downturns between market and portfolios. Yet most
of the observations are in the first quadrant.
The same pattern is observed for the portfolios ranked according to the book-to-market
criterion. Except for the two lowest book-to-market portfolios, the values estimated for β3
are closer to the single-beta estimated values than the values for β1 .
These results confirm the results of Ang and Chen (2002) and the literature on asym-
metric correlations showing that portfolios are more sensitive to market reactions in bear
markets than in bull periods, except for largest capitalization and lower book-to-market
firms .
The estimation results for β2 and β4 require more attention. All β2 are negative, ex-
cept for the two highest portfolios ranked by size and for the two lowest book-to-market
portfolios. However, few estimated betas are significantly different from zero for both sets
of portfolios (3 for the size criterion and 4 for the book-to-market criterion). The situ-
ation is different for β4 for the book-to-market portfolios. They are all negative, except
for the lowest portfolio and are almost always significantly different from zero. Another
noteworthy finding is that the estimated value is about the same, irrespective of the decile
portfolio. A higher book-to-market value provides a refuge in market downturns. This
effect is totally absent for the size portfolios, where no systematic pattern is detected and
almost all estimated betas are statistically zero.

8
4.2 Symmetry and Stability Test Results

Given the large differences between β1 and β3 for size portfolios, it is not surprising that
the test results for the equality of β1 and β3 in the last column of Table 3 indicate a strong
rejection. For all portfolios, we reject the equality at a 0.1% level. For the book-to-market
portfolios, we also reject the equality between β1 and β3 except for the second portfolio.
An assumption often tested in the literature is the asymmetry of market betas between
bull and bear markets. In the context of our four-quadrant model the market beta for bull
markets is equal to β1 + β2 , while the beta for bear markets is equal to β3 + β4 . The last
column of Table 3 reports the results of a test for the equality of these two bear and bull
betas. The equality is strongly rejected for all portfolios. It is interesting to note that the
strong rejection for book-to-market portfolios comes mainly from the β4 coefficient.
One of the reasons we have given for estimating a quadrant-dependent model is to
achieve a greater stability of the betas than with a single beta. Our first stability test, a
classical Chow test, aims at establishing whether each of the four betas can be considered
constant over the first and the second half of the sample. The sample is long enough to
provide samples with two very different histories. Results are reported in Table 4
The number of stable slope parameters for the portfolios ranked by the book-to-equity
value is always higher for the four-quadrant model than for the traditional market model.
It is also interesting to note that the most stable parameters are β1 and especially β3 , where
for eight out of ten portfolios one cannot reject the equality of the parameter over the two
sub-samples. However, we do not reach the same conclusion for the set of portfolios ranked
by their market capitalization. The stability of the two other betas confirms that these
parameters are not significantly different from zero in the two sub-periods.
Given these quite different results for the size and book-to-market portfolios, we estimate
the four-quadrant model for 450 stocks in the S&P500 index 4 and run the same Chow test
over the two equal sub-samples. Results in Table 4 for the individual securities in the
S&P500 index confirm that the four-quadrant model exhibits far more stability than the
market model. About 80% of the β1 and β3 coefficients exhibit stability over the two equal
sub-samples from June 1998 to August 2007. This has important implications for capturing
abnormal returns in event studies. While a single-beta model estimated over a long period
may not be satisfying, a model separating the concomitant upside and downside movements
appears more appropriate for a large majority of securities.
To check the robustness of this stability we split the sample into non-overlapping pe-
riods of 252 days (one year). We then estimate the single-beta and the four-beta models
4
Only 450 shares that make up the S&P index on 23 August 2007, existed 5 years before, at the
beginning of the recorded series, on 22 August 1997

9
for the two sets of Fama-French portfolios over each year and test the equality with the
corresponding parameters in the following year. Results are reported in the A columns of
Table 5. In columns B we report test results for an estimation period of 756 trading days
and an equality test with the following year as before. For both experiments, we indicate
in the table the number of times we cannot reject the equality of the parameters. For the
first experiment, we have a collection of 43 test results, while for the second we have only
14 of them.
Results of the two experiments confirm the previous findings. The slope parameters
in quadrants 1 and 3 are more stable than the single beta for both sets of portfolios.
Moreover, β3 tends to be more stable that β1 . This may suggest that investor reactions
in downturns are more uniform than in upturns, where different waves of optimism may
change the risk exposure. The last line for the individual securities of the S&P500 also
confirms the previous strong finding of stability in average. This is illustrated in Figures
1 and 2 showing the histograms of the non-rejections of the beta stability over one-year
intervals. The figures show that β3 is more stable than β1 since the frequencies of 7 or 8
non-rejections are higher than for β1 . For comparison, we show the same histogram for a
single beta in Figure 3. It shows clearly that the frequencies for 7 or 8 non-rejections are
lower than for the two other betas. We do not report these histograms for the betas of the
two other quadrants, β2 and β4 . Although they show a high frequency for the maximum
number of non-rejections, we know that the stability is around a zero value.
Finally, we illustrate the variation of betas over time for a unique beta and for the first
and third quadrants. The betas are estimated over yearly non-overlapping time periods
and plotted in Figures 4 and 5 for three size and book-to-market portfolios respectively.
The three portfolios represent the lower, median and higher deciles. We can observe that
the estimated values for β, β1 and β3 , show relatively low volatility. The average value
of the betas varies between 0.5 and one depending on the various portfolios. Small-firm
(portfolio 1) and value (portfolio 10) portfolios exhibit more volatility than the large-firm
and growth portfolios.
We do not plot the estimated values for β2 and β4 since they are much more volatile
around a mean value of zero. Therefore, the stability test results reported previously were
misleading since the non rejection of stability simply meant that they were not significantly
different from zero.
Although our test results showed that estimating the betas per quadrant increases the
stability of the estimates, the graphs in Figures 4 and 4 have shown that there remains
some time variation in the estimates of β1 and β3 , the two parameters that appear to be
significantly different from zero in all portfolios. One potential source of variation may be

10
the remaining asymmetry due to coskewness. Therefore, in the next section, we explore an
extended model in which squared market returns are considered along with market returns
in each quadrant.

4.3 Estimates and Tests of the Harvey and Siddique (2000) Model per Quad-
rant

We first estimate the model in (4), which is the model proposed by Harvey and Siddique
(2000) to account for co-skewness in addition to the market beta as a risk factor. Estimation
results are reported in Table 6. The main finding is that all the γ parameters corresponding
to the co-skewness coefficient are highly significant when portfolios are sorted on the book-
to-market criterion. Moreover, the sign is uniformly negative, apart from the first two
portfolios. There is no clear pattern in the magnitude of the coefficients, but it should be
noted that the addition of the co-skewness factor barely affects the β estimates.
Results for the portfolios ranked by size are not as clear as for the book-to-market
portfolios. While it appears that the first two portfolios, the smallest capitalization, exhibit
a negative and significant relationship with squared market returns, the coefficients for the
other portfolios are either not significant or small in magnitude.
Our results are in line with Barone Adesi, Gagliardini, and Urga (2004) for the book-to-
market portfolios. The co-skewness is positive for the lowest book-to-market portfolios and
negative for the others. However, for size portfolios, we only find a negative co-skewness for
the smallest capitalization portfolios. We do not find a positive and significant relationship
with coskewness for the bigger capitalization.5
We now want to check how this co-skewness relationship is behaving in each quadrant.
Estimation results for the size portfolios, reported in Table 7, are quite informative and
rich in implications. First, we now find a clear and significant relationship for the size
portfolios in the first and third quadrants, with co-movements between the market and
the portfolios that are both positive or negative. In the positive quadrant, the estimated
coefficients for γ1 are all positive and large in magnitude, except for the largest portfolio.
We find a similar pattern for the γ3 parameters but the estimated coefficients are much
smaller. The negative relationship for the largest capitalization starts earlier than for the
first quadrant. The high middle-size portfolios exhibit no significant relationship in terms of
co-skewness. None of the estimated coefficients for γ2 and γ4 are significantly different from
zero. Therefore, while the relationship with co-skewness will not appear with a single-beta,
single-gamma relationship, it shows up in a strong and significant way when we partition
the observations. It should be recalled that the data points contained in the first and third
5
The choice of daily data instead of monthly may be the source of the difference.

11
quadrants represent more than 80% of the observations. Therefore the noise added by the
opposite co-movements between the portfolio returns and the market returns was occulting
the co-skewness effect. It should also be noticed that the estimated values for β1 and β3
did not practically change with the addition of squared returns. Moreover, none of the
coefficients estimated for β2 and β4 are now significantly different from zero.
Our findings for portfolios ranked by the book-to-market ratio are even more interesting.
Results are reported in Table 8. They show that most of the co-skewness effect is concen-
trated in the third quadrant, meaning that this particular asymmetry is linked to joint
down movements of portfolios and market. The sign is now everywhere negative except
for the smallest growth portfolios. The effect is also generally increasing in absolute value
from growth to value portfolios. Again the betas remain very close to the betas estimated
without the squared market returns.
The overall effect of the two risks, considered by quadrant, on expected returns is
evaluated in the two next sections.

4.4 Estimates and Tests of Risk Prices

Following the methodology proposed by Fama and MacBeth (1973), we estimate in a sec-
ond pass the prices associated with the market and coskewness risks in a cross-sectional
regression of expected (average) returns on the estimates of the βs and γs. As the num-
ber of portfolios used in the previous analysis is small, we conduct the analysis with the
individual securities.
We report the estimated risk prices over all securities in Table 9. We kept only 450
securities out of the original 500 that constituted the S&P500 on the last date of our sample
and for which we had all quotes. We first regress the average return of each security over the
period 25 June 1998 to 29 December 2006 on the estimated βs and γs over the same period.
Not surprisingly, unconditional estimated risk prices reported in Panels A and B are close
to zero, the R2 increasing slightly with the addition of the coskewness risk measure.
We then run the same regression in each quadrant with the beta and gamma coefficients
estimated by quadrant. The first striking result is the high values of the R2 in the q1 and
q3 quadrants (0.64 and 0.79, and 0.66 and .85 in Panels A and B respectively). As it could
have been anticipated, the highest risk prices ηβ are for the similar comovements in the
asset and the market. Indeed, the asset does not offer any diversification or insurance
value and the market requires more of a premium. In the q4 line, where the asset goes
up when the market goes down, the market does not require any premium. The price is
small and statistically not different from zero. In the q2 line, when the asset goes down
contrary to the market, the price is positive and significant but smaller than in the q1 and

12
q3 quadrants. For the two latter cases, it is interesting to note that the prices for the up
and down movements are not quantitatively different in absolute value.
The same comments remain true when we introduce the price of coskewness risk ηγ .
This can be seen in Panel B of Table 9. The price of beta risk is now even closer for the q1
and the q3 quadrants at an absolute value around 1.3. This is not quite the case for the price
of the coskewness risk. It appears slightly higher for the up movements. This asymmetry
in the relation between returns and returns squared has been widely documented in the
GARCH literature. It comes not only from the risk exposure asymmetry as we have seen
in Table 8, but also from an asymmetric price of risk.
It should be noted that our results are not directly comparable to the results reported
in Pettengill, Sundaram, and Mathur (1995), and in Hung, Shackleton, and Xu (2004).
These two papers also explore the presence of asymmetry in the prices of the beta risk and
the coskewness risk respectively, but they consider only up and down market states and do
not measure the risk differently in these two states. In other words, they impose the same
beta in up and down movements, but allow for different prices of risk in these two states.
However, both their results and our results point to the fact that separating the periods
of positive and negative risk premia for securities is important. We have shown that it is
important not only for the asymmetry in the exposure to risks but also for the different
prices attributed by the market to the coskewness risk in bear and bull markets.
Our findings provide strong support for the CAPM model or its extension with a coskew-
ness risk as long as we look, as suggested by theory, at the same comovements in the market
and the individual securities, which constitute the bulk of the observations. When testing
with all the observations, we obtain the usual result of a zero relation between average
returns and the beta risk measures. The mixed results in studies that do not measure or
price risk differently in distinct phases of the market are due to the noise present in the
opposite movements in individual asset and market returns and to the symmetry assumed
in the risk and price in up and down comovements.

