P. 1
Forecasting Stock Market Crashes

Forecasting Stock Market Crashes

|Views: 19|Likes:
Published by maddierogers
Journal of Financial Economics 61 (2001) 345–381

Forecasting crashes: trading volume, past returns, and conditional skewness in stock prices$
Joseph Chena, Harrison Honga, Jeremy C. Steinb,*
a b

Graduate School of Business, Stanford University, Stanford, CA 94305, USA Department of Economics, Harvard University, Cambridge, MA 02138, USA Received 6 January 2000; accepted 17 July 2000

Abstract We develop a series of cross-sectional regression specifications to forecast skewness in the daily retu
Journal of Financial Economics 61 (2001) 345–381

Forecasting crashes: trading volume, past returns, and conditional skewness in stock prices$
Joseph Chena, Harrison Honga, Jeremy C. Steinb,*
a b

Graduate School of Business, Stanford University, Stanford, CA 94305, USA Department of Economics, Harvard University, Cambridge, MA 02138, USA Received 6 January 2000; accepted 17 July 2000

Abstract We develop a series of cross-sectional regression specifications to forecast skewness in the daily retu

More info:

Published by: maddierogers on Apr 29, 2011
Copyright:Attribution Non-commercial

Availability:

Read on Scribd mobile: iPhone, iPad and Android.
download as PDF, TXT or read online from Scribd
See more
See less

08/13/2012

pdf

text

original

Journal of Financial Economics 61 (2001) 345–381

Forecasting crashes: trading volume,
past returns, and conditional skewness in
stock prices
$
Joseph Chen
a
, Harrison Hong
a
, Jeremy C. Stein
b,
*
a
Graduate School of Business, Stanford University, Stanford, CA 94305, USA
b
Department of Economics, Harvard University, Cambridge, MA 02138, USA
Received 6 January 2000; accepted 17 July 2000
Abstract
We develop a series of cross-sectional regression specifications to forecast skewness in
the daily returns of individual stocks. Negative skewness is most pronounced in stocks
that have experienced (1) an increase in trading volume relative to trend over the prior
six months, consistent with the model of Hong and Stein (NBER Working Paper, 1999),
and (2) positive returns over the prior 36 months, which fits with a number of theories,
most notably Blanchard and Watson’s (Crises in Economic and Financial Structure.
Lexington Books, Lexington, MA, 1982, pp. 295–315) rendition of stock-price bubbles.
Analogous results also obtain when we attempt to forecast the skewness of the
aggregate stock market, though our statistical power in this case is limited. r 2001
Elsevier Science S.A. All rights reserved.
JEL classification: G12; G14
Keywords: Crashes; Trading volume; Skewness
$
We are grateful to the National Science Foundation for research support, and to John
Campbell, Kent Daniel, Ken Froot, Ravi Jagannathan, Phillipe Jorion, Chris Lamoreaux, Ken
Singleton, an anonymous referee, and seminar participants at Arizona, Arizona State, Cornell,
Harvard Business School, Northwestern, Maryland, Stanford, Texas, the UCLA Liquidity
Conference, and the NBER for helpful comments and suggestions. Thanks also to Jun Pan for
generously sharing her option-pricing software with us.
*Corresponding author.
E-mail address: jeremy stein@harvard.edu (J.C. Stein).
0304-405X/01/$ - see front matter r 2001 Elsevier Science S.A. All rights reserved.
PII: S 0 3 0 4 - 4 0 5 X( 0 1 ) 0 0 0 6 6 - 6
1. Introduction
Aggregate stock market returns are asymmetrically distributed. This
asymmetry can be measured in several ways. First, and most simply, the very
largest movements in the market are usually decreases, rather than increases –
that is, the stock market is more prone to melt down than to melt up. For
example, of the ten biggest one-day movements in the S&P 500 since 1947, nine
were declines.
1
Second, a large literature documents that market returns exhibit
negative skewness, or a closely related property, ‘‘asymmetric volatility’’ – a
tendency for volatility to go up with negative returns.
2
Finally, since the crash
of October 1987, the prices of stock index options have been strongly indicative
of a negative asymmetry in returns, with the implied volatilities of out-of-the-
money puts far exceeding those of out-of-the-money calls; this pattern has
come to be known as the ‘‘smirk’’ in index-implied volatilities. (See, e.g., Bates,
1997; Bakshi et al., 1997; and Dumas et al., 1998.)
While the existence of negative asymmetries in market returns is generally
not disputed, it is less clear what underlying economic mechanism these
asymmetries reflect. Perhaps the most venerable theory is based on leverage
effects (Black, 1976; Christie, 1982), whereby a drop in prices raises operating
and financial leverage, and hence the volatility of subsequent returns. However,
it appears that leverage effects are not of sufficient quantitative importance to
explain the data (Schwert, 1989; Bekaert and Wu, 2000). This is especially true
if one is interested in asymmetries at a relatively high frequency, e.g., in daily
data. To explain these, one has to argue that intraday changes in leverage have
a large impact on volatility – that a drop in prices on Monday morning leads to
a large increase in leverage and hence in volatility by Monday afternoon, so
that overall, the return for the full day Monday is negatively skewed.
An alternative theory is based on a ‘‘volatility feedback’’ mechanism. As
developed by Pindyck (1984), French et al. (1987), Campbell and Hentschel
(1992), and others, the idea is as follows: When a large piece of good news
arrives, this signals that market volatility has increased, so the direct positive
effect of the good news is partially offset by an increase in the risk premium. On
the other hand, when a large piece of bad news arrives, the direct effect and the
risk-premium effect now go in the same direction, so the impact of the news is
amplified. While the volatility-feedback story is in some ways more attractive
1
Moreover, the one increase – of 9.10% on October 21, 1987 – was right on the heels of the
20.47% decline on October 19, and arguably represented a correction of the microstructural
distortions that arose on that chaotic day, rather than an independent price change.
2
If, in a discrete-time setting, a negative return in period t raises volatility in period t þ1 and
thereafter, returns measured over multiple periods will be negatively skewed, even if single-period
returns are not. The literature on these phenomena includes Pindyck (1984), French et al. (1987),
Campbell and Hentschel (1992), Nelson (1991), Engle and Ng (1993), Glosten et al. (1993), Braun
et al. (1995), Duffee (1995), Bekaert and Wu (2000), and Wu (2001).
J. Chen et al. / Journal of Financial Economics 61 (2001) 345–381 346
than the leverage-effects story, there are again questions as to whether it has the
quantitative kick that is needed to explain the data. The thrust of the critique,
first articulated by Poterba and Summers (1986), is that shocks to market
volatility are for the most part very short-lived, and hence cannot be expected
to have a large impact on risk premiums.
A third explanation for asymmetries in stock market returns comes from
stochastic bubble models of the sort pioneered by Blanchard and Watson
(1982). The asymmetry here is due to the popping of the bubble – a
low-probability event that produces large negative returns.
What the leverage-effects, volatility-feedback, and bubble theories all have in
common is that they can be cast in a representative-investor framework. In
contrast, a more recent explanation of return asymmetries, Hong and Stein
(1999), argues that investor heterogeneity is central to the phenomenon. The
Hong-Stein model rests on two key assumptions: (1) there are differences of
opinion among investors as to fundamental value, and (2) some – though not
all – investors face short-sales constraints. The constrained investors can be
thought of as mutual funds, whose charters typically prohibit them from taking
short positions; the unconstrained investors can be thought of as hedge funds
or other arbitrageurs.
When differences of opinion are initially large, those bearish investors who
are subject to the short-sales constraint will be forced to a corner solution, in
which they sell all of their shares and just sit out of the market. As a
consequence of being at a corner, their information is not fully incorporated
into prices. For example, if the market-clearing price is $100, and a particular
investor is sitting out, it must be that his valuation is less than $100, but one
has no way of knowing by how much – it could be $95, but it could also be
much lower, say $50.
However, if after this information is hidden, other, previously more-bullish
investors have a change of heart and bail out of the market, the originally
more-bearish group may become the marginal ‘‘support buyers’’ and hence
more will be learned about their signals. In particular, if the investor who was
sitting out at a price of $100 jumps in and buys at $95, this is good news relative
to continuing to sit on the sidelines even as the price drops further. Thus,
accumulated hidden information tends to come out during market declines,
which is another way of saying that returns are negatively skewed.
With its focus on differences of opinion, the Hong-Stein model has
distinctive empirical implications that are not shared by the representative-
investor theories. In particular, the Hong-Stein model predicts that negative
skewness in returns will be most pronounced around periods of heavy trading
volume. This is because – like in many models with differences of opinion –
trading volume proxies for the intensity of disagreement. (See Varian, 1989;
Harris and Raviv, 1993; Kandel and Pearson, 1995; and Odean, 1998a for
other models with this feature.)
J. Chen et al. / Journal of Financial Economics 61 (2001) 345–381 347
When disagreement (and hence trading volume) is high, it is more likely that
bearish investors will wind up at a corner, with their information incompletely
revealed in prices. And it is precisely this hiding of information that sets the
stage for negative skewness in subsequent rounds of trade, when the arrival of
bad news to other, previously more-bullish investors can force the hidden
information to come out.
In this paper, we undertake an empirical investigation that is motivated by
this differences-of-opinion theory. We develop a series of cross-sectional
regression specifications that attempt to forecast skewness in the daily returns
to individual stocks. Thus, when we speak of ‘‘forecasting crashes’’ in the title
of the paper, we are adopting a narrow and euphemistic definition of the word
‘‘crashes,’’ associating it solely with the conditional skewness of the return
distribution; we are not in the business of forecasting negative expected returns.
This usage follows Bates (1991, 1997), who also interprets conditional skewness
– in his case, inferred from options prices – as a measure of crash expectations.
One of our key forecasting variables is the recent deviation of turnover from
its trend. For example, at the firm level, we ask whether the skewness in daily
returns measured over a given six-month period (say, July 1–December 31,
1998) can be predicted from the detrended level of turnover over the prior six-
month period (January 1–June 30, 1998). It turns out that firms that experience
larger increases in turnover relative to trend are indeed predicted to have more
negative skewness; moreover, the effect of turnover is strongly statistically and
economically significant.
In an effort to isolate the effects of turnover, our specifications also include a
number of control variables. These control variables can be divided into two
categories. In the first category are those that, like detrended turnover, capture
time-varying influences on skewness. The most significant variable in this
category is past returns. We find that when past returns have been high,
skewness is forecasted to become more negative. The predictive power is
strongest for returns in the prior six months, but there is some ability to predict
negative skewness based on returns as far back as 36 months. In a similar vein,
glamour stocks – those with low ratios of book value to market value – are also
forecasted to have more negative skewness. (Harvey and Siddique (2000) also
examine how skewness varies with past returns and book-to-market.) These
results can be rationalized in a number of ways, but they are perhaps most
clearly suggested by models of stochastic bubbles. In the context of a bubble
model, high past returns or a low book-to-market value imply that the bubble
has been building up for a long time, so that there is a larger drop when it pops
and prices fall back to fundamentals.
The second category of variables that help to explain skewness are those that
appear to be picking up relatively fixed firm characteristics. For example, it has
been documented by others (e.g., Damodaran, 1987; Harvey and Siddique,
2000) that skewness is more negative on average for large-cap firms – a pattern
J. Chen et al. / Journal of Financial Economics 61 (2001) 345–381 348
that also shows up strongly in our multivariate regressions. We are not aware
of any theories that would have naturally led one to anticipate this finding.
Rather, for our purposes a variable like size is best thought of as an atheoretic
control – it is included in our regressions to help ensure that we do not
mistakenly attribute explanatory power to turnover when it is actually
proxying for some other firm characteristic. Such a control might be redundant
to the extent that detrending the turnover variable already removes firm effects,
but we keep it in to be safe.
In addition to running our cross-sectional regressions with the individual-
firm data, we also experiment briefly with analogous time-series regressions for
the U.S. stock market as a whole. Here, we attempt to forecast the skewness in
the daily returns to the market using detrended market turnover and past
market returns. Obviously, this pure time-series approach entails an enormous
loss in statistical power – with data going back to 1962, we have less than 70
independent observations of market skewness measured at six-month intervals
– which is why it is not the main focus of our analysis. Nevertheless, it is
comforting to note that the qualitative results from the aggregate-market
regressions closely parallel those from the cross-sectional regressions in that
high values of both detrended turnover and past returns also forecast more
negative market skewness. The coefficient estimates continue to imply
economically meaningful effects, although that for detrended turnover is no
longer statistically significant.
While both the cross-sectional and time-series results for turnover are
broadly consistent with the theory we are interested in, we should stress that we
do not at this point view them as a tight test. There are several reasons why one
might wish to remain skeptical. First, beyond the effects of turnover, we
document other strong influences on skewness, such as firm size, that are not
easily rationalized within the context of the Hong-Stein model, and for which
there are no other widely accepted explanations. Second, even if innovations to
trading volume proxy for the intensity of disagreement among investors, they
likely capture other factors as well – such as changes in trading costs – that we
have not adequately controlled for. Finally, and most generally, our efforts to
model the determinants of conditional skewness at the firm level are really
quite exploratory in nature. Given how early it is in this game, we are naturally
reluctant to declare an unqualified victory for any one theory.
The remainder of the paper is organized as follows. In Section 2, we review
in more detail the theoretical work that motivates our empirical specification.
In Section 3, we discuss our sample and the construction of our key variables.
In Section 4, we present our baseline cross-sectional regressions, along with a
variety of sensitivities and sample splits. In Section 5, we consider the
analogous time-series regressions, in which we attempt to forecast the skewness
in aggregate-market returns. In Section 6, we use an option-pricing metric to
evaluate the economic significance of our results. Section 7 concludes.
J. Chen et al. / Journal of Financial Economics 61 (2001) 345–381 349
2. Theoretical background
The model of Hong and Stein (1999), which provides the principal
motivation for our empirical tests, begins with the assumption that there are
two investors, A and B, each of whom receives a private signal about a stock’s
terminal payoff. As a matter of objective reality, each investor’s signal contains
some useful information. However, each of the two investors only pays
attention to their own signal, even if that of the other investor is revealed to
them. This deviation from full Bayesian rationality – which can be thought of
as a form of overconfidence – leads to irreducible differences of opinion about
the stock’s value.
In addition to investors A and B, the model also incorporates a class of fully
rational, risk-neutral arbitrageurs. These arbitrageurs recognize that the best
estimate of the stock’s true value is formed by averaging the signals of A and B.
However, the arbitrageurs may not always get to see both of the signals,
because A and B face short-sales constraints. Importantly, the arbitrageurs
themselves are not short-sales constrained, so they can take infinitely large
positive or negative positions. Perhaps the most natural interpretation of these
assumptions is not to take the short-sales constraint literally – as an absolute
technological impediment to trade – but rather to think of investors A and B as
institutions like equity mutual funds, many of whom are precluded by their
charters or operating policies from ever taking short positions.
3
In contrast, the
arbitrageurs might be thought of as hedge funds who are not subject to such
restrictions.
Even though investors A and B can be said to suffer from behavioral biases
(i.e., overconfidence), the market as a whole is efficient, in the sense of there
being no predictability in returns. This is because of the presence of the risk-
neutral, unconstrained arbitrageurs. Hence, unlike most of the behavioral
finance literature, which relies on limited arbitrage, the model’s only
implications are for the higher-order moments of the return distribution.
There are two trading dates. To see how the model can generate
asymmetries, imagine that at time 1, investor B gets a pessimistic signal, so
that B’s valuation for the stock lies well below A’s. Because of the short-sales
constraint, B will simply sit out of the market, and the only trade will be
between investor A and the arbitrageurs. The arbitrageurs are rational enough
to figure out that B’s signal is below A’s, but they cannot know by how much.
3
In fact, Almazan et al. (1999) document that roughly 70% of mutual funds explicitly state (in
Form N-SAR that they file with the SEC) that they are not permitted to sell short. This is obviously
a lower bound on the fraction of funds that never take short positions. Moreover, Koski and
Pontiff (1999) find that 79% of equity mutual funds make no use whatsoever of derivatives (either
futures or options), suggesting that funds are also not finding synthetic ways to take short
positions.
J. Chen et al. / Journal of Financial Economics 61 (2001) 345–381 350
Thus the market price at time 1 impounds A’s prior information, but does not
fully reflect B’s time-1 signal.
Next, move to time 2, and suppose that A gets a new positive signal. In this
case, A continues to be the more optimistic of the two, so A’s new time-2 signal
is incorporated into the price, while B’s time-1 signal remains hidden. On the
other hand, if A gets a bad signal at time 2, some of B’s previously hidden
information might come out. This is because as A bails out of the market at
time 2, the arbitrageurs learn something by observing if and at what price B
steps in and starts being willing to buy. In other words, there is information in
how B responds to A’s reduced demand for the stock – in whether or not B gets
up off the sidelines and provides buying support. Thus more information
comes out, and variance is greater, when the stock price is falling at time 2, as
opposed to rising. This greater variance on the downside implies that time-2
returns will be negatively skewed.
However, this logic is not sufficient to establish that unconditional returns
(i.e., the average across time 1 and time 2) are negatively skewed. There is a
countervailing positive-skewness effect at time 1, since the most negative draws
of B’s signal are the ones that get hidden from the market at this time. When
A’s and B’s priors are sufficiently close to one another, the positive time-1
skewness can actually overwhelm the negative time-2 skewness, so that returns
are on average positively skewed. Nevertheless, Hong and Stein show that if
the ex ante divergence of opinion (i.e., the difference in priors) between A and B
is great enough, the time-2 effect dominates, and unconditional returns are
negatively skewed. It is this unconditional skewness feature – driven by the
short-sales constraint – that most clearly distinguishes the model of Hong and
Stein from other related models in which pent-up information is revealed
through the trading process (e.g., Grossman, 1988; Genotte and Leland, 1990;
Jacklin et al., 1992; and Romer, 1993). In these other models, returns are on
average symmetrically distributed, albeit potentially quite volatile.
Moreover, the ex ante divergence in priors between A and B – which Hong
and Stein denote by H – not only governs the extent of negative skewness, it
also governs trading volume. In particular, when H is large, trading volume is
unusually high at times 1 and 2. This high trading volume is associated with a
greater likelihood of B moving to the sidelines at time 1, and subsequently
moving off the sidelines at time 2 – precisely the mechanism that generates
negative skewness. Thus the comparative statics properties of the model with
respect to the parameter H lead to the prediction that increases in trading
volume should forecast more negative skewness. This comparative static result
holds regardless of whether unconditional skewness (averaged across different
values of H) is positive or negative, and it forms the basis for our empirical
tests.
In order to isolate this particular theoretical effect, we need to be aware of
other potentially confounding factors. For example, it is well known that
J. Chen et al. / Journal of Financial Economics 61 (2001) 345–381 351
trading volume is correlated with past returns (Shefrin and Statman, 1985;
Lakonishok and Smidt, 1986; Odean, 1998b). And, as noted above, past
returns might also help predict skewness, if there are stochastic bubbles of the
sort described by Blanchard and Watson (1982).
4
Indeed, just such a pattern
has been documented in recent work by Harvey and Siddique (2000). To
control for this tendency, all of our regressions include a number of lags of past
returns on the right-hand side.
In a similar vein, one might also worry about skewness being correlated with
volatility. There are a number of models that can deliver such a correlation; in
the volatility-feedback model of Campbell and Hentschel (1992), for example,
higher levels of volatility are associated with more negative skewness. To the
extent that such an effect is present in our data, we would like to know whether
turnover is forecasting skewness directly – as it should, according to the Hong-
Stein model – or whether it is really just forecasting volatility, which is in turn
correlated with skewness. To address this concern, all of our regressions
include some control for volatility, and we experiment with several ways of
doing this control.
3. Data
To construct our variables, we begin with data on daily stock prices and
monthly trading volume for all NYSE and AMEX firms, from the CRSP daily
and monthly stock files. Our sample period begins in July 1962, which is as far
back as we can get the trading volume data; because our regressions use many
lags, we do not actually begin to forecast returns until December 1965. We do
not include NASDAQ firms because we want to have a uniform and accurate
measure of trading volume, and the dealer nature of the NASDAQ market is
likely to render turnover in its stocks not directly comparable to that of NYSE
and AMEX stocks. We also follow convention and exclude ADRs, REITs,
closed-end funds, primes, and scoresFi.e., stocks that do not have a CRSP
share type code of 10 or 11.
For most of our analysis, we further truncate the sample by eliminating the
very smallest stocks in the NYSE/AMEX universe – in particular, those with a
market capitalization below the 20th percentile NYSE breakpoint. We do so
because our goal is to use trading volume as a proxy for differences of opinion.
Theoretical models that relate trading volume to differences of opinion
typically assume that transactions costs are zero. In reality, variations in
transactions costs are likely to be an important driver of trading volume, and
4
In the model of Coval and Hirshleifer (1998), there is also conditional negative skewness after
periods of positive returns, even though unconditional average skewness is zero.
J. Chen et al. / Journal of Financial Economics 61 (2001) 345–381 352
more so for very small stocks. By eliminating the smallest stocks, we hope to
raise the ratio of signal (differences of opinion) to noise (transactions costs) in
our key explanatory variable. We also report some sensitivities in which the
smallest stocks are analyzed separately (see Table 4 below), and as one would
expect from this discussion, the coefficients on turnover for this subsample are
noticeably smaller.
Our baseline measure of skewness, which we denote NCSKEW, for
‘‘negative coefficient of skewness,’’ is calculated by taking the negative of
(the sample analog to) the third moment of daily returns, and dividing it by
(the sample analog to) the standard deviation of daily returns raised to the
third power. Thus, for any stock i over any six-month period t; we have
NCSKEW
it
¼ À nðn À1Þ
3=2
¸
R
3
it

