Short-term Momentum∗

Mamdouh Medhat Maik Schmeling

October 30, 2019

∗
First version: March 27, 2018. We thank Cliff Asness, Pedro Barroso, John Campbell, Giovanni Cespa,
Zhi Da, Alexander Hillert, Alexandre Jeanneret, Christian Julliard, Tim Krönke, Albert Menkveld, Markus
Nöth, Richard Payne, Lasse Pedersen, Chris Polk, Angelo Ranaldo, Ioanid Roşu, Lucio Sarno, Julian Thimme,
Gyuri Venter, and Michela Verardo. We also thank seminar participants at the Research in Behavioral Fi-
nance Conference, London Empirical Asset Pricing Workshop, Annual Meeting of the Swiss Society for
Financial Market Research, Annual Meeting of the French Finance Association, London Business School
Summer Symposium, and at various universities for their helpful comments. We also thank Chicago Quanti-
tative Alliance (CQA) for the 1st place award at their 2018 Academic Competition. All errors are our own.
Medhat: Cass Business School, City, University of London. Email: [email protected].
Webpage. Schmeling: Goethe University Frankfurt and Centre for Economic Policy Research (CEPR), Lon-
don. Email: [email protected]. Webpage.

Electronic copy available at: https://ssrn.com/abstract=3150525

 Short-term Momentum

Abstract
We document a striking pattern in the cross section of U.S. and international stock
returns: double sorting on the previous month’s return and share turnover results in
significant short-term reversal among low-turnover stocks whereas high-turnover stocks
exhibit short-term momentum. Short-term momentum is as profitable and persistent
as conventional momentum, is exclusively a large-cap phenomenon, and survives trans-
action costs. Using a tractable yet flexible model, we illustrate how our results are
difficult to reconcile with rational expectations but are in line with inefficient use of
the information conveyed by prices. We derive novel predictions about the effects of
liquidity trading and the link to fundamentals for which we find supporting evidence.

JEL Classification: G12, G14

Keywords: Momentum, Reversal, Cursed Expectations Equilibrium, Return Predictability

Electronic copy available at: https://ssrn.com/abstract=3150525

 1. Introduction

A key stylized fact in the asset pricing literature is that stock returns exhibit reversal at
short horizons of one month (Jegadeesh, 1990) but continuation—or momentum—at longer
horizons between two and twelve months (Jegadeesh and Titman, 1993, 2001). In this paper,
we show that reversal and momentum coexist with almost equal strengths at the one-month
horizon. While last month’s thinly-traded stocks exhibit a strong short-term reversal effect,
last month’s heavily-traded stocks exhibit an almost equally strong continuation effect which
we dub short-term momentum. Figure 1 illustrates our main results for the U.S.

Coexistence of reversal and momentum in one-month returns

20

16.44
Average excess return (%, annualized)

10
0

-3.39
-10

-16.92
-20

Short-term reversal* Short-term reversal Short-term momentum

Figure 1: Coexistence of reversal and momentum in one-month returns. This figure shows average
excess returns to three zero-cost long-short strategies that buy the previous month’s winners and short the
previous month’s losers among U.S.stocks: a conventional short-term reversal strategy as in Jegadeesh (1990)
based on the full cross section of stocks (middle bar); a short-term reversal* strategy that only takes positions
in stocks with low turnover in the previous month (left bar); and a short-term momentum strategy that only
takes positions in stocks with high turnover in the previous month (right bar). The strategies are based on
portfolios from univariate- or double sorts into deciles using NYSE breakpoints and are value-weighted and
rebalanced at the end of each month. The performance of the portfolios underlying the short-term reversal*
and short-term momentum strategies is provided in Table I. The sample comprises all common non-financial
shares on the NYSE/Amex/Nasdaq exchanges and covers the period from July 1963 to December 2018.

Electronic copy available at: https://ssrn.com/abstract=3150525

 To obtain our main empirical results, we double sort stocks on last month’s return and
last month’s share turnover (trading volume divided by shares outstanding) using NYSE
breakpoints and form value-weighted portfolios. We find, first, that there is strong short-
term reversal among low-turnover stocks. The short-term reversal* strategy—which buys
last month’s winners and shorts last month’s loser among low-turnover stocks—generates
a negative and significant average return of −16.9% per annum (the left bar in Figure 1).
Second, our key result is that this reversal effect is completely overturned for high-turnover
stocks. The short-term momentum strategy—which buys last month’s winners and shorts
last month’s losers among high-turnover stocks—generates a positive and significant average
return of +16.4% per year (the right bar in Figure 1). We show that neither strategy
is explained by standard factors, including conventional momentum and reversal factors.
We also show that short-term momentum persists for 12 months after formation; that it
is concentrated among the largest, most liquid stocks; and that it survives conservative
estimates of transaction costs. Finally, we show that our main results are not limited to the
U.S. but also hold in an international sample of 22 developed markets.
To explain these results, we turn to equilibrium models of speculative trading that allow
for deviations from rational expectations equilibrium (REE), i.e., that allow for inefficient use
of price-inferred information.1 This class of models has been invoked to explain why trading
volume greatly exceeds what is expected under REE and why patterns in returns and volume
are tightly linked (e.g., Hong and Stein, 2007; French, 2008).2 A tractable setup that allows
for such deviations yet nests REE as a special case is the ‘cursed expectations equilibrium’
(CEE) model due to Eyster, Rabin, and Vayanos (2019). Under CEE, some traders neglect
the informational content of prices and do not fully invert them to uncover others’ private
1
Under REE, traders efficiently condition on prices to estimate each others’ private information, implying
that any speculative trading is due to information asymmetries among traders that persist in equilibrium.
See Grossman (1976), Hellwig (1980), Grossman and Stiglitz (1980), Diamond and Verrecchia (1981), and
Admati (1985) for REE, and see, e.g., Milgrom and Stokey (1982) and Tirole (1982) for trading under REE.
2
A common theme in deviations from REE is that some traders do not efficiently condition on prices. This
causes informationally inefficient prices and increases speculative trading volume relative to REE. Deviations
from REE include noncommon priors (Harrison and Kreps, 1978); traders that are only able to process a
subset of the available information (Hong and Stein, 1999, 2007); traders driven by sentiment (Barberis,
Shleifer, and Vishney, 1998); traders that exhibit certain cognitive biases such as overconfidence (Odean,
1998; Daniel, Hirshleifer, and Subrahmanyam, 1998, 2001; Scheinkman and Xiong, 2003) or limited attention
(Peng and Xiong, 2006; Corwin and Coughenour, 2008); or traders who are ‘dismissive’ and downplay the
precision of each others’ signals (Banerjee, Kaniel, and Kremer, 2009; Banerjee and Kremer, 2010; Banerjee,
2011). Although often behaviorally motivated, such devaitions are not necessarily irrational since traders
may not know how to correctly condition on prices or they may have noncommon priors.

Electronic copy available at: https://ssrn.com/abstract=3150525

information when choosing their demands.3 CEE is both intuitively appealing and backed
up by experimental evidence. However, alternative setups with related deviations from REE
should not change the qualitative conclusions we draw.4 We use the model to compare the
REE and CEE cases in terms of their ability to capture our empirical findings and we derive
additional testable predictions that allow us to distinguish between the two in the data.
We find that information asymmetry alone (the REE case) is insufficient to produce
continuation in returns accompanied by high volume, even at low degrees of nonspeculative
(i.e., liquidity related) trading. In contrast, even a modest degree of informational negligence
(the CEE case) produces high-volume continuation but low-volume reversal—consistent with
Figure 1—as long as the degree of nonspeculative trading is not too high. Importantly,
nonspeculative trading counteracts informational negligence. This has two key implications.
First, it explains why short-term momentum is concentrated among the largest, most liquid
stocks, since trading in smaller and less liquid stocks is to a greater extent driven by liquidity
needs and provision (Avramov, Chordia, and Goyal, 2006; Nagel, 2012). Second, it implies
a testable prediction: time periods with lower degrees of nonspeculative trading should be
associated with stronger short-term momentum but weaker short-term reversal*. To test
this, we exploit a seasonality in the degree of liquidity trading at the turn of the month
driven by institutional cash needs (Etula, Rinne, Suominen, and Vaittinen, 2019). We find
it is strongly borne out in the data, in that excluding the values for the last trading days
from the two sorting variables (which also introduces an implementation lag) leads to much
stronger short-term momentum and much weaker short-term reversal∗ .
The model produces additional testable predictions about the link to fundamentals. Con-
sider a regression of fundamental innovations on the current return and the interaction of the
current return and volume. If all traders are rational (the REE case), the interaction effect
is negative. In contrast, if some traders make inefficient use of price-inferred information
(the CEE case), the interaction effect is positive. Furthermore, in the CEE case, the direct
3
CEE is the financial market’s analog of Eyster and Rabin’s (2005) game theoretic ‘cursed equilibrium’ in
which players fail to fully infer others’ private information from their actions. Cursed equilibrium captures
the psychological principle underlying the winner’s curse in common-value auctions, i.e., the fact that the
average winning bid exceeds the average value of the object being auctioned. The winner underappreciates
that the lower bids placed by others convey that they are more pessimistic about the true value and, hence,
overbids. Fully rational bidders would avoid the winner’s curse by tempering their bids (Thaler, 1988).
4
Eyster, Rabin, and Vayanos show that informational negligence and dismissiveness can have similar
implications for return autocorrelation and expected volume, and, furthermore, that informational negligence
can act as an “enabling bias” that amplifies the effects of overconfidence and other biases on expected volume.

Electronic copy available at: https://ssrn.com/abstract=3150525

effect (of the current return) is negative if information asymmetry is severe, but is otherwise
positive. The data is strongly in favor of the CEE case. In predictive regressions of firms’
fundamental growth rates, the interaction of current one-month returns and turnover has
a positive and significant effect at the one-year horizon and the effect remains positive at
horizons up to five years. In addition, the direct effect (of the current one-month return) is
positive and significant at all horizons, ruling out severe information asymmetry. Coupled
with the fact that short-term momentum persists for 12 months, these results lend credence to
our view that high-turnover momentum is due to inefficient use of price-inferred information.
We finally provide a battery of robustness tests. The most important findings are as
follows. First, constructing short-term momentum with a control for variables related to
analysts’ forecasts leads to similar conclusions as our tests of the model’s predictions. Second,
short-term momentum is neither driven by post-earnings announcement drift nor by industry
momentum. Third, it is immune to the critique by Bandarchuk and Hilscher (2013), who
argue that “characteristic screens” mechanically lead to higher momentum returns. Lastly,
it exhibits far less crash risk than conventional momentum (Daniel and Moskowitz, 2016).

Related literature

An early theoretical literature studies how trading motives affect return autocorrelation.
Campbell, Grossman, and Wang (1993) show that higher volume is associated with stronger
reversal when trading is nonspeculative. This rules out nonspeculative trading as an ex-
planation for the short-term momentum effect we document. Wang (1994) and Llorente,
Michaely, Saar, and Wang (2002) consider speculative trading under REE and predict that
higher volume should be associated with weaker reversal or even continuation among stocks
with severe information asymmetry. While this is seemingly consistent with our short-term
momentum effect, we show that several other key empirical results in our paper are difficult
to reconcile with these models. Hence, we go beyond this earlier literature and allow for
speculative trading under deviations from REE based on the recently proposed CEE model,
and we compare the REE and CEE cases both theoretically and empirically.
In this sense, our paper is related to Banerjee (2011), which compares how ‘disagreement’
(i.e., heterogeneity in beliefs among traders) behaves under REE and under deviations from
REE based on ‘dismissiveness.’ Banerjee concludes that his evidence is broadly consistent

Electronic copy available at: https://ssrn.com/abstract=3150525

with traders conditioning on prices, although with substantial variation across stocks in the
degree to which they do so. Our paper adds to this by theoretically identifying the stocks
where deviations from REE are most likely to be pronounced and by empirically testing this
mechanism’s additional novel predictions on U.S. and international stocks.
Similarly, an early empirical literature studies the relation between volume and short-
horizon return autocorrelation in the cross section, but provides seemingly conflicting ev-
idence. For example, Conrad, Hameed, and Niden (1994) find that volume strengthens
reversal in weekly returns among Nasdaq stocks, whereas Cooper (1999) finds the opposite
result among the largest NYSE/Amex stocks. Our results reconcile these findings since we
show that short-term reversal* is strongest among smaller stocks while short-term momen-
tum is concentrated among larger ones. Llorente et al. (2002) employ ordinary least squares
(OLS) regressions of daily returns at the individual stock level and find that volume weak-
ens reversal among small/illiquid stocks but strengthens it among large/liquid ones. While
their findings seem at odds with ours, it is important to note that we employ value-weighted
portfolios based on NYSE breakpoints, a monthly frequency, and a large sample of U.S. and
international stocks. This renders our approach more comprehensive and less susceptible to
influence from microstructure issues such as bid-ask bounce and nonsynchronous trading.5
Moving away from stock markets, Gargano, Riddiough, and Sarno (2019) find stronger
reversal in daily currency returns among currencies with low volume on the previous day.
While they do not document momentum among currencies with high volume, their findings
on low-volume currency reversal are in line with our findings for U.S. and international stocks.

2. Theoretical results

This section presents our theoretical results. We compare the REE and CEE cases in terms
of their ability to capture our empirical findings and we derive additional testable predictions
that allow us to distinguish between the two in the data. To incorporate asymmetric infor-
mation in our analysis, we assume that only a fraction of traders possess private information.
5
Other related work includes Asness (1995), who double sorts on volume and the previous month’s return
but considers equal-weighted portfolios. He also finds stronger reversal among low-volume stocks but does
not document significant momentum among high-volume ones. Lee and Swaminathan (2000) double sort on
volume and past returns but consider longer formation periods of at least three months as well as equal-
weighted portfolios. Their focus is on how volume affects the link between the value and momentum effects,
and not on the coexistence of reversal and momentum in one-month returns.

Electronic copy available at: https://ssrn.com/abstract=3150525

2.1. Setup

There are two time periods and three dates, t = 0, 1, 2. The market features a riskless and a
risky asset. The riskless asset is in infinite supply and its one-period gross return is one. The
risky asset is in zero net supply. At t = 2, the risky asset pays a dividend, d = d + d + d,
e

where d is a constant while d ∼ N (0, Vd ) and de ∼ N (0, Vde) are independent. Traders can
learn about d but not about d.
e The price of the risky asset at date t is denoted by pt .

All trading is at t = 1. There is a continuum of traders with total mass 1, divided into two
groups: I, the group of relatively informed traders with mass ι ∈ (0, 1), and U , the group of
relatively uninformed traders with mass 1 − ι. Traders are identical within groups. At t = 1,
all traders observe a public signal, s = d + s , where s ∼ N (0, Vs ). In addition, I-traders
observe a private signal, δ = d + δ , where δ ∼ N (0, Vδ ). The public signal ensures that
U -traders’ information set is nonempty when they do not condition on p1 . It also ensures
that I-traders’ information set always nests that of U -traders.
Traders have zero initial endowment of the assets. I-traders receive a random endowment
of zd at t = 2, where z ∼ N (0, Vz ) is observed privately at t = 1.6 The shocks {d , d,
e s , δ , z}

are mutually independent. The endowment shock, z, is equivalent to a shock to the supply
of the risky asset and induces nonspeculative (i.e., liquidity or hedging) motives for trading.
This prevents a fully revealing REE. Its variance, Vz , is a measure of these motives.
Traders have negative exponential (CARA) utility over consumption at t = 2 with abso-
lute risk aversion α. I-traders choose their demand of the risky asset, xI ∈ R, by maximizing

E −e−α(xI (d−p1 )+zd)  {s, δ, z, p1 } .