4.5 Estimated Risk Premia for Size and Book-to-Market Portfolios

A final test of the proposed model consists in assessing whether risk premia observed on
portfolios ranked by size or book-to-market portfolios can be explained by the market and
coskewness premia given their exposure to these risks. With the prices of risk estimated in
the previous section, we can compute the predicted risk premia for the ME and BE sets of
portfolios from their exposures to the market and coskewness risks. These were reported in
Table 6 for the unconditional exposures and Tables 7 and 8 for the conditional exposures.
However, these were estimated over a different sample period than the estimated prices

13
of risk. We reestimated the risk exposures of the portfolios over the same sample period
(from 26 June 1998 to 29 Devember 2006). For space considerations, we do not report
these results which are qualitatively the same and quantitatively close to the 1963-2006
estimates.
From an unconditional point of view it is obvious that we cannot reproduce the average
risk premia given that the estimated risk prices are practically zero. This is illustrated in
Tables 10 and 5. Moreover, the decreasing pattern in risk premia of ME portfolios and the
increasing pattern of risk premia for BE portfolios are not reproduced by the unconditional
market and coskewness model.
Results are much more interesting for the conditional risk premia. For the portfolios
ranked by market capitalization, the estimated premia in the first quadrant exhibit the same
increasing pattern as in the data for the first five portfolios. Afterwards, the risk premium
decreases in the data while it does not in the model except for the sixth portfolio. Results
are similar in the third quadrant where the magnitudes are similar but where premia are
negative. The premia associated with coskewness are sizable only for the smallest portfolios.
The observed premia in quadrants 2 and 4 are much smaller in magnitude. The estimated
premia in quadrant 4 are close to zero for small observed positive premia.
For the portfolios ranked by book-to-market value, the pattern of the observed risk
premia in quadrant 1 is matched perfectly by the estimated premia, although the level is
above the observed one, except for the last two portfolios. It is also the case for quadrant
3 in absolute value. The contribution of coskewness is of second order but it helps fitting
the premia in the first quadrant but not in the third. Comments made for quadrants 2 and
4 for size portfolios apply roughly to book-to-market portfolios.
Overall, we can conclude that the quadrant model produces premia that are close to the
observed ones. The market premium is by far the most important leading us to conclude
that the CAPM is supported by the data once we separate the comovements from the
opposite movements between the market and the individual securities or portfolios. The
coskewness contribution is much smaller in average but Figure 7 shows that the price of
risk of coskewness varies widely over the sample for the q1 and q3 quadrants. Therefore,
if we were to investigate the time variation in the premia it could contribute significantly
in some periods. The time variation in the price of beta risk, plotted in Figure 6, is less
pronounced.

5 Conclusion

In this paper, we have revealed several asymmetries in the measure and the pricing of risk
based on a model that relied more on the concept of dependence than on the concept of

14
correlation. In the relation between returns on individual securities or portfolios and the
market we have distinguished four states, depending on the coincidence and dissidence in
their joint movements. This amounted to running regressions and then computing corre-
lations quadrant by quadrant. This is conceptually different from the studies that have
considered different risk exposures or prices of risk in up and down markets since these
include inverse movements between security returns and market returns that add noise to
the relation. Financial theory considers comovements between asset returns and market
returns and the best way to test these theories is to consider phases where the two move
in the same direction.
Our main findings are that the size and book-to-market portfolios are not exposed
symmetrically to the market risk in up and down phases. The average value for the beta
risk when not distinguishing between the different states is closer to the estimate we get in
the downturns. Portfolios are generally less exposed to this risk in bull markets. However,
the price of risk appears to be quantitatively the same for up and down comovements. This
is not the case for the price of coskewness risk, which is slightly higher in up comovement
phases than in downturns. This combines with very different and asymmetric exposures to
the coskewness risk for the two sets of portfolios. The size portfolios are mainly exposed
to this risk in the positive quadrant, while the book-to-market portfolios are significantly
related to this risk in the negative quadrant. These risk exposures and risk prices translate
into quadrant premia that match the observed patterns of average premia and are not too
far from their levels. The main contributor is the market premium with a smaller but not
negligible role for coskewness.
Another important finding is the greater stability of the risk exposure estimates when
we distinguish between the different phases of the relation between security returns and
market returns. This result is especially significant when conducting event studies since it
affects directly the assessment of abnormal returns through the assumed risk model.

References

Ang, Andrew, and Joseph Chen, 2002, Asymmetric correlations of equity portfolios, Journal
of Financial Economics 63, 443–494.

, and Yuhang Xing, 2006, Downside risk, Review of Financial Studies 19, 1191–1239.

Barberis, Nicholas, Ming Huang, and Tano Santos, 2001, Prospect theory and asset prices,
Quaterly Journal of Economics 116, 1–53.

15
Barone Adesi, G., P. Gagliardini, and G. Urga, 2004, Testing asset pricing models with
coskewness, Journal of Business Economic and Statistics 22, 474–485.

Bollerslev, Tim P., Robert F. Engle, and J. Woolridge, 1988, A capital asset pricing model
with time-varying covariances, Journal of Political Economy 96, 116–131.

Bonomo, M., and R. Garcia, 1994, Can a well-fitted equilibrium asset pricing model produce
mean reversion, Journal of Applied Econometrics 9, 19–29.

Campbell, John Y., and John H. Cochrane, 1999, By force a habit: a consumption-based
explanation of aggregate stock market behavior, Journal of Political Economy 107, 205–
251.

Campbell, John Y., and Jianping Mei, 1993, Where do betas come from? asset price
dynamics and the source of systematic risk, Review of Financial Studies 6, 567–592.