ðn À1Þðn À2Þ
¸
R
2
it

3=2

; ð1Þ
where R
it
represents the sequence of de-meaned daily returns to stock i during
period t; and n is the number of observations on daily returns during the
period. In calculating NCSKEW, as well as any other moments that rely on
daily return data, we drop any firm that has more than five missing
observations on daily returns in a given period. These daily ‘‘returns’’ are,
more precisely, actually log changes in price. We use log changes as opposed to
simple daily percentage returns because they allow for a natural benchmark – if
stock returns were lognormally distributed, then an NCSKEW measure based
on log changes should have a mean of zero. We have also redone everything
with an NCSKEW measure based instead on simple daily percentage returns,
and none of our main results are affected. Using simple percentage returns
instead of log changes does have two (predictable) effects: (1) it makes
returns look more positively skewed on average and (2) it induces a
pronounced correlation between skewness and contemporaneously measured
volatility. However, given that we control for volatility in all of our regression
specifications, using simple percentage returns does not materially alter the
coefficients on turnover and past returns.
Scaling the raw third moment by the standard deviation cubed allows for
comparisons across stocks with different variances; this is the usual normal-
ization for skewness statistics (Greene, 1993). By putting a minus sign in front
of the third moment, we are adopting the convention that an increase in
NCSKEW corresponds to a stock being more ‘‘crash prone’’ – i.e., having a
more left-skewed distribution.
For most of our regressions, the daily firm-level returns that go into the
calculation of the NCSKEW variable are market-adjusted returns – the log
change in stock i less the log change in the value-weighted CRSP index for that
day. However, we also run everything with variations of NCSKEW based on
both (1) excess returns (the log change in stock i less the T-bill return) as well as
J. Chen et al. / Journal of Financial Economics 61 (2001) 345–381 353
(2) beta-adjusted returns. As will be seen, these variations do not make much
difference to our results with NCSKEW.
In addition to NCSKEW, we also work with a second measure of return
asymmetries that does not involve third moments, and hence is less likely to be
overly influenced by a handful of extreme days. This alternative measure,
which we denote by DUVOL, for ‘‘down-to-up volatility,’’ is computed as
follows. For any stock i over any six-month period t; we separate all the days
with returns below the period mean (‘‘down’’ days) from those with returns
above the period mean (‘‘up’’ days), and compute the standard deviation for
each of these subsamples separately. We then take the log of the ratio of (the
sample analog to) the standard deviation on the down days to (the sample
analog to) the standard deviation on the up days. Thus we have
DUVOL
it
¼ log ðn
u
À1Þ
¸
DOWN
R
2
it

ðn
d
À1Þ
¸
UP
R
2
it
¸
; ð2Þ
where n
u
and n
d
are the number of up and down days, respectively. Again, the
convention is that a higher value of this measure corresponds to a more left-
skewed distribution. To preview, our results with NCSKEW and DUVOL are
for the most part quite similar, so it does not appear that they depend on a
particular parametric representation of return asymmetries.
In our regressions with firm-level data, we use nonoverlapping six-month
observations on skewness. In particular, the NCSKEW and DUVOL measures
are calculated using data from either January 1–June 30 or July 1–December 31
of each calendar year. We could alternatively use overlapping data, so that we
would have a new skewness measure every month, but there is little payoff to
doing so, since, as will become clear shortly, we already have more than enough
statistical power as it is. We have, however, checked our results by re-running
everything using different nonoverlapping intervals – e.g., February 1–July 31
and August 1–January 31, March 1–August 31 and September 1–February 28,
etc. In all cases, the results are essentially identical. When we turn to the time-
series regressions with aggregate-market data, statistical power becomes a real
issue, and we use overlapping observations.
The choice of a six-month horizon for measuring skewness is admittedly
somewhat arbitrary. In principle, the effects that we are interested in could be
playing themselves out over a shorter horizon, so that trading volume on
Monday forecasts skewness for the rest of the week, but has little predictive
power beyond that. Unfortunately, the model of Hong and Stein does not give
us much guidance in this regard. Lacking this theoretical guidance, our choice
to use six months’ worth of daily returns to estimate skewness is driven more
by measurement concerns. For example, if we estimated skewness using only
one month’s worth of data, we would presumably have more measurement
error; this is particularly relevant given that a higher-order moment like
J. Chen et al. / Journal of Financial Economics 61 (2001) 345–381 354
skewness is strongly influenced by outliers in the data. The important point to
note, however, is that to the extent that our measurement horizon does not
correspond well to the underlying theory, this should simply blur our ability to
find what the theory predicts – i.e., it should make our tests too conservative.
Besides the skewness measures, the other variables that we use are very
familiar and do not merit much discussion. SIGMA
it
is the standard deviation
of stock i ’s daily returns, measured over the six-month period t: RET
it
is the
cumulative return on stock i; also measured over the six-month period t: When
we compute NCSKEW or DUVOL using either market-adjusted or beta-
adjusted returns, SIGMA and RET are computed using market-adjusted
returns. When we compute NCSKEW or DUVOL using excess returns,
SIGMA and RET are based on excess returns as well.
LOGSIZE
it
is the log of firm i ’s stock market capitalization at the end of
period t: BK/MKT
it
is firm i ’s book-to-market ratio at the end of period t:
LOGCOVER
it
is the log of one plus the number of analysts (from the I/B/E/S
database) covering firm i at the end of period t: TURNOVER
it
is the average
monthly share turnover in stock i; defined as shares traded divided by shares
outstanding over period t:
In our baseline specification, we work with detrended turnover, which we
denote DTURNOVER. The detrending is done very simply, by subtracting
from the TURNOVER variable a moving average of its value over the prior 18
months. Again, the rationale for doing this detrending is that, as a matter of
conservatism, we want to eliminate any component of turnover that can be
thought of as a relatively fixed firm characteristic. This detrending is roughly
analogous to doing a fixed-effects specification in a shorter-lived panel. Since
we have such a long time series, it makes little sense to require that firm effects
be literally constant over the entire sample period. Instead, the detrending
controls for firm characteristics that adjust gradually.
Table 1 presents a variety of summary statistics for our sample. Panel A
shows the means and standard deviations of all of our variables for (1) the full
sample of individual firms, (2) five size-based subsamples, and (3) the market as
a whole, defined as the value-weighted NYSE/AMEX index. (When working
with the market as a whole, all the variables are based on simple excess returns
relative to T-bills.) Panels B and C look at contemporaneous correlations and
autocorrelations, respectively, for the sample of individual firms. In Panels B
and C, as in most of our subsequent regression analysis, we restrict the sample
to those firms with a market capitalization above the 20th percentile NYSE
breakpoint.
One interesting point that emerges from Panel A is that while there is
negative skewness – i.e., positive mean values of NCSKEW and DUVOL – for
the market as a whole, the opposite is true for individual stocks, which are
positively skewed. This discrepancy can in principle be understood within the
strict confines of the Hong-Stein model, since, as noted above, the model
J. Chen et al. / Journal of Financial Economics 61 (2001) 345–381 355
Table 1
Summary statistics
The sample period is from July 1962 to December 1998, except for LOGCOVER
t
, which is measured starting in December 1976. NCSKEW
t
is the
negative coefficient of (daily) skewness, measured using market-adjusted returns in the six-month period t. DUVOL
t
is the log of the ratio of down-day
to up-day standard deviation, measured using market-adjusted returns in the six-month period t. SIGMA
t
is the standard deviation of (daily) market-
adjusted returns measured in the six-month period t. LOGSIZE
t
is the log of market capitalization measured at the end of period t. BK=MKT
t
is the
most recently available observation of the book-to-market ratio at the end of period t. LOGCOVER
t
is the log of one plus the number of analysts
covering the stock at the end of period t. DTURNOVER
t
is average monthly turnover in the six-month period t, detrended by a moving average of
turnover in the prior 18 months. TURNOVER
t
is the average monthly turnover measured in the six-month period t. RET
t
is the market-adjusted
cumulative return in the six-month period t. Size quintiles are determined using NYSE breakpoints.
All firms
Quintile 5
(largest)
firms
Quintile 4
firms
Quintile 3
firms
Quintile 2
firms
Quintile 1
(smallest)
firms
Market
portfolio
Panel A: First and second moments
NCSKEW
t
Mean À0.262 À0.139 À0.155 À0.198 À0.266 À0.362 0.268
Standard dev. 0.939 0.806 0.904 0.923 0.994 0.964 0.735
DUVOL
t
Mean À0.190 À0.128 À0.141 À0.171 À0.213 À0.224 0.172
Standard dev. 0.436 0.364 0.391 0.406 0.437 0.476 0.377
J
.
C
h
e
n
e
t
a
l
.
/
J
o
u
r
n
a
l
o
f
F
i
n
a
n
c
i
a
l
E
c
o
n
o
m
i
c
s
6
1
(
2
0
0
1
)
3
4
5

3
8
1
3
5
6
SIGMA
t
Mean 0.025 0.015 0.017 0.020 0.023 0.034 0.008
Standard dev. 0.018 0.005 0.007 0.008 0.010 0.023 0.003
LOGSIZE
t
Mean 5.177 8.249 6.860 5.924 4.984 3.121 N/A
Standard dev. 2.073 1.035 0.653 0.642 0.656 1.108
BK/MKT
t
Mean 0.983 0.667 0.782 0.824 0.935 1.275 N/A
Standard dev. 14.036 0.472 0.710 0.870 1.197 22.920
LOGCOVER
t
Mean 1.991 3.006 2.512 2.030 1.565 1.140 N/A
Standard dev. 0.840 0.431 0.503 0.563 0.564 0.508
DTURNOVER
t
Mean 0.001 0.000 0.002 0.002 0.001 À0.000 0.002
Standard dev. 0.066 0.039 0.040 0.042 0.046 0.095 0.005
TURNOVER
t
Mean 0.050 0.051 0.056 0.055 0.054 0.043 0.037
Standard dev. 0.075 0.050 0.055 0.060 0.063 0.098 0.022
RET
t
Mean 0.003 0.024 0.015 0.021 0.017 À0.019 0.029
Standard dev. 0.297 0.164 0.202 0.240 0.288 0.372 0.108
No. of obs. 100,898 13,988 14,291 14,727 16,651 41,241 421
J
.
C
h
e
n
e
t
a
l
.
/
J
o
u
r
n
a
l
o
f
F
i
n
a
n
c
i
a
l
E
c
o
n
o
m
i
c
s
6
1
(
2
0
0
1
)
3
4
5

3
8
1
3
5
7
NCSKEW
tÀ1
DUVOL
tÀ1
SIGMA
tÀ1
LOGSIZE
tÀ1
BK=MKT
tÀ1
LOGCOVER
tÀ1
DTURNOVER
tÀ1
TURNOVER
tÀ1
RET
tÀ1
Panel C: Autocorrelations and cross-correlations (using only firms above 20th percentile in size)
NCSKEW
t
0.047 0.059 À0.047 0.063 À0.030 0.056 0.022 0.032 0.043
DUVOL
t
0.061 0.090 À0.109 0.068 À0.011 0.066 0.016 À0.024 0.047
SIGMA
t
À0.008 À0.071 0.715 À0.292 À0.050 À0.218 0.042 0.318 À0.014
LOGSIZE
t
0.049 0.055 À0.342 0.976 À0.182 0.719 0.000 0.093 À0.011
BK/MKT
t
0.022 0.047 À0.067 À0.181 0.782 À0.027 0.022 0.017 À0.080
LOGCOVER
t
0.079 0.098 À0.257 0.736 À0.035 0.852 0.006 0.166 À0.079
DTURNOVER
t
À0.028 À0.028 À0.059 0.009 0.039 0.019 0.381 À0.130 0.119
TURNOVER
t
0.015 À0.052 0.294 0.104 0.029 0.179 0.195 0.781 0.086
RET
t
À0.002 0.006 À0.032 À0.042 0.051 À0.023 À0.013 À0.064 0.030
Table 1 (continued)
NCSKEW
t
DUVOL
t
SIGMA
t
LOGSIZE
t
BK/MKT
t
LOGCOVER
t
DTURNOVER
t
TURNOVER
t
RET
t
Panel B: Contemporaneous correlations (using only firms above 20th percentile in size)
NCSKEW
t
0.875 0.008 0.038 0.311 0.081 0.007 0.028 À0.302
DUVOL
t
À0.076 0.045 0.068 0.100 À0.013 À0.042 À0.371
SIGMA
t
À0.307 À0.056 À0.238 0.130 0.398 0.034
LOGSIZE
t
À0.213 0.729 0.002 0.101 À0.014
BK/MKT
t
À0.026 0.026 0.006 À0.104
LOGCOVER
t
À0.013 0.158 À0.080
DTURNOVER
t
0.376 0.133
TURNOVER
t
0.061
RET
t
J
.
C
h
e
n
e
t
a
l
.
/
J
o
u
r
n
a
l
o
f
F
i
n
a
n
c
i
a
l
E
c
o
n
o
m
i
c
s
6
1
(
2
0
0
1
)
3
4
5

3
8
1
3
5
8
allows for either positive or negative unconditional skewness, depending on
the degree of ex ante investor heterogeneity. In other words, if one is willing
to assume that differences of opinion about the market are on average
more pronounced than differences of opinion about individual stocks, the
model can produce negative skewness for the latter and positive skewness for
the former.
However, it is not clear that such an assumption is empirically defensible. An
alternative interpretation of the data in Table 1A is that even if the Hong-Stein
model provides a reasonable account of skewness in market returns, it must be
missing something when it comes to explaining the mean skewness of
individual stocks. For example, it might be that large positive events like
hostile takeovers (which the theory ignores) impart an added degree of positive
skewness to individual stocks but wash out across the market as a whole. This
view does not imply that we cannot learn something about the theory by
looking at firm-level data; the theory will certainly gain some credence if it does
a good job of explaining cross-sectional variation in skewness, even if it cannot
fit the mean skewness at the firm level. Nevertheless, it is worth emphasizing
the caveat that, without further embellishments, the theory might not provide a
convincing rationale for everything that is going on at the individual stock
level.
The most noteworthy fact in Panel B of Table 1 is the contemporaneous
correlation between our two skewness measures, NCSKEW and DUVOL,
which is 0.88. While these two measures are quite different in their
construction, they appear to be picking up much the same information. Also
worth pointing out is that the correlation between NCSKEW and SIGMA is
less than 0.01, and that between DUVOL and SIGMA is about À0.08; these
low correlations lend some preliminary (and comforting) support to the notion
that forecasting either of our skewness measures is a quite distinct exercise
from forecasting volatility. Panel C documents that, unlike SIGMA – which
has an autocorrelation coefficient of 0.72 – neither of our skewness measures
has much persistence. For NCSKEW the autocorrelation is on the order of
0.05; for DUVOL it is 0.09.
4. Forecasting skewness in the cross-section
4.1. Baseline specification
Table 2 presents our baseline cross-sectional regression specification. We
pool all the data (excluding firms with market capitalization below the 20th
percentile NYSE breakpoint) and regress NCSKEW
it+1
against its own lagged
value, NCSKEW
it
, as well as SIGMA
it
, LOGSIZE
it
, DTURNOVER
it
, and
six lags of past returns, RET
it
yRET
itÀ5
. We also include dummy variables
J. Chen et al. / Journal of Financial Economics 61 (2001) 345–381 359
for each time period t: The regression can be interpreted as an effort to
predict – based on information available at the end of period t – cross-sectional
variation in skewness over period t þ1:
Table 2
Forecasting skewness in the cross-section: pooled regressions
The sample period is from July 1962 to December 1998 and includes only those firms with market
capitalization above the 20th percentile breakpoint of NYSE. The dependent variable is
NCSKEW
tþ1
; the negative coefficient of (daily) skewness in the six-month period t þ1:
NCSKEW
tþ1
is computed based on market-adjusted returns, beta-adjusted returns and simple
excess returns in cols. 1–3, respectively. SIGMA
t
is the (daily) standard deviation of returns in the
six-month period t: LOGSIZE
t
is the log of market capitalization at the end of period t:
DTURNOVER
t
is average monthly turnover in the six-month period t; detrended by a moving
average of turnover in the prior 18 months. RET
t
yRET
tÀ5
are returns in the six-month periods t
through t À5 (these past returns are market adjusted in cols. 1À2 and excess in col. 3). All
regressions also contain dummies for each time period (not shown); t-statistics, which are in
parentheses, are adjusted for heteroskedasticity and serial correlation.
(1) Base case:
market-adjusted
returns
(2) Beta-adjusted
returns
(3) Excess
returns
NCSKEW
t
0.053 0.051 0.052
(7.778) (7.441) (7.920)
SIGMA
t
À4.566 À3.370 À2.701
(À7.180) (À5.242) (À4.706)
LOGSIZE
t
0.037 0.046 0.059
(11.129) (13.465) (19.110)
DTURNOVER
t
0.437 0.364 0.364
(3.839) (3.175) (3.329)
RET
t
0.218 0.197 0.221
(10.701) (9.638) (11.607)
RET
tÀ1
0.082 0.082 0.109
(4.296) (4.220) (6.175)
RET
tÀ2
0.103 0.108 0.089
(5.497) (5.675) (5.149)
RET
tÀ3
0.054 0.067 0.053
(2.830) (3.462) (3.001)
RET
tÀ4
0.062 0.058 0.041
(3.403) (3.133) (2.477)
RET
tÀ5
0.071 0.083 0.092
(3.759) (4.335) (5.257)
No. of obs. 51,426 51,426 51,426
R
2
0.030 0.031 0.082
J. Chen et al. / Journal of Financial Economics 61 (2001) 345–381 360
In column 1, we use market-adjusted returns as the basis for computing the
NCSKEW measure. In column 2 we use beta-adjusted returns, and in column 3
we use simple excess returns. The results are quite similar in all three cases. In
particular, the coefficients on detrended turnover are positive and strongly
statistically significant in each of the three columns, albeit somewhat larger (by
about 20%) in magnitude when market-adjusted returns are used. (We expect
lower coefficient estimates when using simple excess returns as compared to
market-adjusted returns – after all, DTURNOVER is a firm-specific variable,
so it should have more ability to explain skewness in the purely idiosyncratic
component of stock returns.) The past return terms are also always positive
and strongly significant. Thus stocks that have experienced either a surge in
turnover or high past returns are predicted to have more negative skewness –
i.e., to become more crash-prone, all else equal. The coefficient on size is also
positive, suggesting that negative skewness is more likely in large-cap stocks.
As noted above, the findings for past returns and size run broadly parallel to
previous work by Harvey and Siddique (2000). Nevertheless, there are several
distinctions between our results and theirs. To begin, ours are couched in a
multivariate regression framework, while theirs are based on univariate sorts.
But more significantly, our measure of skewness is quite different from theirs,
for two reasons. First, we look at daily returns, while they look at monthly
returns. Second, we look at individual stocks, while they look at portfolios of
stocks. The skewness of a portfolio of stocks is not the same thing as the
average skewness of its component stocks, especially if, as Harvey and
Siddique (2000) stress, coskewness varies systematically with firm character-
istics.
We have done some detailed comparisons to make these latter points
explicit. For 25 portfolios sorted on size and book-to-market, we have
computed both (1) the skewness of monthly portfolio returns, as in Harvey and
Siddique (2000), and (2) the average skewness of daily individual stock returns,
a measure analogous to what we use here. We can then ask the following:
Across the 25 portfolios, what is the correlation of the two skewness measures?
The answer is about 0.22, a relatively low, albeit significantly positive,
correlation. Thus, while it might have been reasonable to conjecture – based on
the prior evidence in Harvey and Siddique (2000) – that our firm-level
NCSKEW variable would also be related to past returns and size, such results
were by no means a foregone conclusion.
As we have already stressed, the positive coefficient on size is not something
one would have necessarily predicted ex ante based on the Hong-Stein model.
Nevertheless, it is possible to come up with rationalizations after the fact.
Suppose that managers can to some extent control the rate at which
information about their firms gets out. It seems plausible that if they uncover
good news, they will disclose all this good news right away. In contrast, if they
are sitting on bad news, they may try to delay its release, with the result that the
J. Chen et al. / Journal of Financial Economics 61 (2001) 345–381 361
bad news dribbles out slowly. This behavior will tend to impart positive
skewness to firm-level returns, and may explain why returns on individual
stocks are on average positively skewed at the same time that market returns
are negatively skewed. Moreover, if one adds the further assumption that it is
easier for managers of small firms to temporarily hide bad news – since they
face less scrutiny from outside analysts than do managers of large firms – the
resulting positive skewness will be more pronounced for small firms. We return
to this idea in Section 4.5 below, and use it to develop some additional testable
implications.
4.2. Robustness
In Table 3 we conduct a number of further robustness checks. Everything
is a variation on column 1 of Table 2, and uses an NCSKEW measure
based on market-adjusted returns. First, in column 1 of Table 3, we truncate
outliers of the NCSKEW variable, setting all observations that are more
than three standard deviations from the mean in any period t to the three-
standard-deviation tail values in that period. This has little impact on
the results, suggesting that they are not driven by a handful of outlier
observations.
In column 2, we replace the DTURNOVER variable with its un-detrended
analog, TURNOVER. This means that we are now admitting into considera-
tion differences in turnover across firms that are not merely temporary
deviations from trend but more in the nature of long-run firm characteristics.
In other words, we are essentially removing our fixed-effect control from the
turnover variable. According to the theory, one might expect that long-run
cross-firm variation in turnover would also predict skewness – some firms
might be subject to persistently large differences in investor opinion, and these
too should matter for return asymmetries. The coefficient estimate on
TURNVOVER in column 2 confirms this notion, roughly doubling in
magnitude from its base-case value. This implies that our fixed-effect approach
of using DTURNOVER instead of TURNOVER everywhere else in the paper
is quite conservative – in doing so, we are throwing out a dimension of the data
that is strongly supportive of the theory.
In columns 3 and 4, we investigate whether our results are somehow tied to
the way that we have controlled for volatility. Recall that the central issue here
is whether DTURNOVER
it
is really forecasting NCSKEW
it+1
directly, or
whether it is instead forecasting SIGMA
it+1
, and showing up in the regression
only because SIGMA
it+1
is correlated with NCSKEW
it+1
. Ideally, we would
like to add a period-t control variable to the regression that is a good forecast
of SIGMA
it+1
, so that we can verify that DTURNOVER
it
is still significant
even after the inclusion of this control. Our use of SIGMA
it
in the base-case
specification can be motivated on the grounds that it is probably the best
J. Chen et al. / Journal of Financial Economics 61 (2001) 345–381 362
univariate predictor of SIGMA
it+1
, given the very pronounced serial
correlation in the SIGMA variable.
But using just one past lag is not necessarily the best way to forecast
SIGMA
it+1
. One can presumably do better by allowing for richer dynamics. In
this spirit, we add in column 3 two further lags of SIGMA (SIGMA
itÀ1
and
SIGMA
itÀ2
) to the base-case specification. These two lags are insignificant, and
hence our coefficients on DTURNOVER
it
as well as on the six RET terms are
virtually unchanged. Motivated by the work of Glosten et al. (1993), who find
that the effect of past volatility on future volatility depends on the sign of the
past return, we also experiment with allowing two coefficients on SIGMA
it
, one
for positive past returns and one for negative past returns. This variation (not
shown in the table) makes no difference to the results.
In column 4 we take our logic one step further. We create a fitted value of
SIGMA
it+1
, which we denote by SIGMAHAT
it+1,
based on the following
information set available in period t: SIGMA
it
, SIGMA
itÀ1
, SIGMA
itÀ2
,
LOGSIZE
it
, DTURNOVER
it
, and RET
it
yRET
itÀ5
. We then replace
SIGMA
it
in the base case with this fitted value of future volatility,
SIGMAHAT
it+1
. This is equivalent to an instrumental-variables regression
in which future volatility SIGMA
it+1
is included on the right-hand side, but is
instrumented using the information available in period t: As can be seen, this
variation leads to almost exactly the same results as in the base case.
Overall, based on the evidence in columns 3 and 4 of Table 3, we conclude
that it is highly unlikely that our base-case success in forecasting NCSKEW
with the DTURNOVER and RET variables arises because these variables are
able to forecast SIGMA. In other words, these variables really appear to be
predicting cross-firm differences in the asymmetry of stock returns, rather than
just differences in volatility.
In column 5, we add the book-to-market ratio, BK/MKT, to the
base-case specification. This variable attracts a significant negative coefficient,
which means that it tells the same story as the past-return terms: glamour
stocks, like those with high past returns, are more crash-prone. However, the
addition of BK/MKT has no impact on the DTURNOVER coefficient.
In column 6, we use the DUVOL measure of return asymmetry as the left-
hand-side variable in place of NCSKEW. Although the difference in units
precludes a direct comparison of the point estimates, the qualitative patterns
are generally the same as in the corresponding specification in column 1 of
Table 2. Indeed, the t-statistic on DTURNOVER is actually a bit higher (4.35
vs. 3.84) as is the R
2
of the regression (6.7% vs. 3.0%).
Finally, in an unreported sensitivity, we check to make sure our results are
robust to how we have modeled the effect of the lagged skewness variable,
NCSKEW
it
. Instead of estimating just one coefficient on NCSKEW
it
, we allow
this effect to be a function of the realization of NCSKEW
it
itself. We
implement this by interacting NCSKEW
it
with five dummy variables,
J. Chen et al. / Journal of Financial Economics 61 (2001) 345–381 363
Table 3
Forecasting skewness in the cross-section: robustness checks
The sample period is from July 1962 to December 1998 and includes only those firms with market capitalization above the 20th percentile breakpoint
of NYSE. In columns 1–5, the dependent variable is NCSKEW
tþ1
; the negative coefficient of (daily) skewness in the six-month period t þ1: In column
6, the dependent variable is DUVOL
tþ1
; the log of the ratio of down-day to up-day standard deviation in the six-month period t þ1: In all columns,
returns are market-adjusted. SIGMA
t
is the standard deviation of (daily) returns in the six-month period t: LOGSIZE
t
is the log of market
capitalization at the end of period t: BK/MKT
t
is the most recently available observation of the book-to-market ratio at the end of period t:
DTURNOVER
t
is average monthly turnover in the six-month period t; detrended by a moving average of turnover in the prior 18 months, except in
column 3, where turnover is not detrended. RET
t
yRET
tÀ5
are returns in the six-month periods t through t À5: SIGMAHAT
tþ1
is the predicted value
of SIGMA
tþ1
calculated from a regression of SIGMA
tþ1
on SIGMA
t
,y, SIGMA
tÀ2
; LOGSIZE
t
, DTURNOVER
t
, and RET
t
...RET
tÀ5
: All
regressions also contain dummies for each time period (not shown); t-statistics, which are in parentheses, are adjusted for heteroskedasticity and serial
correlation.
(1) Outliers
truncated
(2) Turnover
not detrended
(3) More lags
of past volatility
(4) Fitted future
volatility
(5) Book-to-
market
(6) Using
DUVOL
t+1
NCSKEW
t
0.050 0.053 0.053 0.051 0.054 0.096
(DUVOL
t
in col. 6) (8.675) (7.837) (7.663) (7.454) (7.750) (16.627)
SIGMAHAT
tþ1
À6.178
(À7.180)
SIGMA
t
À4.994 À6.618 À3.953 À4.999 À4.956
(À8.938) (À9.822) (À3.751) (À7.552) (À15.698)
SIGMA
tÀ1
À0.460
(À0.384)
SIGMA
tÀ2
À0.367
(À0.353)
LOGSIZE
t
0.035 0.033 0.037 0.034 0.035 0.014
(12.047) (9.980) (10.898) (9.351) (10.095) (9.572)
BK/MKT
t
À0.020
(À3.808)
J
.
C
h
e
n
e
t
a
l
.
/
J
o
u
r
n
a
l
o
f
F
i
n
a
n
c
i
a
l
E
c
o
n
o
m
i
c
s
6
1
(
2
0
0
1
)
3
4
5

3
8
1
3
6
4
DTURNOVER
t
0.375 0.761 0.411 0.387 0.455 0.202
(TURNOVER
t
in col. 2) (3.729) (7.685) (3.459) (3.410) (3.848) (4.346)
RET
t
0.206 0.217 0.218 0.208 0.213 0.142
(11.787) (10.887) (10.761) (10.249) (10.071) (15.810)
RET
tÀ1
0.075 0.071 0.083 0.084 0.081 0.014
(4.587) (3.828) (4.329) (4.428) (4.087) (1.671)
RET
tÀ2
0.100 0.088 0.104 0.106 0.098 0.045
(6.273) (4.734) (5.472) (5.621) (5.038) (5.587)
RET
tÀ3
0.049 0.033 0.054 0.056 0.056 0.009
(3.030) (1.727) (2.819) (2.943) (2.816) (1.131)
RET
tÀ4
0.048 0.041 0.060 0.064 0.051 0.014
(3.084) (2.287) (3.337) (3.523) (2.722) (1.808)
RET
tÀ5
0.057 0.054 0.072 0.073 0.066 0.014
(3.580) (2.923) (3.789) (3.820) (3.324) (1.705)
No. of obs. 51,426 52,229 51,393 51,426 48,630 51,426
R
2
0.039 0.031 0.030 0.030 0.030 0.067
J
.
C
h
e
n
e
t
a
l
.
/
J
o
u
r
n
a
l
o
f
F
i
n
a
n
c
i
a
l
E
c
o
n
o
m
i
c
s
6
1
(
2
0
0
1
)
3
4
5

3
8
1
3
6
5
one corresponding to each quintile of NCSKEW
it
. In other words, we estimate
five separate slope coefficients on lagged skewness, depending on the
quintile that lagged skewness falls in. As it turns out, while there appear to
be some modest nonlinearities in the effect of lagged skewness, these
nonlinearities do not at all impact the coefficients on any of the other variables
of interest.
4.3. Cuts on firm size
In Table 4, we disaggregate our base-case analysis by size. We take the
specification from column 1 of Table 2 and run it separately for five size-based
subsamples, corresponding to quintiles based on NYSE breakpoints. (Recall
that in Tables 2 and 3, we omit the smallest of these five quintiles from
our sample.) Two conclusions stand out. First, as suspected, the coefficient
on DTURNOVER for the smallest category of firms is noticeably lower
than for any other group, albeit still positive.
5
Again, this is probably
because variation in turnover for these tiny firms is driven in large part by
variation in trading costs, whereas our theory requires a good proxy for
differences of opinion. Second, once one moves beyond the smallest quintile,
the coefficients look reasonably stable. There is certainly no hint that the effects
that we are interested in go away for larger firms. Indeed, the highest point
estimate for the DTURNOVER coefficient comes from the next-to-largest
quintile.
The fact that the coefficients on DTURNOVER are robust for large firms is
not surprising in light of the underlying theory. As we have emphasized, the
model of Hong and Stein is not predicated on impediments to arbitrage – it
incorporates a class of fully risk-neutral arbitrageurs who can take infinite long
or short positions. Thus, as long as some investors other than the arbitrageurs
(e.g., mutual funds) continue to be short-sales constrained, the model does not
have the feature that the key effects diminish as one moves to larger stocks,
where arbitrage activity is presumably more efficient. This is in contrast to
behavioral models based on limited arbitrage (e.g., DeLong et al., 1990) whose
implications for return predictability are often thought of as applying more
forcefully to small stocks.
6
5
Also for these smallest firms, the coefficient on SIGMA changes signs, and the coefficients on
the past-return terms are smaller and much less significant. Given the potential distortions
associated with infrequent trading and price discreteness for this group, we are reluctant to hazard
an economic interpretation for these anomalies.
6
Several recent papers find that predictability – based on either ‘‘momentum’’ or ‘‘value’’
strategies – is stronger in small-cap stocks (see, e.g., Fama, 1998; Hong et al., 2000; and Griffin and
Lemmon, 1999).
J. Chen et al. / Journal of Financial Economics 61 (2001) 345–381 366
4.4. Stability over subperiods
In Table 5, we examine the intertemporal stability of our baseline regression,
using a Fama-MacBeth (1973) approach. Specifically, we run a separate, purely
cross-sectional variant of the regression in column 1 of Table 2 for every one of
Table 4
Forecasting skewness in the cross-section: cuts by firm size
The sample period is from July 1962 to December 1998. The dependent variable in all columns is
NCSKEW
tþ1
; the negative coefficient of (daily) skewness in the six-month period t þ1: In all
columns, returns are market-adjusted. SIGMA
t
is the standard deviation of (daily) returns in the
six-month period t: LOGSIZE
t
is the log of market capitalization at the end of period t:
DTURNOVER
t
is average monthly turnover in the six-month period t; detrended by a moving
average of turnover in the prior 18 months. RET
t
yRET
tÀ5
are returns in the six-month periods t
through t À5: All regressions also contain dummies for each time period (not shown); t-statistics
are adjusted for heteroskedasticity and serial correlation. Firm size cuts are based on NYSE
breakpoints.
Quintile 5
(largest)
firms
Quintile 4
firms
Quintile 3
firms
Quintile 2
firms
Quintile 1
(smallest)
firms
NCSKEW
t
0.053 0.059 0.054 0.043 0.045
(3.758) (3.653) (4.341) (3.690) (5.431)
SIGMA
t
À3.043 À4.362 À4.409 À4.062 2.894
(À1.243) (À2.263) (À3.771) (À4.612) (8.793)
LOGSIZE
t
0.009 0.057 0.049 0.105 0.066
(1.021) (1.855) (1.590) (3.639) (8.800)
DTURNOVER
t
0.404 0.637 0.551 0.264 0.079
(1.812) (2.450) (2.554) (1.391) (1.072)
RET
t
0.260 0.335 0.215 0.155 0.010
(5.637) (7.000) (5.359) (4.682) (0.569)
RET
tÀ1
0.047 0.001 0.083 0.134 0.017
(1.009) (0.024) (2.157) (4.269) (1.076)
RET
tÀ2
0.163 0.165 0.104 0.069 0.014
(3.554) (3.726) (2.651) (2.298) (0.816)
RET
tÀ3
0.025 0.078 0.093 0.033 0.028
(0.535) (1.682) (2.334) (1.112) (1.823)
RET
tÀ4
0.162 0.101 0.071 0.006 0.014
(3.637) (2.540) (1.852) (0.215) (0.864)
RET
tÀ5
0.128 0.089 0.134 0.013 À0.010
(2.906) (1.801) (3.503) (0.465) (À0.632)
No. of obs. 12,749 12,520 12,407 13,750 29,165
R
2
0.035 0.030 0.024 0.029 0.028
J. Chen et al. / Journal of Financial Economics 61 (2001) 345–381 367
Table 5
Forecasting skewness in the cross-section: Fama-MacBeth approach
The sample period is from July 1962 to December 1998 and includes only those firms with market
capitalization above the 20th percentile breakpoint of NYSE. The dependent variable is
NCSKEW
tþ1
; the negative coefficient of (daily) skewness in the six-month period t þ1: In all
cases, returns are market-adjusted. The specification is the same as in col. 1 of Table 2. SIGMA
t
is
the standard deviation of (daily) returns in the six-month period t: LOGSIZE
t
is the log of market
capitalization at the end of period t: DTURNOVER
t
is average monthly turnover in the six-month
period t; detrended by a moving average of turnover in the prior 18 months. RET
t
yRET
tÀ5
are
returns in the six-month periods t through t À5: Panel A reports only the coefficient on
DTURNOVER
t
for each period. Panel B reports the mean coefficients for different subperiods, and
the associated t-statistics, based on the time-series standard deviations of the coefficients, and
adjusted for serial correlation.
1960s 1970s 1980s 1990s
Panel A: Period-by-period regressions (12/1965 to 6/1998); coefficient on detrended turnover only
12/1965 0.383 6/1970 0.129 6/1980 1.730 6/1990 1.780
6/1966 1.053 12/1970 0.973 12/1980 0.707 12/1990 À0.194
12/1966 0.248 6/1971 1.145 6/1981 À0.156 6/1991 1.065
6/1967 À0.081 12/1971 0.269 12/1981 À0.757 12/1991 0.058
12/1967 0.201 6/1972 0.955 6/1982 2.738 6/1992 0.835
6/1968 0.468 12/1972 À0.207 12/1982 0.373 12/1992 0.569
12/1968 1.218 6/1973 0.148 6/1983 2.314 6/1993 0.161
6/1969 1.101 12/1973 À0.904 12/1983 0.334 12/1993 0.803
12/1969 0.498 6/1974 2.257 6/1984 À0.751 6/1994 0.459
12/1974 0.579 12/1984 0.545 12/1994 0.372
6/1975 À0.363 6/1985 2.448 6/1995 1.026
12/1975 À0.083 12/1985 À0.182 12/1995 À0.913
6/1976 0.029 6/1986 À0.686 6/1996 À0.631
12/1976 À0.016 12/1986 0.388 12/1996 1.981
6/1977 0.876 6/1987 0.672 6/1997 0.643
12/1977 1.901 12/1987 0.464 12/1997 0.062
6/1978 0.918 6/1988 0.404 6/1998 0.381
12/1978 1.512 12/1988 À0.941
6/1979 1.506 6/1989 0.121
12/1979 0.210 12/1989 À0.038
All
periods
Late
1960s 1970s 1980s 1990s
Panel B: Average coefficients by subperiods
NCSKEW
t
0.063 0.099 0.079 0.064 0.024
(4.880) (2.173) (4.517) (2.707) (1.258)
SIGMA
t
À5.017 À11.577 À9.507 À3.884 2.407
(À2.312) (À2.614) (À3.063) (À1.061) (0.288)
LOGSIZE
t
0.030 0.005 0.040 0.027 0.032
(4.141) (0.222) (2.216) (2.776) (4.200)
J. Chen et al. / Journal of Financial Economics 61 (2001) 345–381 368
the 66 six-month periods in our sample. We then take simple time-averages of
the cross-sectional regression coefficients over various subperiods, and
compute the associated t-statistics based on the time-series properties of the
coefficients (and adjusting for serial correlation). In Panel A of Table 5, we
display the coefficient on DTURNOVER from every one of the 66 regressions.
In Panel B, we show time-averages of all the regression coefficients for the full
sample and for each of four decade-based subperiods: the 1960s, the 1970s, the
1980s, and the 1990s.
The overriding conclusion that emerges from Table 5 is that our results are
remarkably stable over time. For example, the coefficient on DTURNOVER –
which averages 0.532 over the full sample period – reaches a low of 0.486 in the
1980s and a high of 0.592 in the 1970s. Moreover, even taken alone, three of
the four decade-based subperiods produce a statistically significant result for
DTURNOVER.
4.5. Why are small stocks more positively skewed?
One of the most striking patterns that we have documented is that
small stocks are more positively skewed than large stocks. Given that
Table 5 (continued)
All
periods
Late
1960s 1970s 1980s 1990s
DTURNOVER
t
0.532 0.565 0.592 0.486 0.497
(3.981) (2.280) (2.549) (1.372) (2.326)
RET
t
0.249 0.335 0.234 0.229 0.242
(6.614) (1.807) (3.909) (3.663) (2.312)
RET
tÀ1
0.099 0.229 0.026 0.085 0.132
(3.287) (1.684) (0.427) (1.838) (2.711)
RET
tÀ2
0.139 0.100 0.222 0.132 0.071
(4.357) (1.098) (3.452) (2.323) (1.387)
RET
tÀ3
0.082 0.057 0.139 0.017 0.104
(2.555) (0.645) (2.596) (0.341) (1.513)
RET
tÀ4
0.081 0.045 0.091 0.044 0.133
(2.887) (0.390) (1.524) (0.917) (2.453)
RET
tÀ5
0.082 0.139 0.056 0.036 0.136
(1.967) (1.767) (1.014) (0.193) (1.492)
No. of obs. 66 9 20 20 17
J. Chen et al. / Journal of Financial Economics 61 (2001) 345–381 369
this pattern is not clearly predicted by any existing theories (of which we
are aware) we have had to come up with a new hypothesis after the fact in
order to rationalize it. As described above, this hypothesis begins with the
assumption that managers have some discretion over the disclosure of
information, and prefer to announce good news immediately, while allowing
bad news to dribble out slowly. This behavior tends to impart a degree of
positive skewness to returns. Moreover, if there is more scope for such
managerial discretion in small firms – perhaps because they face less scrutiny
from security analysts – then the positive-skewness effect will be more
pronounced in small stocks.
The one satisfying thing about this after-the-fact hypothesis is
that it yields new testable predictions. Specifically, positive skewness
ought to be greater in firms with fewer analysts, after controlling for
size. Table 6 investigates this prediction, taking our baseline specifica-
tions for both NCSKEW and DUVOL, and in each case adding LOGCOVER,
the log of one plus the number of analysts covering the stock. (The
sample period in Table 6 is shorter, since analyst coverage is not
available from I/B/E/S prior to December 1976.) The coefficients on
LOGCOVER have the predicted positive sign, and are strongly statistically
significant.
7
The coefficients on LOGSIZE go down a bit, but remain
significant as well. Nothing else changes.
We do not mean to cast Table 6 as a definitive test of the discretio-
nary-disclosure hypothesis; this idea is outside the main scope of the paper,
and pursuing it more seriously would take us too far afield. Nevertheless,
it is comforting to know that the most obvious auxiliary prediction of the
hypothesis is borne out in the data, and that as a result, we at least have a
plausible explanation for what would otherwise be a puzzling feature of
our data.
5. Forecasting market skewness
We now turn to forecasting skewness in the returns to the aggregate market.
While this is in many ways the more interesting exercise from an economic
viewpoint, our statistical power is severely limited. Thus it may be asking too
much to expect that the results here will be strongly statistically significant in
their own right; rather, one might more reasonably hope that they look
qualitatively similar to those from the cross-sectional regressions.
7
After developing the discretionary-disclosure hypothesis, and running the regressions in
Table 6, we became aware of a closely related working paper by Damodaran (1987). Using data
from 1979 to 1983, he also finds that firms with fewer analysts have more positively skewed returns.
J. Chen et al. / Journal of Financial Economics 61 (2001) 345–381 370
Our definition of the aggregate market is the value-weighted NYSE-AMEX
index, and all returns are excess returns relative to T-bills. To avoid any
temptation to further mine the data, we use essentially the same specification as
Table 6
Forecasting skewness in the cross-section: adding analyst coverage
The sample period is from December 1976 to December 1998 and includes only those firms with
market capitalization above the 20th percentile breakpoint of the NYSE. The dependent variable in
col. 1 is NCSKEW
tþ1
; the negative coefficient of skewness in the six-month period t þ1; and in col.
2 it is DUVOL
tþ1
; the log of the ratio of down-day to up-day standard deviation in the six-month
period t þ1: SIGMA
t
is the standard deviation of (daily) returns in the six-month period t.
LOGSIZE
t
is the log of market capitalization at the end of period t. LOGCOVER
t
is the log of one
plus the number of analysts covering the stock at the end of period t. DTURNOVER
t
is average
monthly turnover in the six-month period t; detrended by a moving average of turnover in the prior
18 months. RET
t
yRET
tÀ5
are returns in the six-month periods t through t À5: All regressions
also contain dummies for each time period (not shown); t-statistics are adjusted for
heteroskedasticity and serial correlation.
(1) Using
NCSKEW
tþ1
measure
(2) Using
DUVOL
tþ1
measure
NCSKEW
t
0.049 0.090
(DUVOL
t
in col.2) (6.649) (13.945)
SIGMA
t
À3.188 À4.022
(À4.366) (À11.586)
LOGSIZE
t
0.032 0.011
(7.992) (6.665)
DTURNOVER
t
0.504 0.207
(3.504) (3.648)
LOGCOVER
t
0.019 0.006
(4.059) (3.288)
RET
t
0.219 0.135
(8.767) (12.599)
RET
tÀ1
0.085 0.010
(3.681) (1.044)
RET
tÀ2
0.100 0.040
(4.410) (4.253)
RET
tÀ3
0.058 0.006
(2.564) (0.683)
RET
tÀ4
0.065 0.012
(3.140) (1.290)
RET
tÀ5
0.055 0.006
(2.493) (0.705)
No. of obs. 40,688 40,688
R
2
0.025 0.051
J. Chen et al. / Journal of Financial Economics 61 (2001) 345–381 371
in our baseline cross-sectional analysis.
8
Specifically, we use all the same right-
hand-side variables, except for LOGSIZE and the time dummies. The
DTURNOVER variable is constructed exactly as before, by detrending
TURNOVER with its own moving average over the prior 18 months.
In an effort to get the most out of the little time-series data that we have, we
now use monthly overlapping observations. (The t-statistics we report are
adjusted accordingly.) This yields a total of 401 observations that can be used
in the regressions. However, a new concern that arises with the time-series
approach is the extent to which our inferences are dominated by the enormous
daily movements during October 1987. To address this concern, we also re-run
our regressions omitting October 1987. This brings us down to 371
observations.
9
The results are summarized in Table 7. In columns 1 and 2 we use the
NCSKEW measure of skewness, and run the regressions with and without
October 1987, respectively. In columns 3 and 4 we use the DUVOL measure of
skewness, and again run the regression with and without October 1987. The
basic story is the same in all four columns. The six past-return terms are always
positive, and many are individually statistically significant. In contrast, the
coefficient on DTURNOVER, while positive in each of the four regressions, is
never statistically significant. Dropping October 1987 seems to increase the
precision of the DTURNOVER coefficient estimate somewhat, but the highest
t-statistic across the four specifications is only 1.15.
Nevertheless, holding statistical significance aside, the point estimates
suggest large quantitative effects relative to the cross-sectional regres-
sions. Indeed, the coefficients on DTURNOVER and the RET terms are
now on the order of ten times bigger than they were in the previous tables.
Thus both turnover and past returns could well be very important for
forecasting the skewness of market returns, but we lack the statistical power to
assert these conclusions – particularly regarding turnover – with much
confidence.
In light of this power problem, we obtained an alternative series on NYSE
volume going back to 1928 from Gallant et al. (1992), who in turn take it from
the S&P Security Price Record. The one drawback with these data is that we
cannot use them to literally calculate turnover, since they give only the number
of shares traded and not the number of shares outstanding. Thus we cannot
8
Harvey and Siddique (1999) build an autoregressive conditional skewness model for aggregate-
market returns; while their specification is very different from that here, it shares the common
element that lagged skewness helps to forecast future skewness.
9
We lose 30 observations because we do not allow any observation on NCSKEW, DUVOL,
SIGMA, or DTURNOVER to enter the regression if it draws on data from October 1987. Because
of the detrending, the DTURNOVER variable in any given month draws on 24 months’ worth of
data.
J. Chen et al. / Journal of Financial Economics 61 (2001) 345–381 372
quite reproduce our baseline specifications for the longer post-1928 sample
period. Nevertheless, using detrended values of raw trading volume to
approximate detrended turnover, we get results for this sample period that
are very similar to those reported in Table 7.
Table 7
Forecasting skewness in the market: time-series regressions
The sample period is from July 1962 to December 1998 and is based on market returns in excess of
the risk-free rate, where the market is defined as the value-weighted portfolio of all NYSE/AMEX
stocks. The dependent variable in cols. 1 and 2 is NCSKEW
tþ1
; the negative coefficient of skewness
in the six-month period t þ1; and in cols. 3 and 4 it is DUVOL
tþ1
; the log of the ratio of down-day
to up-day standard deviation in the six-month period t þ1: SIGMA
t
is the standard deviation of
(daily) market returns in the six-month period t: DTURNOVER
t
is the average monthly turnover
of the market portfolio in the six-month period t; detrended by a moving average of turnover in the
prior 18 months. RET
t
yRET
tÀ5
are returns in the six-month periods t through t À5; t-statistics,
which are in parentheses, are adjusted for heteroskedasticity and serial correlation.
(1) Dep. variable
is NCSKEW
tþ1
(2) Dep. variable
is NCSKEW
tþ1
;
excluding 10/87
(3) Dep. variable
is DUVOL
tþ1
(4) Dep. variable
is DUVOL
tþ1
;
excluding 10/87
NCSKEW
t
0.100 0.123 0.221 0.217
(DUVOL
t
in
col.3 and 4)
(0.855) (1.232) (1.842) (0.844)
SIGMA
t
18.183 13.708 1.196 À3.574
(1.137) (0.749) (0.156) (À0.300)
DTURNOVER
t
6.002 9.349 6.324 9.462
(0.262) (0.828) (0.704) (1.148)
RET
t
2.647 1.809 1.484 1.184
(4.147) (4.406) (4.168) (3.398)
RET
tÀ1
1.585 1.077 0.482 0.332
(3.086) (2.939) (1.481) (1.061)
RET
tÀ2
1.473 0.926 0.554 0.386
(2.242) (1.922) (1.898) (1.357)
RET
tÀ3
0.589 0.443 0.126 0.017
(0.602) (0.734) (0.325) (0.049)
RET
tÀ4
1.283 0.680 0.475 0.287
(2.264) (1.575) (1.726) (0.968)
RET
tÀ5
1.187 0.596 0.686 0.470
(2.288) (1.930) (2.326) (1.753)
No. of obs. 401 371 401 371
R
2
0.265 0.264 0.304 0.274
J. Chen et al. / Journal of Financial Economics 61 (2001) 345–381 373
6. Economic significance of the results: an option-pricing metric
Thus far, we have focused on the statistical significance of our results, and
have not really asked whether they imply magnitudes that are economically
meaningful. Assessing economic significance in the current context is a bit
tricky. The thought experiment that is typically undertaken is something like
this: suppose that the right-hand-side variable of interest – in this case,
DTURNOVER – is shocked by two standard deviations. How much does the
left-hand-side variable – NCSKEW or DUVOL – move? What makes things
difficult here is that most people have little sense for what would constitute an
economically interesting change in NCSKEW or DUVOL.
To help frame things in a way that is hopefully more intuitive, we can
translate statements about NCSKEW into statements about the prices of out-
of-the-money put options. The idea behind our metric can be understood as
follows. Imagine that you are pricing an out-of-the-money put on a stock
whose returns you initially believe to be symmetrically distributed – i.e., a stock
for which you believe that NCSKEW is equal to zero. Now the stock
experiences a surge in turnover. As a result, you revise your forecast of
NCSKEW, using the DTURNOVER coefficient estimate from our regressions.
Given this new forecast of NCSKEW – but holding volatility fixed – by how
much does the value of the put option increase?
To answer this sort of question precisely, we need to (1) find an option-
pricing model that admits skewness in returns and (2) create a mapping from
the parameters of this model to our NCSKEW variable. The model we use is
the stochastic-volatility model of Das and Sundaram (1999), in which the
dynamics of stock prices are summarized by the following two diffusion
equations:
dp
t
¼ a dt þV
1=2
t
dz
1
; ð3Þ
dV
t
¼ kðV
0
ÀV
t
Þdt þZV
1=2
t
dz
2
; ð4Þ
where p
t
is the log of the stock price, a is the expected return on the stock, V
t
is
the current variance, k is the mean reversion parameter for the variance
process, V
0
is the long-run mean level of variance, and Z is the volatility of the
variance process. The two Wiener processes dz
1
and dz
2
are instantaneously
correlated, with a correlation coefficient of r: The parameter r is the one of
central interest for our purposes, as it governs the skewness of stock returns:
when r=0, log returns are symmetrically distributed; when ro0, log returns
are negatively skewed.
In order to map the parameters of the option-pricing model into our
NCSKEW variable, we draw on formulas given by Das and Sundaram that
express the skewness in daily log returns as a function of the diffusion
parameters. If we are willing to fix all the other parameters besides r; these
J. Chen et al. / Journal of Financial Economics 61 (2001) 345–381 374
formulas allow us to ask to what value of r a given value of NCSKEW
corresponds. Once we have obtained the implied value of r in this way, we can
calculate options prices and thereby see the impact of a given value of
NCSKEW.
Table 8 illustrates the results of this exercise. Consider first Panel A,
where the parameters are chosen so as to be reasonable for individual
stocks: k=1, V
0
=0.16, V
t
=0.16, and Z ¼ 0:4: (Setting the variance V to
0.16 corresponds to an annual standard deviation of returns of 40%.) We also
set the stock price P=100, and the riskless rate r=0. We begin with a
hypothetical firm 1, which has symmetrically distributed returns – i.e., it has
NCSKEW=0. This is equivalent to a value of r=0. Next, we take firm 2,
which is identical to firm 1, except that it has a two-standard-deviation higher
value of DTURNOVER. The standard deviation of DTURNOVER (for firms
above the 20th percentile NYSE breakpoint) is 0.042, and from Table 2,
column 1, the coefficient on DTURNOVER is 0.437. Hence the value of
NCSKEW for firm 2 is 0.037 (2 Â0.042 Â0.437=0.037). Using Eq. (21) in Das
and Sundaram (1999, p. 223) this value of skewness in daily returns for firm 2
can be shown to imply r=À0.38, assuming all the other diffusion parameters
stay fixed.
Panel A of Table 8 displays the impact of this change in r for the prices of
six-month European put options. That is, it calculates put prices for both firm
1 (which has NCKSEW=0 and thus r=0) and firm 2 (which has
NCSKEW=0.037 and thus r=À0.38). As can be seen, the impact on put
prices is substantial, particularly if one goes relatively far out-of-the-money.
For example, a put with a strike of 70 is worth 1.20 for firm 1 but 1.44 for firm
2, an increase of 20.14%. Or expressed in a different way, the firm 1 put has a
Black-Scholes (1973) implied volatility of 40.33%, while the firm 2 put has an
implied volatility of 42.50%.
Panel B undertakes a similar experiment to gauge the significance of our
time-series results. We keep all the diffusion parameters the same as in Panel A,
except that we now set V
0
¼ V
t
¼ 0:04; corresponding to an annual standard
deviation of returns of 20%. For the market as a whole, the standard deviation
of DTURNOVER is 0.005 (see Table 1A). Using the coefficient estimate on
DTURNOVER of 6.00 from Table 7, column 1, a two-standard-deviation
shock to DTURNOVER translates into a movement of 0.06 in the NCSKEW
variable. Given the other diffusion parameters, this value of 0.06 for NCSKEW
is equivalent to r=À0.33.
Panel B then compares the prices of six-month European puts across two
regimes, the first with r=0 and the second with r=À0.33. Once again, the
differences appear to be meaningful. For example, a put with a strike price of
85 is worth 0.86 in regime 1 but 1.07 in regime 2, an increase of 24.66%. The
corresponding implied volatilities are 20.36% and 21.84%, respectively. These
results reinforce a point made above: while the time-series estimates are
J. Chen et al. / Journal of Financial Economics 61 (2001) 345–381 375
Table 8
Economic significance of trading volume for skewness in stock returns: an option-pricing metric
Using the stochastic volatility option pricing model (and notation) of Das and Sundaram (1999) we consider what a two-standard-deviation shock in
detrended trading volume implies for the prices of six-month European options.
Panel A: Options on individual stocks
The benchmark parameters are as follows: stock price P=100, interest rate r=0, annualized long-run variance V
0
=0.16, current variance V=0.16,
mean reversion in variance k=1, and volatility of variance Z=0.4. Firm 1 is assumed to have a value of r=0. Firm 2 is assumed to have a value of
r=À0.38. These values of r imply that the difference in daily skewness between Firms 1 and 2 is equivalent to that created by a two-standard-devi-
ation move in the DTURNOVER variable, using our baseline firm-level sample and coefficient estimates from Table 2, col.1.
70 80 90 100 110 120 130
Firm 1: r=0
Six-month European put price 1.197 3.044 6.287 11.082 17.325 24.748 33.044
Black-Scholes implied vol. 40.33% 39.79% 39.50% 39.41% 39.48% 39.67% 39.93%
Firm 2: r=À0.38
Six-month European put price 1.438 3.297 6.419 10.994 17.011 24.282 32.525
Black-Scholes implied vol. 42.50% 41.16% 40.03% 39.10% 38.35% 37.77% 37.34%
Percent increase in put
price: Firm 2 vs. Firm1
20.14% 8.30% 2.09% À0.80% À1.81% À1.88% À1.57%
J
.
C
h
e
n
e
t
a
l
.
/
J
o
u
r
n
a
l
o
f
F
i
n
a
n
c
i
a
l
E
c
o
n
o
m
i
c
s
6
1
(
2
0
0
1
)
3
4
5

3
8
1
3
7
6
Panel B: Options on the market portfolio
The benchmark parameters are as follows: stock price P=100, interest rate r=0, annualized long run variance V
0
=0.04, current variance V=0.04,
mean reversion in variance k ¼ 1; and volatility of variance Z ¼ 0:4: Regime 1 is assumed to have a value of r ¼ 0: Regime 2 is assumed to have a value
of r=À0.33. These values of r imply that the difference in daily skewness between Regimes 1 and 2 is equivalent to that created by a two-standard-
deviation move in the market DTURNOVER variable, using our time-series estimates from Table 7, col. 1.
85 90 95 100 105 110 115
Regime 1: r=0
Six-month European put price 0.859 1.693 3.121 5.330 8.367 12.093 16.298
Black-Scholes implied vol. 20.36% 19.61% 19.09% 18.91% 19.07% 19.49% 20.04%
Regime 2: r=À0.33
Six-month European put price 1.070 1.912 3.258 5.289 8.134 11.755 15.955
Black-Scholes implied vol. 21.84% 20.68% 19.63% 18.76% 18.21% 18.01% 18.10%
Percent increase in put price:
Regime 2 vs. Regime 1
24.66% 12.91% 4.39% À0.77% À2.79% À2.80% À2.10%
J
.
C
h
e
n
e
t
a
l
.
/
J
o
u
r
n
a
l
o
f
F
i
n
a
n
c
i
a
l
E
c
o
n
o
m
i
c
s
6
1
(
2
0
0
1
)
3
4
5

3
8
1
3
7
7
statistically much weaker than those from the cross-section, they are no less
indicative of important economic effects.
Although they are not shown in Table 8, we have also done similar
calculations to measure the economic significance of our results for past returns.
As it turns out, the quantitative influence of past returns on skewness is stronger
than that of trading volume. With individual stocks, a shock of 40% to the RET
variable in the most recent six-month period (note from Table 1 that 40% is
approximately a two-standard-deviation shock for a firm in the next-to-largest
quintile) translates into a movement of 0.087 in the NCSKEW variable. This in
turn is equivalent to r going from zero to À0.920, which causes the put with the
70 strike to rise in value from 1.20 to 1.73, a 44.86% increase.
In the case of the aggregate market, the coefficients on past returns suggest
effects on skewness that are so large that they cannot generally be captured in
the context of a pure diffusion model like that of Das and Sundaram (1999).
For example, even if the RET variable has moved by only 7% in the last six
months, one has to adjust r from zero to À0.978 to reflect the corresponding
predicted change in NCSKEW. Given that the standard deviation of RET for
the market is about 11%, this means that we cannot even capture a one-
standard-deviation shock to six-month returns without violating the constraint
that the absolute value of r not exceed one. Rather, we are left to conclude
that, for the market as a whole, large movements in past returns give rise to
conditional negative skewness so substantial that it cannot be adequately
represented in terms of a pure diffusion process – one would instead need some
type of mixed jump-diffusion model.
7. Conclusions
Three robust findings about conditional skewness emerge from our analysis
of individual stocks. In the cross-section, negative skewness is greater in stocks
that (1) have experienced an increase in trading volume relative to trend over
the prior six months, (2) have had positive returns over the prior 36 months,
and (3) are larger in terms of market capitalization. The first two results also
have direct analogs in the time-series behavior of the aggregate market, though
the statistical power of our tests in this case (especially with respect to trading
volume) is quite limited.
Let us try to put each of these findings into some perspective. The first,
regarding trading volume, is the most novel, and is the one we were looking for
based on a specific theoretical prediction from the model of Hong and Stein
(1999). Clearly, our results here are supportive of the theory. At the same time,
this does not mean that there are not other plausible interpretations. While we
have attempted to control for some of the most obvious alternative stories, no
doubt others can be thought up. This caveat would seem to be particularly
J. Chen et al. / Journal of Financial Economics 61 (2001) 345–381 378
relevant given that there has been so little research to date on conditional
skewness at the individual stock level.
The second and third findings, having to do with the effects of past returns
and size on skewness, do not speak directly to predictions made by the Hong-
Stein model. Rather, these variables are included in our regressions because
prior work (Harvey and Siddique, 2000) suggests that they might enter
significantly, and we want to be careful to isolate the effects of trading volume
from other factors.
Having verified the importance of past returns, we have found it helpful to
think about it in terms of models of stochastic bubbles, such as that developed by
Blanchard and Watson (1982). However, we would stop well short of claiming to
have strong evidence in favor of the existence of bubbles. Indeed, there is a large
body of research from the 1980s (see, e.g., reviews by West, 1988; Flood and
Hodrick, 1990) that focuses on a very different set of implications of bubble
models – having to do with the relation between prices and measures of
fundamentals such as dividends – and tends to reach mostly skeptical conclusions
on this question. Rather, the more modest statement to be made is that previous
research has not examined the implications of bubble models for conditional
skewness, and that on this one score, the bubble models look pretty good.
With respect to the third finding – that small-cap stocks are more positively
skewed than large-cap stocks – we are not even aware of an existing theory that
provides a simple explanation. Instead, we have developed an informal
hypothesis after the fact, based on the ideas that (1) managers prefer to disclose
good news right away, while dribbling bad news out slowly, and (2) managers
of small companies have more scope for hiding bad news from the market in
this way. This discretionary-disclosure hypothesis in turn yields the further
prediction that, controlling for size, positive skewness ought to be more
pronounced in stocks with fewer analysts – a prediction which is clearly
supported in the data.
A fair criticism of this whole line of discussion is that we have three main
empirical results, and a different theoretical story for each: the Hong-Stein
(1999) model to explain the effect of turnover on skewness, stochastic bubbles to
explain the effect of past returns, and discretionary disclosure to explain the
effect of size. This lack of unity is unsatisfying, and it serves to further
underscore our previous caveat about the extent to which one should at this
point consider any single theory to be strongly supported by the data. A natural
challenge for future work in this area is to come up with a parsimonious model
that captures these three patterns in a more integrated fashion.
References
Almazan, A., Brown, B.C., Carlson, M., Chapman, D.A., 1999. Why constrain your mutual fund
manager?. Working paper, University of Texas, Austin.
J. Chen et al. / Journal of Financial Economics 61 (2001) 345–381 379
Bakshi, G., Cao, C., Chen, Z., 1997. Empirical performance of alternative option pricing models.
Journal of Finance 52, 2003–2049.
Bates, D., 1991. The crash of ’87: was it expected? The evidence from options markets. Journal of
Finance 46, 1009–1044.
Bates, D., 1997. Post-’87 crash fears in S&P 500 futures options. NBER working paper
5894.
Bekaert, G., Wu, G., 2000. Asymmetric volatility and risk in equity markets. Review of Financial
Studies 13, 1–42.
Black, F., 1976. Studies of stock price volatility changes. Proceedings of the 1976 Meetings of the
American Statistical Association, Business and Economical Statistics Section, 177–181.
Black, F., Scholes, M., 1973. The pricing of options and corporate liabilities. Journal of Political
Economy 81, 637–659.
Blanchard, O.J., Watson, M.W., 1982. Bubbles, rational expectations, and financial markets. In:
Wachtel, P. (Ed.), Crises in Economic and Financial Structure. Lexington Books, Lexington,
MA, pp. 295–315.
Braun, P.A., Nelson, D.B., Sunier, A.M., 1995. Good news, bad news, volatility, and betas. Journal
of Finance 50, 1575–1603.
Campbell, J.Y., Hentschel, L., 1992. No news is good news: an asymmetric model of changing
volatility in stock returns. Journal of Financial Economics 31, 281–318.
Christie, A.A., 1982. The stochastic behavior of common stock variances – value, leverage and
interest rate effects. Journal of Financial Economics 10, 407–432.
Coval, J.D., Hirshleifer, D., 1998. Trading-generated news, sidelined investors, and conditional
patterns in security returns. Working paper, University of Michigan Business School, Ann
Arbor.
Damodaran, A., 1987. Information bias: measures and implications. Working paper, New York
University.
Das, S.R., Sundaram, R.K., 1999. Of smiles and smirks: a term structure perspective. Journal of
Financial and Quantitative Analysis 34, 211–239.
DeLong, J.B., Shleifer, A., Summers, L.H., Waldmann, R.J., 1990. Noise trader risk in financial
markets. Journal of Political Economy 98, 703–738.
Duffee, G.R., 1995. Stock returns and volatility: a firm-level analysis. Journal of Financial
Economics 37, 399–420.
Dumas, B., Fleming, J., Whaley, R.E., 1998. Implied volatility functions: empirical tests. Journal of
Finance 53, 2059–2106.
Engle, R.F., Ng, V.K., 1993. Measuring and testing the impact of news on volatility. Journal of
Finance 48, 1749–1778.
Fama, E.F., 1998. Market efficiency, long-term returns, and behavioral finance. Journal of
Financial Economics 49, 283–306.
Fama, E.F., MacBeth, J.D., 1973. Risk, return and equilibrium: empirical tests. Journal of Political
Economy 81, 607–636.
Flood, R.P., Hodrick, R.J., 1990. On testing for speculative bubbles. Journal of Economic
Perspectives 4, 85–101.
French, K.R., Schwert, G.W., Stambaugh, R.F., 1987. Expected stock returns and volatility.
Journal of Financial Economics 19, 3–29.
Gallant, A.R., Rossi, P.E., Tauchen, G., 1992. Stock prices and volume. Review of Financial
Studies 5, 199–242.
Genotte, G., Leland, H., 1990. Market liquidity, hedging and crashes. American Economic Review
80, 999–1021.
Glosten, L., Jagannathan, R., Runkle, D.E., 1993. On the relation between the expected
value and the volatility of the nominal excess return on stocks. Journal of Finance 48,
1779–1801.
J. Chen et al. / Journal of Financial Economics 61 (2001) 345–381 380
Greene, W.H., 1993. Econometric Analysis. Macmillan, New York.
Griffin, J.M., Lemmon, M.L., 1999. Does book-to-market equity proxy for distress risk or over-
reaction? Working paper, Arizona State University, Tempe, AZ.
Grossman, S.J., 1988. An analysis of the implications for stock and futures price volatility of
program trading and dynamic hedging strategies. Journal of Business 61, 275–298.
Harris, M., Raviv, A., 1993. Differences of opinion make a horse race. Review of Financial Studies,
473–506.
Harvey, C.R., Siddique, A., 1999. Autoregressive conditional skewness. Journal of Financial and
Quantitative Analysis 34, 465–487.
Harvey, C.R., Siddique, A., 2000. Conditional skewness in asset pricing tests. Journal of Finance
55, 1263–1295.
Hong, H., Lim, T., Stein, J.C., 2000. Bad news travels slowly: size, analyst coverage and the
profitability of momentum strategies. Journal of Finance 55, 65–295.
Hong, H., Stein, J. C., 1999. Differences of opinion, rational arbitrage and market crashes. NBER
Working paper.
Jacklin, C.J., Kleidon, A.W., Pfleiderer, Paul., 1992. Underestimation of portfolio insurance and
the crash of October 1987. Review of Financial Studies 5, 35–63.
Kandel, E., Pearson, N.D., 1995. Differential interpretation of public signals and trade in
speculative markets. Journal of Political Economy 103, 831–872.
Koski, J.L., Pontiff, J., 1999. How are derivatives used? Evidence from the mutual fund industry.
Journal of Finance 54, 791–816.
Lakonishok, J., Smidt, S., 1986. Volume for winners and losers: taxation and other motives for
stock trading. Journal of Finance 41, 951–974.
Nelson, D., 1991. Conditional heteroskedasticity in asset returns: a new approach. Econometrica
59, 347–370.
Odean, T., 1998a. Volume, volatility, price and profit when all traders are above average. Journal
of Finance 53, 1887–1934.
Odean, T., 1998b. Are investors reluctant to realize their losses? Journal of Finance 53, 1775–1798.
Pindyck, R.S., 1984. Risk, inflation, and the stock market. American Economic Review 74,
334–351.
Poterba, J.M., Summers, L.H., 1986. The persistence of volatility and stock market fluctuations.
American Economic Review 76, 1142–1151.
Romer, D., 1993. Rational asset-price movements without news. American Economic Review 83,
1112–1130.
Schwert, G.W., 1989. Why does stock market volatility change over time? Journal of Finance 44,
1115–1153.
Shefrin, H., Statman, M., 1985. The disposition to sell winners too early and ride losers too long:
theory and evidence. Journal of Finance 40, 777–790.
Varian, H.R., 1989. Differences of opinion in financial markets. In: Stone, C. (Ed.), Financial Risk:
Theory, Evidence and Implications: Proceedings of the 11th Annual Economic Policy
Conference of the Federal Reserve Bank of St. Louis. Kluwer Academic Publishers, Boston,
pp. 3–37.
West, K.D., 1988. Bubbles, fads and stock price volatility tests: a partial evaluation. Journal of
Finance 43, 639–656.
Wu, G., 2000. The determinants of asymmetric volatility. Review of Financial Studies, in
preparation.
J. Chen et al. / Journal of Financial Economics 61 (2001) 345–381 381

346

J. Chen et al. / Journal of Financial Economics 61 (2001) 345–381

1. Introduction Aggregate stock market returns are asymmetrically distributed. This asymmetry can be measured in several ways. First, and most simply, the very largest movements in the market are usually decreases, rather than increases – that is, the stock market is more prone to melt down than to melt up. For example, of the ten biggest one-day movements in the S&P 500 since 1947, nine were declines.1 Second, a large literature documents that market returns exhibit negative skewness, or a closely related property, ‘‘asymmetric volatility’’ – a tendency for volatility to go up with negative returns.2 Finally, since the crash of October 1987, the prices of stock index options have been strongly indicative of a negative asymmetry in returns, with the implied volatilities of out-of-themoney puts far exceeding those of out-of-the-money calls; this pattern has come to be known as the ‘‘smirk’’ in index-implied volatilities. (See, e.g., Bates, 1997; Bakshi et al., 1997; and Dumas et al., 1998.) While the existence of negative asymmetries in market returns is generally not disputed, it is less clear what underlying economic mechanism these asymmetries reflect. Perhaps the most venerable theory is based on leverage effects (Black, 1976; Christie, 1982), whereby a drop in prices raises operating and financial leverage, and hence the volatility of subsequent returns. However, it appears that leverage effects are not of sufficient quantitative importance to explain the data (Schwert, 1989; Bekaert and Wu, 2000). This is especially true if one is interested in asymmetries at a relatively high frequency, e.g., in daily data. To explain these, one has to argue that intraday changes in leverage have a large impact on volatility – that a drop in prices on Monday morning leads to a large increase in leverage and hence in volatility by Monday afternoon, so that overall, the return for the full day Monday is negatively skewed. An alternative theory is based on a ‘‘volatility feedback’’ mechanism. As developed by Pindyck (1984), French et al. (1987), Campbell and Hentschel (1992), and others, the idea is as follows: When a large piece of good news arrives, this signals that market volatility has increased, so the direct positive effect of the good news is partially offset by an increase in the risk premium. On the other hand, when a large piece of bad news arrives, the direct effect and the risk-premium effect now go in the same direction, so the impact of the news is amplified. While the volatility-feedback story is in some ways more attractive
1 Moreover, the one increase – of 9.10% on October 21, 1987 – was right on the heels of the 20.47% decline on October 19, and arguably represented a correction of the microstructural distortions that arose on that chaotic day, rather than an independent price change. 2 If, in a discrete-time setting, a negative return in period t raises volatility in period t þ 1 and thereafter, returns measured over multiple periods will be negatively skewed, even if single-period returns are not. The literature on these phenomena includes Pindyck (1984), French et al. (1987), Campbell and Hentschel (1992), Nelson (1991), Engle and Ng (1993), Glosten et al. (1993), Braun et al. (1995), Duffee (1995), Bekaert and Wu (2000), and Wu (2001).

J. Chen et al. / Journal of Financial Economics 61 (2001) 345–381

347

than the leverage-effects story, there are again questions as to whether it has the quantitative kick that is needed to explain the data. The thrust of the critique, first articulated by Poterba and Summers (1986), is that shocks to market volatility are for the most part very short-lived, and hence cannot be expected to have a large impact on risk premiums. A third explanation for asymmetries in stock market returns comes from stochastic bubble models of the sort pioneered by Blanchard and Watson (1982). The asymmetry here is due to the popping of the bubble – a low-probability event that produces large negative returns. What the leverage-effects, volatility-feedback, and bubble theories all have in common is that they can be cast in a representative-investor framework. In contrast, a more recent explanation of return asymmetries, Hong and Stein (1999), argues that investor heterogeneity is central to the phenomenon. The Hong-Stein model rests on two key assumptions: (1) there are differences of opinion among investors as to fundamental value, and (2) some – though not all – investors face short-sales constraints. The constrained investors can be thought of as mutual funds, whose charters typically prohibit them from taking short positions; the unconstrained investors can be thought of as hedge funds or other arbitrageurs. When differences of opinion are initially large, those bearish investors who are subject to the short-sales constraint will be forced to a corner solution, in which they sell all of their shares and just sit out of the market. As a consequence of being at a corner, their information is not fully incorporated into prices. For example, if the market-clearing price is $100, and a particular investor is sitting out, it must be that his valuation is less than $100, but one has no way of knowing by how much – it could be $95, but it could also be much lower, say $50. However, if after this information is hidden, other, previously more-bullish investors have a change of heart and bail out of the market, the originally more-bearish group may become the marginal ‘‘support buyers’’ and hence more will be learned about their signals. In particular, if the investor who was sitting out at a price of $100 jumps in and buys at $95, this is good news relative to continuing to sit on the sidelines even as the price drops further. Thus, accumulated hidden information tends to come out during market declines, which is another way of saying that returns are negatively skewed. With its focus on differences of opinion, the Hong-Stein model has distinctive empirical implications that are not shared by the representativeinvestor theories. In particular, the Hong-Stein model predicts that negative skewness in returns will be most pronounced around periods of heavy trading volume. This is because – like in many models with differences of opinion – trading volume proxies for the intensity of disagreement. (See Varian, 1989; Harris and Raviv, 1993; Kandel and Pearson, 1995; and Odean, 1998a for other models with this feature.)

glamour stocks – those with low ratios of book value to market value – are also forecasted to have more negative skewness. but they are perhaps most clearly suggested by models of stochastic bubbles. These control variables can be divided into two categories. the effect of turnover is strongly statistically and economically significant. Damodaran. we are not in the business of forecasting negative expected returns.. capture time-varying influences on skewness. For example. 1997). with their information incompletely revealed in prices. 2000) that skewness is more negative on average for large-cap firms – a pattern . it has been documented by others (e. In the first category are those that. We find that when past returns have been high. moreover. at the firm level. we undertake an empirical investigation that is motivated by this differences-of-opinion theory. inferred from options prices – as a measure of crash expectations. high past returns or a low book-to-market value imply that the bubble has been building up for a long time. but there is some ability to predict negative skewness based on returns as far back as 36 months. previously more-bullish investors can force the hidden information to come out. (Harvey and Siddique (2000) also examine how skewness varies with past returns and book-to-market. We develop a series of cross-sectional regression specifications that attempt to forecast skewness in the daily returns to individual stocks. like detrended turnover. / Journal of Financial Economics 61 (2001) 345–381 When disagreement (and hence trading volume) is high. In the context of a bubble model. July 1–December 31. we ask whether the skewness in daily returns measured over a given six-month period (say. 1998) can be predicted from the detrended level of turnover over the prior sixmonth period (January 1–June 30. it is more likely that bearish investors will wind up at a corner. Chen et al. Thus. when the arrival of bad news to other. 1987. The most significant variable in this category is past returns. It turns out that firms that experience larger increases in turnover relative to trend are indeed predicted to have more negative skewness. who also interprets conditional skewness – in his case. In an effort to isolate the effects of turnover. we are adopting a narrow and euphemistic definition of the word ‘‘crashes.g. This usage follows Bates (1991.’’ associating it solely with the conditional skewness of the return distribution. For example. so that there is a larger drop when it pops and prices fall back to fundamentals. And it is precisely this hiding of information that sets the stage for negative skewness in subsequent rounds of trade. skewness is forecasted to become more negative. our specifications also include a number of control variables.348 J. Harvey and Siddique. The second category of variables that help to explain skewness are those that appear to be picking up relatively fixed firm characteristics.) These results can be rationalized in a number of ways. 1998). One of our key forecasting variables is the recent deviation of turnover from its trend. In this paper. The predictive power is strongest for returns in the prior six months. when we speak of ‘‘forecasting crashes’’ in the title of the paper. In a similar vein.

this pure time-series approach entails an enormous loss in statistical power – with data going back to 1962. Given how early it is in this game.S. Here. stock market as a whole. While both the cross-sectional and time-series results for turnover are broadly consistent with the theory we are interested in. Finally. Chen et al. and for which there are no other widely accepted explanations. even if innovations to trading volume proxy for the intensity of disagreement among investors. In Section 3. they likely capture other factors as well – such as changes in trading costs – that we have not adequately controlled for. we attempt to forecast the skewness in the daily returns to the market using detrended market turnover and past market returns. we use an option-pricing metric to evaluate the economic significance of our results. First. In addition to running our cross-sectional regressions with the individualfirm data. we have less than 70 independent observations of market skewness measured at six-month intervals – which is why it is not the main focus of our analysis. along with a variety of sensitivities and sample splits.J. Obviously. We are not aware of any theories that would have naturally led one to anticipate this finding. Such a control might be redundant to the extent that detrending the turnover variable already removes firm effects. we also experiment briefly with analogous time-series regressions for the U. Nevertheless. The remainder of the paper is organized as follows. In Section 4. beyond the effects of turnover. such as firm size. for our purposes a variable like size is best thought of as an atheoretic control – it is included in our regressions to help ensure that we do not mistakenly attribute explanatory power to turnover when it is actually proxying for some other firm characteristic. Section 7 concludes. In Section 5. we document other strong influences on skewness. Rather. we are naturally reluctant to declare an unqualified victory for any one theory. In Section 6. we review in more detail the theoretical work that motivates our empirical specification. and most generally. Second. in which we attempt to forecast the skewness in aggregate-market returns. that are not easily rationalized within the context of the Hong-Stein model. we consider the analogous time-series regressions. our efforts to model the determinants of conditional skewness at the firm level are really quite exploratory in nature. we should stress that we do not at this point view them as a tight test. / Journal of Financial Economics 61 (2001) 345–381 349 that also shows up strongly in our multivariate regressions. There are several reasons why one might wish to remain skeptical. The coefficient estimates continue to imply economically meaningful effects. it is comforting to note that the qualitative results from the aggregate-market regressions closely parallel those from the cross-sectional regressions in that high values of both detrended turnover and past returns also forecast more negative market skewness. although that for detrended turnover is no longer statistically significant. we present our baseline cross-sectional regressions. but we keep it in to be safe. . In Section 2. we discuss our sample and the construction of our key variables.

However. A and B. investor B gets a pessimistic signal. In addition to investors A and B. This deviation from full Bayesian rationality – which can be thought of as a form of overconfidence – leads to irreducible differences of opinion about the stock’s value. This is obviously a lower bound on the fraction of funds that never take short positions. Because of the short-sales constraint. and the only trade will be between investor A and the arbitrageurs. Almazan et al. The arbitrageurs are rational enough to figure out that B’s signal is below A’s. each investor’s signal contains some useful information. but they cannot know by how much. unconstrained arbitrageurs. each of whom receives a private signal about a stock’s terminal payoff. which provides the principal motivation for our empirical tests. Moreover. the arbitrageurs might be thought of as hedge funds who are not subject to such restrictions. unlike most of the behavioral finance literature. many of whom are precluded by their charters or operating policies from ever taking short positions. suggesting that funds are also not finding synthetic ways to take short positions. This is because of the presence of the riskneutral. / Journal of Financial Economics 61 (2001) 345–381 2. the model also incorporates a class of fully rational. As a matter of objective reality. Hence. (1999) document that roughly 70% of mutual funds explicitly state (in Form N-SAR that they file with the SEC) that they are not permitted to sell short.3 In contrast.350 J. However. the arbitrageurs themselves are not short-sales constrained. the market as a whole is efficient. even if that of the other investor is revealed to them. the model’s only implications are for the higher-order moments of the return distribution. Importantly. so they can take infinitely large positive or negative positions. overconfidence). begins with the assumption that there are two investors. each of the two investors only pays attention to their own signal. the arbitrageurs may not always get to see both of the signals. These arbitrageurs recognize that the best estimate of the stock’s true value is formed by averaging the signals of A and B.e. . because A and B face short-sales constraints. Koski and Pontiff (1999) find that 79% of equity mutual funds make no use whatsoever of derivatives (either futures or options). so that B’s valuation for the stock lies well below A’s. Chen et al. in the sense of there being no predictability in returns. There are two trading dates. 3 In fact. risk-neutral arbitrageurs. To see how the model can generate asymmetries. B will simply sit out of the market.. Theoretical background The model of Hong and Stein (1999). which relies on limited arbitrage. Even though investors A and B can be said to suffer from behavioral biases (i. Perhaps the most natural interpretation of these assumptions is not to take the short-sales constraint literally – as an absolute technological impediment to trade – but rather to think of investors A and B as institutions like equity mutual funds. imagine that at time 1.

There is a countervailing positive-skewness effect at time 1. the ex ante divergence in priors between A and B – which Hong and Stein denote by H – not only governs the extent of negative skewness. Next. when the stock price is falling at time 2. there is information in how B responds to A’s reduced demand for the stock – in whether or not B gets up off the sidelines and provides buying support. some of B’s previously hidden information might come out.g. the arbitrageurs learn something by observing if and at what price B steps in and starts being willing to buy. Jacklin et al.. It is this unconditional skewness feature – driven by the short-sales constraint – that most clearly distinguishes the model of Hong and Stein from other related models in which pent-up information is revealed through the trading process (e. In this case. 1993). when H is large. while B’s time-1 signal remains hidden. In order to isolate this particular theoretical effect. In particular. it also governs trading volume. trading volume is unusually high at times 1 and 2.. In these other models. Hong and Stein show that if the ex ante divergence of opinion (i.. Chen et al.e. and suppose that A gets a new positive signal. However. so A’s new time-2 signal is incorporated into the price. Genotte and Leland. and variance is greater. the positive time-1 skewness can actually overwhelm the negative time-2 skewness.J. this logic is not sufficient to establish that unconditional returns (i.. and Romer. 1990. since the most negative draws of B’s signal are the ones that get hidden from the market at this time. This high trading volume is associated with a greater likelihood of B moving to the sidelines at time 1. 1992. if A gets a bad signal at time 2. On the other hand. returns are on average symmetrically distributed. the average across time 1 and time 2) are negatively skewed. so that returns are on average positively skewed. as opposed to rising. This comparative static result holds regardless of whether unconditional skewness (averaged across different values of H) is positive or negative. we need to be aware of other potentially confounding factors. This greater variance on the downside implies that time-2 returns will be negatively skewed. move to time 2. 1988.e. A continues to be the more optimistic of the two. albeit potentially quite volatile. When A’s and B’s priors are sufficiently close to one another. Moreover. Grossman. and subsequently moving off the sidelines at time 2 – precisely the mechanism that generates negative skewness. Nevertheless. and it forms the basis for our empirical tests. the time-2 effect dominates. and unconditional returns are negatively skewed. but does not fully reflect B’s time-1 signal. the difference in priors) between A and B is great enough. / Journal of Financial Economics 61 (2001) 345–381 351 Thus the market price at time 1 impounds A’s prior information. it is well known that . In other words. This is because as A bails out of the market at time 2. Thus the comparative statics properties of the model with respect to the parameter H lead to the prediction that increases in trading volume should forecast more negative skewness. For example. Thus more information comes out.

There are a number of models that can deliver such a correlation. Our sample period begins in July 1962. Theoretical models that relate trading volume to differences of opinion typically assume that transactions costs are zero. To the extent that such an effect is present in our data. past returns might also help predict skewness. all of our regressions include some control for volatility. / Journal of Financial Economics 61 (2001) 345–381 trading volume is correlated with past returns (Shefrin and Statman. To address this concern. Chen et al.. higher levels of volatility are associated with more negative skewness. which is as far back as we can get the trading volume data.e. according to the HongStein model – or whether it is really just forecasting volatility. which is in turn correlated with skewness. and we experiment with several ways of doing this control. and scoresFi. 4 . and In the model of Coval and Hirshleifer (1998). we further truncate the sample by eliminating the very smallest stocks in the NYSE/AMEX universe – in particular. we would like to know whether turnover is forecasting skewness directly – as it should. We do not include NASDAQ firms because we want to have a uniform and accurate measure of trading volume. there is also conditional negative skewness after periods of positive returns. 1986. even though unconditional average skewness is zero. Lakonishok and Smidt.4 Indeed. Data To construct our variables. primes. closed-end funds. In reality. variations in transactions costs are likely to be an important driver of trading volume. we do not actually begin to forecast returns until December 1965. Odean. And. one might also worry about skewness being correlated with volatility. we begin with data on daily stock prices and monthly trading volume for all NYSE and AMEX firms. 1998b). for example. and the dealer nature of the NASDAQ market is likely to render turnover in its stocks not directly comparable to that of NYSE and AMEX stocks. stocks that do not have a CRSP share type code of 10 or 11. We do so because our goal is to use trading volume as a proxy for differences of opinion. those with a market capitalization below the 20th percentile NYSE breakpoint.352 J. because our regressions use many lags. 1985. as noted above. from the CRSP daily and monthly stock files. In a similar vein. For most of our analysis. REITs. just such a pattern has been documented in recent work by Harvey and Siddique (2000). We also follow convention and exclude ADRs. all of our regressions include a number of lags of past returns on the right-hand side. if there are stochastic bubbles of the sort described by Blanchard and Watson (1982). To control for this tendency. 3. in the volatility-feedback model of Campbell and Hentschel (1992).

as well as any other moments that rely on daily return data. for ‘‘negative coefficient of skewness. Using simple percentage returns instead of log changes does have two (predictable) effects: (1) it makes returns look more positively skewed on average and (2) it induces a pronounced correlation between skewness and contemporaneously measured volatility.’’ is calculated by taking the negative of (the sample analog to) the third moment of daily returns. more precisely. In calculating NCSKEW. We use log changes as opposed to simple daily percentage returns because they allow for a natural benchmark – if stock returns were lognormally distributed. However. and n is the number of observations on daily returns during the period. the daily firm-level returns that go into the calculation of the NCSKEW variable are market-adjusted returns – the log change in stock i less the log change in the value-weighted CRSP index for that day. actually log changes in price. We also report some sensitivities in which the smallest stocks are analyzed separately (see Table 4 below). / Journal of Financial Economics 61 (2001) 345–381 353 more so for very small stocks.e. Scaling the raw third moment by the standard deviation cubed allows for comparisons across stocks with different variances. For most of our regressions. given that we control for volatility in all of our regression specifications. 1993). we also run everything with variations of NCSKEW based on both (1) excess returns (the log change in stock i less the T-bill return) as well as . the coefficients on turnover for this subsample are noticeably smaller. we hope to raise the ratio of signal (differences of opinion) to noise (transactions costs) in our key explanatory variable. and as one would expect from this discussion.J. We have also redone everything with an NCSKEW measure based instead on simple daily percentage returns. Our baseline measure of skewness. for any stock i over any six-month period t. this is the usual normalization for skewness statistics (Greene. ð1Þ it it where Rit represents the sequence of de-meaned daily returns to stock i during period t. However. Chen et al. Thus. and dividing it by (the sample analog to) the standard deviation of daily returns raised to the third power. By putting a minus sign in front of the third moment.. having a more left-skewed distribution. we drop any firm that has more than five missing observations on daily returns in a given period. we have  X 3=2  X  NCSKEWit ¼ À nðn À 1Þ3=2 R3 ðn À 1Þðn À 2Þ R2 . which we denote NCSKEW. then an NCSKEW measure based on log changes should have a mean of zero. using simple percentage returns does not materially alter the coefficients on turnover and past returns. These daily ‘‘returns’’ are. and none of our main results are affected. By eliminating the smallest stocks. we are adopting the convention that an increase in NCSKEW corresponds to a stock being more ‘‘crash prone’’ – i.

but has little predictive power beyond that.g. Chen et al. The choice of a six-month horizon for measuring skewness is admittedly somewhat arbitrary. but there is little payoff to doing so. For example. however.’’ is computed as follows. the effects that we are interested in could be playing themselves out over a shorter horizon. In particular. Again. our results with NCSKEW and DUVOL are for the most part quite similar. and hence is less likely to be overly influenced by a handful of extreme days. In all cases. respectively. As will be seen. we also work with a second measure of return asymmetries that does not involve third moments. we separate all the days with returns below the period mean (‘‘down’’ days) from those with returns above the period mean (‘‘up’’ days). for ‘‘down-to-up volatility. March 1–August 31 and September 1–February 28. We have. This alternative measure. the convention is that a higher value of this measure corresponds to a more leftskewed distribution. our choice to use six months’ worth of daily returns to estimate skewness is driven more by measurement concerns. ð2Þ ðnd À 1Þ DOWN UP where nu and nd are the number of up and down days. this is particularly relevant given that a higher-order moment like . we use nonoverlapping six-month observations on skewness. Unfortunately. statistical power becomes a real issue. In addition to NCSKEW. if we estimated skewness using only one month’s worth of data. For any stock i over any six-month period t. which we denote by DUVOL. When we turn to the timeseries regressions with aggregate-market data. we already have more than enough statistical power as it is. so that we would have a new skewness measure every month. etc. checked our results by re-running everything using different nonoverlapping intervals – e. as will become clear shortly. February 1–July 31 and August 1–January 31.354 J. the model of Hong and Stein does not give us much guidance in this regard. In our regressions with firm-level data. In principle. To preview. / Journal of Financial Economics 61 (2001) 345–381 (2) beta-adjusted returns. the results are essentially identical. since. Lacking this theoretical guidance. Thus we have ( !)  X X 2 2 DUVOLit ¼ log ðnu À 1Þ Rit Rit . We could alternatively use overlapping data. so it does not appear that they depend on a particular parametric representation of return asymmetries. these variations do not make much difference to our results with NCSKEW. so that trading volume on Monday forecasts skewness for the rest of the week. and compute the standard deviation for each of these subsamples separately. We then take the log of the ratio of (the sample analog to) the standard deviation on the down days to (the sample analog to) the standard deviation on the up days.. and we use overlapping observations. the NCSKEW and DUVOL measures are calculated using data from either January 1–June 30 or July 1–December 31 of each calendar year. we would presumably have more measurement error.

we want to eliminate any component of turnover that can be thought of as a relatively fixed firm characteristic.e. (2) five size-based subsamples. / Journal of Financial Economics 61 (2001) 345–381 355 skewness is strongly influenced by outliers in the data. In Panels B and C. the other variables that we use are very familiar and do not merit much discussion. is that to the extent that our measurement horizon does not correspond well to the underlying theory.e. it makes little sense to require that firm effects be literally constant over the entire sample period. Table 1 presents a variety of summary statistics for our sample. SIGMA and RET are based on excess returns as well. defined as the value-weighted NYSE/AMEX index. positive mean values of NCSKEW and DUVOL – for the market as a whole. Chen et al. LOGSIZEit is the log of firm i ’s stock market capitalization at the end of period t: BK/MKTit is firm i ’s book-to-market ratio at the end of period t: LOGCOVERit is the log of one plus the number of analysts (from the I/B/E/S database) covering firm i at the end of period t: TURNOVERit is the average monthly share turnover in stock i. Instead. When we compute NCSKEW or DUVOL using excess returns. which are positively skewed. Again. which we denote DTURNOVER.. and (3) the market as a whole. Since we have such a long time series. This detrending is roughly analogous to doing a fixed-effects specification in a shorter-lived panel. SIGMA and RET are computed using market-adjusted returns.. (When working with the market as a whole. defined as shares traded divided by shares outstanding over period t: In our baseline specification. as noted above. This discrepancy can in principle be understood within the strict confines of the Hong-Stein model. the model . also measured over the six-month period t: When we compute NCSKEW or DUVOL using either market-adjusted or betaadjusted returns. as in most of our subsequent regression analysis. however. all the variables are based on simple excess returns relative to T-bills. respectively. the rationale for doing this detrending is that. we work with detrended turnover. the detrending controls for firm characteristics that adjust gradually. we restrict the sample to those firms with a market capitalization above the 20th percentile NYSE breakpoint. measured over the six-month period t: RETit is the cumulative return on stock i. The important point to note. for the sample of individual firms. as a matter of conservatism. Besides the skewness measures.) Panels B and C look at contemporaneous correlations and autocorrelations. Panel A shows the means and standard deviations of all of our variables for (1) the full sample of individual firms. the opposite is true for individual stocks. it should make our tests too conservative. One interesting point that emerges from Panel A is that while there is negative skewness – i. this should simply blur our ability to find what the theory predicts – i. by subtracting from the TURNOVER variable a moving average of its value over the prior 18 months. since. The detrending is done very simply. SIGMAit is the standard deviation of stock i ’s daily returns.J.

RETt is the market-adjusted cumulative return in the six-month period t. which is measured starting in December 1976. measured using market-adjusted returns in the six-month period t. SIGMAt is the standard deviation of (daily) marketadjusted returns measured in the six-month period t. À0. measured using market-adjusted returns in the six-month period t.171 0.139 0. BK=MKTt is the most recently available observation of the book-to-market ratio at the end of period t. Chen et al.476 0. TURNOVERt is the average monthly turnover measured in the six-month period t.362 0.172 0.964 À0.437 À0.190 0.923 À0.128 0. Size quintiles are determined using NYSE breakpoints. Quintile 5 (largest) firms Quintile 4 firms Quintile 3 firms Quintile 2 firms Quintile 1 (smallest) firms Market portfolio All firms Panel A: First and second moments NCSKEWt Mean Standard dev.262 0.213 0. / Journal of Financial Economics 61 (2001) 345–381 Table 1 Summary statistics The sample period is from July 1962 to December 1998.266 0. DTURNOVERt is average monthly turnover in the six-month period t.356 J. LOGSIZEt is the log of market capitalization measured at the end of period t.141 0. DUVOLt is the log of the ratio of down-day to up-day standard deviation.904 À0.436 À0.806 À0.406 À0. NCSKEWt is the negative coefficient of (daily) skewness.198 0.939 À0.224 0.364 À0. DUVOLt Mean Standard dev.994 À0.377 .391 À0.155 0. detrended by a moving average of turnover in the prior 18 months.268 0.735 0. except for LOGCOVERt. LOGCOVERt is the log of one plus the number of analysts covering the stock at the end of period t.

249 1. of obs. LOGSIZEt Mean Standard dev.642 0.924 0.001 0.018 5.055 0.054 0.202 14.017 0.983 14.503 0.177 2.030 0.029 0. Chen et al.710 2.055 0.040 0.023 0.036 1.431 0.988 0.008 0.512 0.073 0.007 6. BK/MKTt Mean Standard dev.010 4.564 0.050 0.508 À0.472 3.563 0. 0. RETt Mean Standard dev.039 0.121 1.984 0.991 0.037 0. No.653 0.019 0.000 0. TURNOVERt Mean Standard dev.060 0.046 0.005 0.920 1.291 0.727 0.015 0.656 0.000 0.043 0.372 41.005 8.035 0.824 0.108 1.002 0.002 0.020 0.164 13.860 0.288 16.275 22.025 0.023 3.056 0.006 0.022 0.651 0.066 0.015 0.075 0.001 0.021 0.SIGMAt Mean Standard dev. DTURNOVERt Mean Standard dev.063 0.034 0.870 2.003 0.002 0.017 0.003 N/A J.098 À0.840 0.782 0.008 5. / Journal of Financial Economics 61 (2001) 345–381 357 N/A N/A 0.935 1.050 0.051 0.140 0.240 14.297 100.024 0.197 1.095 0.667 0.565 0.241 0.108 421 .898 0. LOGCOVERt Mean Standard dev.042 0.

023 0.068 À0.257 À0.022 0. Chen et al.061 DUVOLtÀ1 SIGMAtÀ1 LOGSIZEtÀ1 BK=MKTtÀ1 LOGCOVERtÀ1 DTURNOVERtÀ1 TURNOVERtÀ1 RETtÀ1 Panel C: Autocorrelations and cross-correlations (using only firms above 20th percentile in size) NCSKEWt DUVOLt SIGMAt LOGSIZEt BK/MKTt LOGCOVERt DTURNOVERt TURNOVERt RETt 0.006 À0.976 À0.104 À0.043 0.045 À0.030 .047 0.027 0.119 0.342 À0.090 À0.049 0.068 À0.032 0. / Journal of Financial Economics 61 (2001) 345–381 NCSKEWt DUVOLt SIGMAt LOGSIZEt BK/MKTt LOGCOVERt DTURNOVERt TURNOVERt RETt NCSKEWtÀ1 0.056 À0.166 À0.130 0.056 0.852 0.719 À0.026 0.050 À0.022 0.019 0.130 0.086 0.729 À0.181 0.294 À0.195 À0.008 À0.028 À0.292 0.039 0.158 À0.047 À0.066 À0.042 À0.063 0.022 0.218 0.307 0.238 0.017 0.014 0.032 À0.016 0.028 À0.213 0.007 À0.035 0.076 0.398 0.302 À0.008 0.013 0.736 0.009 0.067 À0.042 0.047 0.052 0.079 0.055 0.782 À0.318 0.011 À0.098 À0.014 À0.093 0.013 0.011 À0.061 À0.000 0.079 À0.080 0.064 0.371 0.100 À0.101 À0.059 0.071 0.002 0.015 À0.047 À0.182 0.133 0.059 0.038 0.781 À0.030 À0.028 0.002 0.109 0.006 À0.311 0.875 0.376 0.381 0.715 À0.006 0.358 Table 1 (continued) NCSKEWt DUVOLt SIGMAt LOGSIZEt BK/MKTt LOGCOVERt DTURNOVERt TURNOVERt RETt Panel B: Contemporaneous correlations (using only firms above 20th percentile in size) J.034 0.104 0.024 0.179 À0.026 À0.013 0.081 0.080 À0.051 0.042 À0.029 0.

We pool all the data (excluding firms with market capitalization below the 20th percentile NYSE breakpoint) and regress NCSKEWit+1 against its own lagged value. and six lags of past returns. Baseline specification Table 2 presents our baseline cross-sectional regression specification. it must be missing something when it comes to explaining the mean skewness of individual stocks.01. We also include dummy variables . For NCSKEW the autocorrelation is on the order of 0. The most noteworthy fact in Panel B of Table 1 is the contemporaneous correlation between our two skewness measures. Forecasting skewness in the cross-section 4. In other words. without further embellishments. depending on the degree of ex ante investor heterogeneity. as well as SIGMAit.1. Nevertheless. While these two measures are quite different in their construction. Chen et al. these low correlations lend some preliminary (and comforting) support to the notion that forecasting either of our skewness measures is a quite distinct exercise from forecasting volatility.05. DTURNOVERit. the theory will certainly gain some credence if it does a good job of explaining cross-sectional variation in skewness.88. they appear to be picking up much the same information. even if it cannot fit the mean skewness at the firm level. it is not clear that such an assumption is empirically defensible. RETit yRETitÀ5. 4.J.09.72 – neither of our skewness measures has much persistence. However. NCSKEWit. An alternative interpretation of the data in Table 1A is that even if the Hong-Stein model provides a reasonable account of skewness in market returns. NCSKEW and DUVOL. if one is willing to assume that differences of opinion about the market are on average more pronounced than differences of opinion about individual stocks. This view does not imply that we cannot learn something about the theory by looking at firm-level data. it is worth emphasizing the caveat that. for DUVOL it is 0. and that between DUVOL and SIGMA is about À0. the theory might not provide a convincing rationale for everything that is going on at the individual stock level. Panel C documents that. it might be that large positive events like hostile takeovers (which the theory ignores) impart an added degree of positive skewness to individual stocks but wash out across the market as a whole. LOGSIZEit.08. unlike SIGMA – which has an autocorrelation coefficient of 0. For example. / Journal of Financial Economics 61 (2001) 345–381 359 allows for either positive or negative unconditional skewness. Also worth pointing out is that the correlation between NCSKEW and SIGMA is less than 0. the model can produce negative skewness for the latter and positive skewness for the former. which is 0.

053 (3. t-statistics.426 0.370 (À5.082 (4.462) 0.110) 0. Chen et al. (1) Base case: market-adjusted returns NCSKEWt SIGMAt LOGSIZEt DTURNOVERt RETt RETtÀ1 RETtÀ2 RETtÀ3 RETtÀ4 RETtÀ5 No.441) À3.051 (7.053 (7. are adjusted for heteroskedasticity and serial correlation.218 (10.041 (2.706) 0.220) 0.092 (5.403) 0.129) 0.759) 51.638) 0.778) À4.046 (13.103 (5.465) 0.329) 0.920) À2.058 (3.257) 51. All regressions also contain dummies for each time period (not shown).242) 0.175) 0.180) 0.089 (5.052 (7. detrended by a moving average of turnover in the prior 18 months.426 0.566 (À7. The dependent variable is NCSKEWtþ1 .030 (2) Beta-adjusted returns 0.054 (2. RETtyRETtÀ5 are returns in the six-month periods t through t À 5 (these past returns are market adjusted in cols. 1–3.197 (9.675) 0.108 (5.839) 0.497) 0.071 (3. beta-adjusted returns and simple excess returns in cols.067 (3.175) 0.059 (19. SIGMAt is the (daily) standard deviation of returns in the six-month period t: LOGSIZEt is the log of market capitalization at the end of period t: DTURNOVERt is average monthly turnover in the six-month period t. of obs.082 (4.701 (À4.437 (3. respectively.426 0.149) 0. 1À2 and excess in col.360 J.701) 0.082 for each time period t: The regression can be interpreted as an effort to predict – based on information available at the end of period t – cross-sectional variation in skewness over period t þ 1: .607) 0. which are in parentheses. the negative coefficient of (daily) skewness in the six-month period t þ 1: NCSKEWtþ1 is computed based on market-adjusted returns.221 (11.477) 0. / Journal of Financial Economics 61 (2001) 345–381 Table 2 Forecasting skewness in the cross-section: pooled regressions The sample period is from July 1962 to December 1998 and includes only those firms with market capitalization above the 20th percentile breakpoint of NYSE.031 (3) Excess returns 0.001) 0.335) 51.133) 0.296) 0.364 (3.109 (6.083 (4. 3).062 (3. R2 0.830) 0.364 (3.037 (11.

a measure analogous to what we use here. the findings for past returns and size run broadly parallel to previous work by Harvey and Siddique (2000).e. they may try to delay its release. suggesting that negative skewness is more likely in large-cap stocks. Nevertheless. such results were by no means a foregone conclusion. as in Harvey and Siddique (2000). For 25 portfolios sorted on size and book-to-market. what is the correlation of the two skewness measures? The answer is about 0. and (2) the average skewness of daily individual stock returns. The coefficient on size is also positive. correlation. In particular. (We expect lower coefficient estimates when using simple excess returns as compared to market-adjusted returns – after all. so it should have more ability to explain skewness in the purely idiosyncratic component of stock returns. ours are couched in a multivariate regression framework. But more significantly. while they look at portfolios of stocks. it is possible to come up with rationalizations after the fact. Thus. for two reasons. First. the positive coefficient on size is not something one would have necessarily predicted ex ante based on the Hong-Stein model. while they look at monthly returns. we look at daily returns.) The past return terms are also always positive and strongly significant. As noted above. to become more crash-prone. especially if. It seems plausible that if they uncover good news. To begin. We have done some detailed comparisons to make these latter points explicit. if they are sitting on bad news. we look at individual stocks. while theirs are based on univariate sorts. as Harvey and Siddique (2000) stress. and in column 3 we use simple excess returns. while it might have been reasonable to conjecture – based on the prior evidence in Harvey and Siddique (2000) – that our firm-level NCSKEW variable would also be related to past returns and size. Second. a relatively low.J. our measure of skewness is quite different from theirs. The skewness of a portfolio of stocks is not the same thing as the average skewness of its component stocks. As we have already stressed. The results are quite similar in all three cases.22. all else equal. the coefficients on detrended turnover are positive and strongly statistically significant in each of the three columns. In contrast. DTURNOVER is a firm-specific variable. We can then ask the following: Across the 25 portfolios. we use market-adjusted returns as the basis for computing the NCSKEW measure. albeit significantly positive. we have computed both (1) the skewness of monthly portfolio returns. albeit somewhat larger (by about 20%) in magnitude when market-adjusted returns are used. / Journal of Financial Economics 61 (2001) 345–381 361 In column 1. there are several distinctions between our results and theirs. Chen et al.. In column 2 we use beta-adjusted returns. they will disclose all this good news right away. with the result that the . Suppose that managers can to some extent control the rate at which information about their firms gets out. coskewness varies systematically with firm characteristics. Nevertheless. Thus stocks that have experienced either a surge in turnover or high past returns are predicted to have more negative skewness – i.

TURNOVER. we are essentially removing our fixed-effect control from the turnover variable. we truncate outliers of the NCSKEW variable. In other words. one might expect that long-run cross-firm variation in turnover would also predict skewness – some firms might be subject to persistently large differences in investor opinion. In columns 3 and 4. This has little impact on the results. and showing up in the regression only because SIGMAit+1 is correlated with NCSKEWit+1. This means that we are now admitting into consideration differences in turnover across firms that are not merely temporary deviations from trend but more in the nature of long-run firm characteristics. Moreover. and use it to develop some additional testable implications. or whether it is instead forecasting SIGMAit+1. and uses an NCSKEW measure based on market-adjusted returns.2. if one adds the further assumption that it is easier for managers of small firms to temporarily hide bad news – since they face less scrutiny from outside analysts than do managers of large firms – the resulting positive skewness will be more pronounced for small firms. Chen et al. roughly doubling in magnitude from its base-case value. 4. Recall that the central issue here is whether DTURNOVERit is really forecasting NCSKEWit+1 directly. suggesting that they are not driven by a handful of outlier observations. we investigate whether our results are somehow tied to the way that we have controlled for volatility. We return to this idea in Section 4. we are throwing out a dimension of the data that is strongly supportive of the theory. in column 1 of Table 3.5 below. Everything is a variation on column 1 of Table 2. so that we can verify that DTURNOVERit is still significant even after the inclusion of this control. This behavior will tend to impart positive skewness to firm-level returns. / Journal of Financial Economics 61 (2001) 345–381 bad news dribbles out slowly. This implies that our fixed-effect approach of using DTURNOVER instead of TURNOVER everywhere else in the paper is quite conservative – in doing so. In column 2. Ideally. Robustness In Table 3 we conduct a number of further robustness checks. First. The coefficient estimate on TURNVOVER in column 2 confirms this notion.362 J. and may explain why returns on individual stocks are on average positively skewed at the same time that market returns are negatively skewed. setting all observations that are more than three standard deviations from the mean in any period t to the threestandard-deviation tail values in that period. we replace the DTURNOVER variable with its un-detrended analog. we would like to add a period-t control variable to the regression that is a good forecast of SIGMAit+1. Our use of SIGMAit in the base-case specification can be motivated on the grounds that it is probably the best . According to the theory. and these too should matter for return asymmetries.

given the very pronounced serial correlation in the SIGMA variable. based on the following information set available in period t: SIGMAit. are more crash-prone. SIGMAitÀ1. we also experiment with allowing two coefficients on SIGMAit. Finally. These two lags are insignificant. this variation leads to almost exactly the same results as in the base case. we allow this effect to be a function of the realization of NCSKEWit itself. Overall. DTURNOVERit. SIGMAitÀ2. This variation (not shown in the table) makes no difference to the results. Chen et al. This variable attracts a significant negative coefficient. In column 5. but is instrumented using the information available in period t: As can be seen. we conclude that it is highly unlikely that our base-case success in forecasting NCSKEW with the DTURNOVER and RET variables arises because these variables are able to forecast SIGMA. (1993). Motivated by the work of Glosten et al. we check to make sure our results are robust to how we have modeled the effect of the lagged skewness variable. Although the difference in units precludes a direct comparison of the point estimates. the qualitative patterns are generally the same as in the corresponding specification in column 1 of Table 2. the addition of BK/MKT has no impact on the DTURNOVER coefficient. we add the book-to-market ratio. 3. one for positive past returns and one for negative past returns.0%). .J. Instead of estimating just one coefficient on NCSKEWit.7% vs. these variables really appear to be predicting cross-firm differences in the asymmetry of stock returns. based on the evidence in columns 3 and 4 of Table 3. / Journal of Financial Economics 61 (2001) 345–381 363 univariate predictor of SIGMAit+1. We then replace SIGMAit in the base case with this fitted value of future volatility. the t-statistic on DTURNOVER is actually a bit higher (4. In this spirit. One can presumably do better by allowing for richer dynamics. NCSKEWit. rather than just differences in volatility. we add in column 3 two further lags of SIGMA (SIGMAitÀ1 and SIGMAitÀ2) to the base-case specification. In column 6. Indeed. We implement this by interacting NCSKEWit with five dummy variables. in an unreported sensitivity. SIGMAHATit+1. we use the DUVOL measure of return asymmetry as the lefthand-side variable in place of NCSKEW. who find that the effect of past volatility on future volatility depends on the sign of the past return. In other words.35 vs. 3. and RETit yRETitÀ5. LOGSIZEit. which means that it tells the same story as the past-return terms: glamour stocks.84) as is the R2 of the regression (6. which we denote by SIGMAHATit+1. to the base-case specification. In column 4 we take our logic one step further. We create a fitted value of SIGMAit+1. and hence our coefficients on DTURNOVERit as well as on the six RET terms are virtually unchanged. This is equivalent to an instrumental-variables regression in which future volatility SIGMAit+1 is included on the right-hand side. like those with high past returns. However. BK/MKT. But using just one past lag is not necessarily the best way to forecast SIGMAit+1.