  

For a given χ ∈ [0, 1], U -traders choose their demand, xU ∈ R, by maximizing

1−χ  −αx (d−p1 )  χ

E −e−αxU (d−p1 )  {s, p1 }
 
E −e U s ,

which is a weighted geometric average of their expected utility given {s, p1 } or just s. When
χ = 0, U -traders have fully rational expectations. When χ = 1, U -traders fully neglect the
informational content of p1 . Intermediate χ captures partial informational negligence.
6
The assumption that it is only the informed (I) traders that receive an endowment shock is standard in
models of asymmetric information. See, e.g., Vayanos and Wang (2012) and the discussion therein.

Electronic copy available at: https://ssrn.com/abstract=3150525

2.2. Equilibrium

Our setup implies that p2 = d and p0 = d, whereas p1 is determined in equilibrium by market

clearing, ιxI + (1 − ι)xU = 0. We show in Appendix IA.1.1 that our setup features a unique
linear equilibrium in which p1 is available in closed form. Let VbdI be I-traders’ posterior
variance of d . Similarly, let VbdU,s,p and VbdU,s be U -traders’ posterior variances of d in the
fully rational and fully negligent cases, and set VdU = (1 − χ)VbdU,s,p + χVbdU,s . Then

 
p1 = d + κVbdI + (1 − κ)VdU Vs−1 s + κVbdI Vδ−1 δ + (1 − κ)(1 − χ)VbdU,s,p Vp−1 pe − αVd ιz . (1)
| {z } | {z } | {z }
Public information Private and price-inferred information Risk

I
Vde +Vbd
Here, pe = δ − α Vb I Vδ z is the signal inferred from the price and Vp is the variance of
d

its error; Vd is the posterior variance of d given the information set of the (hypothetical)
‘average’ trader; and κ = ιVd (Vde + VbdI )−1 ∈ (0, 1) is the weight on I-traders’ posteriors. The
price reflects the ‘average’ trader’s dividend estimate minus the risk premium he demands for
holding the asset. Under REE (χ = 0), the price optimally weights pe by its relative quality.
When χ = 1, the price gives zero weight to pe. More generally, we show in Appendix IA.1.1
that whenever χ > 0 the price ‘underreacts’ to pe by giving it too little weight relative to REE.
Underreaction induces positive autocorrelation in returns. We show in Appendix IA.1.2 that
returns are always negatively autocorrelated under REE, but are positively autocorrelated
when χ > 0 given sufficiently low Vz .

2.3. Predicting returns with current returns and volume

We are interested in predicting the second-period return, r2 = p2 − p1 , given the first-period

return, r1 = p1 − p0 , and the first-period trading volume, v1 . Trading volume is the average
across traders of the absolute changes in demands. Since traders have zero initial endowment,

v1 = ι |xI | + (1 − ι) |xU | = 2ι |xI | . (2)

Figure 2 illustrates how expected volume and return autocorrelation vary with informa-
tional negligence (χ), nonspeculative trading motives (Vz ), and information asymmetry. We
Vd
increase information asymmetry by decreasing the public-signal quality, λs = Vd +Vs
, for a

Electronic copy available at: https://ssrn.com/abstract=3150525

 A Varying χ for V z = 0.52 B Varying χ for V z = 0.52
0.5 0.4
χ=1
0.4 0.3

Return autocorrelation
χ=1

Expected volume
0.3 χ = 0.5 0.2
χ = 0.5
0.2 0.1
← Δχ > 0 ← Δχ > 0
0.1 0.0
χ=0
χ=0
← λδ ← λδ
0.0 -0.1

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

λs λs

C Varying V z for χ = 0.5 D Varying V z for χ = 0.5

0.5 0.4

0.4 0.3 V z = 0.252

Return autocorrelation
V z = 12
Expected volume

0.3 V z = 0.52 0.2

← ΔV z > 0
← ΔV z > 0
2
V z = 0.52
0.2 V z = 0.25 0.1

V z = 12
0.1 0.0

← λδ ← λδ
0.0 -0.1

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

λs λs

Figure 2: Expected volume and return autocorrelation. This figure shows expected volume (panels
A and C) and return autocorrelation (panels B and D) as functions of informational negligence (χ), nonspec-
ulative trading motives (Vz ), and information asymmetry (lower λs for a fixed λδ ). The fixed parameters are
ι = 0.5, α = 1, Vd = 12 , Vde = 0.52 , and λδ = 0.75.

Vd
fixed level of I-traders’ private-signal quality, λδ = Vd +Vδ
. The fixed parameters are set to
arbitrary values, although the patterns are qualitatively robust to varying these parameters.
In the figure, higher χ increases expected volume and induces more positive autocor-
relation in returns. Higher Vz also increases expected volume, but induces more negative
autocorrelation in returns. U -traders’ underappreciation of the price-inferred signal, pe, gen-
erates volume by causing the price to underreact to I-investors’ private information, which
implies continuation in returns. I-traders’ nonspeculative trading motives generate volume
by causing the price to deviate from the fundamental value in manner that is ‘exogenous’ to
information, which implies reversal in returns. While some information asymmetry is needed
to get nontrivial effects of χ and Vz , the effects of λs itself are less clear cut. Under REE,
lower λs decreases expected volume and induces more negative autocorrelation in returns.
When χ > 0, however, lower λs amplifies the effect of χ but at a decreasing rate as Vz
increases. These patterns will help with the interpretation of our main theoretical result.

Electronic copy available at: https://ssrn.com/abstract=3150525

Proposition 1 (Predicting returns with current returns and volume).

1. The conditional expectation of r2 given {r1 , v1 } is given by

E [ r2 | {r1 , v1 }] = tanh(γvr v1 r1 )βv v1 , (3)

1 COV[r1 ,xI ] 1 COV[r2 ,xI ]

where γvr = 2ι V[r1 ]V[xI ]−COV[r1 ,xI ]2
> 0 and βv = 2ι V[xI ]
.

2. Eq. (3) can be approximated as

E [ r2 | {r1 , v1 }] = ψr r1 + ψvr v1 r1 (4)

where ψr = −γvr βv v1 2 , ψvr = 2γvr βv v1 , and where v1 is expected volume.

3. Under REE (χ = 0), it holds that ψr ≥ 0 and ψvr ≤ 0 for any Vz ≥ 0, with equalities
for Vz = 0. When χ > 0, it holds that ψr < 0 and ψvr > 0 given sufficiently low Vz .

The non-linear form of Eq. (3) complicates comparative statics. Hence, we work with
Eq. (4), which has the form of a bivariate linear regression with r1 and the volume-return
interaction, v1 r1 , as predictors of r2 . Volume does not have a direct effect because it is
independent of the direction of trade. As such, ψr is the return autoregression when volume
is zero and ψvr is its change as volume increases. Note that ψr and ψvr have opposite signs.
∂E[ r2 | r1 ,v1 ]
Since ∂r1
= ψr + ψvr v1 , the case of ψr < 0 and ψvr > 0 corresponds to our main
empirical results: expected and realized returns are negatively related when volume is low
(v1 < − ψψvrr = v1
2
), but positively related when volume is high (v1 > v1
2
). Part 3 shows that
REE produces the opposite of our main empirical results, in that ψr ≥ 0 and ψvr ≤ 0.
Figure 3 illustrates how ψr and ψvr vary with χ, Vz , and λs . The fixed parameter values
are as in Figure 2, and the results are again qualitatively robust to varying these parameters.
In the REE case (χ = 0; panel A), ψvr becomes more negative and ψr becomes more positive
as Vz increases or λs decreases. Since there is no underreaction, any ‘rational volume’ is
a result of a nonzero endowment shock, z, which also drives the error in the price-inferred
signal, pe. Higher volume is thus associated with stronger reversal under REE, so that ψvr < 0
and ψr > 0. Higher Vz amplifies the reversal effect, and so does lower λs under REE (see
panel B of Figure 2), leading to a more negative ψvr and a more positive ψr .

Electronic copy available at: https://ssrn.com/abstract=3150525

A χ = 0 and V z ∈ {0.252, 0.52, 12} B χ = 0.5 and V z ∈ {0.252, 0.52, 12} C χ = 1 and V z ∈ {0.252, 0.52, 12}
2 2 2
ψrv
ψr
← ΔVz > 0 ← ΔVz > 0
1 1 1
← ΔVz > 0

0 0 0

-1 ← ΔVz > 0 -1 ← ΔVz > 0 -1

← ΔVz > 0

← λδ ← λδ ← λδ
-2 -2 -2

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

λs λs λs

Figure 3: Predicting returns with current returns and volume. This figure shows the coefficients
from E [ r2 | {r1 , v1 }] = ψr r1 + ψvr v1 r1 as functions of informational negligence (χ), nonspeculative trading
motives (Vz ), and information asymmetry (lower λs for a fixed λδ ). The fixed parameters are ι = 0.33, α =
1, Vd = 12 , Vde = 0.52 , and λδ = 0.75.

The CEE case (χ > 0) reproduces our main empirical results as long as Vz is not too
high relative to χ. Specifically, ψvr > 0 and ψr < 0 already at moderate χ given moderate
Vz (χ = 0.5; panel B). The same is true at high Vz given either sufficiently low λs or fully
negligent U -traders (panels B and C). Whenever U -traders’ underappreciation of the price-
inferred signal is not overpowered by I-traders’ nonspeculative trading motives, volume is
predominantly a result of underreaction. In these cases, higher volume is associated with
stronger continuation, so that ψvr > 0 and ψr < 0. Lower Vz or λs amplifies the continuation
effect (see panel D of Figure 2), leading to a more positive ψvr and a more negative ψr .

2.4. Predicting fundamentals with current returns and volume

Our second theoretical result concerns the predictability of the innovation in the dividend
(or ‘fundamental’), d − d, by the current return-volume pair.

Proposition 2 (Predicting fundamentals with current returns and volume).

The conditional expectation of d − d given {r1 , v1 } can be approximated as

  
E d − d  {r1 , v1 } = ψer r1 + ψevr v1 r1 , (5)

1 COV[d,r1 ]V[xI ]−COV[r1 ,xI ]COV[d,xI ]

where ψer = βer − γvr βev v1 2 and ψevr = 2γvr βev v1 , with βer = 2ι V[r1 ]V[xI ]−COV[r1 ,xI ]2
1 COV[d,xI ]V[r1 ]−COV[r1 ,xI ]COV[d,r1 ]
and βev = 2ι V[r1 ]V[xI ]−COV[r1 ,xI ]2
.

10

Electronic copy available at: https://ssrn.com/abstract=3150525

A χ = 0 and V z ∈ {0.252, 0.52, 12} B χ = 0.5 and V z ∈ {0.252, 0.52, 12} C χ = 1 and V z ∈ {0.252, 0.52, 12}
2 2 2
~
ψ rv
← ΔVz > 0 ~
ψ r ← ΔVz > 0 ← ΔVz > 0
1 1 1

0 0 0

← ΔVz > 0 ← ΔVz > 0

-1 ← ΔVz > 0 -1 -1

← λδ ← λδ ← λδ
-2 -2 -2

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

λs λs λs

Figure 4: Predicting
  fundamentals
 with current returns and volume. This figure shows the coeffi-
cients from E d − d  {r1 , v1 } = ψer r1 + ψevr v1 r1 as functions of informational negligence (χ), nonspeculative
trading motives (Vz ), and information asymmetry (lower λs for a fixed λδ ). The fixed parameters are
ι = 0.33, α = 1, Vd = 12 , Vde = 0.52 , and λδ = 0.75.

Eq. (5) also has the form of a bivariate linear regression with r1 and v1 r1 as predictors.
Figure 4 illustrates how ψer and ψevr vary with χ, Vz , and λs . The REE case (panel A)
produces ψer > 0 and ψevr < 0. In addition, ψer becomes more positive and ψevr becomes
more negative as Vz increases or λs decreases. When volume is zero, the price-inferred signal
reveals I-traders’ private signal. Under REE, U -traders fully appreciate this information, so
that r1 = E[d − d | {s, δ}] in this case and therefore ψer > 0. Higher Vz or lower λs increases
the precision of this estimate when volume is zero, implying a more positive ψer . Conversely,
since any ‘rational volume’ is a result of a nonzero endowment shock, the current return
becomes a more noisy estimate as volume increases, implying ψevr < 0. Higher Vz or lower λs
amplifies this noise when volume is nonzero, implying a more negative ψevr .
Allowing for deviations from REE produces somewhat different patterns for ψer but very
different patterns for ψevr . For the direct effect, ψer > 0 when λs is not too low, but ψer < 0
when λs is sufficiently low. When χ > 0 and volume is zero, there is a zero combined effect
of underreaction and any endowment shock, so that r1 reflects the public signal, s. Hence,
ψer > 0 when λs is not too low, but ψer < 0 when λs is sufficiently low, since then s is essentially
noise. For the interaction effect, ψevr > 0 for moderate χ given moderate Vz (panel B). The
same is true at high Vz given either sufficiently low λs or fully negligent U -traders (panels B
and C). When Vz is not too high relative to χ, higher volume means greater underreaction,
hence ψevr > 0. Lower Vz or λs amplifies underreaction, so that ψevr becomes more positive.

11

Electronic copy available at: https://ssrn.com/abstract=3150525

2.5. Testable predictions

Our analysis of the relation between volume and return autocorrelation has two implications.
First, information asymmetry alone is insufficient to generate low-volume reversal and high-
volume momentum. In fact, under REE, information asymmetry increases the discrepancy
between the model and our main empirical results. Second, inefficient use of price-inferred
information and nonspeculative trading motives drive the relation between volume and return
autocorrelation in opposite directions. We thus have our first testable prediction.

Prediction 1. If there is inefficient use of price-inferred information, a decrease in the

degree of nonspeculative (i.e., liquidity related) trading in the cross section or over time is
associated with stronger high-volume continuation but weaker low-volume reversal.

There is much evidence that trading in smaller stocks is to a greater extent driven by
liquidity needs and provision—and is thus to a greater extent nonspeculative—than is trading
in larger stocks (see, e.g., Avramov et al., 2006; Nagel, 2012; Hendershott and Seasholes,
2014, and the references therein).7 As such, a move from smaller to larger stocks maps to a
cross-sectional decrease in nonspeculative trading. The cross-sectional part of Prediction 1
therefore captures our empirical finding that short-term momentum is stronger among larger
and more liquid stocks, while short-term reversal∗ is stronger among smaller and less liquid
ones. We will test the time-series part by exploiting a seasonality in the degree of liquidity
trading at the turn of the month driven by institutional cash needs (e.g., Etula et al., 2019).
Our analysis of the predictability of fundamental innovations by a given return-volume
pair leads to our second testable prediction.

Prediction 2. Consider a regression of fundamental innovations on the current return and

the current volume-return interaction. If all traders are rational, the interaction effect is
negative and the direct effect is positive. If there is inefficient use of price-inferred infor-
mation, the interaction effect is positive, while the direct effect is (i) negative if information
asymmetry is severe, but (ii) otherwise positive.

Hence, in predictive regressions of fundamental growth rates, the sign of the interaction
7
Hendershott and Seasholes, for instance, find that “trades by competing market makers have the best
ability to predict the returns of small stocks after controlling for specialists’ inventories and past returns.
Specialist inventories have the best ability to predict the returns of large stocks.” (p. 141).

12

Electronic copy available at: https://ssrn.com/abstract=3150525

effect allows us to distinguish between the CEE and REE cases. Furthermore, in the CEE
case, the sign of the direct effect sheds light on the severity of information asymmetry.

2.6. Model discussion

Before proceeding to the empirics, we briefly discuss the model and our theoretical results.
While we adopt the CEE model due to its flexibility and tractability, it is important to
note that the exact specification of the model can vary substantially without qualitatively
affecting the conclusion we draw as long as two assumptions are satisfied. First, there can be
deviations from REE. Second, there are nonspeculative trading motives. In particular, related
deviations from REE (e.g., ‘dismissiveness’) and other motives for nonspeculative trading
(e.g., alternative investment opportunities) should lead to qualitatively similar predictions.
For our purpose, the main limitations of the CEE model are two-fold. First, the static
(two-period) setup. Second, the single risky asset. The static setup allows us to solve for the
equilibrium in closed form and develop the intuition more transparently. The single risky asset
avoids overly complicating the analysis by introducing vector and matrix notation. While
beyond the scope of this paper, generalizing the model to a dynamic, multi-asset setup while
maintaining tractability could, for instance, be achieved by assuming a ‘factor structure’ and
either (i) a finite horizon or (ii) an infinite horizon with overlapping generations.8
Proposition 1 shows that we always have ψvr ≤ 0 under REE. This differs from Wang
(1994) but is line with Llorente et al. (2002) since we also do not assume that private informa-
tion is long lived. However, while we find that ψvr becomes more negative with information
asymmetry under REE, Llorente et al. (2002) find that it becomes less negative. As Llorente
et al. explain, this follows from their assumption that nonspeculative trading is not ‘exoge-
nous’ to information about the risky asset—as we assume—but is due to hedging a non-traded
asset which is correlated with the risky asset (see their footnotes 7 and 9, p. 1016). We do
not assume long-lived private information or ‘endogenous’ nonspeculative trading in order to
eliminate other, potentially confounding effects that make the intuition less transparent.
8
A factor structure allows cross-sectional implications to be derived from the time-series properties of
asset prices (see, e.g., Watanabe, 2008; Van Nieuwerburgh and Veldkamp, 2010). Banerjee et al. (2009)
consider a multi-asset, finite-horizon setup and provide sufficient conditions for deviations from REE to
generate positively autocorrelated multi-period returns. Banerjee (2011) considers a multi-asset, infinite
horizon setup with overlapping generations and discusses the conditions under which deviations from REE
lead to similar predictions for returns and volume across the multiple equilibria that arise in such a setting.

13

Electronic copy available at: https://ssrn.com/abstract=3150525

 3. Main results

This section presents our main results. We first show that double sorting stocks on the pre-
vious month’s return and turnover results in significant short-term reversal for low-turnover
stocks but almost equally strong short-term momentum for high-turnover stocks. We then
show that standard risk-factors cannot account for either effect and that the returns to the
short-term momentum strategy persist for 12 months after formation. Finally, we show that
short-term momentum is concentrated among the largest, most liquid stocks.

3.1. Data and key variables

Our main sample consists of firms with monthly market data from CRSP and annual ac-
counting data from Compustat. We consider common shares (CRSP’s SHRCD 10 and 11)
traded on the NYSE, Amex, and Nasdaq exchanges. Following Fama and French (1993,
2015), we impose a six month lag between annual accounting data and subsequent returns
to avoid look-ahead bias. Hence, if a firm’s fiscal year ends in December of calendar year
t − 1, we assume that this data is publicly available at the end of June of calendar year t.
Following Hou, Xue, and Zhang (2015, 2018), we exclude financial firms, although retaining
these has no impact our results. Our main sample covers July 1963 to December 2018.
Our goal is to characterize how current returns depend on past short-term performance
and past trading volume. We measure short-term performance using the return over the
previous month (r1,0 ), similar to the conventional short-term reversal effect (Jegadeesh, 1990).
To be consistent, we use the same one-month horizon for trading volume. Following Lo and
Wang (2000), we measure the latter using share turnover (TO 1,0 ), i.e., the total number
of trades (CRSP’s VOL times 100 to get actual values) deflated by the number of shares
outstanding (CRSP’s SHROUT times 1000 to get actual values). Following Gao and Ritter
(2010), we adjust the trading volume of NASDAQ stocks prior to 2004 to ensure comparability
with NYSE/AMEX stocks, although our results are insensitive to this adjustment.9 We later
show that our results are robust to measuring past performance and turnover over horizons
of up to 6 months, but that the one-month horizon produces the strongest results.

9
Prior to February 2001, we divide NASDAQ volume by 2.0. From February 2001 to December 2001, we
divide by 1.8. From January 2002 to December 2003, we divide by 1.6. From January 2004 and onwards,
NASDAQ volume no longer differs from NYSE/AMEX volume, and we apply no adjustment.

14

Electronic copy available at: https://ssrn.com/abstract=3150525

 Table I: Double sorts on one-month returns and turnover. This table shows portfolios double sorted on previous month’s return (r1,0 ) and
turnover (TO 1,0 ). We use conditional sorts into deciles based on NYSE breakpoints, first on r1,0 and then on TO 1,0 . Portfolios are value-weighted and
rebalanced at the end of each month. The table also shows the performance of long-short strategies across the deciles. Test-statistics (in parentheses)
are adjusted for heteroscedasticity and autocorrelation. Time-series averages of the portfolio characteristics are provided in Table IA.1 in the Appendix.
The sample excludes financial firms. Data are monthly and cover July 1963 to December 2018, except when applying the q-factors, which are available
from January 1967.

r1,0 deciles r1,0 strategies

Low 2 3 4 5 6 7 8 9 High E[re ] αFF6 αq

TO 1,0 deciles
Portfolio excess return
Low 1.28 1.23 0.99 0.85 0.70 0.80 0.59 0.74 0.26 −0.14 −1.41 −1.45 −1.43
(−7.13) (−6.19) (−5.95)
2 1.54 1.22 0.98 0.99 1.05 0.98 0.70 0.69 0.57 0.35 −1.19 −1.31 −1.34
(−4.61) (−4.04) (−4.21)
3 1.71 1.53 0.96 1.11 0.99 0.94 0.75 0.60 0.64 0.36 −1.34 −1.62 −1.66
(−5.02) (−5.61) (−4.87)
4 1.51 1.35 1.43 0.98 1.10 1.07 0.81 0.83 0.64 0.65 −0.85 −1.02 −0.91

15
(−3.63) (−4.15) (−2.92)
5 1.11 1.10 1.26 1.17 1.10 1.00 0.60 0.92 0.90 0.66 −0.45 −0.63 −0.51
(−1.94) (−2.29) (−1.54)
6 1.26 1.38 1.40 1.14 1.00 1.14 1.12 1.19 0.78 0.67 −0.59 −0.60 −0.41
(−2.50) (−2.35) (−1.20)
7 1.39 1.06 1.12 1.22 0.84 1.05 0.83 0.98 0.96 0.73 −0.67 −0.85 −0.96
(−2.52) (−2.55) (−2.37)
8 0.92 1.17 1.25 0.99 1.12 1.02 1.02 0.85 0.82 1.15 0.23 0.13 0.21
(0.85) (0.47) (0.60)
9 0.71 1.37 1.29 1.24 1.21 1.22 1.00 1.12 1.02 0.75 0.05 0.00 0.19
(0.21) (0.01) (0.55)
High 0.00 0.83 1.14 1.08 1.03 0.78 1.01 1.16 0.99 1.36 1.37 1.37 1.65
(4.74) (4.22) (4.47)

Electronic copy available at: https://ssrn.com/abstract=3150525

TO 1,0 strategies
E[re ] −1.28 −0.41 0.15 0.23 0.33 −0.01 0.42 0.42 0.73 1.50
(−5.04) (−1.75) (0.59) (0.80) (1.34) (−0.05) (1.58) (1.54) (2.76) (5.46)
αFF6 −1.26 −0.31 0.18 0.17 0.29 −0.02 0.34 0.57 0.81 1.56
(−4.71) (−1.33) (0.79) (0.73) (1.32) (−0.09) (1.48) (2.12) (3.29) (5.34)
αq −1.39 −0.14 0.21 0.05 0.33 0.02 0.28 0.42 0.91 1.70
(−4.98) (−0.57) (0.82) (0.18) (1.39) (0.07) (1.17) (1.41) (3.15) (5.07)
3.2. Double sorts on previous month’s return and turnover

Table I shows portfolios double sorted on previous month’s return (r1,0 ) and turnover (TO 1,0 ).
We use conditional decile sorts based on NYSE breakpoints, first on r1,0 and then on TO 1,0 .
Hence, we first look at winners vs. losers and then look at low vs. high turnover among
winners and losers, which is in line with the insight from Proposition 1. Using deciles allows
us to account for nonlinearities while maintaining a reasonable number of stocks per portfolio.
We later use coarser quintile double sorts when we control for size.10 Portfolios are value-
weighted and rebalanced at the end of each month. The table shows each portfolio’s average
excess return as well as the performance of high-minus-low strategies within each decile.
Abnormal returns are relative to Fama and French’s (2015) five-factor model including the
momentum factor (FF6) as well as Hou, Xue, and Zhang’s (2015) q-factor model.11 Test
statistics are adjusted for heteroscedasticity and autocorrelation (Newey and West, 1987).
The double sorts reveal a striking pattern in one-month returns. The long-short strategy
that buys last month’s winners and shorts last month’s losers within the lowest turnover decile
yields −1.41% per month with t = −7.13, which is strong evidence of short-term reversal.
In the following, we label this strategy “short-term reversal*” (STREV*) to distinguish it
from the conventional short-term reversal strategy. In stark contrast, the long-short strategy
that buys last month’s winners and shorts last month’s losers within the highest turnover
decile yields 1.37% per month with t = 4.74, which suggests a strong short-term continuation
effect. Consequently, we will in the following refer to this strategy as “short-term momentum”
(STMOM).12 The abnormal returns of the STMOM and STREV* strategies relative to the
FF6 and q-factor models are as large and as strong as the strategies’ average excess returns.
Table IA.1 in the appendix shows time-series averages of the portfolios’ characteristics.
Among low-turnover stocks, the average r1,0 is −14% for losers and +19% for winners, while
the average TO 1,0 is just 2% in both cases. Among high-turnover stocks, the average r1,0 is
10
We use conditional sorts because independent double sorts and NYSE breakpoints result in a few empty
before July 1969. Nonetheless, Table IA.3 in the Appendix shows that very similar results hold for indepen-
dent double sorts from July 1969. Sorting first on returns and then on turnover produces the largest spreads
in holding period returns, but the results are qualitatively very similar if the sorting order is reversed.
11
The FF6 factor returns are from Ken French’s Data Library. The q-factor returns were kindly provided
by Lu Zhang and are available from January 1967.
12
For comparison, the conventional STREV strategy (from a univariate decile sort on r1,0 ) yields −0.28%
with t = −1.68 over our sample. The corresponding conventional momentum strategy (from a univariate
decile sort on prior 12-2 month return) yields 1.21% per month with t = 4.77.

16

Electronic copy available at: https://ssrn.com/abstract=3150525

−18% for losers and +28% for winners, while the average TO 1,0 is 43% for losers and 47% for
winners. According to part 3 of Proposition 1, when there is inefficient use of price-inferred
information (the CEE case) and trading is not overwhelmingly nonspeculative, low turnover
is associated with reversal while high turnover is associated with continuation. We see this
as the negative and significant average return to the STREV* strategy but the positive and
significant average return to the STMOM strategy.
The portfolios’ average market capitalizations are generally ‘inverse U-shaped’ along both
dimensions of the double sort. Stocks in the highest turnover decile are larger than those
in the lowest turnover decile, but there is little difference in size between winners and losers
within the highest and lowest turnover deciles. Accordingly, the average Spearman’s rank
correlation between size and each of TO 1,0 and r1,0 are 29% and 12%. Hence, STMOM
trades in larger stocks compared to STREV*, but neither strategy’s performance is driven
by a particular size tilt.13 Section 3.5 explores the strategies’ link to size in greater detail.

3.3. Factor exposures and abnormal returns

Table II shows results from time-series regressions of the returns to the STMOM and STREV*
strategies. The explanatory variables are the FF6 factors as well as the conventional 1-month
short-term reversal factor (STREV); the 60-13 month long-term reversal factor (LTREV; see
De Bondt and Thaler, 1985); and Pástor and Stambaugh’s (2003) traded liquidity factor
(PSLIQ).14 Table IA.2 in the appendix shows the corresponding results using the q-factors.
The first three specifications show that STMOM has a negative market loading and a
marginally significant positive loading on the conventional momentum factor, but that it
does not load significantly on the remaining FF6 factors. Controlling for the conventional
reversal factors increases STMOM’s abnormal return to 2.21% per month (t = 7.54). Spec-
ifications four to six show that STREV* has only modest loadings on the FF6 factors and
that controlling for the conventional reversal factors does not explain its returns. Neither
strategy loads significantly on PSLIQ.15
13
Table IA.4 in the Appendix shows similar average characteristics for independent double sorts.
14
The returns to the conventional short- and long-term reversal factors are from Ken French’s data library.
The returns to the traded liquidity factor are from Lubos Pástor’s website.
15
The STMOM strategy is not within the univariate span of the MOM factor (abnormal return of 1.13%
with t = 3.36). The converse is also true. The STREV* strategy is not within the univariate span of the
standard STREV factor (abnormal return of −0.91% with t = −4.78), but the standard STREV factor is, in
fact, within the univariate span of the STREV* strategy (abnormal return of 0.03% with t = 0.31).

17

Electronic copy available at: https://ssrn.com/abstract=3150525

Table II: Short-term momentum’s factor exposures and abnormal returns. This table shows time-
series regression results for the short-term momentum (STMOM) strategy from Table I. The explanatory
variable are the factors from Fama and French’s (2015) five-factor model in addition to the momentum factor
(MOM), the two reversal factors (STREV and LTREV), and the traded liquidity factor (PSLIQ). The table
also shows the corresponding results for the short-term reversal* (STREV*) strategy. Test-statistics (in
parentheses) are adjusted for heteroscedasticity and autocorrelation. Data are monthly and cover July 1963
to December 2018, except for specifications employing PSLIQ, which is available from January 1968.

Intercepts, slopes, and test-statistics (in parentheses)

from time-series regressions of the form yt = α + β 0 Xt + t

Short-term momentum Short-term reversal*

Independent (1) (2) (3) (4) (5) (6) (7) (8)

variable

Intercept 1.37 1.56 1.37 2.21 −1.41 −1.29 −1.45 −0.93

(4.74) (5.54) (4.22) (7.54) (−7.13) (−6.57) (−6.19) (−4.43)
MKT −0.38 −0.35 −0.14 −0.25 −0.16 −0.01
(−4.53) (−3.97) (−1.96) (−3.67) (−2.87) (−0.17)
SMB 0.00 0.04 −0.30 −0.19
(−0.03) (0.44) (−2.59) (−2.40)
HML 0.03 0.08 0.00 0.11
(0.12) (0.45) (−0.04) (0.91)
RMW −0.40 −0.34 −0.09 −0.14
(−1.56) (−2.80) (−0.86) (−1.29)
CMA 0.18 −0.20 0.05 0.02
(0.66) (−0.98) (0.24) (0.07)
MOM 0.33 0.05 0.30 0.13
(1.89) (0.54) (4.05) (2.80)
STREV −1.51 −0.97
(−14.38) (−11.11)
LTREV 0.34 −0.16
(2.68) (−1.12)
PSLIQ 0.08 −0.06
(0.97) (−0.99)
Adj. R2 3.8% 6.8% 34.0% 3.6% 9.8% 35.1%

3.4. Persistence and historical performance

According to part 3 of Proposition 1, high-turnover momentum is due to some traders un-

derappreciating the informational content of prices, whereas low-turnover reversal is due to
liquidity trading. We would thus expect persistent post-formation drift for STMOM since
it takes time for private information to be fully impounded in prices, but a comparably
short-lived drift for STREV* since liquidity trading should be corrected fairly quickly.
The left panel of Figure 5 shows the two strategies’ post-formation returns. In line with
our reasoning, the average returns to STREV* become indistinguishable from zero just three
months after formation. In contrast, there is much stronger drift in the returns to STMOM,
which persist for 12 months after formation. The persistence of STMOM is remarkably similar
to that for conventional momentum strategies (see, e.g., Jegadeesh and Titman, 2001). It

18

Electronic copy available at: https://ssrn.com/abstract=3150525

 A B
8 1000
Short-term momentum Short-term momentum
Average cumulative sum of excess returns (%) Short-term reversal* Short-term reversal*
Short-term reversal

Cumulative sum of excess returns (%)

6 600

4
200

2
-200

0
-600

-2

-1000

1 6 12 18 24 1970 1980 1990 2000 2010 2020

Months after portfolio formation

Figure 5: Short-term momentum’s persistence and historical performance. Panel A shows the
average cumulative sums of post-formation excess returns to each of the short-term momentum (STMOM) and
short-term reversal* (STREV*) strategies along with 95% confidence bands. Panel B shows a time-series plot
of cumulative sums of excess returns to STMOM and STREV* as well as a conventional short-term reversal
strategy. Data are monthly and cover July 1963 to December 2018.

means that traders can build up positions in STMOM relatively slowly and, consequently,
reduce trading costs by rebalancing less frequently.
The right panel shows a time-series plot of cumulative sums of excess returns to the two
strategies. STREV* returns are remarkably consistent until 2011, but have slightly tapered
off more recently. STMOM delivered low returns during 1975-79, but has otherwise delivered
consistently positive returns with no other subsample driving its performance.

3.5. Short-term momentum and size

The average market capitalizations in Table IA.1 in the appendix show that STMOM tends
to trade in larger stocks compared to STREV*. In this subsection, we show that constructing
the strategies with an explicit control for size reveals a much stronger result: while STREV*
is indeed strongest among small-caps, STMOM is exclusively a large-caps phenomenon.
The strategies are from 2 × 2 × 2 triple sorts on size, the previous month’s return (r1,0 ),
and the previous month’s turnover (TO 1,0 ). We use conditional sorts; first on size, then
on returns, and then on turnover. The breakpoint for size is the median for NYSE stocks,
while the breakpoints for returns and turnover are the 20th and 80th percentiles for NYSE
stocks. The latter sorts are coarser than the ones in Table I to ensure sufficiently diversified
portfolios. Portfolios are value-weighted and rebalanced at the end of each month.

19

Electronic copy available at: https://ssrn.com/abstract=3150525

Table III: Short-term momentum conditional on size. This table shows time-series regression results
for short-term momentum (STMOM) strategies constructed within small- and large-cap stocks. It also shows
the corresponding results for short-term reversal* (STREV*) strategies. The strategies are from 2 × 2 × 2
triple-sorts on size and the previous month’s return (r1,0 ) and turnover (TO 1,0 ). We use conditional sorts,
first on size, then on returns, and then on turnover. The breakpoint for size is the median for NYSE stocks,
while the breakpoints for returns and turnover are the 20th and 80th percentiles for NYSE stocks. Portfolios
are value-weighted and rebalanced at the end of each month. Test-statistics (in parentheses) are adjusted for
heteroscedasticity and autocorrelation. Data are monthly and cover July 1963 to December 2018, except for
specifications employing PSLIQ, which is available from January 1968.

Intercepts, slopes, and test-statistics (in parentheses)

from time-series regressions of the form yt = α + β 0 Xt + t

Short-term momentum Short-term reversal*

Independent (1) (2) (3) (4) (5) (6)

variable

Panel A: Large-caps
Intercept 0.62 0.52 1.18 −0.84 −0.92 −0.46
(3.01) (2.25) (6.96) (−5.45) (−6.38) (−2.90)
FF6 Yes Yes Yes Yes
STREV, LTREV, and PSLIQ Yes Yes
Adj. R2 11.6% 50.2% 6.8% 42.9%

Panel B: Small-caps
Intercept 0.20 0.03 0.64 −1.41 −1.51 −1.09
(0.89) (0.12) (3.32) (−9.03) (−7.59) (−8.00)
FF6 Yes Yes Yes Yes
STREV, LTREV, and PSLIQ Yes Yes
Adj. R2 12.0% 54.3% 14.6% 51.7%

Table III shows the strategies’ performance. STREV* yields −0.84% per month among
large-caps (t = −5.45) but a considerably larger −1.41% per month among small-caps (t =
−9.03) and similarly for the abnormal returns. This is not surprising given that reversal is
often attributed to liquidity trading, which is known to be more pronounced among smaller
stocks (see, e.g., Nagel, 2012). STMOM, on the other hand, yields a significant 0.62% per
month (t = 3.01) among large-caps but an insignificant 20 basis points per month among
small-caps and similarly for the abnormal returns.
As argued above, the results of Table III are in line Prediction 1, since a move from
smaller to larger stocks maps to a cross sectional decrease in the degree of nonspeculative
(i.e., liquidity related) trading. On a more practical note, these results suggest that STMOM
is much easier to implement than is STREV*. STMOM is concentrated among large-caps,
which are cheaper to trade, and it exhibits post-formation drift for up to 12 months, allowing
traders to slowly rebalance their portfolios.

20

Electronic copy available at: https://ssrn.com/abstract=3150525

3.6. Cross-sectional regressions to predict returns

Tables I–III illustrate our main results using corner portfolios. Here, we use regressions to
show that the same results hold for the entire cross section of stocks.
Table IV shows Fama and MacBeth (1973) cross-sectional regressions of monthly returns
on previous month’s return (r1,0 ), turnover (TO 1,0 ), and their interaction (TO 1,0 × r1,0 ). As
such, these regressions closely mimic the approximate relation in Eq. (4). We control for prior
12-2 month performance (r12,2 ), cash-based operating profitability (COP/A−1 ), investment
(Inv ), book-to-market (log(B/M )), and Size (log of market capitalization as of prior June).16
To mitigate the influence of small stocks, we use weighted least squares (WLS) with market
capitalization as weights and perform the regressions separately for all-but-microcaps and
microcaps. To mitigate outliers and ease interpretation, independent variables are trimmed
at the 1st and 99th percentiles and then standardized by their cross-sectional average and
standard deviation. The interaction is the product of the standardized variables.
The first three specifications show that the univariate effect of r1,0 is negative and sig-
nificant, albeit much stronger among microcaps. They also show that the univariate effect
of the interaction is positive, albeit only significant among all-but-microcaps. The signs of
the effects are consistent with the model’s CEE case (part 3 of Proposition 1) and their rela-
tive strengths across size groups is consistent with Prediction 1. Specifications four to seven
show that employing the main variables together generally strengthens their effects. The
last specification shows that adding controls only strengthens these conclusions. In terms of
economic significance, the t-statistic on the slope of TO 1,0 × r1,0 among all-but-microcaps is
two-and-half times that on the variable associated with conventional momentum (r12,2 ).
16
Following Fama and French (2015), book equity, B, is shareholder’s equity plus deferred taxes minus
preferred stock. In the definition of B, shareholder’s equity is SEQ. If SEQ is missing, we substitute it by
common equity (CEQ) plus preferred stock (defined below), or else by total assets minus total liabilities
(AT − LT). Deferred taxes is deferred taxes and investment tax credits (TXDITC) or else deferred taxes
and/or investment tax credit (TXDB and/or ITCB). Finally, preferred stock is redemption value (PSTKR)
or else liquidating value (PSTKL) or else carrying value (PSTK). B/M is book equity divided by 6 months
lagged equity market capitalization (CRSP’s PRC times SHROUT), where the lagging is to avoid taking
unintentional positions in conventional momentum. Inv equals A/A−1 − 1, i.e., the year-over-year percentage
change in total assets. Following Ball, Gerakos, Linnainmaa, and Nikolaev (2016), COP is total revenue
(REVT) minus cost of goods sold (COGS), minus selling, general, and administrative expenses (XSGA),
plus R&D expenditures (XRD, zero if missing), minus the change in accounts receivable (RECT), minus
the change in inventory (INVT), minus the change in prepaid expenses (XPP), plus the change in deferred
revenue (DRC + DRLT), plus the change in trade accounts payable (AP), plus the change in accrued expenses
(XACC). All changes are annual changes and missing changes are set to zero.

21

Electronic copy available at: https://ssrn.com/abstract=3150525

Table IV: Cross-sectional regressions to predict returns. This table shows Fama and MacBeth (1973)
cross-sectional regressions of firms’ monthly returns on previous month’s return (r1,0 ), previous month’s
turnover (TO 1,0 ), and their interaction (TO 1,0 ×r1,0 ). Regressions are estimated using weighted least squares
(WLS) with market equity as weights. Independent variables are trimmed at the 1st and 99th percentiles, then
standardized by their cross-sectional average and standard deviation. The interaction is the product of the
standardized variables. Test statistics are adjusted for heteroscedasticity and autocorrelation. Microcaps are
firms with a market capitalization below the 20th percentile of the monthly market capitalization distribution
based on NYSE breakpoints.
The controls are prior 12-2 months performance (r12,2 ), cash-based operating profitability-to-lagged
assets (COP/A−1 ), investment (Inv ), book-to-market equity (log(B/M ), where M is market equity as of
prior December), and Size (log of market capitalization as of prior June). The sample excludes financial
firms and firms with negative book equity. Data are monthly and cover July 1963 to December 2018.

Average slopes (×100) and test-statistics (in parentheses)

from monthly WLS regressions of the form rit = β 0t Xit + it

Independent (1) (2) (3) (4) (5) (6) (7) (8)

variables

Panel A: All-but-microcaps

r1,0 −0.12 −0.17 −0.17 −0.21 −0.42

(−2.09) (−2.92) (−3.25) (−3.84) (−8.22)
TO 1,0 × r1,0 0.17 0.19 0.15 0.16 0.22
(5.82) (6.65) (5.70) (5.74) (8.20)
TO 1,0 0.06 0.07 0.05 0.07 −0.04
(0.68) (0.75) (0.61) (0.78) (−0.62)
r12,2 0.28
(3.28)
COP/A−1 0.27
(3.96)
Inv −0.15
(−3.50)
log(B/M ) 0.14
(2.39)
Size −0.08
(−1.91)
Adj. R2 2.2% 0.6% 3.1% 2.7% 5.0% 3.6% 5.4% 11.5%

Panel B: Microcaps

r1,0 −0.46 −0.61 −0.49 −0.59 −0.74

(−6.00) (−7.23) (−6.10) (−7.26) (−8.71)
TO 1,0 × r1,0 0.03 0.18 0.05 0.19 0.19
(0.99) (6.58) (1.82) (7.09) (6.62)
TO 1,0 0.03 0.05 −0.01 0.01 0.00
(0.46) (0.94) (−0.12) (0.22) (0.08)
r12,2 0.44
(6.31)
COP/A−1 0.36
(7.68)
Inv −0.27
(−6.47)
log(B/M ) 0.30
(4.30)
Size −0.14
(−2.03)
Adj. R2 0.8% 0.2% 0.9% 1.0% 1.7% 1.2% 1.9% 4.3%

22

Electronic copy available at: https://ssrn.com/abstract=3150525

3.7. International evidence

Our final main result is out-of-sample evidence for the coexistence of reversal and momentum
at the one-month horizon using international stock market data.
The international sample is from the Compustat Global Securities database and com-
prises the 22 developed markets considered by Fama and French (2017): Australia, Austria,
Belgium, Canada, Denmark, Finland, France, Germany, Great Britain, Greece, Hong Kong,
Italy, Ireland, Japan, the Netherlands, New Zealand, Norway, Portugal, Singapore, Spain,
Sweden, and Switzerland. We consider common shares (TPCI equal to zero) and compute
all returns and market values in U.S. dollars. The period is January 1993 to December 2018.
Table V shows international portfolios double sorted on previous month’s return (r1,0 )
and turnover (TO 1,0 ). Following Asness, Frazzini, and Pedersen (2017), we use independent
double sorts to form country-specific portfolios that are value-weighted and rebalanced at the
end of each month, and then weight each country’s portfolio by the country’s total market
capitalization for the previous month to form each international portfolio.
The international STREV* strategy yields an extremely large −3.25% per month (t =
−6.36). This is over twice that of its U.S. counterpart, presumably because of the on av-
erage less liquid markets outside the U.S. For comparison, the international counterpart of
the conventional STREV strategy yields −0.37% per month with t = −1.29 (untabulated).
Despite this extreme reversal among low-turnover stocks, high-turnover stocks still exhibit
strong short-term momentum, as the international STMOM strategy yields a highly signif-
icant 1.39% per month (t = 3.65), which is comparable to its U.S. counterpart. Moreover,
the two strategies’ abnormal returns relative to Fama and French’s (2017) developed markets
6-factor model are about as large and strong as their average returns. Figure IA.1 in the
Appendix shows that the returns to the international STMOM strategy persist for 24 months
after formation, while those to the international STREV* strategy stagnate after 3 months.
Figure 6 shows the annualized Sharpe Ratios to the country-specific STMOM and STREV*
strategies. The STMOM strategy generates positive average returns in all 22 market. The
STREV* strategy generates negative average returns in all countries but the U.K.

23

Electronic copy available at: https://ssrn.com/abstract=3150525

 Table V: Double sorts on one-month returns and turnover: International evidence. This table shows international portfolios from double
sorts on previous month’s return (r1,0 ) and turnover (TO 1,0 ). We use independent double sorts to form country-specific portfolios that are value-weighted
and rebalanced at the end of each month. We then weight each country’s portfolio by the country’s total market capitalization for the previous month to
form each international portfolio. All returns and market values are in U.S. dollars and excess returns are above the monthly U.S. T-bill rate. The table
also shows the raw and risk-adjusted returns to long-short strategies within the deciles, where risk-adjustment is relative to Fama and French’s (2017)
developed markets five-factor model including the momentum factor (DMFF6). Test-statistics (in parentheses) are adjusted for heteroscedasticity and
autocorrelation. Data are monthly and cover January 1993 to December 2018.

r1,0 deciles r1,0 strategies

Low 2 3 4 5 6 7 8 9 High E[re ] αDMFF6

TO 1,0 deciles
Portfolio excess return
Low 1.28 0.57 0.42 −0.14 0.21 −0.43 −0.27 −0.55 −1.06 −1.97 −3.25 −3.06
(−6.36) (−5.83)
2 1.66 0.69 0.25 0.21 0.17 0.09 0.11 0.02 −0.40 −0.88 −2.53 −2.09
(−4.63) (−2.90)
3 1.02 0.59 0.39 0.24 0.29 0.06 −0.31 −0.03 −0.57 −1.37 −2.40 −2.09
(−5.16) (−4.50)

24
4 0.63 0.87 0.44 0.30 0.39 0.41 0.20 −0.10 −0.18 −0.52 −1.14 −0.74
(−2.73) (−1.71)
5 1.27 0.75 0.31 0.37 0.27 0.43 0.60 0.33 0.31 −0.54 −1.81 −1.47
(−4.13) (−2.24)
6 0.46 0.62 0.38 0.47 0.61 0.27 0.46 0.32 −0.20 −0.15 −0.61 −0.10
(−1.58) (−0.22)
7 0.91 0.62 0.04 0.51 −0.11 0.42 0.39 0.20 0.21 0.08 −0.83 −0.93
(−2.01) (−1.65)
8 0.38 0.46 0.24 0.33 0.33 0.25 0.25 0.81 0.40 0.37 −0.01 −0.04
(−0.03) (−0.08)
9 −0.06 0.16 0.38 0.36 −0.27 0.52 0.26 0.63 0.38 0.26 0.32 0.46
(1.13) (1.41)
High −1.03 −1.30 −0.63 −0.21 −1.04 −0.11 −0.22 −0.24 0.10 0.36 1.39 1.36

Electronic copy available at: https://ssrn.com/abstract=3150525

(3.65) (3.45)

TO 1,0 strategies
E[re ] −2.30 −1.87 −1.05 −0.07 −1.25 0.32 0.05 0.31 1.16 2.33
(−4.74) (−4.69) (−2.08) (−0.14) (−3.17) (0.76) (0.10) (0.64) (2.08) (3.77)
αDMFF6 −2.75 −2.11 −1.42 −0.43 −1.52 0.07 −0.43 −0.09 0.78 1.67
(−5.97) (−4.82) (−2.74) (−0.92) (−3.25) (0.18) (−0.94) (−0.19) (1.38) (2.85)
 Coexistence of reversal and momentum in one-month returns: International evidence

1
Sharpe Ratio (annualized)

-1

Short-term momentum
-2 Short-term reversal*

AUS PRT BEL SGP DNK DEU NZL CAN SWE NOR HKG JPN AUT GBR ESP IRL ITA CHE NLD FRA FIN GRC

Figure 6: International short-term momentum returns. This figure shows the annualized Sharpe
Ratios of the country-specific international STMOM and STREV* strategies. The strategies trade country-
specific portfolios from independent double sorts into deciles based on the previous month’s return and
turnover. Portfolios are value-weighted and rebalanced at the end of each month. All returns and market
values are in U.S. dollars. Data are monthly and cover January 1993 to December 2018.

4. Testing the model’s predictions

We now test the predictions that allow us to distinguish between the CEE and REE cases
based on (i) the effects of liquidity trading and (ii) the link to fundamentals. Our results
are strongly in favor of CEE. They are also at odds with severe information asymmetry.
In addition, we go beyond the model and show that constructing short-term momentum
strategies with a control for variables related to analysts’ forecasts leads to similar conclusions.

4.1. Purging liquidity trades

According to Prediction 1, if there is inefficient use of price-inferred information, a decrease

in the degree nonspeculative trading should lead to stronger high-turnover momentum but
weaker low-turnover reversal. Tables III and IV provide cross-sectional evidence for this pre-
diction by varying market capitalization. Here, we provide time-series evidence by exploiting
a seasonality in the degree of liquidity trading at the turn of the month.
Previous studies document that many liquidity driven trades occur at the end of the
month due to institutional cash needs (see Etula et al., 2019, and the references therein).
We therefore argue that time-series variation in the degree nonspeculative trading can be

25

Electronic copy available at: https://ssrn.com/abstract=3150525

Table VI: Short-term momentum and end-of-month effects. This table shows the performance of
zero-cost long-short strategies that buy the previous month’s winners and sell the previous month’s losers
among stocks with different turnover. Portfolios are from double sorts on the previous month’s return
and turnover. We use conditional sorts into deciles based on NYSE breakpoints, first on returns and then
on turnover. Portfolios are value-weighted and rebalanced at the end of each month. In panel A, the
sorting variables exclude their end-of-month values (r1,0−EOM and TO 1,0−EOM ) measured at the month’s
last three trading days. In Panel B, the sorting sorting variables are just the end-of-month values (rEOM
and TO EOM ) measured at the month’s last three trading days. Test-statistics (in parentheses) are adjusted
for heteroscedasticity and autocorrelation. The sample excludes financial firms. Data are monthly and cover
July 1963 to December 2018, except when applying the q-factors, which are available from January 1967.

Panel A Panel B
Performance of r1,0−EOM strategies Performance of rEOM strategies
within TO 1,0−EOM deciles within TO EOM deciles

TO 1,0−EOM decile E[re ] αFF6 αq TO EOM decile E[re ] αFF6 αq

Low −0.41 −0.49 −0.41 Low −2.36 −2.45 −2.56

(−2.16) (−2.58) (−1.75) (−11.51) (−11.13) (−11.59)
2 −0.38 −0.48 −0.48 2 −1.72 −1.80 −1.81
(−1.64) (−1.98) (−1.76) (−7.12) (−6.61) (−6.55)
3 −0.28 −0.47 −0.52 3 −1.73 −1.90 −2.04
(−1.22) (−1.83) (−1.56) (−6.17) (−6.96) (−6.55)
4 −0.21 −0.24 −0.11 4 −1.66 −1.75 −1.77
(−0.81) (−0.75) (−0.29) (−6.06) (−5.60) (−5.64)
5 0.39 0.30 0.46 5 −1.72 −1.99 −1.92
(1.40) (0.90) (1.35) (−6.78) (−6.75) (−6.77)
6 0.22 0.08 0.16 6 −1.86 −1.97 −2.06
(0.89) (0.27) (0.39) (−7.45) (−7.00) (−7.56)
7 0.27 0.33 0.50 7 −1.16 −1.17 −1.20
(1.00) (0.91) (1.29) (−4.86) (−4.54) (−4.42)
8 0.38 0.34 0.53 8 −1.67 −1.72 −1.79
(1.26) (1.11) (1.35) (−5.60) (−5.01) (−4.58)
9 0.72 0.58 0.72 9 −1.16 −1.29 −1.29
(2.89) (2.09) (2.13) (−4.74) (−4.91) (−4.41)
High 1.83 1.80 2.10 High −0.65 −0.86 −0.87
(5.52) (5.22) (5.16) (−2.25) (−2.63) (−2.41)

obtained by focusing on returns and turnover at the end of the month. As such, we repeat
our benchmark double sorts from Table I, but we now exclude the last three trading days for
each month from the sorting variables. Hence, if month t has d trading days, the sorting
variables in month t, used to predict returns for month t + 1, will be based on the returns
and turnover on days 1, 2, . . . , d − 3 of month t. We consider only firm-months with at least
15 non-missing daily returns and turnover. Note that this procedure also introduces a three
day implementation lag, which makes the resulting strategies easier to implement in practice.
Panel A of Table VI reports results for this exercise. There is a clear shift towards stronger
high-turnover momentum and weaker low-turnover reversal compared to the benchmark re-
sults in Table I. The STMOM strategy now yields 1.83% per month with t = 5.52 (compared
to 1.37% with t = 4.74 in the benchmark case). In addition, there is a significant momentum

26

Electronic copy available at: https://ssrn.com/abstract=3150525

effect in the 9th turnover decile. The STREV* strategy, on the other hand, now yields just
−0.41% per month (compared to −1.41% per month in the benchmark case). The abnormal
returns tell a very similar story.
A complementary version of this test is to sort on only the end-of-month values, i.e.,
we now sort on returns and turnover occurring on days d − 2, d − 1, and d of the previous
month. Based on the same reasoning, the last days in a month are likely to be dominated by
liquidity trades, and we would thus expect to find stronger low-turnover reversal but weaker
high-tunover momentum. Panel B of Table VI shows that this is indeed the case. The
STREV* strategy now yields a very large −2.36% per month with t = −11.51. Furthermore,
the STMOM strategy’s average return is now negative: −0.65% per month.17
To show that the last trading days in a month are indeed special, we repeat the tests of
table Table VI on different days during the month. Panel A of Figure 7 shows the performance
of STMOM and STREV* strategies formed on day d based on prior return and turnover
over days d − 20 to d − 3. Skipping 3 days only significantly improves the performance
of STMOM when these days lie at the end of the month. Furthermore, the performance
of the corresponding STREV* strategies is generally weaker around the end of the month.
Panel B of the figure shows the performance of strategies formed on day d based only on
prior return and turnover over days d, d − 1, and d − 2. Sorting on the prior 3 days only
significantly improves the performance of STREV* when these days lie at the end of the
month. Furthermore, the corresponding STMOM strategies generate large negative returns
around the end of the month.
Taken together, these results provide strong time-series evidence for Prediction 1 and
suggest that the performance of STMOM and STREV* does indeed relate to the degree of
nonspeculative (i.e., liquidity related) trading. Traders implementing STMOM in the U.S.
or internationally should skip the last days of each month to purge liquidity trades from the
return and turnover signals. Conversely, traders and researchers looking into returns from
liquidity provision should pay attention to STREV* near the end of the month.

17
Table IA.5 in the Appendix shows that results of Table VI also hold in our international sample. In
untabulated tests, we find very similar results when focusing on the last 5 (instead of 3) trading days for
month t and, furthermore, that these results are robust to also excluding the first 1-3 trading days from the
return for month t + 1.

27

Electronic copy available at: https://ssrn.com/abstract=3150525

 A Sorts on day d based on prior return and turnover over days d − 20 to d − 3
3 (i.e., skipping days d , d − 1, and d − 2)
Average excess return over days d + 1 to d + 21 (%)

Short-term momentum
Short-term reversal*
2
1
0
-1

3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21

Day of rebalancing during the month (d )

B Sorts on day d based on prior return and turnover over days d , d − 1, and d − 2
1
Average excess return over days d + 1 to d + 21 (%)

0
-1
-2
-3

3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21

Day of rebalancing during the month (d )

Figure 7: Short-term momentum and end-of-month effects. This figure shows the average excess
returns to STMOM and STREV* strategies formed on different days (d) during the month. Panel A shows
strategies formed on day d based on prior return and turnover over days d − 20 to d − 3, skipping the most
recent three days (i.e., days d − 2, d − 1, and d). Panel B shows strategies formed on day d based only on prior
return and turnover over the most recent three days (i.e., days d − 2, d − 1, and d). We use conditional sorts
into deciles based on NYSE breakpoints, first on returns and then on turnover. Portfolios are value-weighted
and returns are calculated over days d + 1 to d + 21. The sample excludes financial firms. Data are daily and
cover end of July 1963 to end of December 2018.

28

Electronic copy available at: https://ssrn.com/abstract=3150525

Table VII: Cross-sectional regressions to predict fundamentals. This table shows Fama and MacBeth
cross-sectional regressions of firms’ 1-, 3-, and 5-year growth in gross profits (Panel A) and earnings (Panel
B) on the previous month’s return and turnover, their interaction (TO 1,0 × r1,0 ), and controls. All variables
are measured at the end of firms’ fiscal years. Regressions are estimated using weighted least squares (WLS)
with market equity as weights. Dependent and independent variables are trimmed at the 1st and 99th
percentiles. Independent variables are standardized by their cross-sectional average and standard deviation.
The interaction is the product of the standardized variables. Test-statistics (in parentheses) are adjusted for
heteroscedasticity and autocorrelation.
Controls are prior 12-2 month performance (r12,2 ), gross profits-to-assets (GP/A), earnings-to-book
equity (IB/B ), dividends and repurchases-to-book equity (Div/B, where Div is DVC + PRSTKCC, both set
to zero if missing), investment (Inv ), book-to-market equity (log(B/M ), where M is fiscal-year-end market
capitalization), and Size (log of fiscal-year-end market capitalization). The sample excludes financial firms
and firms with negative book equity. Data are annual and cover 1963 to 2018.

Average slopes (×100) and test-statistics (in parentheses)

from WLS regressions of the form yit = β 0t Xit + it

1-year 3-year 5-year

Independent (1) (2) (3) (4) (5) (6) (7) (8) (9)
variables

GP i,t+τ −GP it
Panel A: Dependent variable is growth in gross profits, yit = Ait

TO 1,0 × r1,0 0.53 0.19 0.14 0.79 0.14 0.14 1.20 0.37 0.37
(5.76) (2.54) (2.76) (2.54) (0.52) (0.72) (2.51) (1.02) (1.49)
r1,0 0.79 0.69 1.24 0.89 1.48 0.96
(5.15) (9.57) (3.67) (5.02) (2.79) (3.27)
TO 1,0 0.76 −0.03 1.90 −0.18 2.87 −0.39
(4.30) (−0.27) (2.49) (−0.45) (2.14) (−0.48)
r12,2 1.55 2.05 2.71
(8.92) (5.60) (5.62)
GP/A 1.94 6.33 11.65
(11.06) (8.01) (7.73)
IB/B −1.13 −2.55 −3.74
(−5.53) (−3.86) (−3.19)
Div/B −0.50 −1.90 −3.56
(−8.91) (−8.52) (−8.87)
Inv 1.51 2.86 5.00
(7.86) (7.17) (5.76)
log(B/M ) −1.76 −5.36 −9.55
(−18.08) (−14.71) (−16.47)
Size −0.20 −1.37 −2.90
(−2.22) (−4.42) (−4.36)
Adj. R2 1.1% 3.5% 29.0% 0.9% 2.9% 31.5% 0.6% 2.5% 35.3%

IB i,t+τ −IB it
Panel B: Dependent variable is growth in earnings, yit = Bit

TO 1,0 × r1,0 0.54 0.28 0.26 0.30 0.00 0.11 0.49 0.22 0.20
(4.37) (2.83) (2.50) (1.44) (0.00) (0.66) (2.11) (1.10) (1.41)
r1,0 0.85 0.80 1.00 0.78 0.86 0.41
(4.87) (5.63) (4.79) (3.58) (2.70) (1.41)
TO 1,0 0.06 −0.41 −0.13 −0.59 0.03 −0.78
(0.38) (−2.78) (−0.27) (−1.36) (0.06) (−1.60)
r12,2 2.07 1.10 0.94
(9.48) (2.89) (5.44)
GP/A 0.89 1.35 1.83
(4.17) (2.75) (3.20)
IB/B −9.06 −12.23 −14.13
(−4.26) (−3.94) (−3.20)
Div/B 0.43 0.31 −0.02
(5.84) (1.41) (−0.06)
Inv −0.64 −0.65 −0.10
(−2.65) (−1.72) (−0.17)
log(B/M ) −1.62 −3.28 −4.78
(−7.92) (−10.65) (−11.79)
Size 1.00 0.99 0.92
(2.85) (1.21) (1.06)
Adj. R2 0.8% 2.3% 16.7% 0.8% 2.6% 14.7% 0.8% 2.2% 14.7%

29

Electronic copy available at: https://ssrn.com/abstract=3150525

4.2. Cross-sectional regressions to predict fundamentals

We test Prediction 2 using Fama and MacBeth regressions of firms’ 1-, 3-, and 5-year growth
in fundamentals on monthly returns, turnover, and their interaction (TO 1,0 × r1,0 ). Panel A
of Table VII shows results for gross profits (REVT − COGS) while Panel B shows results for
earnings (IB). We control for current gross profits-to-assets (GP/A), earnings-to-book equity
(IB/B), dividends and repurchases-to-book equity (Div/B ), prior 12-2 month performance
(r12,2 ), investment (Inv ), book-to-market equity (log(B/M )), and Size (log of equity market
capitalization). All independent variables are measured at the end of firms’ fiscal years. We
use weighted least squares (WLS) with market equity as weights (the results are even stronger
with OLS). To mitigate outliers, dependent and independent variables are trimmed at the 1st
and 99th percentiles. Independent variables are standardized by their cross-sectional average
and standard deviation. The interaction is the product of the standardized variables.
According to Prediction 2, the interaction should have a positive effect if there is ineffi-
cient use of price-inferred information (the CEE case), but a negative effect if all traders are
rational (the REE case). The results are strongly in favor of the CEE case. At the 1-year
horizon, the slope of the interaction is positive and significant with or without controls for
both gross profits and earnings. At the 3- and 5-year horizons, the slope on the interaction
loses its significance when employed alongside the controls, but, importantly, it never be-
comes negative (as would have been the case under REE). Moreover, according to the CEE
case of Prediction 2, the direct effect of the current return is negative if there is severe infor-
mation asymmetry, but is otherwise positive. The results strongly reject severe information
asymmetry, as the slope on the previous month’s return is positive in all specifications, and
also highly significant except for 5-year earnings growth when employed with controls.

4.3. Short-term momentum and analysts’ forecasts

Until now, our tests have been guided by the model’s predictions—about the effects of liq-
uidity trading and the link to fundamentals—and have provided evidence in line with the
CEE case and at odds with severe information asymmetry. We now go beyond the model and
show that constructing short-term momentum strategies with a control for variables related
to analysts’ forecasts leads to similar conclusions.

30

Electronic copy available at: https://ssrn.com/abstract=3150525

Table VIII: Short-term momentum conditional on analysts’ forecast variables. This table shows
the performance of short-term momentum (STMOM) strategies constructed conditional on analysts’ fore-
cast variables. It also shows the corresponding results for short-term reversal* (STREV*) strategies. The
strategies are from 2 × 2 × 2 triple sorts on the conditioning variable, the previous month’s return (r1,0 ),
and the previous month’s turnover (TO 1,0 ). We use conditional sorts, first on the conditioning variable,
then on returns, and then on turnover. The breakpoints are the 20th and 80th percentiles for NYSE stocks.
Portfolios are value-weighted and rebalanced at the end of each month. Test-statistics (in parentheses) are
adjusted for heteroscedasticity and autocorrelation.
All analysts’ forecast variables are for the month closest to—but preceding—the month in which a firm
announces its quarterly earnings (Compustat’s RDQ). NtAnalysts is the number of analysts following a stock.
NtAnalysts ⊥ Mt is the residual from monthly cross-sectional regressions of log(NtAnalysts ) on log(Mt ), where
Mt is market equity, estimated using weighted least squares (WLS) with market equity as weights. EPS d t is
analysts’ forecasts of firms’ earnings-per-share. The sample excludes financial firms. Data are monthly and
cover January 1985 to December 2018, where the start date is due to availability of analyst forecast data.

Raw and risk-adjusted performance

of conditional strategies

Short-term momentum Short-term reversal*

Conditioning variable E[re ] αFF6 E[re ] αFF6

(1) NtAnalysts High 1.14 1.12 −0.58 −0.59

(3.20) (2.84) (−2.41) (−2.31)
Low 0.62 0.50 −0.88 −0.86
(1.52) (1.18) (−3.69) (−2.58)

(2) NtAnalysts ⊥ Mt High 0.84 0.99 −0.41 −0.70

(2.50) (2.68) (−1.27) (−1.98)
Low 0.47 0.52 −0.97 −0.94
(1.15) (1.14) (−3.21) (−2.71)

(3) std(EPS
d t) High 1.38 1.28 −0.28 −0.23
(2.69) (2.31) (−0.99) (−0.79)
Low 0.32 0.39 −1.13 −1.10
(1.00) (1.32) (−4.29) (−3.20)

(4) std(EPS
d t )/Pt High 1.08 1.16 −0.95 −1.24
(2.06) (2.17) (−2.51) (−3.08)
Low 0.01 0.21 −1.27 −1.26
(0.03) (0.71) (−5.45) (−5.54)

(5) std(EPS
d t )/|med(EPS
d t )| High 0.97 1.07 −0.74 −0.91
(2.24) (1.98) (−1.65) (−1.93)
Low −0.25 −0.39 −1.45 −1.61
(−0.70) (−1.07) (−4.69) (−4.89)

(6) d t ) − min(EPS
max(EPS d t) High 1.49 1.53 −0.40 −0.38
(3.20) (2.53) (−1.17) (−1.19)
Low −0.13 −0.03 −1.16 −1.12
(−0.36) (−0.08) (−4.14) (−2.74)

Following the literature, we use analysts’ forecasts data from the Institutional Brokers’
Estimate System (I/B/E/S). We use only unadjusted forecast data to mitigate the reporting
inaccuracies, rounding errors, and look-ahead-biases identified in previous studies (see, e.g.,
Diether, Malloy, and Scherbina, 2002). Our tests employ earnings-per-share (EPS) forecasts
for the month closest to—but preceding—the month in which a firm announces its quarterly
earnings (Compustat’s RDQ). The sample period is January 1985 to December 2018.

31

Electronic copy available at: https://ssrn.com/abstract=3150525

 The number of analysts covering a stock is a common proxy for the degree and quality of
‘information production’ about the stock.18 Citing preceding evidence, Llorente et al. (2002)
argue that higher analyst coverage should be associated with lower information asymmetry
(see also Hong, Lim, and Stein, 2000). Table VIII shows that short-term momentum is
concentrated among stocks with high analyst coverage—i.e., those with presumably low in-
formation asymmetry—which is further evidence against severe information asymmetry. The
table’s first specification shows the performance of STMOM strategies from 2 × 2 × 2 triple
quintile sorts that condition on analyst coverage. STMOM only yields significant returns
among stocks with high coverage, while STREV* is strongest among stocks with low cover-
age. The table’s second specification shows that these results are not mechanically driven
by the fact that larger firms tend to be covered by more analysts: conditioning on analyst
coverage orthogonalized with respect to size yields qualitatively similar results. On the other
hand, these results are in line with the CEE model if some traders have a greater tendency
to neglect the informational content of prices for stocks covered by more analysts.
Following Diether et al. (2002), numerous papers use measures of the dispersion in an-
alysts’ forecasts as proxies for heterogenous beliefs—or ‘disagreements’—among traders.19
While this interpretation comes with many caveats, there is some consensus that forecast
dispersion can at least be viewed as positively correlated with the underlying, unobservable
heterogeneity in beliefs (see the discussions in, e.g., Verardo 2009 and Banerjee 2011). Since
short-term momentum is nonexistent among stocks with low coverage (presumably those with
high information asymmetry), it is interesting to examine whether it is instead concentrated
among stocks with high forecast dispersion (presumably those with strong disagreements
among traders). The remaining specifications in Table VIII show that this is indeed the
case, irrespective of the dispersion measure: STMOM is concentrated among high-dispersion
stocks. Furthermore, STREV* is stronger among low-dispersion stocks. These results are in
line with the CEE model if analysts’ forecast dispersion is positively correlated with the ex-
tent to which traders underappreciate what prices convey about others’ private information.
18
Brennan and Subrahmanyam (1995) find that higher analyst coverage is associated with lower ‘adverse
selection costs’ in the sense of Kyle (1985). Easley, O’Hara, and Paperman (1998) find that higher analyst
coverage is associated with a lower ‘probability of informed trading.’
19
See, e.g., Johnson (2004), Verardo (2009), Banerjee (2011), Cen, Wei, and Yang (2017), Cujean and
Hasler (2017), and Loh and Stulz (2018), among many others

32

Electronic copy available at: https://ssrn.com/abstract=3150525

 5. Robustness

This section provides robustness tests of our main results. We show that short-term momen-
tum survives conservative transaction costs; that it is not driven by post-earnings announce-
ment drift, industry momentum, or volatility; that it exhibits far less crash risk compared to
conventional momentum; and that it extends beyond the one-month horizon we focus on.

5.1. Transaction costs

Previous studies document that strategies based on the previous month’s return have high
turnover and may thus be relatively expensive to trade (see, e.g., Da, Liu, and Schaumburg
2012 and Novy-Marx and Velikov 2015). Here, we show that the profitability of short-term
momentum persists after transaction costs.
Following Koijen, Moskowitz, Pedersen, and Vrugt (2018) and Bollerslev, Hood, Huss,
and Pedersen (2018), we measure a strategy’s average transaction costs using the average
turnover and ‘half-spreads’ (one-half of the proportional bid-ask spread) of the underlying
portfolios. While half-spreads are simpler than the ‘effective bid-ask spread’ proposed by
Hasbrouck (2009) and employed by Novy-Marx and Velikov (2015), the resulting costs are
still conservative, especially for large institutional traders. Half-spreads assume market orders
and immediate liquidity demand—instead of the more efficient limit orders—and are thus
likely to overstate the actual average costs associated with implementing a strategy.20
Fix a portfolio. Let wi,t−1 ≥ 0 be the weight of stock i in the sort at the end of month
t − 1, and let rit be the stock’s return over month t. Portfolio turnover (PTO) in month t is

1 X  
PTOt = (1 + rit )wi,t−1 − wit , (6)
2 i

where the sum is taken over all stocks in the portfolio in months t − 1 or t and is divided by
2 to avoid double counting buys and sells. The product (1 + rit )wi,t−1 is the weight of stock
20
Novy-Marx and Velikov (2015) state that the effective bid-ask spread “does not account for the price
impact of large trades [...] and should thus be interpreted as the costs faced by a small liquidity demander”
but that it is “nevertheless conservative (i.e., an upper-bound), because it assumes market orders” (p. 108).
The half-spread can be interpreted in a similar way. Using proprietary data, Frazzini, Israel, and Moskowitz
(2015) find that the actual trading costs faced by a large institutional trader are an order of magnitude smaller
than those estimated for the average trader, and state that “this is because a large institutional trader [...]
often trades within the spread, using limit orders and tries to supply rather than demand liquidity” (p. 20).

33

Electronic copy available at: https://ssrn.com/abstract=3150525

Table IX: Short-term momentum and transaction costs. This table shows strategy performance after
transaction costs. It shows, for each strategy, the average gross excess return (i.e., before costs), the average
turnover and half-spreads for the long and short sides, the average total transaction cost, and, finally, the
average net excess return and net FF6 abnormal return. ‘Turnover’ is the time-series average of one-half of
the sum of absolute monthly changes in portfolio weights (Eq. (6)). ‘Half-spread’ is the time-series average of
the portfolios’ monthly value-weighted half-spreads. ‘Cost’ (in % per month) is the sum of average turnover
times average half-spread for the long and short sides. ‘EOM’ are the 3 trading-days at the end of the previous
month. ‘Large-caps’ are stocks with a market capitalization above the monthly NYSE median. Test-statistics
(in parentheses) are adjusted for heteroscedasticity and autocorrelation. Data are monthly and cover July
1963 to December 2018, but the half-spreads are computed starting from 1983 due to data availability.

Long side Short side

Strategy e
E[rgross ] Turnover Half-spread Turnover Half-spread Cost e ]
E[rnet αFF6
net

Short-term momentum 1.37 0.89 0.17 0.90 0.24 0.37 1.00 1.00
(4.74) (3.47) (3.09)

Short-term reversal* −1.41 0.96 0.84 0.94 1.22 1.94 – –

(−7.13)

Short-term momentum 1.83 0.89 0.18 0.91 0.24 0.38 1.45 1.42
(excluding EOM) (5.52) (4.38) (4.12)

Short-term momentum 0.77 0.83 0.08 0.84 0.10 0.15 0.62 0.59
(large-caps, excluding EOM) (3.90) (3.14) (2.78)

i just before rebalancing at the end of month t. We set wi,t−1 = 0 if stock i is not in the
portfolio in month t − 1, and similarly for wit . We measure portfolio half-spread (PHS) in
any given month as the value-weighted average of 12 (ASK − BID)/( 21 (ASK + BID)).
Table IX shows strategy performance after transaction costs. The benchmark STMOM
strategy (Table I) turns over 89% and 90% of its long and short sides on average each month,
yet these incur modest average half-spreads of 0.17% and 0.24% per month because they tend
to trade in large stocks. Hence, its average monthly cost is 0.89 × 0.17 + 0.90 × 0.24 = 0.37%,
implying a net average return of 1.00% per month with t = 3.47. Its net abnormal return
is equally large and strong. The STREV* strategy turns over more often and incurs higher
half-spreads because it tend to trade in small stocks. Its average monthly cost of 1.94%
therefore subsumes its average gross return when judged by our conservative approach.
We also consider the STMOM strategy purged of liquidity trades (Panel A; Table VI),
which has a built-in implementation lag of 3 days. It incurs about the same average cost as
the benchmark STMOM strategy but generates a considerably higher average gross return.
Its average net return is therefore a considerably higher 1.45% per month with t = 4.38.21
21
For comparison, the conventional momentum strategy incurs an average cost of 0.26% per month and its
net return is 0.95% per month (t = 3.74). Moskowitz and Grinblatt’s (1999) industry momentum strategy
incurs an average cost of 0.33% per month and its net return is 0.64% per month (t = 2.98). The average
cost of the conventional STREV strategy is 0.56% per month, which subsumes its average gross return.

34

Electronic copy available at: https://ssrn.com/abstract=3150525

 Lastly, we consider a large-caps version of the STMOM strategy purged of liquidity trades,
i.e., constructed as in Table III but excluding the last 3 trading days from the return and
turnover signals. The strategy yields 0.77% per month with t = 3.90 before transaction costs.
Its average monthly cost is just 0.15% because its long and short sides turn over less often
and incur substantially lower half-spreads than the other STMOM strategies. The strategy
thus yields 0.62% per month with t = 3.14 after transaction costs and a similar net abnormal
return. It does so despite trading exclusively in the 50% largest, most liquid stocks.22

5.2. Post-earnings announcement drift

Firms’ quarterly earnings announcement are known to generate large share turnover and
persistent subsequent price drifts (see, e.g., Hong and Stein 2007 for a review, as well as
Medhat and Schmeling 2018). Nonetheless, the results presented here show that earnings
announcers and post-earnings announcement drift (PEAD) are not driving our main results.
Table IA.6 in the Appendix repeats the double sorts in Table I but excludes firms whose
most recent earnings announcement date (Compustat’s RDQ) fell in the previous month.
With the exclusion of earnings announcers, STREV* yields −1.77% per month (t = 6.07)
while STMOM yields 1.26% per month (t = 2.65) with similar abnormal returns over the
period from January 1972 for which we have quarterly earnings announcement dates.

5.3. Industry momentum

Moskowitz and Grinblatt (1999) find that industry momentum is strongest at the one month
horizon (see also Asness, Porter, and Stevens, 2000). Note that this is consistent with Predic-
tion 1 if averaging returns within an industry is an alternative way of decreasing the degree
of nonspeculative (i.e., liquidity related) trading, thus leading to one-month momentum in
industries despite one-month reversal in individual stocks.23 Nonetheless, the results pre-
22
The strategy’s long and short sides trade in an average of 39 and 37 stocks with average market capital-
izations of 3.4 and 3.2 billion dollars. It is therefore comparable in terms of average market capitalization to
the “Fortune 500” strategies considered by Novy-Marx (2013). In December of 2018 (the end of our sample),
the long and short sides had average market capitalizations of 8.2 and 12.3 billion dollars and value-weighted
average half-spreads of just 2 and 1 basis points.
23
Moskowitz and Grinblatt (1999) explicitly entertain this possibility: “One possible explanation for the
discrepancy between short-term (one-month) reversals for individual stocks and short-term continuations for
industries is that the one-month return reversal for individual stocks is generated by microstructure effects
(such as bid-ask bounce and liquidity effects), which are alleviated by forming industry portfolios.” (p. 1274).

35

Electronic copy available at: https://ssrn.com/abstract=3150525

sented here show that industry momentum does not drive the profitability of our short-term
momentum strategy.
Table IA.7 in the Appendix shows spanning tests employing our STMOM strategy and a
one-month industry momentum (IMOM) strategy. We sort the 49 industries from Fama and
French (1997), excluding financials, on their previous month’s value-weighted average return
and form a ‘winner’ and a ‘loser’ industry portfolio each containing 5 industries. Portfolios
are value-weighted and rebalanced at the end of each month. The IMOM strategy is long
the winner- and short the loser industry portfolio. The spanning tests show that STMOM is
not within the span of IMOM, with or without controlling for the FF6 or the q-factors. On
the other hand, IMOM is in fact within the combined span of STMOM and the q-factors.
In untabulated tests, we find that reproducing our double sorts in Table I using industry-
demeaned values for the previous month’s return and turnover has no material effect on our
results. The STREV* strategy yields −1.99% per month (t = −10.08) while the STMOM
strategy yields 1.15% per month (t = 3.93) with similar abnormal returns. That is, the
profitability of short-term momentum remains largely undiminished when the strategy is
constructed in a way that explicitly sorts against the one-month industry momentum signal.

5.4. Crash risk

Daniel and Moskowitz (2016) show that the conventional momentum strategy can experience
infrequent but persistent strings of negative returns—or crashes—that are contemporaneous
with market rebounds, in that conventional momentum “will have significant negative market
exposure following bear markets precisely when the market swings upward” (p. 229). Here,
we show that short-term momentum exhibits far less crash risk than conventional momentum.
Table IA.8 in the Appendix presents the results. Panel A shows that STMOM exhibits a
mild negative skew and a moderate kurtosis, both of which are similar to those observed for
the market but orders of magnitude lower than those observed for conventional momentum.
It also shows that while conventional momentum exhibits negative and significant coskewness
(Harvey and Siddique, 2000) and downside beta (Henriksson and Merton, 1981), both are
small in magnitude and statistically insignificant for STMOM.

36

Electronic copy available at: https://ssrn.com/abstract=3150525

 Panel B shows results from Daniel and Moskowitz’s market-timing regressions,

e
rte = (α0 + αB IB,t−1 ) + (β0 + (βB + βB,U IU,t )IB,t−1 ) rmt + t , (7)

where rte is a strategy’s excess return, rmt

e
is the market excess return, IB,t−1 is an ex-ante 24-
month ‘bear market’ indicator, and IU,t is a contemporaneous 1-month ‘up-market’ indicator.
Here, βB captures the difference in beta during bear markets, while βB,U capture the difference
in beta in rebounds after bear markets.
For conventional momentum, βb0 = 0.17 (t = 2.43), βbB = −0.69 (t = −2.37), and
βbB,U = −0.40 (t = −1.24) over our sample period. In a bear market, conventional momentum
has a beta of βb0 + βbB = −0.31 when the contemporaneous market return is negative but a
beta of βb0 + βbB + βbB,U = −0.71 when the contemporaneous market return is positive. For
STMOM, βb0 = −0.25 (t = −2.72), βbB = −0.77 (t = −2.99), and βbB,U = 0.81 (t = 1.96).
As such, its bear market beta is −1.02 when the contemporaneous market return is negative
but just −0.21 when the contemporaneous market return is positive. That is, STMOM is a
much better hedge in bear markets and has much less negative market exposure in rebounds.

5.5. Volatility and residual turnover

Bandarchuk and Hilscher (2013) show that “characteristic-screens” lead to elevated profits
for conventional momentum strategies because they identify stocks with more volatile past
returns. In particular, they show that the part of share turnover that is unrelated to volatility
no longer has the ability to elevate the profits of conventional momentum strategies. Here,
we show that volatility is not driving our main results.
Table IA.9 in the Appendix shows the performance of strategies that buy the previous
month’s winners and sell the previous month’s losers among stocks with different values
for the previous month’s ‘residual turnover’ (RTO 1,0 ; Panel A) and volatility (σ1,0 ; Panel
B). The underlying portfolios are from double sorts as in Table I, except that the previous
month’s turnover is replaced by either RTO 1,0 or σ1,0 . RTO 1,0 is the residual from cross-
sectional regressions of the previous month’s share turnover on σ1,0 , estimated using WLS
with market equity as weights, and σ1,0 is the standard deviation of the previous month’s
daily stock returns using a minimum of 15 daily observations. Panel A shows that using

37

Electronic copy available at: https://ssrn.com/abstract=3150525

residual turnover does not materially alter our main results: the performance of the resulting
STREV* and STMOM strategies is very similarly to that of their benchmark counterparts
in Table I. Panel B shows that while there is strong one-month reversal among low-volatility
stocks, there is no one-month continuation among high-volatility stocks. In untabulated tests,
we find very similar results using idiosyncratic volatility relative to the FF6 or q-factors.

5.6. Longer formation periods

Our main results use past short-term performance and turnover measured over the previous
month because the conventional short-term reversal effect is based on a one-month formation
period (Jegadeesh, 1990). Here, we show that our results are robust to using longer formation
periods up to 6 months, although the one-month horizon produces the strongest results.
Table IA.10 in the Appendix shows the performance of alternative short-term momen-
tum and short-term reversal* strategies constructed as in Table I, except that the sorting is
on cumulative return and average monthly turnover for the previous 2, . . . , 6 months. The
alternative short-term momentum strategies generate significant average returns between
0.99% and 1.24% per month, although lower and statistically weaker than the 1.37% for the
benchmark strategy. The average returns to the alternative short-term reversal∗ strategies
are decreasing with the formation period and insignificant at the 5 month horizon. Each
alternative strategy is within the univariate span of the corresponding benchmark strategy.

6. Conclusion

Reversal and momentum coexist with almost equal strengths at the one-month horizon.
While last month’s low-turnover stocks exhibit a strong short-term reversal effect (−16.9%
per anmum), last month’s high-turnover stocks exhibit an almost equally strong continuation
effect (+16.4% per annum) which we dub short-term momentum. We show that short-term
momentum is not explained by standard risk factors; that it persists for 12 months after
formation; that it is concentrated among the largest, most liquid stocks; and that it survives
conservative estimates of transaction costs. Finally, we show that our main results are not
limited to the U.S. but also hold in an international sample of 22 developed markets.

38

Electronic copy available at: https://ssrn.com/abstract=3150525

 Our results are difficult to reconcile with the asymmetric information hypothesis of Wang
(1994) and Llorente et al. (2002), which counterfactually predicts that short-term momentum
should be stronger among smaller and less liquid stocks. Instead, we use the ‘cursed expec-
tations equilibrium’ (CEE) model due to Eyster, Rabin, and Vayanos (2019) to show that
our results are inconsistent with rational expectations equilibrium (REE) but are in line with
inefficient use of price-inferred information. In addition, we derive novel testable predictions
that allow us to distinguish the REE and CEE cases in the data. While we adopt the CEE
model due to its flexibility and tractability, it is important to note that the exact specification
of the model can vary substantially without qualitatively affecting the conclusions we draw.
Under CEE, a decrease in the degree of nonspeculative (i.e., liquidity driven) trading
is associated with stronger continuation among high-turnover stocks. This explains why
short-term momentum is concentrated among the largest, most liquid stocks, since trading
in smaller and less liquid stocks is to a greater extent driven by liquidity needs and provision.
In addition, it implies that time periods with lower degrees of nonspeculative trading should
be associated with stronger short-term momentum. We test this prediction by exploiting a
seasonality in the degree of liquidity trading at the turn of the month driven by institutional
cash needs (Etula et al., 2019). We find it is strongly borne out in the data, in that excluding
the values for the last trading days from the two sorting variables (which also introduces an
implementation lag) leads to much stronger short-term momentum.
The model produces additional testable predictions regarding the signs of the effects in
a regression of fundamental innovations on current returns and the interaction of current
returns and volume. Under REE, the interaction effect is negative. However, under CEE,
the interaction effect is positive. Furthermore, under CEE, the direct effect (of the current
return) is negative if information asymmetry is severe, but is otherwise positive. We use
predictive regressions of firms’ fundamental growth rates to show that the data is strongly
in favor of the CEE case and at odds with severe information asymmetry.
We also provide a battery of robustness tests. Most importantly, constructing short-term
momentum with a control for variables related to analysts’ forecasts leads to similar conclu-
sions as our tests of the model’s predictions. In addition, we show that short-term momentum
is not driven by post-earnings announcement drift, industry momentum, or volatility. Lastly,
we show that it exhibits far less crash risk than does conventional momentum.

39

Electronic copy available at: https://ssrn.com/abstract=3150525

 Our findings have a number of implications. First, they help reconcile the seemingly
conflicting evidence in the earlier literature (Conrad et al., 1994; Cooper, 1999). Second,
they present independent evidence for inefficient use of price-inferred information in U.S. and
international stocks. Compared to previous literature that attempts to distinguish the REE
and non-REE cases (e.g., Banerjee, 2011), our tests are novel, as they are based on double
sorts on the previous month’s return and turnover, and, furthermore, predictive regressions
of firms’ fundamental growth rates. Third, while conventional short-term reversal strategies
deliver insignificant returns (Hou et al., 2015, 2018), our short-term reversal* strategy—which
conditions on low turnover in the previous month—delivers highly significant returns and is
a cleaner measure of reversal due to nonspeculative trading. Fourth, our findings challenge
the key stylized that momentum in the cross section of monthly stock returns is confined to
horizons of 12-2 months. Our short-term momentum strategy demonstrates the existence of
significant and persistent continuation at the one-month horizon, albeit confined to stocks
with high turnover in the previous month. According to our model, the returns to these
stocks reflect inefficient use of price-inferred information. Finally, on a more practical level,
short-term momentum is both statistically and economically significant—especially among
large-caps—and is not spanned by standard risk factors. Hence, it offers diversification
benefits relative to standard return factors both in the U.S. and internationally.

40

Electronic copy available at: https://ssrn.com/abstract=3150525

 References
Admati, A. R. (1985). A Noisy Rational Expectations Equilibrium for Multi-Asset Securities Markets.
Econometrica 53 (3), 629–658.

Asness, C. S. (1995). The Power of Past Stock Returns to Explain Future Stock Returns. Working Paper .

Asness, C. S., A. Frazzini, and L. H. Pedersen (2017). Quality Minus Junk. Working Paper.

Asness, C. S., R. B. Porter, and R. Stevens (2000). Predicting Stock Returns Using Industry-Relative Firm
Characteristics. Working Paper (AQR Capital Management).

Avramov, D., T. Chordia, and A. Goyal (2006). Liquidity and Autocorrelations in Individual Stock Returns.
Journal of Finance 61 (5), 2365–2394.

Ball, R., J. Gerakos, J. T. Linnainmaa, and V. V. Nikolaev (2016). Accruals, cash flows, and operating
profitability in the cross section of stock returns. Journal of Financial Economics 121, 28–45.

Bandarchuk, P. and J. Hilscher (2013). Sources of Momentum Profits: Evidence on the Irrelevance of
Characteristics. Review of Finance 17, 809–845.

Banerjee, S. (2011). Learning from Prices and the Dispersion in Beliefs. Review of Financial Studies 24 (9),
3025–3068.
Banerjee, S., R. Kaniel, and I. Kremer (2009). Price Drift as an Outcome of Differences in Higher-Order
Beliefs. Review of Financial Studies 22 (9), 3707–3734.

Banerjee, S. and I. Kremer (2010). Disagreement and Learning: Dynamic Patterns of Trade. Journal of
Finance 65 (4), 1269–1302.

Barberis, N., A. Shleifer, and R. Vishney (1998). A model of investor sentiment. Journal of Finanial
Economics 49 (3), 307–343.

Bollerslev, T., B. Hood, J. Huss, and L. H. Pedersen (2018). Risk Everywhere: Modeling and Managing
Volatility. Review of Financial Studies 31 (7), 2729–2773.

Brennan, M. J. and A. Subrahmanyam (1995). Investment analysis and price formation in securities markets.
Journal of Financial Economics 38 (3), 361–381.

Campbell, J. Y., S. J. Grossman, and J. Wang (1993). Trading volume and serial correlation in stock returns.
Quarterly Journal of Economics 108, 905–939.

Cen, L., K. C. J. Wei, and L. Yang (2017). Disagreement, Underreaction, and Stock Returns. Management
Science 63 (4), 1214–1231.

Conrad, J. S., A. Hameed, and C. Niden (1994). Volume and Autocovariances in Short-Horizon Individual
Security Returns. Journal of Finance 49 (4), 1305–1329.

Cooper, M. (1999). Filter Rules Based on Price and Volume in Individual Security Overreaction. Review of
Financial Studies 12 (4), 901–935.

Corwin, S. A. and J. F. Coughenour (2008). Limited Attention and the Allocation of Effort in Securities
Trading. Journal of Finance 63 (6), 3031–3067.

Cujean, J. and M. Hasler (2017). Why Does Return Predictability Concentrate in Bad Times? Journal of
Finance 72 (6), 2717–2758.

Da, Z., Q. Liu, and E. Schaumburg (2012). A Closer Look at the Short-Term Return Reversal. Management
Science 60 (3), 658–674.

Daniel, K., D. Hirshleifer, and A. Subrahmanyam (1998). Investor Psychology and Security Market under-
and Overreactions. Journal of Finance 53 (6), 1839–1885.

41

Electronic copy available at: https://ssrn.com/abstract=3150525

Daniel, K., D. Hirshleifer, and A. Subrahmanyam (2001). Overconfidence, Arbitrage, and Equilibrium Asset
Pricing. Journal of Finance 56 (3), 921–965.

Daniel, K. and T. J. Moskowitz (2016). Momentum Crashes. Journal of Financial Economics 122 (1),
221–247.
De Bondt, W. and R. Thaler (1985). Does the stock market overreact? Journal of Finance 40, 793–805.

Diamond, D. W. and R. E. Verrecchia (1981). Information Aggregation in a Noisy Rational Expectations

Economy. Journal of Financial Economics 9 (3), 221–235.

Diether, K. B., C. J. Malloy, and A. Scherbina (2002). Differences of Opinion and the Cross Section of Stock
Returns. Journal of Finance 57 (5), 2113–2141.

Easley, D., M. O’Hara, and J. Paperman (1998). Financial analysts and information-based trade. Journal of
Financial Markets 1 (2), 175–201.

Etula, E., K. Rinne, M. Suominen, and L. Vaittinen (2019, May). Dash for Cash: Monthly Market Impact
of Institutional Liquidity Needs. Review of Financial Studies Forthcoming.

Eyster, E. and M. Rabin (2005). Cursed Equilibrium. Econometrica 73 (5), 1623–1672.

Eyster, E., M. Rabin, and D. Vayanos (2019). Financial Markets where Traders Neglect the Informational
Content of Prices. Journal of Finance 71 (1), 371–399.

Fama, E. F. and K. R. French (1993). Common Risk Factors in the Returns on Stocks and Bonds. Journal
of Financial Economics 33, 3–56.

Fama, E. F. and K. R. French (1997). Industry cost of equity. Journal of Financial Economics 43 (2),
153–193.
Fama, E. F. and K. R. French (2015). A five-factor asset pricing model. Journal of Financial Economics 1,
1–22.
Fama, E. F. and K. R. French (2017). International tests of a five-factor asset pricing model. Journal of
Financial Economics 123 (1), 441–463.

Fama, E. F. and J. D. MacBeth (1973). Risk, return, and equilibrium: Empirical tests. Journal of Political
Economy 81, 607–636.

Frazzini, A., R. Israel, and T. J. Moskowitz (2015). Trading Costs of Asset Pricing Anomalies. Working
Paper .

French, K. R. (2008). Presidential address: The cost of active investing. Journal of Finance 63 (4), 1537–1573.

Gao, X. and J. Ritter (2010). The Marketing of Seasoned Equity Offerings. Journal of Financial Eco-
nomics 97 (1), 33–52.

Gargano, A., S. J. Riddiough, and L. Sarno (2019). Foreign Exchange Volume. Working Paper.

Grossman, S. (1976). On the Efficiency of Competitive Stock Markets Where Trades Have Diverse Informa-
tion. Journal of Finance 31 (2), 573–585.

Grossman, S. and J. E. Stiglitz (1980). On the Impossibility of Informationally Efficient Markets. American
Economic Review 70 (3), 393–408.

Harrison, M. and D. M. Kreps (1978). Speculative Investor Behavior in a Stock Market with Heterogeneous
Expectations. The Quarterly Journal of Economics 92 (2), 323–336.

Harvey, C. R. and A. Siddique (2000). Conditional Skewness in Asset Pricing Tests. Journal of Finance 55 (3),
1263–1296.
Hasbrouck, J. (2009). Trading Costs and Returns for U.S. Equities: Estimating Effective Costs from Daily
Data. Journal of Finance 64 (3), 1445–1477.

42

Electronic copy available at: https://ssrn.com/abstract=3150525

Hellwig, M. (1980). On the aggregation of information in competitive markets. Journal of Economic The-
ory 22 (3), 477–498.

Hendershott, T. and M. S. Seasholes (2014). Liquidity provision and stock return predictability. Journal of
Banking and Finance 45, 140–151.

Henriksson, R. D. and R. C. Merton (1981). On Market Timing and Investment Performance. II. Statistical
Procedures for Evaluating Forecasting Skills. Journal of Business 54 (4), 513–533.

Hong, H., T. Lim, and J. C. Stein (2000). Bad News Travels Slowly: Size, Analyst Coverage, and the
Profitability of Momentum Strategies. Journal of Finance 55 (1), 265–295.

Hong, H. and J. C. Stein (1999). A Unified Theory of Underreaction, Momentum Trading, and Overreaction
in Asset Markets. Journal of Finance 54 (6), 2143–2184.

Hong, H. and J. C. Stein (2007). Disagreement and the Stock Market. Journal of Economic Perspec-
tives 21 (2), 109–128.

Hou, K., C. Xue, and L. Zhang (2015). Digesting Anomalies: An Investment Approach. Review of Financial
Studies 28 (3), 650–705.

Hou, K., C. Xue, and L. Zhang (2018). Replicating Anomalies. Review of Financial Studies Studies Forth-
coming.

Jegadeesh, N. (1990). Evidence of predictable behavior of security returns. Journal of Finance 45, 881–898.

Jegadeesh, N. and S. Titman (1993). Returns to buying winners and selling losers: Implications for stock
market efficiency. Journal of Finance 48 (1), 65–91.

Jegadeesh, N. and S. Titman (2001). Profitability of momentum strategies: An evaluation of alternative

explanations. Journal of Finance 56, 699–720.

Johnson, T. C. (2004). Forecast Dispersion and the Cross Section of Expected Returns. Journal of Fi-
nance 59 (5), 1957–1978.

Koijen, R., T. J. Moskowitz, L. H. Pedersen, and E. B. Vrugt (2018). Carry. Journal of Financial Eco-
nomics 127, 197–225.

Kyle, A. S. (1985). Continuous auctions and insider trading. Econometrica 53 (6), 1315–1335.

Lee, C. M. C. and B. Swaminathan (2000). Price momentum and trading volume. Journal of Finance 55,
2017–2069.
Llorente, G., R. Michaely, G. Saar, and J. Wang (2002). Dynamic volume-return relation of individual stocks.
Review of Financial Studies 15, 1005–1047.

Lo, A. W. and J. Wang (2000). Trading Volume: Definitions, Data Analysis, and Implications of Portfolio
Theory. Review of Financial Studies 13 (2), 257–300.

Loh, R. K. and R. M. Stulz (2018). Is Sell-Side Research More Valuable in Bad Times? Journal of
Finance 73 (3), 959–1013.

Medhat, M. and M. Schmeling (2018). Dissecting announcement returns. Working Paper.

Milgrom, P. and N. Stokey (1982). Information, Trade and Common Knowledge. Journal of Economic
Theory 26 (1), 17–27.

Moskowitz, T. J. and M. Grinblatt (1999). Do Industries Explain Momentum? Journal of Finance 54 (4),
1249–1290.
Nagel, S. (2012). Evaporating liquidity. Review of Financial Studies 25, 2005–2039.

Newey, W. K. and K. D. West (1987). A Simple, Positive Semi-Definite, Heteroskedasticity and Autocorre-
lation Consistent Covariance Matrix. Econometrica 55, 703–708.

43

Electronic copy available at: https://ssrn.com/abstract=3150525

Novy-Marx, R. (2013). The Other Side of Value: The Gross Profitability Premium. Journal of Financial
Economics 108, 1–28.

Novy-Marx, R. and M. Velikov (2015). A Taxonomy of Anomalies and Their Trading Costs. Review of
Financial Studies 29 (1), 104–147.

Odean, T. (1998). Volume, Volatility, Price, and Profit When All Traders Are Above Average. Journal of
Finance 53 (6), 1887–1934.

Pástor, L. and R. F. Stambaugh (2003). Liquidity Risk and Expected Stock Returns. Journal of Political
Economy 111 (3), 642–685.

Peng, L. and W. Xiong (2006). Investor attention, overconfidence and category learning. Journal of Financial
Economics 80 (3), 563–602.

Scheinkman, J. A. and W. Xiong (2003). Overconfidence and Speculative Bubbles. Journal of Political
Economy 111 (6), 1183–1219.

Thaler, R. (1988). Anomalies: The Winner’s Curse. Journal of Economic Perspectives 2 (1), 191–202.

Tirole, J. (1982). On the Possibility of Speculation under Rational Expectations. Econometrica 50 (5),
1163–1181.
Van Nieuwerburgh, S. and L. Veldkamp (2010). Information Acquisition and Under-Diversification. Review
of Economic Studies 77 (2), 779–805.

Vayanos, D. and J. Wang (2012). Liquidity and Asset Returns Under Asymmetric Information and Imperfect
Competition. Review of Financial Studies 25 (5), 1340–1365.

Verardo, M. (2009). Heterogeneous Beliefs and Momentum Profits. Journal of Financial and Quantitative
Analysis 44 (4), 795–822.

Wang, J. (1994). A model of competitive stock trading volume. Journal of Political Economy 102, 127–168.

Watanabe, M. (2008). Price Volatility and Investor Behavior in an Overlapping Generations Model with
Information Asymmetry. Journal of Finance 63 (1), 229–272.

44

Electronic copy available at: https://ssrn.com/abstract=3150525

 Internet Appendix for

Short-term Momentum
(not for publication)

IA – 1

Electronic copy available at: https://ssrn.com/abstract=3150525

 IA.1. Derivations and proofs

IA.1.1. Model solution

The traders’ optimal demands of the risky asset are

d + E[d | {s, δ, p1 }] − p1 d + (1 − χ)E[d | {s, p1 }] + χE[d | s] − p1

xI = −z and xU =
. (IA.1)
α(Vde + V[d | {s, δ, p1 }]) α Vde + (1 − χ)V[d | {s, p1 }] + χV[d | s]

I-traders receive both the public and private signals, s and δ, so p1 is redundant in their information
set. Their posteriors about d are

1
= VbdI and E[d | {s, δ}] = VbdI Vs−1 s + Vδ−1 δ .


V[d | {s, δ}] = 1 1 1
Vd + Vs + Vδ

U -traders observe s and put a weight of χ on their beliefs conditional on s alone and a weight of
1 − χ on their beliefs conditional on s and p1 . Their posteriors conditional on s alone are

1
V[d | s] = 1 1 = VbdU,s and E[d | s] = VbdU,s Vs−1 s.
Vd + Vs

For their posteriors conditional on s and p1 , U -traders conjecture the linear form

p1 = A0 + As s + Aδ δ + Az z, (IA.2)

and invert it to infer an additional unbiased signal about d ,

p1 − A0 − As s
pe = = δ + Az
e = d + p ,
Aδ

Az e2 Vz . Hence,
e ∼ N (0, Vp ) with Vp = Vδ + A
where A
e=
Aδ and where p = δ + Az

1
= VbdU,s,p and E[d | {s, p1 }] = VbdU,s,p Vs−1 s + Vp−1 pe .


V[d | {s, p1 }] = 1 1 1
Vd + Vs + Vp

Given I- and U -traders’ posteriors, the market clearing condition becomes

 
p1
ιz = ι d+E[d | {s,δ}]
I) + (1 − ι) d+(1−χ)E[d | {s,p 1 }]+χE[d | s]
U,s,p U,s − ι
I + 1−ι
U,s,p U,s .
α(Vde +Vbd α(Vde +(1−χ)Vbd +χVbd ) α Vde +Vbd Vde +(1−χ)Vbd +χVbd

IA – 2

Electronic copy available at: https://ssrn.com/abstract=3150525

  −1
Setting VdU = (1 − χ)VbdU,s,p + χVb U,s and Vd = ι
I + 1−ι
U
, we have
Vde +Vbd Vde +Vd

 
p1 = d + κVbdI + (1 − κ)VdU Vs−1 s + κVbdI Vδ−1 δ + (1 − κ)(1 − χ)VbdU,s,p Vp−1 pe − αVd ιz,

ι 1−ι
where κ = I Vd ∈ (0, 1) and therefore 1 − κ = V .
U d
Matching coefficients in Eq. (IA.2),
Vde+Vbd Vde +Vd

A0 = d, (IA.3)
 
As = κVbdI + (1 − κ)VdU Vs−1 ∈ (0, 1), (IA.4)

Aδ = κVbdI Vδ−1 + (1 − κ)(1 − χ)VbdU,s,p Vp−1 ∈ (0, 1), (IA.5)

Az = (1 − κ)(1 − χ)VbdU,s,p Vp−1 A

e − ιαVd . (IA.6)

bI
Solving the equation A
e= Az e = −α Vde +Vd Vδ < 0, and therefore Az < 0.
finally yields A
Aδ bI Vd
Note that Vd is the weighted harmonic mean of the posterior variances of d across traders and
that we can write Vd = κ(Vde + VbdI ) + (1 − κ)(Vde + VdU ). More generally, if we define the operators

 
Eχ [ · ] = κE[ · | {s, δ}] + (1 − κ) (1 − χ)E[ · | {s, p1 }] + χE[ · | s] , (IA.7)
 
Vχ [ · ] = κV[ · | {s, δ}] + (1 − κ) (1 − χ)V[ · | {s, p1 }] + χV[ · | s] , (IA.8)

then we can write p1 = Eχ [ d ] − αVχ [ d ]ιz. The equilibrium price thus reflects the dividend’s
posterior expectation and variance from the viewpoint of the (hypothetical) ‘average’ trader who
clears the market. The ‘average’ trader puts a weight of κ on the beliefs of I-traders, a weight
of 1 − κ on the beliefs of U -traders, and, among U -traders, a weight of 1 − χ on the beliefs that
condition on p1 and a weight of χ on the beliefs that do not condition on p1 .
The equilibrium features underreaction to private information whenever χ > 0. To see this, let
Ape = (1 − κ)(1 − χ)VbdU,s,p Vp−1 ∈ (0, 1) be the dependence of p1 on pe. Then
 
∂Ape ∂κ
= −VbdU,s,p Vp−1 (1 − χ) + (1 − κ) < 0,
∂χ ∂χ

∂κ ι b U,s −Vb U,s,p

because ∂χ = 1−ι (1 − κ)2 V > 0. Hence, Ape decreases monotonically towards zero with χ,
V e +Vb I
d d
is maximal for χ = 0 (the REE case), and is minimal for χ = 1 (the full negligence case). As a
result, for any χ > 0, the price underreacts to private information relative to the REE case.

IA – 3

Electronic copy available at: https://ssrn.com/abstract=3150525

IA.1.2. Informational negligence and the autocorrelation in returns

Let r2 = d − p1 and r1 = p1 − d be the second- and first-period returns. We now show that return
autocorrelation is always negative under REE but is positive when χ > 0 given sufficiently low Vz .
We start by writing p1 in terms of the independent shocks. Using Eq. (IA.2) and the price
coefficients {A0 , As , Aδ , Az }, we can write

p1 = A0 + Ad d + As s + Aδ δ + Az z, (IA.9)

where the only new coefficient, Ad = As + Aδ , is given by

 
Ad = κVbdI Vs−1 + Vδ−1 + (1 − κ) (1 − χ)VbdU,s,p Vs−1 + Vp−1 + χVbdU,s Vs−1 . (IA.10)

Note that Ad ∈ (0, 1) since it is a weighted average of the posterior precisions of d .
Using Eqs. (IA.7)-(IA.8), we have

h  i
E[r2 | r1 ] = E[d − p1 | p1 ] = E d − Eχ [ d ] + αVχ [ d ]ιz  p1


h  i
= αVd ι E [z | p1 ] + E d − κE[ d | s, δ] − (1 − κ)E[ d | s, p1 ]  p1

h  i
+ χ(1 − k)E E [d | s, p1 ] − E [d | s]  p1


COV[z, p1 ] h  i
= αVd ι (p1 − d) + E d − E[ d | s, p1 ]  p1

V[p1 ]
 
+χ(1 − k) VbdU,s,p Vs−1 E [s | p1 ] + Vp−1 E [e
p | p1 ] − VbdU,s Vs−1 E [s | p1 ]

αVd ιAz Vz + χ(1 − κ)∆U

= r1 ,
V[r1 ]

where have repeatedly applied the law of iterated expectations and defined

  
A2z
∆U = VbdU,s,p Vs−1 (Ad Vd + As Vs ) + Vp−1 Ad Vd + Aδ Vδ + Aδ V z − VbdU,s Vs−1 (Ad Vd + As Vs ).

Since r2 and r1 are jointly normal, we must have COV[r2 , r1 ] = αVd ιAz Vz + χ(1 − κ)∆U . Note
that αVd ιAz Vz ≤ 0 for Vz ≥ 0 with equality for Vz = 0, but that ∆U > 0 for any Vz . Hence,
COV[r2 , r1 ] ≤ 0 for Vz ≥ 0 under REE (χ = 0) with equality for Vz = 0. If χ > 0, we have
COV[r2 , r1 ] > 0 for Vz = 0. By continuity, χ > 0 implies COV[r2 , r1 ] > 0 given sufficiently low Vz .

IA – 4

Electronic copy available at: https://ssrn.com/abstract=3150525

IA.1.3. An auxiliary result for conditional expectations given returns and volume

This appendix states and proves an auxiliary result about the conditional expectation of a normal
variable, X1 , given another normal variable, X2 , as well as the absolute value of a third normal
variable, |X3 |. We will use this result in the proofs of Propositions 1 and 2. Given a matrix and a
column vector, M and v, we use the brackets notation, as in M [j:k,l:m] or v [j] , to denote subelements
of M and v. In particular, M [j,k] and v [j] are scalars while M [j,k:l] is a row vector.

Lemma 1 (Wang, 1994). Let (X1 , X2 , X3 )0 be a 3-dimensional vector of jointly normally distributed
random variables with mean zero and covariance matrix Σ, and define W = |X3 |. Then

 
E X1  {X2 , W } = {x2 , w} = β [1] x2 − tanh Γ[1,2] wx2 β [2] w
  
(IA.11)

−1
for any (x2 , w) ∈ R × (0, ∞), where β = Σ[1,2:3] Γ and Γ = Σ[2:3,2:3] .

Proof of Lemma. Fix a point (x1 , x2 , w) ∈ R2 ×(0, ∞). Then the distribution function of (X1 , X2 , W )
is given by

FX1 ,X2 ,W (x1 , x2 , w) = P (X1 ≤ x1 , X2 ≤ x2 , W ≤ w)

= P (X1 , X2 , X3 ) ∈ (−∞, x1 ) × (−∞, x2 ) × (−w, w)
Z x1 Z x2 Z w
= φ(x̃1 , x̃2 , x3 ; 0, Σ) dx3 dx̃2 dx̃1 ,
−∞ −∞ −w

where φ( · ) is the density of the multivariate normal distribution. Let

 
1 0 0
 
J = 0 1 0  ,
 
 
0 0 −1

and note that Ω = JΣJ 0 is the covariance matrix of (X1 , X2 , −X3 )0 . Then (X1 , X2 , W ) has density

∂ 3 FX1 ,X2 ,W
fX1 ,X2 ,W (x1 , x2 , w) = = φ (x1 , x2 , w ; 0, Σ) + φ (x1 , x2 , −w ; 0, Σ)
∂w∂x2 ∂x1
= φ (x1 , x2 , w ; 0, