Realized Skewness and Future Stock Returns: The Role of Information Youngmin Choi and Suzanne S. Lee 1 April 25, 2014 Abstract This paper examines the relationship between realized daily skewness and future stock returns and investigates the impact of information releases on that relationship. We find that there exists a negative relationship between realized daily skewness and subsequent stock returns when there is no high-impact information release, but that the relationship becomes positive if the realized skewness is associated with such releases. We show that the profitability of a zero-investment portfolio can be enhanced by incorporating this positive relationship in the presence of high-impact information into investment strategies. As the positive relationship mainly results from riskier stocks with volume increase in the presence of information releases, we offer an explanation for this finding based on the divergence of investors’ opinion. JEL classification codes: G14, G17 Key words: realized skewness, return predictability, equity market, high frequency data, investors opinion 1 Both authors are with the Scheller College of Business, Georgia Institute of Technology, 800 West Peachtree Street NW, Atlanta, GA 30308. We would like to thank seminar participants at Georgia Tech for their comments and suggestions. All comments are welcome and any errors are ours. Please send any correspondence to Youngmin Choi at [email protected] or Suzanne S. Lee at [email protected].
48
Embed
Realized Skewness and Future Stock Returns: The … › 6fbd › bd39afc09d20d...Realized Skewness and Future Stock Returns: The Role of Information Youngmin Choi and Suzanne S. Lee1
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Realized Skewness and Future Stock Returns:
The Role of Information
Youngmin Choi and Suzanne S. Lee1
April 25, 2014
Abstract
This paper examines the relationship between realized daily skewness and future stock returns and
investigates the impact of information releases on that relationship. We find that there exists a negative
relationship between realized daily skewness and subsequent stock returns when there is no high-impact
information release, but that the relationship becomes positive if the realized skewness is associated with
such releases. We show that the profitability of a zero-investment portfolio can be enhanced by
incorporating this positive relationship in the presence of high-impact information into investment
strategies. As the positive relationship mainly results from riskier stocks with volume increase in the
presence of information releases, we offer an explanation for this finding based on the divergence of
investors’ opinion.
JEL classification codes: G14, G17
Key words: realized skewness, return predictability, equity market, high frequency data, investors opinion
1 Both authors are with the Scheller College of Business, Georgia Institute of Technology, 800 West Peachtree
Street NW, Atlanta, GA 30308. We would like to thank seminar participants at Georgia Tech for their comments
and suggestions. All comments are welcome and any errors are ours. Please send any correspondence to Youngmin
Many studies in finance focus on the higher moments of asset return distribution and their
important implications in asset pricing. A large stream of literature has been devoted to investigating the
cross-sectional relationship between higher moments and stock returns. In particular, the skewness that
measures the asymmetry of return distribution has been an interesting topic for research. Several studies
theoretically predict and empirically confirm the negative relationship between skewness and subsequent
stock returns (see Arditti (1967), Kraus and Litzenberger (1976), Mitton and Vorkink (2007), Barberis
and Huang (2008), Zhang (2005), Kelly (2012), Conrad, Dittmar, and Ghysels (2013), and Amaya,
Christoffersen, Jacobs, and Vasquez (2013), among many others).
Despite such intriguing results on the negative relationship, important questions have yet to be
answered in this literature, including: Does this negative relationship hold in every circumstance, and
what determines this relationship? We are motivated to answer these questions because a large body of
literature suggests the importance of information releases on future stock return predictability. For
example, recent studies document the finding that stock returns exhibit momentum when price changes
are accompanied by information releases, while they exhibit return reversal if price changes are not
accompanied by information.2 Given this literature, we investigate the role of high-impact information
releases in explaining this relationship between the realized skewness and future stock returns. We define
high-impact information releases as those firm-level information releases that tend to generate (or be
associated with) unusually large price changes, namely jumps in asset prices. We consider this type of
information releases because it is well documented in the asset pricing literature that the skewness of
return distribution is better captured by asset pricing models that incorporate jumps.3 Hence, we expect
that examining the aforementioned relationship in the presence (or absence) of high-impact information
releases will allow us to discover the answer to our questions.
2 See Savor (2012), Chan (2003), Pritamani and Singal (2001), and Tetlock (2010) for references. These papers
argue that investors under-react to information-related price changes, but overreact to other shocks. 3 See Merton (1976) and Bakshi, Cao, and Chen (1997), among many others.
2
In measuring skewness in stock returns, we apply high-frequency data to estimate the realized
daily skewness, following Amaya, Christoffersen, Jacobs, and Vasquez (2013), who use this method to
present a negative relationship between realized skewness and next week’s stock returns. Another reason
why we use the realized daily skewness of a stock is because we need a daily-level skewness measure to
incorporate the impact of information releases. We consider the relationship between this skewness
measure and subsequent returns over the next one to 20 trading days in relation to high-impact
information releases. To mitigate the concern regarding the impact of microstructure noise in high-
frequency data on skewness inference, we choose our sample period from recent years (January 2005-
December 2010). This is because more recent data have far fewer infrequent trading or illiquidity
problems that may generate unnecessary noise in the high-frequency data. In estimating the effect of high-
impact information releases on subsequent stock price dynamics, we employ the Fama-McBeth (1973)
regression method.
In selecting high-impact information releases, we depend on a recent study by Lee (2012), who
used high-frequency data to identify important information releases that are more likely than others to
generate jumps in prices in individual U.S. stock markets. In particular, we select two highly influential
firm-specific information events such as earnings announcements and analyst recommendation releases.
This choice of high-impact information events is also motivated by the statistical limiting property,
whereby the realized skewness measure based on the sum of cubed returns is likely to capture only the
jump components of return distribution if we use high frequency data for estimation (see Jacod (2012)).
In addition, these two types of information are the well-documented types of information that generate
momentum or drift in stock returns after the announcements.
By including analyst recommendation reports as well as earning announcements in our
information criteria, we are able to provide extensive coverage of firm-related news. While earnings
announcements by a firm provide clear information on the cash flows of the firm, analyst
recommendation reports cover not only cash-related news but also other firm-specific news such as
additional refined and comprehensive information on the firm. To our knowledge, no study in the current
literature has examined how these high-impact news releases affect the relationship between the realized
skewness capturing jumps in returns and subsequent return patterns.