Campbell, John Y., Christopher Polk, and Tuomo Vuolteenaho, 2005, Growth or glamour?
fundamentals and systematic risk in stock returns, Discussion paper NBER.

Campbell, John Y., and Tuomo Vuolteenaho, 2004, Bad beta, good beta, American Eco-
nomic Review 94, 1249–1275.

Cecchetti, S. G., P. Lam, and N. C. Mark, 1990, Evaluating empirical tests of asset pricing
models: Alternative interpretations, American Economic Review 80, 48–51.

Chabi-Yo, F., R. Garcia, and E. Renault, 2008, State-dependence can explain the risk
aversion puzzle, Review of Financial Studies 21, 973–1011.

Cho, Young-Hye, and Robert F. Engle, 1999, Time-varying betas and asymmetric effects
of news: Empirical analysis of blue chip stocks, Working Paper 7330 National Bureau of
Economic Research 1050 Massachussetts Avenue Cambridge, MA 02138.

Chou, Ray, Robert F. Engle, and Alex Kane, 1992, Measuring risk aversion from excess
returns and a stock index, Journal of Econometrics 52, 201–224.

Denuit, Michel, and Olivier Scaillet, 2004, Nonparametric tests for positive quadrant de-
pendence, Joural of Financial Econometrics 2, 422–450.

Fama, E., and J. MacBeth, 1973, Risk return and equilibrium: empirical tests, Journal of
Political Economy 81, 607–636.

Ferson, Wayne, and Campbell R. Harvey, 1991, The variation of economic risk premiums,
The Journal of Political Economy 99, 285–315.

16
Ferson, Wayne E., and Campbell R. Harvey, 1999, Conditioning variables and the cross
section of stock returns, The Journal of Finance 54, 1325–1360.

Garcia, Ren, and Georges Tsafack, 2006, Dependence structure and extreme comovements
in international equity and bond markets, Discussion paper CIRANO.

Ghysels, Eric, 1998, On stable factor structure in the pricing of risk: Do time-varying betas
help or hurt?, The Journal of Finance 53, 549–573.

Harvey, C.R., and A. Siddique, 2000, Coniditional skewness in asset pricing tests, The
Journal of Finance 55, 1263–1295.

Harvey, Campbell R., 1991, The world price of covariance risk, The Journal of Finance 46,
111–157.

Hung, D., M. Shackleton, and X. Xu, 2004, Capm, higher co-moment and factor models of
uk stock returns, Journal of Business Finance and Accounting 31, 87–112.

Jokung, Octave, and Jean-Christophe Meyfredi, 2004, Improving the market model: The
four-state alternative, in 7th SGF Conference of the Swiss Society for Financial Market
Research.

Kraus, Alan, and Robert Litzenberger, 1976, Skewness preference and the valuation of risk
assets, The Journal of Finance 31, 1085–1100.

Longin, Franois, and Bruno Solnik, 2001, Extreme correlation of international equity mar-
kets, The Journal of Finance 56, 649–676 Ils trouvent en utilisant EVT, que la correlation
n’est pas lie la volatilit du march mais la tendance de celui-ci. La corrlation semble
n’augmenter qu’en priode de march baissier (mais pas en priode haussire).

Maheu, John M., and Thomas H. McCurdy, 2000, Identifying bull and bear markets in
stock returns, Journal of Business &Economic Statistics 18, 100–112.

Norsworthy, N.R.and Gorener, R., R.E. Schuler, I.W. Morgan, and D. Li, 2001, Asset
pricing and prospect theory, Rensselaer Polytechnic Institute.

Norsworthy, N.R.and Gorener, R., R.E. Schuler, I.W. Morgan, and Ding Li, 2003, Ex-
pected utility, prospect theory and asset pricing, Discussion Paper 08-03-01 Rensselaer
Polytechnic Institute Financial Tehcnology.

Pettengill, Glen M., Sridhar Sundaram, and Ike Mathur, 1995, The conditional relation
between beta and returns, Journal of Financial and Quantitative Analysis 30, 101–116.

17
Shanken, Jay, 1992, On the estimation of the beta-pricing model, Review of Financial
Studies 5, 1129–1140.

Turner, Christopher, Richard Startz, and Charles R. Nelson, 1989, A markov model of
heteroskedasticity, risk and learning in stock market, Journal of Financial Economics
25, 3–22.

Wang, Kevin Q., 2003, Asset pricing wih a conditioning information: A new test, The
Journal of Finance 58, 161–196.

18
 Table 1: Summary statistics for the market index and equity portfolios
This table displays the summary statistics of the market index and the equity portfolios
daily returns obtained from K. French’s web site. Data are sampled from 1 July 1963
to 29 December 2006, resulting in 10951 observations for the market index and for each
portfolio. Panel A presents summary statistics for the 10 portfolios formed on their market
capitalization and panel B, those for the 10 portfolios formed on their book-to-market value.
Q1 corresponds to the first quartile of the distributions, Q3 to the third quartile and Kurt
to the excess kurtosis. Results are given in percentage.
Min Q1 Median Q3 Max Mean stdvev Skew Kurt
Market Index −17.16 −0.41 0.05 0.46 8.63 0.02 0.88 −0.74 18.12

Panel A. Size Portfolios (equally weighted)

1 −10.12 −0.28 0.09 0.40 6.79 0.03 0.74 −1.26 14.56
2 −11.31 −0.35 0.09 0.47 7.27 0.03 0.88 −0.83 9.68
3 −10.71 −0.38 0.09 0.49 6.59 0.03 0.89 −0.67 7.30
4 −12.63 −0.38 0.08 0.49 7.87 0.03 0.90 −0.65 9.12
5 −12.02 −0.38 0.08 0.50 7.62 0.03 0.90 −0.55 8.03
6 −13.51 −0.39 0.07 0.49 7.83 0.03 0.87 −0.62 10.24
7 −14.36 −0.39 0.07 0.48 8.28 0.03 0.87 −0.65 12.35
8 −15.83 −0.40 0.06 0.48 8.94 0.03 0.89 −0.64 14.37
9 −17.76 −0.41 0.05 0.46 9.98 0.02 0.88 −0.77 20.93
10 −19.73 −0.46 0.03 0.50 8.37 0.02 0.95 −0.72 21.22