051 (7.014 (9. are adjusted for heteroskedasticity and serial correlation.999 (À7.938) À6.956 (À15.033 (9.353) 0.RETtÀ5 : All regressions also contain dummies for each time period (not shown).572) 0..384) À0. DTURNOVERt. Chen et al. SIGMAtÀ2 . In columns 1–5. returns are market-adjusted.037 (10.053 (7.837) (3) More lags of past volatility 0. 6) SIGMAHATtþ1 SIGMAt SIGMAtÀ1 SIGMAtÀ2 LOGSIZEt BK/MKTt 0.751) À0. except in column 3.035 (10.034 (9. the negative coefficient of (daily) skewness in the six-month period t þ 1: In column 6.822) À3.953 (À3. SIGMAt is the standard deviation of (daily) returns in the six-month period t: LOGSIZEt is the log of market capitalization at the end of period t: BK/MKTt is the most recently available observation of the book-to-market ratio at the end of period t: DTURNOVERt is average monthly turnover in the six-month period t.627) J. t-statistics.180) À4. the dependent variable is NCSKEWtþ1 .808) 0.367 (À0.020 (À3.675) (2) Turnover not detrended 0.047) 0.178 (À7. RETtyRETtÀ5 are returns in the six-month periods t through t À 5: SIGMAHATtþ1 is the predicted value of SIGMAtþ1 calculated from a regression of SIGMAtþ1 on SIGMAt.994 (À8. LOGSIZEt.053 (7.460 (À0.663) (4) Fitted future volatility 0.552) À4.050 (8.351) 0.035 (12. the dependent variable is DUVOLtþ1 .618 (À9.054 (7.096 (16. the log of the ratio of down-day to up-day standard deviation in the six-month period t þ 1: In all columns.. (1) Outliers truncated NCSKEWt (DUVOLt in col.364 Table 3 Forecasting skewness in the cross-section: robustness checks The sample period is from July 1962 to December 1998 and includes only those firms with market capitalization above the 20th percentile breakpoint of NYSE.980) À4. detrended by a moving average of turnover in the prior 18 months. which are in parentheses. and RETt.095) À0.698) (5) Book-tomarket 0. / Journal of Financial Economics 61 (2001) 345–381 .454) À6.y.898) 0.750) (6) Using DUVOLt+1 0. where turnover is not detrended.

051 (2.014 (1.848) 0.734) 0.671) 0.206 (11.142 (15.229 0.213 (10.073 (3.075 (4.729) 0.249) 0.054 (2.039 0.083 (4.009 (1. of obs. Chen et al.048 (3.621) 0.828) 0.287) 0.038) 0.057 (3.041 (2.049 (3.375 (3.030 0.030) 0.426 0.630 0.100 (6.411 (3.587) 0.523) 0.071 (3.580) 51.081 (4.087) 0.810) 0.045 (5.705) 51.426 0.820) 51.071) 0.056 (2.472) 0.887) 0.217 (10.131) 0.014 (1.587) 0.031 0.387 (3.060 (3.685) 0.098 (5.727) 0.054 (2.324) 48.943) 0.030 0.273) 0.033 (1.722) 0.787) 0.106 (5.819) 0.218 (10.808) 0.064 (3.202 (4.104 (5.030 0.067 J.761) 0.072 (3.208 (10.789) 51.066 (3.410) 0.088 (4.816) 0. R 2 0.337) 0.426 0. / Journal of Financial Economics 61 (2001) 345–381 365 .329) 0.428) 0.923) 52.056 (2.393 0.DTURNOVERt (TURNOVERt in col.455 (3.014 (1.346) 0. 2) RETt RETtÀ1 RETtÀ2 RETtÀ3 RETtÀ4 RETtÀ5 No.084) 0.084 (4.761 (7.459) 0.

albeit still positive. As we have emphasized. Thus. 6 Several recent papers find that predictability – based on either ‘‘momentum’’ or ‘‘value’’ strategies – is stronger in small-cap stocks (see. / Journal of Financial Economics 61 (2001) 345–381 one corresponding to each quintile of NCSKEWit. we disaggregate our base-case analysis by size. the coefficients look reasonably stable.g. 4. There is certainly no hint that the effects that we are interested in go away for larger firms. we are reluctant to hazard an economic interpretation for these anomalies.366 J. 1998. The fact that the coefficients on DTURNOVER are robust for large firms is not surprising in light of the underlying theory. Indeed.. and the coefficients on the past-return terms are smaller and much less significant. the model does not have the feature that the key effects diminish as one moves to larger stocks. whereas our theory requires a good proxy for differences of opinion.. Hong et al.5 Again.g.) Two conclusions stand out. Fama. 2000. In other words. Second. the coefficient on DTURNOVER for the smallest category of firms is noticeably lower than for any other group. we estimate five separate slope coefficients on lagged skewness. 1990) whose implications for return predictability are often thought of as applying more forcefully to small stocks. the model of Hong and Stein is not predicated on impediments to arbitrage – it incorporates a class of fully risk-neutral arbitrageurs who can take infinite long or short positions. This is in contrast to behavioral models based on limited arbitrage (e. where arbitrage activity is presumably more efficient. as suspected. the highest point estimate for the DTURNOVER coefficient comes from the next-to-largest quintile. the coefficient on SIGMA changes signs. these nonlinearities do not at all impact the coefficients on any of the other variables of interest. depending on the quintile that lagged skewness falls in.. We take the specification from column 1 of Table 2 and run it separately for five size-based subsamples. corresponding to quintiles based on NYSE breakpoints. 1999). (Recall that in Tables 2 and 3.g. and Griffin and Lemmon. as long as some investors other than the arbitrageurs (e. Given the potential distortions associated with infrequent trading and price discreteness for this group. once one moves beyond the smallest quintile.3. e. this is probably because variation in turnover for these tiny firms is driven in large part by variation in trading costs. . DeLong et al. As it turns out. we omit the smallest of these five quintiles from our sample. First.. Chen et al.6 5 Also for these smallest firms.. Cuts on firm size In Table 4. while there appear to be some modest nonlinearities in the effect of lagged skewness. mutual funds) continue to be short-sales constrained.

554) 0.014 (0. Chen et al.053 (3.045 (5.010 (À0.690) À4.021) 0.335 (7.128 (2.215) 0.816) 0.535) 0.078 (1.014 (0.028 4.047 (1.906) 12.758) À3.010 (0.341) À4.450) 0.134 (4. R2 0.404 (1.059 (3.855) 0.503) 12.001 (0.554) 0.864) À0.726) 0. SIGMAt is the standard deviation of (daily) returns in the six-month period t: LOGSIZEt is the log of market capitalization at the end of period t: DTURNOVERt is average monthly turnover in the six-month period t.298) 0.101 (2.653) À4.771) 0.359) 0.013 (0.155 (4.035 Quintile 4 firms 0.076) 0. Stability over subperiods In Table 5.750 0.083 (2.852) 0. of obs.054 (4.165 0.264 (1.112) 0.024 Quintile 2 firms 0. Specifically.029 Quintile 1 (smallest) firms 0.894 (8. The dependent variable in all columns is NCSKEWtþ1 .639) 0.749 0.465) 13.105 (3.025 (0. t-statistics are adjusted for heteroskedasticity and serial correlation.362 (À2. we run a separate.033 (1.024) 0.043 (3.263) 0.823) 0.043 (À1.682) 0.000) 0.632) 29.801) 12. purely cross-sectional variant of the regression in column 1 of Table 2 for every one of .260 (5.431) 2. using a Fama-MacBeth (1973) approach.800) 0.551 (2. the negative coefficient of (daily) skewness in the six-month period t þ 1: In all columns.612) 0.243) 0.793) 0.071 (1.682) 0.069 (2. RETtyRETtÀ5 are returns in the six-month periods t through t À 5: All regressions also contain dummies for each time period (not shown). we examine the intertemporal stability of our baseline regression.269) 0.066 (8.104 (2.030 Quintile 3 firms 0.134 (3.334) 0.072) 0.4.637) 0.540) 0.651) 0.590) 0.163 (3.162 (3.017 (1. returns are market-adjusted.391) 0. Firm size cuts are based on NYSE breakpoints.006 (0.049 (1.093 (2.062 (À4.215 (5.637) 0.569) 0.165 (3.009 (1.079 (1. detrended by a moving average of turnover in the prior 18 months.812) 0.009) 0.089 (1.028 (1. Quintile 5 (largest) firms NCSKEWt SIGMAt LOGSIZEt DTURNOVERt RETt RETtÀ1 RETtÀ2 RETtÀ3 RETtÀ4 RETtÀ5 No.J.407 0.157) 0.409 (À3. / Journal of Financial Economics 61 (2001) 345–381 Table 4 Forecasting skewness in the cross-section: cuts by firm size 367 The sample period is from July 1962 to December 1998.520 0.637 (2.057 (1.

368 J.730 0.464 0.757 2.145 0.517) À9.955 À0. the negative coefficient of (daily) skewness in the six-month period t þ 1: In all cases.061) 0.040 (2.026 À0. detrended by a moving average of turnover in the prior 18 months.913 À0.062 0.835 0.381 All periods 1990s Panel B: Average coefficients by subperiods NCSKEWt 0.334 À0.258) 2. returns are market-adjusted.161 0.207 0.099 (2.577 (À2.383 1.173) À11.024 (1.372 1.776) 0.182 À0.498 6/1970 12/1970 6/1971 12/1971 6/1972 12/1972 6/1973 12/1973 6/1974 12/1974 6/1975 12/1975 6/1976 12/1976 6/1977 12/1977 6/1978 12/1978 6/1979 12/1979 0. The specification is the same as in col. RETtyRETtÀ5 are returns in the six-month periods t through t À 5: Panel A reports only the coefficient on DTURNOVERt for each period.901 0.507 (À3. and the associated t-statistics.404 À0.121 À0.803 0.248 À0. 1960s 1970s 1980s 1990s Panel A: Period-by-period regressions (12/1965 to 6/1998). SIGMAt is the standard deviation of (daily) returns in the six-month period t: LOGSIZEt is the log of market capitalization at the end of period t: DTURNOVERt is average monthly turnover in the six-month period t.941 0.079 (4.257 0.210 Late 1960s 6/1980 12/1980 6/1981 12/1981 6/1982 12/1982 6/1983 12/1983 6/1984 12/1984 6/1985 12/1985 6/1986 12/1986 6/1987 12/1987 6/1988 12/1988 6/1989 12/1989 1970s 1.156 À0.314 0.643 0.904 2.569 0.373 2. coefficient on detrended turnover only 12/1965 6/1966 12/1966 6/1967 12/1967 6/1968 12/1968 6/1969 12/1969 0.016 0.218 1.030 (4.038 1980s 6/1990 12/1990 6/1991 12/1991 6/1992 12/1992 6/1993 12/1993 6/1994 12/1994 6/1995 12/1995 6/1996 12/1996 6/1997 12/1997 6/1998 1.780 À0.876 1.579 À0.407 (0.269 0.672 0.200) SIGMAt LOGSIZEt . and adjusted for serial correlation.981 0.880) À5.884 (À1.222) 0.707 À0.686 0.448 À0.194 1.081 0.973 1.459 0. based on the time-series standard deviations of the coefficients. Panel B reports the mean coefficients for different subperiods.506 0.288) 0.101 0.388 0. Chen et al.918 1.141) 0.512 1.017 (À2.363 À0.027 (2.065 0.063 (4.029 À0.468 1.063) 0.545 2.005 (0.201 0.631 1. The dependent variable is NCSKEWtþ1 .064 (2.216) 0.751 0. / Journal of Financial Economics 61 (2001) 345–381 Table 5 Forecasting skewness in the cross-section: Fama-MacBeth approach The sample period is from July 1962 to December 1998 and includes only those firms with market capitalization above the 20th percentile breakpoint of NYSE.614) 0. 1 of Table 2.312) 0.083 0.053 0.129 0.058 0.707) À3.148 À0.738 0.032 (4.

645) 0.549) 0. the 1980s. In Panel B. Given that .133 (2. Chen et al.357) 0.017 (0.323) 0.014) 20 1980s 0.335 (1.104 (1.532 over the full sample period – reaches a low of 0.592 in the 1970s.887) 0.807) 0.098) 0.044 (0.967) 66 Late 1960s 0.132 (2.663) 0.326) 0.497 (2. Moreover.492) 17 369 RETt RETtÀ1 RETtÀ2 RETtÀ3 RETtÀ4 RETtÀ5 No.5.222 (3.085 (1.909) 0.592 (2.684) 0. 4.056 (1.524) 0.081 (2. we show time-averages of all the regression coefficients for the full sample and for each of four decade-based subperiods: the 1960s.193) 20 1990s 0.057 (0.091 (1.099 (3. / Journal of Financial Economics 61 (2001) 345–381 Table 5 (continued) All periods DTURNOVERt 0.341) 0.390) 0.229 (3.372) 0. the 1970s.427) 0.139 (2.555) 0. and compute the associated t-statistics based on the time-series properties of the coefficients (and adjusting for serial correlation).917) 0.139 (4. The overriding conclusion that emerges from Table 5 is that our results are remarkably stable over time.071 (1.387) 0.242 (2.036 (0.767) 9 1970s 0.452) 0. We then take simple time-averages of the cross-sectional regression coefficients over various subperiods.136 (1.532 (3.565 (2.486 in the 1980s and a high of 0.614) 0.596) 0. the 66 six-month periods in our sample.280) 0. Why are small stocks more positively skewed? One of the most striking patterns that we have documented is that small stocks are more positively skewed than large stocks.082 (1.026 (0.132 (2.229 (1.234 (3. In Panel A of Table 5.453) 0.838) 0. even taken alone.981) 0.100 (1. and the 1990s.249 (6.045 (0. three of the four decade-based subperiods produce a statistically significant result for DTURNOVER.082 (2.513) 0. we display the coefficient on DTURNOVER from every one of the 66 regressions.J.711) 0.312) 0. of obs.139 (1.486 (1.287) 0. the coefficient on DTURNOVER – which averages 0. For example.

. Thus it may be asking too much to expect that the results here will be strongly statistically significant in their own right.) The coefficients on LOGCOVER have the predicted positive sign. Specifically. Chen et al. if there is more scope for such managerial discretion in small firms – perhaps because they face less scrutiny from security analysts – then the positive-skewness effect will be more pronounced in small stocks. this idea is outside the main scope of the paper. and prefer to announce good news immediately. while allowing bad news to dribble out slowly. 5. Forecasting market skewness We now turn to forecasting skewness in the returns to the aggregate market. we at least have a plausible explanation for what would otherwise be a puzzling feature of our data. The one satisfying thing about this after-the-fact hypothesis is that it yields new testable predictions. As described above. positive skewness ought to be greater in firms with fewer analysts. 7 After developing the discretionary-disclosure hypothesis. Nevertheless. he also finds that firms with fewer analysts have more positively skewed returns. (The sample period in Table 6 is shorter.370 J. we became aware of a closely related working paper by Damodaran (1987). and that as a result. our statistical power is severely limited. / Journal of Financial Economics 61 (2001) 345–381 this pattern is not clearly predicted by any existing theories (of which we are aware) we have had to come up with a new hypothesis after the fact in order to rationalize it. Moreover. one might more reasonably hope that they look qualitatively similar to those from the cross-sectional regressions. but remain significant as well. and are strongly statistically significant. Table 6 investigates this prediction. this hypothesis begins with the assumption that managers have some discretion over the disclosure of information. rather. and pursuing it more seriously would take us too far afield. taking our baseline specifications for both NCSKEW and DUVOL. We do not mean to cast Table 6 as a definitive test of the discretionary-disclosure hypothesis.7 The coefficients on LOGSIZE go down a bit. it is comforting to know that the most obvious auxiliary prediction of the hypothesis is borne out in the data. This behavior tends to impart a degree of positive skewness to returns. and running the regressions in Table 6. Using data from 1979 to 1983. While this is in many ways the more interesting exercise from an economic viewpoint. after controlling for size. since analyst coverage is not available from I/B/E/S prior to December 1976. and in each case adding LOGCOVER. the log of one plus the number of analysts covering the stock. Nothing else changes.

RETtyRETtÀ5 are returns in the six-month periods t through t À 5: All regressions also contain dummies for each time period (not shown). of obs.945) À4. 1 is NCSKEWtþ1 .992) 0.140) 0.135 (12.025 Our definition of the aggregate market is the value-weighted NYSE-AMEX index. 2 it is DUVOLtþ1 .767) 0. and all returns are excess returns relative to T-bills.012 (1. LOGCOVERt is the log of one plus the number of analysts covering the stock at the end of period t.705) 40.090 (13.586) 0.219 (8.290) 0.288) 0. t-statistics are adjusted for heteroskedasticity and serial correlation.032 (7. To avoid any temptation to further mine the data. DTURNOVERt is average monthly turnover in the six-month period t.665) 0.599) 0.649) À3.188 (À4.253) 0.683) 0. LOGSIZEt is the log of market capitalization at the end of period t.681) 0. Chen et al.044) 0.504) 0.085 (3.564) 0. detrended by a moving average of turnover in the prior 18 months.049 (6. (1) Using NCSKEWtþ1 measure NCSKEWt (DUVOLt in col. The dependent variable in col.493) 40.2) SIGMAt LOGSIZEt DTURNOVERt LOGCOVERt RETt RETtÀ1 RETtÀ2 RETtÀ3 RETtÀ4 RETtÀ5 No.011 (6.688 0.040 (4.100 (4.410) 0.059) 0.006 (0.065 (3.504 (3.648) 0.006 (0.006 (3. / Journal of Financial Economics 61 (2001) 345–381 Table 6 Forecasting skewness in the cross-section: adding analyst coverage 371 The sample period is from December 1976 to December 1998 and includes only those firms with market capitalization above the 20th percentile breakpoint of the NYSE. and in col.051 0.022 (À11. the negative coefficient of skewness in the six-month period t þ 1.J.019 (4.366) 0.055 (2. we use essentially the same specification as . the log of the ratio of down-day to up-day standard deviation in the six-month period t þ 1: SIGMAt is the standard deviation of (daily) returns in the six-month period t. R 2 (2) Using DUVOLtþ1 measure 0.010 (1.688 0.058 (2.207 (3.

The six past-return terms are always positive. Chen et al. The one drawback with these data is that we cannot use them to literally calculate turnover. In light of this power problem. since they give only the number of shares traded and not the number of shares outstanding. respectively. it shares the common element that lagged skewness helps to forecast future skewness. except for LOGSIZE and the time dummies. and again run the regression with and without October 1987. Thus we cannot 8 Harvey and Siddique (1999) build an autoregressive conditional skewness model for aggregatemarket returns. and run the regressions with and without October 1987. the point estimates suggest large quantitative effects relative to the cross-sectional regressions. We lose 30 observations because we do not allow any observation on NCSKEW. we also re-run our regressions omitting October 1987. The basic story is the same in all four columns. a new concern that arises with the time-series approach is the extent to which our inferences are dominated by the enormous daily movements during October 1987. The DTURNOVER variable is constructed exactly as before. the DTURNOVER variable in any given month draws on 24 months’ worth of data. To address this concern. Because of the detrending. (1992). while their specification is very different from that here. SIGMA. the coefficient on DTURNOVER.) This yields a total of 401 observations that can be used in the regressions. In an effort to get the most out of the little time-series data that we have. while positive in each of the four regressions. Dropping October 1987 seems to increase the precision of the DTURNOVER coefficient estimate somewhat. the coefficients on DTURNOVER and the RET terms are now on the order of ten times bigger than they were in the previous tables. This brings us down to 371 observations. / Journal of Financial Economics 61 (2001) 345–381 in our baseline cross-sectional analysis.8 Specifically. DUVOL. who in turn take it from the S&P Security Price Record. but the highest t-statistic across the four specifications is only 1. we now use monthly overlapping observations. In contrast. In columns 1 and 2 we use the NCSKEW measure of skewness. (The t-statistics we report are adjusted accordingly. we obtained an alternative series on NYSE volume going back to 1928 from Gallant et al.15. Thus both turnover and past returns could well be very important for forecasting the skewness of market returns. Indeed. In columns 3 and 4 we use the DUVOL measure of skewness. Nevertheless.9 The results are summarized in Table 7.372 J. holding statistical significance aside. we use all the same righthand-side variables. 9 . by detrending TURNOVER with its own moving average over the prior 18 months. or DTURNOVER to enter the regression if it draws on data from October 1987. is never statistically significant. However. and many are individually statistically significant. but we lack the statistical power to assert these conclusions – particularly regarding turnover – with much confidence.