Using a full sample of firm-day observations, we first examine the unconditional relationship
between realized daily skewness and subsequent stock returns over the next 1, 5, 10, and 20 trading days.
We confirm the existing evidence of a negative relationship between realized daily skewness and future
stock returns, which is consistent with Amaya, Christoffersen, Jacobs, and Vasquez (2013). However, we
find that this negative relationship no longer holds when we consider subsamples accompanied by the
aforementioned high-impact information releases. Specifically, in the subsample with information
releases, all subsequent returns over the next 1, 5, 10, and 20 trading days are positively related to the
realized daily skewness, and this positive relationship is statistically significant at the 1% level for all
cases of selected horizons.
When we examine the other subsamples without high-impact information releases, the well-
documented negative relationship between skewness and subsequent stock returns holds. In other words,
we can interpret this as investors accepting lower future returns from investments with higher positive
skewness and require higher returns from investments with negative skewness in the absence of high-
impact information releases. This is consistent with Arditti (1967), who considers the negative skewness
of a stock as a proxy for the riskiness of that stock. However, when the realized daily skewness of a stock
is associated with high-impact information on the firm, our findings suggest that the interpretation of
skewness as a risk measure turns out not to be valid. This outcome is found after controlling for the usual
informative variables such as size factor, value factor, momentum factor, and trading volume, which are
documented as well-known predictors of cross-sectional stock returns in existing literature (Fama and
French (1993), Jegadeesh and Titman (1993), and Lee and Swaminathan 2000).4
4 We also confirm our finding that different patterns between realized skewness and future stock returns hold
when we alter the frequency of analysis to the weekly level, following Amaya, Christoffersen, Jacobs, and Vasquez (2013).
4
We also examine the relationship between realized daily skewness and subsequent returns using
separate subsamples of positive skewness and negative skewness in relation to the high-impact
information releases. We find that the positive relationship in the presence of information releases is
stronger for the subsample with negative skewness than that for the subsample with positive skewness. In
other words, the momentum patterns appear to be stronger for the stocks with negative skewness
accompanied by high-impact information releases. It is also noteworthy that we observed the existing
negative relationship in the absence of information releases mainly for the stocks with positive skewness.
This evidence is consistent with the prediction by Mitton and Vorkink (2007) and Barberis and Huang
(2008), who demonstrate that assets with greater skewness have lower returns.
As regards implications of our results, we examine the profitability of an investment strategy
using these distinctive findings. First, we sort stocks based on their realized daily skewness and measure
the performance of quintile portfolios over the next 10 and 20 day. To examine the effect of the high-
impact information releases, we separate the full sample into a sample of stocks that experience high-
impact information releases and a sample of stocks that do not experience any such releases. We confirm
that the strong negative relationship between realized daily skewness and future stock returns prevails
mostly in the sample without information releases. We again find stronger positive relationship between
realized daily skewness and future stock returns in the sample with information, and the abnormal return
from a zero-investment strategy that is long the highest quintile portfolio and short the lowest quintile
portfolio, based on the realized skewness and the presence of information, provides statistically
significant returns even when we take into account other important risk factors. Furthermore, we provide
evidence that the profitability of a zero-investment portfolio can be enhanced by incorporating the
presence of high-impact information into investment strategies.
Finally, we offer explanations for the relationship between skewness and subsequent return that
depends on the presence of information releases. Since we expect the release of high-impact information
to influence the opinion of investors in the financial markets, we test how realized skewness of stock
returns is related to measures of divergence of investors’ opinion. Using several measures of divergence
5
of opinion, daily turnover, market-adjusted turnover, unexplained trading volume, or change in trading
volume, we find that realized skewness is negatively related to these measures when there are high-impact
information releases, while it is positively related to divergence measures when there are no such
information releases. In addition, we find that measures of divergence of investors’ opinion increase in
high-impact information sample compared to no-information sample. Consistent with Miller (1977) and
Diether, Malloy, and Scherbina (2002), our findings suggest that riskier stocks with greater divergence in
investors’ opinion around information releases have lower subsequent returns. Investors will not be
rewarded for bearing risk (measured by the negative realized skewness) if there are high-impact
information releases because prices reflect the more optimistic valuation of investors who hold these risky
assets and suffer losses in the subsequent periods.
Overall, this paper contributes to several different stems of literature in finance. Our study is the
first in the literature to document the finding that the existing negative relationship between skewness and
subsequent stock holds only in the absence of high impact information releases, while the relationship
becomes positive in the presence of such information releases. Our finding is unique in that it is based on
realized daily skewness using high-frequency data, which can capture different part of the returns
distribution at the micro level. Moreover, our study contributes to the literature on information and return
predictability by showing that return reversal patterns is stronger for stock with positive skewness in the
absence of information releases, while momentum pattern is stronger for stocks with negative skewness
around information releases. Furthermore, we offer an explanation for the new finding of a positive
relationship in the presence of high-impact information releases. We also contribute to the literature on
divergence in investors’ opinion by providing new evidence that the skewness is negatively related to the
increase in divergence of opinion only around the information releases. We believe that this negative
skewness may have contributed to the evidence of lower returns documented in the related literature, such
as Diether, Malloy, and Scherbina (2002).
6
The remainder of this paper is organized as follows. Section 2 presents the data and sample
construction. Section 3 presents empirical results of estimations, and Section 4 offers a potential
explanation for the findings. Finally, Section 5 concludes.
2. Data and Sample Description
We analyze all stocks listed in the Trade and Quote (TAQ) database. The sample period for our
analysis is from January 3, 2005 to December 31, 2010. We choose to consider this recent sample period
to mitigate concerns regarding the impact of market microstructure noise in high-frequency data on
inference by higher moments as data from recent years have much less infrequency trading and fewer
illiquidity problems that may generate unnecessary noise in the data.5 To calculate higher moments, we
record prices of all stocks in the TAQ database at five minute intervals starting from 09:30 EST until
16:00 EST and construct the five-minute returns as the difference between log prices with five-minute
marks, as in Andersen, Bollerslev, Diebold, and Ebens (2001). To ensure sufficient liquidity and filter out
possible noise in high-frequency data, we require a stock to have at least 80 transactions per day and use
the recorded price close to the five-minute interval time grid. In addition to this requirement, we further
exclude stocks that have a price of less than five dollars.