Panel B. Book-to-Market portfolios (equally weighted)

1 −16.68 −0.54 0.03 0.57 10.48 0.02 1.08 −0.26 10.15
2 −17.46 −0.47 0.04 0.52 9.40 0.02 0.97 −0.50 14.25
3 −18.27 −0.44 0.03 0.49 8.31 0.02 0.92 −0.82 19.26
4 −18.69 −0.42 0.04 0.48 7.76 0.02 0.89 −0.92 21.58
5 −17.79 −0.41 0.03 0.46 7.73 0.02 0.86 −0.92 21.68
6 −15.62 −0.39 0.04 0.46 8.07 0.03 0.83 −0.69 16.02
7 −18.23 −0.39 0.05 0.46 7.88 0.03 0.82 −1.03 26.18
8 −17.36 −0.38 0.05 0.47 8.90 0.03 0.80 −0.95 25.29
9 −17.59 −0.40 0.06 0.49 7.56 0.04 0.86 −0.97 20.84
10 −16.19 −0.44 0.05 0.54 7.19 0.04 0.92 −0.55 12.20

19
 Table 2: Number of observations per quadrant
This table reports the number of observations in each quadrant after having partitioned
the daily returns of each portfolio over the four possible states. State 1 corresponds to a
positive market return for both the market index and the portfolio , state 2 to a positive
return for the market index and a loss for the portfolio, State 3 to a negative return for
both the market and the portfolio and Sate 4 to a positive return on the portfolio during
a market loss.
State 1 State 2 State 3 State 4
Panel A. Portfolios sorted by size
1 4847.00 993.00 3717.00 1394.00
2 4956.00 884.00 3925.00 1186.00
3 5066.00 774.00 4036.00 1075.00
4 5029.00 811.00 4079.00 1032.00
5 5125.00 715.00 4162.00 949.00
6 5134.00 706.00 4247.00 864.00
7 5224.00 616.00 4347.00 764.00
8 5273.00 567.00 4486.00 625.00
9 5333.00 507.00 4629.00 482.00
10 5354.00 486.00 4825.00 286.00

Panel B. Portfolios sorted by book-to-market ratio

1 5158.00 682.00 4611.00 500.00
2 5163.00 677.00 4585.00 526.00
3 5079.00 761.00 4494.00 617.00
4 5086.00 754.00 4442.00 669.00
5 4971.00 869.00 4351.00 760.00
6 5045.00 795.00 4351.00 760.00
7 5010.00 830.00 4297.00 814.00
8 4972.00 868.00 4267.00 844.00
9 4942.00 898.00 4230.00 881.00
10 4834.00 1006.00 4140.00 971.00

20
 Table 3: Parameter estimates of the Market Model and the 4-Quadrant Model.
The first five columns show the estimate of the beta as well as the four conditional betas for all the
portfolios. The last two columns report the results of an asymptotic F-test of an equal slope in state 1
and in state 3 ( β1 and β3 ) or in bull (β1 + β2 ) and bear (β3 + β4 ) markets. * (resp. ** and ***) indicates
significant result at the 5% confidence level (resp. 1% and .1%).
β β1 β2 β3 β4 β1 = β3 β1 + β2 = β3 + β4
Panel A. Portfolios sorted by size
1 0.6145∗∗∗ 0.3994∗∗∗ −0.1190∗ 0.6551∗∗∗ 0.0031 353.96∗∗∗ 524.84∗∗∗
2 0.8085∗∗∗ 0.6128∗∗ −0.1303∗ 0.8047∗∗∗ −0.0035 190.79∗∗∗ 450.03∗∗∗
3 0.8572∗∗∗ 0.6926 ∗∗∗ −0.1144 0.8269∗∗∗ 0.0283 106.17∗∗∗ 426.85∗∗∗
4 0.8823∗∗∗ 0.7349∗∗∗ −0.1631∗ 0.8546∗∗∗ 0.0322 95.13∗∗∗ 426.16∗∗∗
5 0.9061∗∗∗ 0.7856∗∗∗ −0.1663 0.8760∗∗∗ 0.1053∗ 61.79∗∗∗ 345.04∗∗∗
6 0.9002∗∗∗ 0.7959∗∗∗ −0.0972 0.8818∗∗∗ 0.0550 80.71∗∗∗ 442.58∗∗∗
7 0.9186∗∗∗ 0.8390∗∗∗ −0.1850 0.9056∗∗∗ −0.0148 56.13∗∗∗ 545.75∗∗∗
8 0.9577∗∗∗ 0.8973 ∗∗∗ −0.1995 0.9414∗∗∗ −0.0709 31.71∗∗∗ 3395.07∗∗∗
9 0.9621∗∗∗ 0.9272∗∗∗ 0.0344 0.9424∗∗∗ 0.0179 5.94∗ 277.15∗∗∗
10 1.0550 ∗∗∗ 1.0599 ∗∗∗ 0.0246 1.0074∗∗∗ −0.1197∗∗ 79.24∗∗∗ 435.48∗∗∗