217 (0. the log of the ratio of down-day to up-day standard deviation in the six-month period t þ 1: SIGMAt is the standard deviation of (daily) market returns in the six-month period t: DTURNOVERt is the average monthly turnover of the market portfolio in the six-month period t. excluding 10/87 0.708 (0.184 (3.647 (4.596 (1.443 (0.855) 18. the negative coefficient of skewness in the six-month period t þ 1. variable is NCSKEWtþ1 .589 (0.147) 1.221 (1. variable is DUVOLtþ1 . 3 and 4 it is DUVOLtþ1 .842) 1. t-statistics.844) À3. of obs.926 (1.3 and 4) SIGMAt DTURNOVERt RETt RETtÀ1 RETtÀ2 RETtÀ3 RETtÀ4 RETtÀ5 No.148) 1.386 (1. R 2 (2) Dep.749) 9.930) 371 0. The dependent variable in cols.274 0.357) 0.156) 6.809 (4. (1) Dep.968) 0.326) 401 0.475 (1.406) 1. 1 and 2 is NCSKEWtþ1 .123 (1.602) 1.100 (0.484 (4. variable is NCSKEWtþ1 NCSKEWt (DUVOLt in col.265 quite reproduce our baseline specifications for the longer post-1928 sample period. and in cols.554 (1.939) 0.680 (1.481) 0.828) 1.017 (0.325) 0.196 (0.704) 1. are adjusted for heteroskedasticity and serial correlation.283 (2.462 (1.349 (0.686 (2.242) 0.898) 0. Chen et al. which are in parentheses.574 (À0. RETtyRETtÀ5 are returns in the six-month periods t through t À 5.332 (1.049) 0.187 (2.264 (3) Dep.086) 1.753) 371 0.137) 6. detrended by a moving average of turnover in the prior 18 months.398) 0. Nevertheless.287 (0.922) 0.300) 9.232) 13.473 (2.264) 1.304 (4) Dep.470 (1. .061) 0. we get results for this sample period that are very similar to those reported in Table 7.J. excluding 10/87 0.726) 0. where the market is defined as the value-weighted portfolio of all NYSE/AMEX stocks.288) 401 0.168) 0. variable is DUVOLtþ1 0.126 (0.077 (2.324 (0.183 (1. using detrended values of raw trading volume to approximate detrended turnover.262) 2.575) 0. / Journal of Financial Economics 61 (2001) 345–381 Table 7 Forecasting skewness in the market: time-series regressions 373 The sample period is from July 1962 to December 1998 and is based on market returns in excess of the risk-free rate.734) 0.585 (3.002 (0.482 (1.

Economic significance of the results: an option-pricing metric Thus far. and have not really asked whether they imply magnitudes that are economically meaningful. Vt is the current variance. we have focused on the statistical significance of our results. Chen et al. with a correlation coefficient of r: The parameter r is the one of central interest for our purposes. k is the mean reversion parameter for the variance process.. The idea behind our metric can be understood as follows. If we are willing to fix all the other parameters besides r. The two Wiener processes dz1 and dz2 are instantaneously correlated. 1=2 ð3Þ dz2 . How much does the left-hand-side variable – NCSKEW or DUVOL – move? What makes things difficult here is that most people have little sense for what would constitute an economically interesting change in NCSKEW or DUVOL. In order to map the parameters of the option-pricing model into our NCSKEW variable. these .e. we need to (1) find an optionpricing model that admits skewness in returns and (2) create a mapping from the parameters of this model to our NCSKEW variable. / Journal of Financial Economics 61 (2001) 345–381 6. ð4Þ dVt ¼ kðV0 À Vt Þdt þ ZVt where pt is the log of the stock price. To help frame things in a way that is hopefully more intuitive. The model we use is the stochastic-volatility model of Das and Sundaram (1999). we can translate statements about NCSKEW into statements about the prices of outof-the-money put options. As a result. Now the stock experiences a surge in turnover. as it governs the skewness of stock returns: when r=0. when ro0. Given this new forecast of NCSKEW – but holding volatility fixed – by how much does the value of the put option increase? To answer this sort of question precisely. you revise your forecast of NCSKEW. V0 is the long-run mean level of variance. and Z is the volatility of the variance process. Imagine that you are pricing an out-of-the-money put on a stock whose returns you initially believe to be symmetrically distributed – i. log returns are symmetrically distributed. Assessing economic significance in the current context is a bit tricky. log returns are negatively skewed. a stock for which you believe that NCSKEW is equal to zero. in which the dynamics of stock prices are summarized by the following two diffusion equations: dpt ¼ a dt þ Vt 1=2 dz1 .374 J. a is the expected return on the stock. we draw on formulas given by Das and Sundaram that express the skewness in daily log returns as a function of the diffusion parameters. DTURNOVER – is shocked by two standard deviations. The thought experiment that is typically undertaken is something like this: suppose that the right-hand-side variable of interest – in this case. using the DTURNOVER coefficient estimate from our regressions.

223) this value of skewness in daily returns for firm 2 can be shown to imply r=À0.437=0. Panel B then compares the prices of six-month European puts across two regimes. which has symmetrically distributed returns – i.86 in regime 1 but 1. Once we have obtained the implied value of r in this way. and the riskless rate r=0.005 (see Table 1A). except that we now set V0 ¼ Vt ¼ 0:04.) We also set the stock price P=100. the firm 1 put has a Black-Scholes (1973) implied volatility of 40.36% and 21. Panel A of Table 8 displays the impact of this change in r for the prices of six-month European put options.07 in regime 2.20 for firm 1 but 1.66%. a two-standard-deviation shock to DTURNOVER translates into a movement of 0. the standard deviation of DTURNOVER is 0. Or expressed in a different way. where the parameters are chosen so as to be reasonable for individual stocks: k=1. corresponding to an annual standard deviation of returns of 20%. and from Table 2. This is equivalent to a value of r=0. For example. The standard deviation of DTURNOVER (for firms above the 20th percentile NYSE breakpoint) is 0. column 1. it calculates put prices for both firm 1 (which has NCKSEW=0 and thus r=0) and firm 2 (which has NCSKEW=0.06 for NCSKEW is equivalent to r=À0. Using the coefficient estimate on DTURNOVER of 6. we take firm 2. Vt =0.e. Next. We keep all the diffusion parameters the same as in Panel A.042. Using Eq. Panel B undertakes a similar experiment to gauge the significance of our time-series results. an increase of 24. an increase of 20.38. p. Chen et al. except that it has a two-standard-deviation higher value of DTURNOVER. a put with a strike price of 85 is worth 0.16 corresponds to an annual standard deviation of returns of 40%. the differences appear to be meaningful. Once again. For the market as a whole.037). and Z ¼ 0:4: (Setting the variance V to 0.84%. which is identical to firm 1. the impact on put prices is substantial.037 and thus r=À0. assuming all the other diffusion parameters stay fixed.J. the coefficient on DTURNOVER is 0. Table 8 illustrates the results of this exercise. respectively. That is.037 (2 Â 0. V0 =0. the first with r=0 and the second with r=À0. Consider first Panel A.33.33.14%.16. we can calculate options prices and thereby see the impact of a given value of NCSKEW.00 from Table 7.16.38).042 Â 0. a put with a strike of 70 is worth 1.33%. Given the other diffusion parameters.06 in the NCSKEW variable.44 for firm 2. particularly if one goes relatively far out-of-the-money. this value of 0.. Hence the value of NCSKEW for firm 2 is 0. while the firm 2 put has an implied volatility of 42. The corresponding implied volatilities are 20. These results reinforce a point made above: while the time-series estimates are . We begin with a hypothetical firm 1. As can be seen. (21) in Das and Sundaram (1999. For example. it has NCSKEW=0.437.50%. / Journal of Financial Economics 61 (2001) 345–381 375 formulas allow us to ask to what value of r a given value of NCSKEW corresponds. column 1.

Firm 1 is assumed to have a value of r=0. current variance V=0.88% 32.80% 17.1.79% 6. Percent increase in put price: Firm 2 vs.77% À1.438 42.10% À0.525 37. Firm 2 is assumed to have a value of r=À0.082 39.16.044 39.011 38. Firm 2: r=À0.376 J.325 39.38.38 Six-month European put price Black-Scholes implied vol.419 40.044 39.287 39. Panel A: Options on individual stocks The benchmark parameters are as follows: stock price P=100.67% 33.14% 3. col. and volatility of variance Z=0.93% 80 90 100 110 120 130 .03% 2.50% 11.50% 20. These values of r imply that the difference in daily skewness between Firms 1 and 2 is equivalent to that created by a two-standard-deviation move in the DTURNOVER variable.33% 3.4.282 37. / Journal of Financial Economics 61 (2001) 345–381 Table 8 Economic significance of trading volume for skewness in stock returns: an option-pricing metric Using the stochastic volatility option pricing model (and notation) of Das and Sundaram (1999) we consider what a two-standard-deviation shock in detrended trading volume implies for the prices of six-month European options.09% 10.16.297 41.30% 6.34% À1.748 39. using our baseline firm-level sample and coefficient estimates from Table 2. Firm1 1.57% 1.35% À1.81% 24.41% 17. Chen et al. interest rate r=0.16% 8.48% 24. mean reversion in variance k=1. 70 Firm 1: r=0 Six-month European put price Black-Scholes implied vol. annualized long-run variance V0=0.197 40.994 39.

21% À2.68% 12.912 20.39% 5.298 20. / Journal of Financial Economics 61 (2001) 345–381 377 . These values of r imply that the difference in daily skewness between Regimes 1 and 2 is equivalent to that created by a two-standarddeviation move in the market DTURNOVER variable.258 19. annualized long run variance V0=0. Chen et al.49% 16. Regime 1 1.07% 12.91% 8. 1.36% 1.33.84% 24.755 18.134 18.76% À0.66% 1.33 Six-month European put price Black-Scholes implied vol.289 18.070 21.330 18.04.04. using our time-series estimates from Table 7. mean reversion in variance k ¼ 1. 85 Regime 1: r=0 Six-month European put price Black-Scholes implied vol.61% 3.955 18.77% 8.121 19. current variance V=0.04% 90 95 100 105 110 115 J. interest rate r=0.91% 3. col.859 20.367 19.80% 15.093 19.10% 0.Panel B: Options on the market portfolio The benchmark parameters are as follows: stock price P=100.09% 5.79% 11. and volatility of variance Z ¼ 0:4: Regime 1 is assumed to have a value of r ¼ 0: Regime 2 is assumed to have a value of r=À0.01% À2.693 19.63% 4. Regime 2: r=À0.10% À2. Percent increase in put price: Regime 2 vs.

978 to reflect the corresponding predicted change in NCSKEW. we have also done similar calculations to measure the economic significance of our results for past returns.73. no doubt others can be thought up. Although they are not shown in Table 8. we are left to conclude that. This caveat would seem to be particularly . regarding trading volume.86% increase. As it turns out. The first. In the case of the aggregate market. a 44. even if the RET variable has moved by only 7% in the last six months. Rather. While we have attempted to control for some of the most obvious alternative stories.087 in the NCSKEW variable. this does not mean that there are not other plausible interpretations. large movements in past returns give rise to conditional negative skewness so substantial that it cannot be adequately represented in terms of a pure diffusion process – one would instead need some type of mixed jump-diffusion model. 7. for the market as a whole. Clearly. Chen et al. For example. is the most novel. In the cross-section. the coefficients on past returns suggest effects on skewness that are so large that they cannot generally be captured in the context of a pure diffusion model like that of Das and Sundaram (1999). Conclusions Three robust findings about conditional skewness emerge from our analysis of individual stocks. (2) have had positive returns over the prior 36 months. this means that we cannot even capture a onestandard-deviation shock to six-month returns without violating the constraint that the absolute value of r not exceed one. and (3) are larger in terms of market capitalization. which causes the put with the 70 strike to rise in value from 1. negative skewness is greater in stocks that (1) have experienced an increase in trading volume relative to trend over the prior six months. the quantitative influence of past returns on skewness is stronger than that of trading volume. though the statistical power of our tests in this case (especially with respect to trading volume) is quite limited. our results here are supportive of the theory. At the same time. they are no less indicative of important economic effects. This in turn is equivalent to r going from zero to À0. one has to adjust r from zero to À0. Let us try to put each of these findings into some perspective.920. a shock of 40% to the RET variable in the most recent six-month period (note from Table 1 that 40% is approximately a two-standard-deviation shock for a firm in the next-to-largest quintile) translates into a movement of 0. The first two results also have direct analogs in the time-series behavior of the aggregate market. Given that the standard deviation of RET for the market is about 11%. and is the one we were looking for based on a specific theoretical prediction from the model of Hong and Stein (1999). With individual stocks.378 J.20 to 1. / Journal of Financial Economics 61 (2001) 345–381 statistically much weaker than those from the cross-section.

. With respect to the third finding – that small-cap stocks are more positively skewed than large-cap stocks – we are not even aware of an existing theory that provides a simple explanation. controlling for size. This discretionary-disclosure hypothesis in turn yields the further prediction that. we have developed an informal hypothesis after the fact. these variables are included in our regressions because prior work (Harvey and Siddique. the more modest statement to be made is that previous research has not examined the implications of bubble models for conditional skewness. we have found it helpful to think about it in terms of models of stochastic bubbles. we would stop well short of claiming to have strong evidence in favor of the existence of bubbles. Indeed. and (2) managers of small companies have more scope for hiding bad news from the market in this way. References Almazan. A fair criticism of this whole line of discussion is that we have three main empirical results. Having verified the importance of past returns. stochastic bubbles to explain the effect of past returns. and it serves to further underscore our previous caveat about the extent to which one should at this point consider any single theory to be strongly supported by the data. / Journal of Financial Economics 61 (2001) 345–381 379 relevant given that there has been so little research to date on conditional skewness at the individual stock level. Rather. e.. D.. Flood and Hodrick. and discretionary disclosure to explain the effect of size. while dribbling bad news out slowly.. and a different theoretical story for each: the Hong-Stein (1999) model to explain the effect of turnover on skewness. having to do with the effects of past returns and size on skewness. The second and third findings.A. A natural challenge for future work in this area is to come up with a parsimonious model that captures these three patterns in a more integrated fashion. do not speak directly to predictions made by the HongStein model. and we want to be careful to isolate the effects of trading volume from other factors. the bubble models look pretty good. B. M.g. . A. Rather. positive skewness ought to be more pronounced in stocks with fewer analysts – a prediction which is clearly supported in the data. there is a large body of research from the 1980s (see. However. Instead.C. 1999.. such as that developed by Blanchard and Watson (1982). Chen et al. University of Texas. Working paper. 2000) suggests that they might enter significantly. based on the ideas that (1) managers prefer to disclose good news right away. Brown. This lack of unity is unsatisfying. and that on this one score. Chapman.J. Carlson. Why constrain your mutual fund manager?. 1988. 1990) that focuses on a very different set of implications of bubble models – having to do with the relation between prices and measures of fundamentals such as dividends – and tends to reach mostly skeptical conclusions on this question. reviews by West. Austin.

1–42. G. Hentschel. Summers. The pricing of options and corporate liabilities. 177–181.. Braun. Working paper. 1976. Tauchen. P.F.D. Shleifer. Journal of Finance 53.R.F. 1973. Journal of Political Economy 81..W. and betas... Z.R. 1997. J. A.. R. 1982. Post-’87 crash fears in S&P 500 futures options. 2000.. 1992. 1999. Information bias: measures and implications. French.. Crises in Economic and Financial Structure. B. Working paper.F.. R.B. H. Black. 999–1021. 1997. Cao. Blanchard. 1779–1801. and conditional patterns in security returns. 1993. E. Journal of Finance 46. Coval. D. Nelson. Market liquidity..380 J. 1998. Trading-generated news. Proceedings of the 1976 Meetings of the American Statistical Association. Campbell. 1995.F. 295–315. Studies of stock price volatility changes. Risk. Bates. Wu. No news is good news: an asymmetric model of changing volatility in stock returns. 1995. Flood. F..R. V. and behavioral finance. Good news. A. Christie. P. M.. 407–432. Journal of Economic Perspectives 4. Stock returns and volatility: a firm-level analysis. In: Wachtel. Asymmetric volatility and risk in equity markets. 1993. J. Scholes. P. 85–101.. Lexington.. 2059–2106. Journal of Finance 48... Runkle. New York University. Noise trader risk in financial markets. F.J. G. O.. Sunier. Damodaran. A. 211–239. R. Journal of Finance 52. 2003–2049. A. Bubbles. Journal of Political Economy 81. Duffee. The stochastic behavior of common stock variances – value.. Whaley... MA. D. L. On testing for speculative bubbles. Fama. pp. leverage and interest rate effects. Waldmann. G. 637–659.. rational expectations. sidelined investors.. G. Measuring and testing the impact of news on volatility. Hodrick. 283–306.. K.. NBER working paper 5894. Engle. J.. Business and Economical Statistics Section. L. American Economic Review 80. R. 1998. Ng. 1575–1603. R.... Stambaugh. R. S. 399–420.. 1987.. C. Fleming. return and equilibrium: empirical tests. Journal of Financial Economics 19. Empirical performance of alternative option pricing models. Hirshleifer. Schwert. 1973. Gallant. Journal of Political Economy 98. J. J. Market efficiency. Leland...D.. Watson. / Journal of Financial Economics 61 (2001) 345–381 Bakshi. 1987. D. Expected stock returns and volatility. (Ed.A. Bekaert. Journal of Financial and Quantitative Analysis 34. G.M.J.. Chen et al. 1992. Journal of Financial Economics 49. D. Journal of Finance 48. E. 1009–1044. A. Journal of Financial Economics 10. Das. Black. Journal of Finance 50.. 703–738. Fama. Journal of Financial Economics 37. Lexington Books.B. Stock prices and volume... Review of Financial Studies 5. 1990.Y. Chen. D. On the relation between the expected value and the volatility of the nominal excess return on stocks. 607–636. L.. Dumas. Of smiles and smirks: a term structure perspective. R.E.). G. University of Michigan Business School.. 1991. The crash of ’87: was it expected? The evidence from options markets. long-term returns. DeLong....A.J. .K. R. Journal of Financial Economics 31.. Ann Arbor. Rossi. 1998.. G. 3–29... 1749–1778. M. MacBeth.P. 199–242. 1990.. Bates. Jagannathan. 1982. and financial markets.. 1990.H. volatility..E.. Review of Financial Studies 13. bad news.E. Implied volatility functions: empirical tests.W. Sundaram. Genotte.R. 281–318. hedging and crashes. Glosten.K.

Odean. 1989.. Tempe. 1998a. Journal of Finance 55. Pfleiderer. Hong. K.W. The persistence of volatility and stock market fluctuations. A. Bad news travels slowly: size. Koski. Hong. Why does stock market volatility change over time? Journal of Finance 44. H. 35–63. Odean. 1995....R. 65–295..J. 1999.R.. in preparation.. Smidt. 1985. 1992. Conditional heteroskedasticity in asset returns: a new approach.M.. M. .. inflation. (Ed. Boston. Poterba.. Paul. Lim. Are investors reluctant to realize their losses? Journal of Finance 53. C. M.. Grossman. The determinants of asymmetric volatility.. Journal of Finance 40. Shefrin. D. 1989. 2000.. Stein. Review of Financial Studies 5. Journal of Finance 54. An analysis of the implications for stock and futures price volatility of program trading and dynamic hedging strategies. and the stock market. rational arbitrage and market crashes. 1986. Volume.. Pindyck. A... Harvey. Econometric Analysis. Rational asset-price movements without news.. Differential interpretation of public signals and trade in speculative markets. 1115–1153.. 1993. J. W. C. analyst coverage and the profitability of momentum strategies. A. Risk.S. Statman. 1984. Conditional skewness in asset pricing tests. Siddique. J.. 473–506. Nelson. American Economic Review 83. 1999. Econometrica 59. 639–656. Wu.. American Economic Review 74. Underestimation of portfolio insurance and the crash of October 1987. 3–37. 1988. J.C. Lemmon. Journal of Business 61. 1142–1151. 1988.. Arizona State University. S. Pontiff. J.). Evidence and Implications: Proceedings of the 11th Annual Economic Policy Conference of the Federal Reserve Bank of St. T.L. How are derivatives used? Evidence from the mutual fund industry. Does book-to-market equity proxy for distress risk or overreaction? Working paper. H. Journal of Political Economy 103. 2000.H. Review of Financial Studies.. Journal of Finance 41. price and profit when all traders are above average. Journal of Financial and Quantitative Analysis 34. T.D. Harris. 347–370. Journal of Finance 55. Varian. The disposition to sell winners too early and ride losers too long: theory and evidence. 1993. 1887–1934.M. 275–298. 1998b. 777–790.W. Griffin. 1991. 1112–1130. 1986.. 1999. G. J. New York. Harvey. Kluwer Academic Publishers. 1263–1295. Kandel. Differences of opinion make a horse race.. A. Bubbles. M.D.. Differences of opinion. C. Chen et al.J. Siddique. Autoregressive conditional skewness. Differences of opinion in financial markets. West. 1993... R. AZ.H.R.. 334–351.. Review of Financial Studies.. Louis... Summers.. Pearson.J. 2000. H. C. 1999. J. J. Macmillan. NBER Working paper. 465–487. Stein. Kleidon. L. American Economic Review 76. C. Financial Risk: Theory. D.. S. Jacklin. Romer. 1775–1798. In: Stone. 951–974. Journal of Finance 43... G. pp. Lakonishok. E.. H. T. Journal of Finance 53. 791–816. Volume for winners and losers: taxation and other motives for stock trading. Raviv. 831–872. / Journal of Financial Economics 61 (2001) 345–381 381 Greene. Schwert. fads and stock price volatility tests: a partial evaluation.L. volatility. N.

You're Reading a Free Preview

Download
scribd
/*********** DO NOT ALTER ANYTHING BELOW THIS LINE ! ************/ var s_code=s.t();if(s_code)document.write(s_code)//-->