We also obtain data on market capitalizations and trading volumes from the Center for Research
in Security Prices (CRSP) database. We also use data from the CRSP database to compute several
measures of divergence in investors’ opinion such as daily turnover, market-adjusted turnover, or
unexplained trading volume. Our accounting data, such as book values of individual firms, are obtained
from the COMPUSTAT database. For returns over horizons beyond one trading day, instead of using
high-frequency returns from the TAQ database, we use daily returns from the CRSP database for
corresponding firms and dates and calculate returns over longer horizons. Given all these filtering
5 The robustness of the realized higher moment inference to noise is also verified by simulation in Amaya et al.
(2013).
7
requirements, the total number of companies covered in our sample varies, ranging from 3,443 to 6,387
per year, depending on changes in market conditions over our sample period.
As regards the five-minute price records, we define the five-minute log return as the difference
between log prices observed at five-minute intervals. The the l-th intraday return for the k-th firm on day t
is first constructed by
��,�,� = log��,�,�� � − log��,�,��, (1)
where ��,�,� is the l-th intraday price of the k-th firm observed on day t, and n is the number of intraday
returns observations in a day. Since we record five-minute prices starting from 09:30 EST to 16:00 EST,
for each day, we have � = 78. Following Amaya, Christoffersen, Jacobs, and Vasquez (2013), who
document the negative relationship between realized skewness and the cross-section of equity returns on a
weekly horizon, we apply the similar methodology to construct other realized higher moments such as
realized daily skewness and realized daily kurtosis. That is, we construct realized daily skewness (RDSk,t)
for the k-th firm on day t by
����,� = √�∑ ��,�,������
����,���
, (2)
where the realized daily variance (����,�) for the k-th firm on day t is used to standardize the sum of
cubed intraday 5 minute returns and is calculated with the sum of squared intraday 5-minute returns as in
����,� =���,�,���
���, (3)
following the well-known realized variance literature (Andersen and Bollerslev (1998) and Andersen,
Bollerslev, Diebold and Labys (2003)). A measure of realized daily kurtosis (����,�) for the k-th firm on
day � is computed as
����,� =�∑ ��,�,� ��������,��
. (4)
8
Table 1 presents the time-series summary statistics of means and medians by year for variables
considered in this paper. This table shows that the number of stocks in our sample decreased significantly
during the period of financial crisis. For example, during the year 2006, the number of stocks included is
6,387, while this number is 3,703 in the year 2008, during which our sample selection criterion filters out
noisier data. We also observe much higher average realized daily variance of 0.30% in the year of 2008
when compared to 0.09% in the years 2006 or 2007. The realized daily skewness is negative in general.
The realized daily kurtosis is shown not to vary as much as volatility. Overall, the distribution of asset
returns using high-frequency data shows time-varying variance, negatively skewed tails, and high kurtosis.
The other control variables for our regression analysis, such as the ratio (in percentage) of trading volume
to total shares of outstanding and momentum returns also show some time variation by year, as expected.
We also observe a recent finding in the literature on asset pricing showing the devastating performance of
momentum strategy after the financial crisis.6
To examine first the relational patterns between higher moments and stock returns without
conditioning on the presence of information, we construct portfolios by sorting stocks into deciles based
on realized daily volatility, realized daily skewness, and realized daily kurtosis. The sorting procedure is
repeated every day during our sample period between January 2005 and December 2010. Table 2 shows
the equal-weighted characteristics of these decile portfolios. Panel A, B, and C present time-series
averages for portfolio characteristics sorted by realized daily volatility, skewness, and kurtosis,
respectively. Since we are interested in the relationship between realized moments and subsequent returns,
we report averages for the decile portfolio returns over the subsequent month in next 20-day return. In
particular, Panel B indicates that while contemporaneous return is positively related to realized skewness,
returns over the subsequent month tend to be negatively related to realized skewness, which is consistent
with the findings of Amaya, Christoffersen, Jacobs, and Vasquez (2013). However, we do not observe
any distinct relational pattern between realized kurtosis and subsequent future returns, which is
inconsistent with Amaya, Christoffersen, Jacobs, and Vasquez (2013). Even when we conduct the same
6 See Daniel, Jaganathan, and Kim (2012) and Daniel and Moskowitz (2013).
9
analysis with a weekly sorting frequency, which is identical to the approach of Amaya, Christoffersen,
Jacobs, and Vasquez (2013), the general relational pattern between realized skewness and next week’s
return remains similar.
We confirm existing evidence and find that the negative relational pattern between realized
skewness and subsequent return does not depend on the choice of sample period. As we shall see in
Section 3, the negative relationship between realized daily skewness and future stock returns is
statistically significant when we perform the Fama-MacBeth regressions. Although Table 2 shows the
results with the return calculated over the next month for the purpose of data description, we also confirm
that this negative relationship holds for subsequent returns over several holding periods of one to 20
trading days. We will get back to this result in our main analysis in Section 3.
In selecting high-impact information releases, we depend on a recent study by Lee (2012), who
used high frequency data to identify important information releases that are more likely than others to
generate large intraday price changes (namely jumps in prices) in individual U.S. stock markets. In
particular, we choose to use the two most influential information events documented therein, namely
earnings announcements and analyst recommendation releases. This choice is also motivated by the
statistical limiting property whereby the realized skewness measure (based on the sum of cubed returns) is
likely to capture only the jump component of return distribution if high frequency data are used for
skewness estimation (see Jacod (2012)). The other reason why we select these two types of information is
that these naturally cover most firm-related news about a firm. While earnings announcements provide
valuable information about cash flow, analyst reports offer extensive coverage of a wider array of
information (see Savor (2012)). Thus, by including not only earnings announcements but also analyst
reports on a firm, we expect to consider a comprehensive set of information about the firm.