Panel B. Portfolios sorted by book-to-market ratio

1 1.1632∗∗∗ 1.1337∗∗∗ 0.0122 1.0688∗∗∗ 0.0292 48.28∗∗∗ 328.58∗∗∗
2 1.0489∗∗∗ 1.0166 ∗∗∗ 0.0749 0.9816∗∗∗ −0.0187 17.8∗∗∗ 385.87∗∗∗
3 0.9763∗∗∗ 0.9216∗∗∗ −0.0455 0.9326∗∗∗ −0.1398 1.66 671.17∗∗∗
4 0.9379∗∗∗ 0.8708 ∗∗∗ −0.1578 ∗∗ 0.9097∗∗∗ −0.1480∗ 17.52∗∗∗ 788.36∗∗∗
5 0.8789∗∗∗ 0.8022 ∗∗∗ −0.0845 ∗ 0.8460∗∗∗ −0.0802∗ 20.82∗∗∗ 972.43∗∗∗
6 0.8545∗∗∗ 0.7759∗∗∗ −0.1034∗ 0.8100∗∗∗ −0.0921∗∗ 12.82∗∗∗ 800.24∗∗∗
7 0.8271∗∗∗ 0.7289 ∗∗∗ −0.0587 0.7995∗∗∗ −0.1713∗∗∗ 49.87∗∗∗ 1145.30∗∗∗
8 0.7926∗∗∗ 0.6877∗∗∗ −0.0872 0.7521∗∗∗ −0.1055∗ 41.10∗∗∗ 1024.5∗∗∗
9 0.8392∗∗∗ 0.6972 ∗∗∗ −0.1244 ∗∗ 0.8008∗∗∗ −0.0791∗ 81.64∗∗∗ 703.24∗∗∗
10 0.8646 ∗∗∗ 0.7053 ∗∗∗ −0.0463 0.7764∗∗∗ −0.1345∗ 27.49∗∗∗ 602.75∗∗∗

Table 4: Stability of the market model and the 4-Quadrant model for the various size and
book-to-market portfolios and for the individual components of the S&P500 index

β β1 β2 β3 β4
Size 3 3 7 2 7
Book-to-market 1 6 5 8 4
S&P500 Securities 183 381 414 347 423
Note: 50 stocks of the S&P500 have been deleted as their time series did not cover our full period of test.

21
 Table 5: Number of stable parameters for both the size and book-to-market criteria.
Panel A corresponds to a one-year period of estimation and one-year period for validation,
while Panel B corresponds to a three-year period of estimation and one-year period for
validation. a For this portfolio, the number of stable conditional betas is reduced by the fact
that the quadrant four contains to few observations to validate the stability between the two
sub-periods under consideration during 3 periods for Panel A and 1 period for Panel B.
β β1 β2 β3 β4
A B A B A B A B A B
1 17 4 26 5 38 12 26 3 42 13
2 18 4 23 6 38 11 25 6 42 13
3 22 6 21 5 39 12 31 6 42 13
4 21 4 24 7 40 10 32 7 42 13
Size 5 21 6 26 7 42 13 30 7 42 13
6 19 6 26 8 42 13 31 7 42 13
7 26 5 30 7 40 11 30 5 42 13
8 17 4 24 9 40 12 28 8 42 13
9 24 5 30 6 42 12 35 12 42 13
10a 17 6 20* 6* 39* 12* 38* 4* 39* 12*
1 12 3 29 7 42 13 26 5 40 12
2 17 5 23 6 42 13 29 7 42 13
3 21 3 28 5 42 13 33 9 42 13
4 19 3 31 8 42 13 24 7 42 13
Book-to-market 5 20 6 23 6 42 13 28 7 42 13
6 20 5 24 9 40 11 29 6 42 13
7 18 5 27 7 42 13 24 5 42 13
8 18 3 29 6 42 13 29 7 42 13
9 15 1 22 5 42 13 26 3 41 12
10 16 2 24 8 42 13 28 7 42 12

22
Table 6: Estimation results for the skewness model in equation (4) for the portfolios sorted
by size and book-to-market criteria

Size Book-to-market
β γ β γ
∗∗∗ ∗∗∗
1 0.6061 −1.3755 1.1699 1.0799∗∗∗
∗∗∗
2 0.8050 −0.5736 1.0505∗∗∗ 0.2609∗
3 0.8575∗∗∗ 0.0450 0.9736 ∗∗∗
−0.4320∗∗∗
4 0.8820∗∗∗ −0.0562 0.9333∗∗∗ −0.7586∗∗∗
5 0.9077∗∗∗ 0.2647 0.8743∗∗∗ −0.7624∗∗∗
6 0.9004∗∗∗ 0.0379 0.8521∗∗∗ −0.4021∗∗∗
7 0.9185∗∗∗ −0.0033 0.8201∗∗∗ −1.1360∗∗∗
8 0.9575∗∗∗ −0.0373 0.7867∗∗∗ −0.9612∗∗∗
9 0.9605∗∗∗ −0.2544 0.8323∗∗∗ −1.1258∗∗∗
10 1.0544∗∗∗ −0.1001 0.8610∗∗∗ −0.5990∗∗∗

23
 Table 7: Estimates of the Harvey and Siddique’s model by quadrant for the portfolios sorted by size
This table shows for each size portfolio, arranged in columns, the estimates of equation (4) on each of the four quadrants. *
(resp. ** and ***) corresponds to a significant value at the 5% confidence level (resp. 1% and .1%).

1 2 3 4 5 6 7 8 9 10
β1 0.3373∗∗∗ 0.5492∗∗∗ 0.6397∗∗∗ 0.6416∗∗∗ 0.6983∗∗∗ 0.704∗∗∗ 0.7558∗∗∗ 0.8323∗∗∗ 0.8679∗∗∗ 1.0846∗∗∗
β2 0.1161 0.0288 0.0589 0.3657 0.0691 0.3741

24
−0.1188 −0.0507 −0.2158 −0.1009
β3 0.659∗∗∗ 0.8456∗∗∗ 0.8991∗∗∗ 0.8949∗∗∗ 0.9352∗∗∗ 0.9129∗∗∗ 0.9226∗∗∗ 0.9443∗∗∗ 0.9056∗∗∗ 0.9564∗∗∗
β4 0.0172 0.121 0.1806 0.2746 0.374 0.1726 0.0666 0.2603 0.2478 0.0051
γ1 2.1239∗∗∗ 2.1679∗∗∗ 1.8153∗∗∗ 3.1909∗∗∗ 2.9954∗∗∗ 3.147∗∗∗ 2.8566∗∗∗ 2.2334∗∗∗ 2.042∗∗∗ −0.8491∗∗∗
γ2 −0.0086 −5.6046 7.1783 −27.1107 −20.9933 0.4155 −31.4866 −116.7997 −6.8345 −111.9928
γ3 0.0771 0.8036∗∗∗ 1.4261∗∗∗ 0.7988∗∗∗ 1.1797∗∗∗ 0.623∗∗∗ 0.3434 0.0587 −0.7578∗∗∗ −1.0683∗∗∗
γ4 0.8853 11.5554 16.1117 29.0023 39.6794 15.9999 8.1227 50.5913 46.1392 8.6955
 Table 8: Estimates of the Harvey and Siddique’s model by quadrant for the portfolios sorted by book-to-market value
This table shows for each book-to-market portfolio, arranged in columns, the estimates of equation (4) on each of the four
quadrants. * (resp. ** and ***) corresponds to a significant value at the 5% confidence level (resp. 1% and .1%).