In order for us to examine how these high-impact news releases affect the relationship between
skewness capturing jumps in prices and the subsequent return dynamic, we create separate subsamples
with and without these influential firm-specific information releases. Since the chosen information events
are firm-specific, we first collect all the firm-date observations for both information events from the
10
International Brokers’ Estimation System (I/B/E/S) from January 2005 to December 2010. Although we
use high-frequency data to calculate realized skewness and other higher moments, our analysis does not
depend on time-stamps at intraday level of these information releases. This is because our study considers
how realized daily skewness accompanied by information releases at the daily level affect returns over the
next one to 20 trading days. Hence, the intraday time-stamp delay concern raised by Bradley, Clarke, Lee,
and Ornthanalai (2014) is not an issue for our study. On the other hand, several studies report that these
information events tend to occur overnight.7 When firm-related events occur during the day, analysts
usually examine the issues during business hours and release their reports on the same day or one day
after the event. In some cases, analysts can release their reports about a firm before the issue is released to
the public in order to offer their expectations. For all of these reasons, when we create our dummy
variable for information release dates, we cover one day before and after information release dates which
are recorded in the I/B/E/S database.
Table 3 present the time-series summary statistics (by year) for all firms in our full sample, firms
with earnings announcements, firms with analyst recommendation releases, firms with both types of
information releases, and firms without the information releases. For each year, we list the total number
of all firm-trading day observations and average firm sizes. Roughly, we have more than four million
firm-date observations in our full sample. We then separately report those statistics for the subsamples
accompanied by earnings announcements and those accompanied by analyst recommendation releases.
Since we are not interested in the separate impact of earnings and analyst recommendation news events
on the relationship between realized skewness and subsequent stock returns, we combine the two
subsamples and call it the “Information” sample, with the rest of the sample becoming the “No-
information” sample. The fact that earnings announcements and analyst recommendation releases can
confound each other is not a concern in this study as we are interested in how these types of influential
information releases affect the subsequent stock return patterns. It is noteworthy that the Information
7 Altınkılıç and Hansen (2009) report that the percentage of overnight recommendation releases is 61%. Bradley,
Clarke, Lee, and Ornthanalai (2014) report that this percentage is over 70% and earnings releases occur overnight in over 80% of cases after the year 2003.
11
sample includes larger firms than the No-information sample, with the average size of firms included in
the Information sample being $8,761.97 million while that for the No-information sample is $5,214.45
million. Although we find this difference in average firm size between the two subsamples, we show later
that the determinant of different relational patterns is not purely due to the size issue.
3. Realized Skewness and Future Returns
3.1. The role of information releases on the relationship between skewness and future
returns
In this section, we assess the cross-sectional relationship between the realized skewness and
future stock returns by conducting various cross-sectional regression analyses with several control
variables. Then, in order to examine how high-impact information releases affect the relationship, we take
two approaches. Using the subsamples (the Information and No-information samples) presented in Table
3, we first conduct the regression analyses separately and examine the differences. The second approach
is to examine the interaction term of daily realized skewness with a dummy variable to identify the
additional impact of the presence of information releases using the full sample. Both approaches offer
consistent results, indicating that the presence of high-impact information matters for the relationship.
Specifically, the regression specification we consider first is as follows:
JK<�9-:�B=56+@"�,�L = 9� + M N��,�N� + M�N��,�NO (13)
where ��,� is the standard deviation of the residuals from the regression, calculated over the model’s
estimation period [t-52, t-5], and the plus and minus superscriptions on the absolute valued returns
indicate when returns were positive or negative.
10 Another possible measure of divergence of opinion in investors might be the dispersion of analyst forecasts (see
Diether, Malloy, and Scherbina (2002)).
11 This definition of daily turnover is equivalent to the volume percentage used as one of control variables in
equation (5) and (6). In this section, we follow the related literature on divergence of investors’ opinion to call it daily turnover.
26
Comparing Panels B and C of Table 11, the divergence in investors’ opinion drastically increases
on the days when there are high-impact information releases. We show that this increase is statistically
significant at the 1% level for all the divergence measures under consideration. While the negative
relationship between realized skewness and subsequent stock returns has been explained by the skewness
preference framework, we argue that the positive relationship between realized skewness and future stock
returns can be explained by the prediction of Miller (1977), who suggests that the increase in divergence
in investors’ opinion may lead a stock to be overvalued because the short-sale constraint prevents the
negative opinion of investors from being revealed in the stock markets. Miller further argues that the
riskier class of stocks might have lower expected returns if the riskiest stocks experience greater
divergence in investors’ opinion.
To link our findings to the prediction of Miller (1977) and test how the realized skewness of stock
returns is related to divergence of opinion, we first examine a contemporaneous relationship between our
realized skewness measure and several measures of divergence in investors’ opinion with ordinary least
squares (OLS) regressions. Panels A, B, and C of Table 12 report the estimated coefficients of OLS
regressions of realized skewness on divergence in investors’ opinion measures for full sample, a sample
with information releases, and a sample without information releases, respectively. Since we conjecture
that the divergence in investors’ opinion around information releases is one of the determinants that lead
to different relationships between realized daily skewness and future returns of a stock, we study how
divergence in investors’ opinion is related to realized daily skewness with and without information
releases.
In Panels B and C of Table 12, the results of the regression analysis of realized daily skewness on
the divergence of opinion show the opposite sign of the coefficient estimates for all measures of
divergence of opinion. Specifically, it illustrates that our measure of realized skewness is negatively and
significantly related to all the measures of divergence of opinion when there are high-impact information
releases, while it is positively related to those divergence measures when there are no information releases.
In other words, in the presence of high-impact information releases, the stocks experiencing the increase
27
in divergence in investors’ opinion tend to have more negatively skewed distributions, which make these
stocks riskier to investors. In turn, our findings suggest that those riskier stocks with the greater
divergence in investors’ opinion generate lower future stock returns, which is consistent with the
prediction of (Miller (1977)).
In particular, this provides an explanation for our finding, presented in Table 6, which shows that
the significantly positive relationship between skewness and future return is mainly found in the negative
skewness sample in the presence of information releases. This explanation also supports the evidence
presented in Panel B of Table 7, which shows that the lowest quintile portfolio from the information
sample offers especially low return in the subsequent period. The stocks in the lowest quintile portfolio
with high-impact information releases constitute the riskiest group of stocks (measured by their negative
realized daily skewness) with highly divergent opinion of investors (due to releases of high-impact
information).12
It is noteworthy that the negative relationship is mainly observed with positive skewness and the
positive relationship with negative skewness, both suggesting lower returns in the subsequent period.
Here, notice that both positive and negative skewness is likely to increase the idiosyncratic volatility of
stock returns due to the presence of extreme movements in either direction. In other words, this increased
idiosyncratic volatility is negatively related to subsequent stock returns with or without high impact
information releases. Our unreported analysis also suggests that the increase in the individual stock’s
volatility level is even higher in our Information sample, which only includes the times of firm-specific
information releases. We believe that our new evidence of a positive relationship between realized
skewness and subsequent return (in addition to the negative relationship already documented in the
12 Chen, Hong, and Stein (2001) show that negative skewness over a six-month period is most pronounced in
stocks that experienced an increase in trading volume in relation to trends over the past six months. Although we also consider volume and skewness, our skewness measure is calculated for one day using intraday data, and our evidence in Table 12 shows a daily contemporaneous relationship with or without information releases, while theirs purpose is to forecast skewness for the subsequent half year without taking into account information releases.
28
literature) contributes to the idiosyncratic volatility puzzle documented in Ang, Hodrick, Xing, and Zhang
(2006).
5. Conclusion
The role of skewness in stock returns in asset pricing has been investigated by many researchers.
It is well documented that the skewness of stock returns is considered one of the risk measures for
expected subsequent returns. Using high-frequency data capturing finer market movements up to intraday
level, we show that highly influential firm-specific information releases (which tend to generate extreme
market volatility) can change conventional views of realized skewness as a risk measure.
First, we show that in general, there exists a negative relationship between realized daily
skewness and subsequent returns over 1 to 20 trading days. We then document that this existing
relationship between realized daily skewness and future return is reversed when we take into account the
presence of high-impact information releases about firms. Specifically, by separately analyzing the stocks
experiencing high-impact information releases and those without information, we discover that there
exists a strong positive relationship between realized daily skewness and subsequent returns when high-
impact information releases occur and a strong negative relationship when there are no influential
information releases. In other words, when there are high-impact information releases generating unusual
market reactions, the risk-return compensation framework does not appear to work. More interestingly,
we find the negatively skewed stocks tend to be the significant driver for the positive relationship, while
positively skewed stocks tend to be the significant driver for the negative relationship.
Second, we document how our finding of a distinctive pattern in the relationship between realized
daily skewness and subsequent returns (conditional on the presence of high-impact information) explains
the cross-section of stock returns even after taking into account common risk factors, including market
excess returns, the size factor, the value factor, and the momentum factor. We also show that this anomaly
might be exploited to produce an abnormal return. Interestingly, our evidence suggests that we might be
able to enhance the profitability of the zero-investment trading strategy by taking account the distinctive
29
relational patterns between realized daily skewness and subsequent returns depending on the presence of
information. Consistent with our finding that the positive relationship between realized daily skewness
and subsequent returns results from negatively skewed stocks, we are also able to significantly enhance
the profitability of the zero-investment trading strategy if we use extremely skewed (measured by realized
daily skewness) in constructing the portfolio. The increase of abnormal returns by using extreme samples
is significant and the gains from the information sample dominate those from the no-information sample
as we focus on more extreme samples.
The different patterns in the relationship depending on the presence (or absence) of high-impact
information are not driven by small firms. While we observe a stronger positive relationship for small
firms, the positive relationship in large firm remains statistically significant. Another robustness check is
for a time horizon over which we compute realized skewness using high-frequency data. When we
examine the same issue with realized weekly skewness instead of realized daily skewness (which we used
in this study), we find that the negative relationship in the no-information sample still holds strongly,
while the positive relationship in the information sample disappears. One of the reasons for this change in
the relationship in the information sample is that the high-impact information considered in this paper
includes earnings announcements and analyst recommendation reports, which are generally reflected in
abnormal return distribution soon (within a day) after they are released.
After we document our new finding of a positive relationship between realized daily skewness
and subsequent stock return when there are high-impact information releases, we offer a possible
explanation for this phenomenon using divergence in investors’ opinion around information releases. We
demonstrate that measures of divergence in investors’ opinion used in this paper dramatically increase
when there are information releases. Furthermore, we find that the divergence in opinion is negatively
related to realized daily skewness when it is accompanied by high-impact information releases, while it is
positively related to realized daily skewness when there are no information releases. In other words, we
note that the increase in the measures of divergence in investors’ opinion is associated with more
negatively skewed stocks, which we find are the main source of the positive relationship between
30
skewness and future returns. This implies that these riskier stocks tend to experience greater divergence in
investors’ opinion when there are high-impact information releases and tend to be over-valued (or
associated with lower subsequent returns), which is consistent with the prediction of Miller (1977) that
risky stocks will not be rewarded with high expected returns if they experience divergent investors’
opinion because negative opinions among investors is not revealed due to short-sale constraints.
31
References
Altınkılıç, Oya, and Robert S Hansen, 2009. On the information role of stock recommendation
revisions, Journal of Accounting and Economics 48, 17-36. Amaya, Diego, Peter Christoffersen, Kris Jacobs, and Aurelio Vasquez, 2013. Does realized
skewness predict the cross-section of equity returns? CREATES Research Paper. Andersen, Torben G, and Tim Bollerslev, 1998. Answering the skeptics: Yes, standard volatility
models do provide accurate forecasts, International economic review 39 (4), 885-905. Andersen, Torben G, Tim Bollerslev, Francis X Diebold, and Heiko Ebens, 2001. The
distribution of realized stock return volatility, Journal of Financial Economics 61, 43-76. Andersen, Torben G, Tim Bollerslev, Francis X Diebold, and Paul Labys, 2003. Modeling and
forecasting realized volatility, Econometrica 71, 579-625. Ang, Andrew, Robert J Hodrick, Yuhang Xing, and Xiaoyan Zhang, 2006. The cross-section of
volatility and expected returns, Journal of Finance 61, 259-299. Arditti, Fred D, 1967. Risk and the required rate on equity, Journal of Finance, 22 (1), 19-36. Bakshi, G, Charles Cao, and Zhiwu Chen, 1997. Empirical performance of alternative option
pricing models, Journal of Finance 52 (5), 2003-2049. Barberis, Nicholas, and Ming Huang, 2008. Stocks as lotteries: The Implications of probability
weighting for security prices, American Economic Review 98, 2066-2100. Boyer, Brian, Todd Mitton, and Keith Vorkink, 2010. Expected idiosyncratic skewness, Review
of Financial Studies 23 (1), 169-202. Bradley, Daniel, Jonathan Clarke, Suzanne Lee, and Chayawat Ornthanalai, 2014. Are analysts’
recommendations informative? Intraday evidence on the impact of time stamp delays, forthcoming in Journal of Finance.