1 2 3 4 5 6 7 8 9 10
β1 1.0389∗∗∗ 0.9953∗∗∗ 0.9172∗∗∗ 0.8929∗∗∗ 0.769∗∗∗ 0.7603∗∗∗ 0.6999∗∗∗ 0.6119∗∗∗ 0.6875∗∗∗ 0.7198∗∗∗

25
β2 −0.0123 0.2855 0.1464 0.0497 −0.1354 −0.0389 −0.0684 −0.1453 0.0648 0.0678
β3 1.099∗∗∗ 0.9622∗∗∗ 0.8711∗∗∗ 0.8379∗∗∗ 0.7637∗∗∗ 0.7631∗∗∗ 0.7022∗∗∗ 0.6551∗∗∗ 0.7119∗∗∗ 0.7196∗∗∗
β4 0.0062 0.139 0.1197 −0.1129 −0.1361 −0.1597 −0.1267 −0.234∗∗∗ −0.001 0.0125
γ1 3.2346∗∗∗ 0.7262 0.1503 −0.7512 1.1276 0.529 0.9964 2.6005∗∗∗ 0.3284 −0.4929
γ2 4.7093 −41.3436 −23.2763 −22.6478 4.1595 −6.2902 0.3901 2.2979 −17.0688 −8.1338
γ3 0.622∗∗∗ −0.3986∗∗∗ −1.2559∗∗∗ −1.4625∗∗∗ −1.6685∗∗∗ −0.9497∗∗∗ −1.9571∗∗∗ −1.9512∗∗∗ −1.786∗∗∗ −1.1358∗∗∗
γ4 −2.4522 16.4862 22.3444 2.7766 −3.6654 −5.1353 2.887 −7.5595 5.8731 11.2049
 Table 9: Estimates of risk prices based on the set of all securities.
Each column gives the risk price corresponding to each factor. Two regressions were esti-
mated. The first one, corresponding to results given in Panel A, only gives the risk prices
of the beta risk in each quadrant. In Panel B, we add to the beta factor the coskewness
factor. Results have been multiplied by 100.
ηβ ηγ Adj.R2
Panel A
Unc. 0.0029 -0.0009
q1 1.3141*** 0.6382
q2 0.692*** 0.1276
q3 -1.4337*** 0.7926
q4 0.1233 0.0000
Panel B
Unc. 0.0063 −0.0011 0.0056
q1 1.3382*** 3.4647*** 0.6632
q2 0.7914*** 1.5821*** 0.1624
q3 -1.3668*** 2.9409*** 0.8555
q4 0.0813 -0.3407 0.0009

26
 Table 10: Risk premia for the portfolios sorted by size.
The first two columns are obtained by multiplying the estimated price of risk by the indi-
vidual exposures to the two risk factors (β and η). The last column provides the observed
average portfolio return over the period. Results are given in percentage.
λβ λη λ β + λη ri
1 0.0032 0.0059 0.0091 0.0441
2 0.0050 0.0045 0.0095 0.0373
3 0.0056 0.0027 0.0084 0.0345
4 0.0058 0.0016 0.0074 0.0245
5 0.0062 0.0012 0.0074 0.0238
6 0.0058 0.0006 0.0064 0.0261
7 0.0057 0.0010 0.0068 0.0197
8 0.0061 0.0003 0.0063 0.0325
9 0.0058 −0.0001 0.0059 0.0310
10 0.0061 −0.0005 0.0056 0.0043

27
 Table 11: Risk Premia for the portfolios sorted by the book-to-market ratio.
The first two columns are obtained by multiplying the estimated price of risk by the
individual exposures to the two risk factors (β and η). The last column provides the
observed average portfolio return over the period. Results are given in percentage.
λβ λη λ β + λη ri
1 0.0069 −0.0007 0.0062 0.0022
2 0.0060 −0.0005 0.0056 0.0141
3 0.0054 0.0002 0.0056 0.0207
4 0.0053 0.0005 0.0058 0.0229
5 0.0048 0.0001 0.0050 0.0288
6 0.0049 0.0009 0.0058 0.0134
7 0.0043 0.0013 0.0056 0.0230
8 0.0040 0.0019 0.0059 0.0282
9 0.0046 0.0020 0.0067 0.0279
10 0.0042 0.0007 0.0050 0.0269