Carhart, Mark M, 1997. On persistence in mutual fund performance, Journal of Finance 52, 57-82.
Chan, Wesley S, 2003. Stock price reaction to news and no-news: Drift and reversal after headlines, Journal of Financial Economics 70, 223-260.
Chen, Joseph, Harrison Hong, and Jeremy C Stein, 2001. Forecasting crashes: Trading volume, past returns, and conditional skewness in stock prices, Journal of Financial Economics 61, 345-381.
Conrad, Jennifer, Robert F Dittmar, and Eric Ghysels, 2013. Ex ante skewness and expected stock returns, Journal of Finance 68, 85-124.
Conrad, Jennifer S, Allaudeen Hameed, and Cathy Niden, 1994. Volume and auto-covariances in short-horizon individual security returns, Journal of Finance 49, 1305-1329.
Daniel, Kent, Ravi Jaganathan, and Soohun Kim, 2012, Tail risk in momentum strategy returns, NBER Working Paper Series.
Daniel, Kent and Tobias J Moskowitz, 2013, Momentum crashes, Swiss Finance Institute Research Paper Series.
Diether, Karl B, Christopher J Malloy, and Anna Scherbina, 2002. Differences of opinion and the cross-section of stock returns, Journal of Finance 57, 2113-2141.
Fama, Eugene F, and Kenneth R French, 1992. The cross-section of expected stock returns, Journal of Finance 47, 427-465.
Fama, Eugene F, and Kenneth R French, 1993. Common risk factors in the returns on stocks and bonds, Journal of Financial Economics 33, 3-56.
32
Fama, Eugene F, and James D MacBeth, 1973. Risk, return, and equilibrium: Empirical tests, Journal of Political Economy, 607-636.
Garfinkel, Jon A, 2009. Measuring investors' opinion divergence, Journal of Accounting
Research 47, 1317-1348. Garfinkel, Jon A, and Jonathan Sokobin, 2006. Volume, opinion divergence, and returns: A
study of post–earnings announcement drift, Journal of Accounting Research 44, 85-112. Harvey, Campbell R., and Akhtar Siddique, 2000. Conditional skewness in asset pricing tests.
Journal of Finance, 55 (3), 1263-1295. Jacod, Jean, 2012. Statistics and high frequency data, in Mathieu Kessler, Alexander Lindner,
and Michael Sorensen (eds.), Statistical methods for stochastic differential equations, 191-310, Boca Raton, FL: Chapman & Hall/CRC.
Jegadeesh, Narasimhan, and Sheridan Titman, 1993. Returns to buying winners and selling losers: Implications for stock market efficiency, Journal of Finance 48, 65-91.
Kelly, Bryan, 2012. Tail risk and asset prices, Chicago Booth Research Paper. Kraus, Alan, and Robert H Litzenberger, 1976. Skewness preference and the valuation of risk
assets, Journal of Finance 31, 1085-1100. Lee, Charles, and Bhaskaran Swaminathan, 2000. Price momentum and trading volume, Journal
of Finance 55, 2017-2069. Lee, Suzanne S, 2012. Jumps and information flow in financial markets, Review of Financial
Studies 25, 439-479. Lucca, David, and Emanuel Moench, 2011. The pre-FOMC announcement drift, FRB of New
York Staff Report. Merton, Robert C., 1976. Option pricing when underlying stock returns are discontinuous.
Journal of Financial Economics, 3, 125-144. Miller, Edward M, 1977. Risk, uncertainty, and divergence of opinion, Journal of Finance 32,
1151-1168. Mitton, Todd, and Keith Vorkink, 2007. Equilibrium under-diversification and the preference for
skewness, Review of Financial Studies 20, 1255-1288. Pritamani, Mahesh, and Vijay Singal, 2001. Return predictability following large price changes
and information releases, Journal of Banking & Finance 25, 631-656. Savor, Pavel, 2012. Stock returns after major price shocks: The impact of information, Journal of
Financial Economics 106 (3), 635-659. Savor, Pavel, and Mungo Wilson, 2013. How much do investors care about macroeconomic risk?
Evidence from scheduled economic announcements, Journal of Financial and
Quantitative Analysis 48 (2), 343-375. Tetlock, Paul C, 2010. Does public financial news resolve asymmetric information? Review of
Financial Studies 23, 3520-3557. Zhang, Yijie, 2005. Individual skewness and the cross-section of average stock returns, Yale
University Working Paper.
33
Figure 1. Univariate analysis for realized daily skewness and subsequent stock returns: Effect of high-impact information releases
This figure presents the relationship between realized daily skewness and subsequent stock returns (10-trading day cumulative return, ������,����). The figure is generated using 10,000 random
samples from original data for better presentation. The left panel presents the relationship between realized daily skewness and next 10-trading day returns when high-impact information about the
firm is released. The middle panel shows the relationship no high-impact information is released. The rightmost figure shows the relationship for the full sample.
34
Table 1. Summary statistics of base sample
This table provides time-series summary statistics of means and medians for the sample used in this paper. The first column in the table, Number of stocks, shows the numbers of stocks included in a
full sample of individual stocks, which is extracted from the TAQ database and merged with the CRSP daily stock file. The second column, Daily return, lists average daily returns of stocks computed
from the CRSP daily file. Realized daily variance, Realized daily skewness, and Realized daily kurtosis are computed using 5-minute returns from the TAQ database according to equations (1), (2),
and (3), respectively. ME ($ mil) represents market capitalization calculated using the closing market prices of firms. Vol pct is computed as the percentage of trading volume relative to total shares
outstanding. Mom is the cumulative return over 11 months, from previous 12 month to previous 2 month, as of day t. In 5-minute return and higher moment calculations, stocks with a price lower than
$5 are excluded. To be included in the sample, a stock must have at least 80 transactions within a trading day.
Number of Stocks Daily return Realized daily variance Realized daily skewness Realized daily kurtosis ME ($mil) Vol pct Mom
Mean Median Mean Median Mean Median Mean Median Mean Median Mean Median Mean Median
Table 2. Characteristics of decile portfolios sorted by realized higher moments
This table presents the cross-sectional characteristics of portfolios sorted by realized higher moments. All stocks in the sample are ranked by their
higher moments and assigned to deciles. The sorting procedure is repeated for every day from January 2005 to December 2010. Panel A, Panel B,
and Panel C present the results of sorted portfolios based on realized volatility, realized skewness and realized kurtosis for a corresponding day,
respectively. The first row of each panel, Next 20-day return, shows the average returns (in basis points) of sorted portfolios over the next 20 trading
days. The remaining characteristics of decile portfolios are computed for daily sorted portfolios. Size, BE/ME, Price, Intraday transactions, and
Number of stocks are the average market capitalization (in millions of dollars), average book-to-market equity ratio, average price of stocks in each
sorted portfolio, average intraday transactions per day, and average number of stocks included in daily decile portfolios, respectively.
Panel A: Characteristics of portfolios sorted by realized daily volatility
Number of stocks 298 298 298 298 298 298 298 298 298 298
36
Table 3. Time-series distribution of firm-specific information releases
This table describes our firm-date observations with and without firm-specific information releases over our sample period (from 2005 to 2010). The first two columns, Full sample, list the number of
all firm-date observations and average size of firms included in the sample for each year. The next two columns include information about the subsamples that includes firm-date observations
accompanied by two different types of firm-specific information releases. The Earnings announcements column shows the number of firm-date observations accompanied by earnings announcements
for corresponding firms (from the I/B/E/S database). The Analyst recommendations column shows the total number of subsample for all observations accompanied by one or more analyst
recommendation reports on a corresponding firm (from the I/B/E/S database). The two columns in Information sample show the observations accompanied by either earnings announcements or
analyst recommendations for a corresponding firm. No-information sample shows the subsample that includes observations that are not accompanied by the two types of firm-specific information
releases we consider in this study (i.e., earnings announcement and analyst recommendations).
Full sample Earnings announcements Analyst recommendations Information sample No-information sample
Obs Mean firm size (mil) Obs Mean firm size (mil) Obs Mean firm size (mil) Obs Mean firm size (mil) Obs Mean firm size (mil)
Table 4. Relationship between realized skewness, subsequent returns, and information releases: Information vs. no-information sample
This table presents the effect of information releases on the relationship between realized daily skewness and subsequent returns by separating all observations into an information sample and a no-
information sample. The table reports coefficient estimates of the following regression: ���,��,��� = + ����,� + ���,� + �,��,���, where �,��,��� is an error term.
Panel A provides results of the regression analysis for all firm-date observations. Panels B and C show the results of the regression analysis for the subsample with at least one of two firm-specific
information releases and the subsample without any of firm-specific information release, respectively. ���,��,��� is the cumulative return of the k-th stock over a period starting on day t+i and
ending on day t+j. ���,� is the realized daily skewness of the k-th stock returns on day � computed using 5-minute returns from the TAQ database. �,� includes several control variables for the k-th
firm on day t, including log size of the firm (log(ME)), log book-to-market ratio (log(BM) computed from COMPUSTAT), the cumulative return over 11 months, from previous 12 month to previous
2 month, as of day t (Mom), and the percentage of trading volume to total shares outstanding (vol pct). All parameters are computed using the Fama-MacBeth cross-sectional regression method and
the numbers in parentheses are t-statistics with robust standard errors. The sample period of analysis is from January 2005 to December 2010.
Panel A: Full sample Panel B: Information sample Panel C: No-information sample
*, **, and *** indicate statistical significance at the 10%, 5%, and 1% levels, respectively.
41
Table 7. Cross-section of stock returns and realized daily skewness
This table shows the raw returns and alpha of quintile portfolios formed based on realized daily skewness. Raw returns (in bps) are computed for quintile portfolios sorted on realized daily skewness
over 10-trading days and 20-trading days after sorting in the first and the second part of each panel, respectively. The results presented in CAPM, 3-Factor, and 4-Factor are the intercepts from time-
series regressions of the returns on quintile portfolios on market excess returns, three factors (market excess returns, size factor, and value factor), and four factors (market excess returns, size factor,
value factor, and momentum factor), respectively. All factors used in the analysis are drawn from the website of Kenneth French. Panel A shows the results for the full sample, Panel B for the
information sample, and Panel C for the no-information sample. The numbers in parentheses are t-statistics.
Panel A: Full sample Panel B. Information sample Panel C. No-information sample
*, **, and *** indicate statistical significance at the 10%, 5%, and 1% levels, respectively.
42
Table 8. Calendar-time zero investment portfolios
This table presents the profitability of zero-investment portfolios based on our finding of a relationship between realized daily skewness and subsequent stock returns.
The table reports coefficient estimates for the following regressions: ��,������� = ���� + ��������,��� + ��,���
���� where ��,���
���� is an error term.
The dependent variables in this analysis are the returns of zero-investment portfolios. The explanatory variables, ��,���, include common risk factors, including the market excess return (MKT), the size factor (SMB), the value
factor (HML), and the momentum factor (UMD). Each panel provides the results for the returns over the next 5-, 10-, and 20-trading days of holding periods after forming the portfolios. Panel A provides the result for the
zero-investment strategy that takes long positions on stocks that experienced negative realized daily skewness and short positions on stocks that experienced positive realized daily skewness without taking account of the
presence of information. Panel B provides the results of a zero-investment strategy that take a long position on the No-information sample and a short position on the Information sample. In panel C, the dependent variable is
the return of a zero-investment portfolio that takes long positions on stocks that experienced negative realized daily skewness with high-impact information releases and short positions on stocks that experienced positive
realized daily skewness with high-impact information releases. Panel D provides the results using as the dependent variable the return on the zero-investment strategy that takes long positions on stocks that experienced
negative realized daily skewness and short positions on stocks that experienced positive realized daily skewness without any high-impact information releases. To show the possible profitability, Panel E and F provide results
of zero-investment portfolios using the top and the bottom percentile portfolios sorted on the realized daily skewness and the presence or the absence of high-impact information releases, respectively. The numbers in
parentheses are t-statistics.
Panel A. Portfolio formed regardless of information releases
Panel B. Combined portfolio with Information and No-information samples
Model parameters ���� � !" �# $ �% & �' (
���� � !" �# $ �% & �' (
5-days Est 5.0318** 0.0686*** 7.9144** -0.0107
t-stat (2.35) (4.19) (2.35) (-0.48)
5-days Est 5.0131** 0.0735*** 0.0131 -0.0237 7.5516** -0.0241 0.0707** 0.0203
*, **, and *** indicate statistical significance at the 10%, 5%, and 1% levels, respectively.
45
Table 10. Result with weekly horizon
This table presents the robustness of our main findings to different estimation horizon for higher moments. The table reports the coefficient estimates for the following regression specification:
���,)�� = + ���*+,,) + �-��.�/,) + �0�1�2�,) + ���,) + �,)��, where �,)�� is an error term.
���,)�� is the subsequent weekly return of the k-th stock after investing in week w. �*+,,) , ��.�/,) and �1�2�,) are the realized volatility, realized skewness, and realized kurtosis of the k-th
stock returns over week w. Three higher moments are computed as the average of realized higher moments over a weekly horizon. Each week /, from Wednesday to the following Tuesday, each
variable is computed for each firm using 5-minute return data from the TAQ database. �,) includes several control variables for the k-th firm over week /, including log (Intra-transactions) is the log
of the average number of intraday transactions on each trading day, log(BM) is the log of book-to-market ratio, and log(Size) is the log of market capitalization.. The coefficient estimates are
computed using the Fama-MacBeth cross-sectional regression method. The numbers in parentheses are t-statistics with robust standard errors. The sample period of analysis is from January 2005 to
December 2010.
Panel A. Full sample Panel B. Information sample Panel C. No-information sample
*, **, and *** indicate statistical significance at the 10%, 5%, and 1% levels, respectively.
46
Table 11. Measures of divergence in investors’ opinion
This table provides the summary statistics for several measures of divergence in investors’ opinion considered in this paper. We report the number of observations (Obs), average (Mean), and standard
deviations (Std.) for the five different measures of divergence in investors’ opinion in the first, second, and third columns of each panel, respectively. In Panel A, the summary statistics for the full
sample are provided. We further separate our sample into a sample with information releases and a sample without information release. Panel B reports the statistics for divergence measures for the
sample with information releases, and Panel C reports those without information release. In Panel D, we further provide the differences in the divergence measures between Panel B and Panel C along
with their statistical significance. Details about the measures of divergence in investors’ opinion are provided in the main text and also in Garfinkel (2009). Daily turnover is computed as the ratio of
the firm’s daily trading volume on day t to total number of shares outstanding on day t. Market-adjusted daily turnover is the difference between a firm’s daily turnover and daily market-wide turnover
on day t (computed from market-wide trading volume and market-wide shares outstanding and adjusted with a 180-day median). Change in market-adjusted turnover is the difference in market-
adjusted turnover between day t and day t-1. Unexpected trading volume is a standardized prediction error from a regression of trading volume on the absolute value of returns of the k-th firm using
the equation (11), (12), and (13). Change in trading volume is the difference in trading volume between day t and day t-1.
Panel A: Full sample
Panel B: Sample with information
Panel C: Sample without information Panel D. (Information) – (No-information)
Variables Obs Mean Std.
Obs Mean Std.
Obs Mean Std. Difference
Daily turnover 2,551,563 0.0107 0.0197
313,654 0.0166 0.0262
2,237,909 0.0099 0.0185 0.0067***
Market-adjusted turnover 2,551,563 0.0009 0.0195
313,654 0.0062 0.0259
2,237,909 0.0001 0.0183 0.0061***
Change in market-adjusted turnover 2,551,563 0.0002 0.0223
*, **, and *** indicate statistical significance at the 10%, 5%, and 1% levels, respectively.
47
Table 12. Determinants of information-related skewness: Divergence in investors’ opinion
This table presents the relationship between realized daily skewness and the measures of divergence in investors’ opinion. The table reports coefficient estimates for the following ordinary least
squares regression in each subsample: ���,� = + ��34�25�67�,� + �,� where �,� is an error term.
That is, a dependent variable of the regression is a realized daily skewness (���,�) of the k-th firm on day t, and the explanatory variables �34�25�67�,� is selected from the measures of divergence
in investors’ opinion such as daily turnover, market-adjusted daily turnover, change in market-adjusted turnover, unexplained trading volume, and change in trading volume of the k-th firm on day t.
The first four measures of divergence in investors’ opinion divergence are documented in Garfinkel (2009). The realized daily skewness (���,�) for the k-th firm on day t is computed using 5-minute
returns from the TAQ database. Daily turnover is computed as the ratio of the firm’s daily trading volume on day t to total number of shares outstanding on day t. Market-adjusted daily turnover is the
difference between a firm’s daily turnover and daily market-wide turnover on day t (computed from market-wide trading volume and market-wide shares outstanding and adjusted with a 180-day
median). Change in market-adjusted turnover is the difference in market-adjusted turnover between day t and day t-1. Unexpected trading volume is a standardized prediction error from a regression
of trading volume on the absolute value of returns of the k-th firm using the equation (11), (12), and (13). Change in trading volume is the difference in trading volume between day t and day t-1.
Panel A shows the regression results for the full sample and includes all realized daily skewness, Panel B shows the subsample of realized daily skewness accompanied by firm-specific information
releases, and Panel C shows the subsample unaccompanied by any firm-specific information releases. The numbers in parentheses are t-statistics with robust standard errors.