28
 Table 12: Decomposition of the total risk premia for the portfolios sorted by size.
This table provides the risk premia corresponding to the conditional risk factors β and γ
and their sum. The last column reports the observed average returns of the portfolios.
λβ λη λ β + λη ri
q1
1 0.4605 −0.0634 0.3971 0.6492
2 0.9261 −0.1821 0.7440 0.9747
3 1.0350 −0.1068 0.9282 1.0392
4 1.1004 −0.0938 1.0066 1.0563
5 1.2016 −0.0699 1.1317 1.0923
6 1.0053 0.0855 1.0908 0.9648
7 1.0948 0.0191 1.1139 0.9337
8 1.2137 0.0226 1.2363 0.9471
9 1.2819 −0.0413 1.2406 0.8813
10 1.2428 0.0896 1.3324 0.8728
q2
1 −0.0681 0.0302 −0.0379 −0.2833
2 0.0092 0.0184 0.0276 −0.3210
3 −0.0027 0.0719 0.0692 −0.3172
4 0.0176 0.0391 0.0567 −0.2767
5 0.0402 0.0359 0.0761 −0.3022
6 0.0660 0.0064 0.0724 −0.2314
7 0.0057 0.0306 0.0363 −0.1969
8 0.0159 0.0262 0.0420 −0.1549
9 0.1016 −0.0538 0.0478 −0.0835
10 0.1419 −0.0798 0.0621 −0.0998
q3
1 −0.2732 −0.3067 −0.5799 −0.7255
2 −0.7354 −0.1464 −0.8818 −1.0160
3 −0.9769 −0.0515 −1.0283 −1.0761
4 −1.0792 0.0028 −1.0763 −1.0783
5 −1.1797 0.0027 −1.1770 −1.1175
6 −1.1701 −0.0055 −1.1757 −0.9953
7 −1.1337 −0.0409 −1.1746 −0.9506
8 −1.2337 −0.0103 −1.2440 −0.9701
9 −1.1542 −0.0310 −1.1851 −0.8893
10 −1.2479 −0.0429 −1.2908 −0.9011
q4
1 0.0002 0.0053 0.0055 0.3479
2 0.0002 0.0032 0.0033 0.4168
3 0.0099 −0.0048 0.0051 0.4315
4 0.0045 −0.0190 −0.0145 0.3833
5 0.0141 −0.0079 0.0061 0.3737
6 0.0017 0.0241 0.0258 0.2739
7 −0.0056 0.0073 0.0016 0.2594
8 −0.0041 −0.0593
29 −0.0634 0.2603
9 −0.0283 0.0989 0.0705 0.1502
10 0.0293 −0.1151 −0.0859 0.1259
Table 13: Decomposition of the total risk premia for the portfolios sorted by the book-to-
market ratio.
This table provides the risk premia corresponding to the conditional risk factors β and γ
and their sum. The last column reports the observed average returns of the portfolios.
λβ λη λ β + λη ri
q1
1 1.2248 0.2119 1.4366 1.0278
2 1.3257 −0.0273 1.2984 0.9661
3 1.1549 −0.0343 1.1206 0.8919
4 1.2079 −0.1158 1.0922 0.8883
5 0.8928 0.0625 0.9552 0.8407
6 1.0651 −0.0815 0.9835 0.8376
7 0.7862 0.0376 0.8238 0.7833
8 0.7787 −0.0308 0.7479 0.7378
9 0.9035 −0.0791 0.8244 0.8622
10 0.7816 −0.0336 0.7480 0.8535
q2
1 0.1083 −0.0404 0.0679 −0.1511
2 0.1975 −0.1026 0.0949 −0.1623
3 −0.0102 0.0481 0.0379 −0.1989
4 −0.2222 0.1990 −0.0232 −0.2201
5 −0.0986 0.1094 0.0108 −0.1973
6 −0.0801 0.0898 0.0097 −0.2159
7 −0.0224 0.0064 −0.0160 −0.1914
8 −0.0669 0.0381 −0.0288 −0.2403
9 −0.1008 0.1123 0.0115 −0.2693
10 0.0397 −0.0136 0.0261 −0.3129
q3
1 −1.3486 −0.0678 −1.4165 −1.0675
2 −1.0998 −0.0704 −1.1702 −0.9557
3 −0.9331 −0.0990 −1.0321 −0.8724
4 −0.9159 −0.0847 −1.0006 −0.8767
5 −0.8750 −0.0615 −0.9365 −0.8515
6 −0.9114 −0.0600 −0.9714 −0.8527
7 −0.8144 −0.0464 −0.8608 −0.8101
8 −0.7009 −0.0833 −0.7842 −0.7752
9 −0.7376 −0.1204 −0.8580 −0.8700
10 −0.8371 0.0301 −0.8070 −0.8510
q4
1 0.0544 −0.4763 −0.4219 0.2732
2 0.0116 −0.0504 −0.0387 0.1997
3 −0.0041 −0.0352 −0.0393 0.2457
4 −0.0368 0.0550 0.0182 0.2634
5 −0.0208 0.0348 0.0140 0.2841
6 −0.0375 0.0791 0.0416 0.2646
7 −0.0261 30
0.0225 −0.0036 0.3128
8 −0.0235 0.0410 0.0175 0.3463
9 −0.0006 −0.0019 −0.0026 0.3337
10 −0.0199 0.0090 −0.0109 0.3892
Figure 1: Histogram of frequencies of non-rejections of stability over eight years for the 450
securities of the S&P500 index - Results for β1 .

31
Figure 2: Histogram of frequencies of non-rejections of stability over eight years for the 450
securities of the S&P500 index - Results for β3 .

32
Figure 3: Histogram of frequencies of non-rejections of stability over eight years for the 450
securities of the S&P500 index - Results for a single β.

33
 2
1 #1
βs

β
-1

β1
β3
-2

0 10 20 30 40
#5
Index
2
1
βs

β
-1

β1
β3
-2

0 10 20 30 40
# 10
Index
2
1
βs

β
-1

β1
β3
-2

0 10 20 30 40

Index

Figure 4: Evolution of of β, β1 and β3 estimated for the three Market Capitalization

portfolios (Small, Medium, Big) on rolling windows of 252 trading days.

34
 2
1
#1
βs

β
-1

β1
β3
-2

0 10 20 30 40
#5
Index
2
1
βs

β
-1

β1
β3
-2

0 10 20 30 40
# 10
Index
2
1
βs

β
-1

β1
β3
-2

0 10 20 30 40

Index

Figure 5: Evolution of β, β1 and β3 estimated for three different portfolios (Low, Medium,
High) formed on the Book-to-Equity ratio, on non overlapping windows of 252 trading
days.

35
Figure 6: Time variation in the price of market risk

36
Figure 7: Time variation in the price of coskewness risk

37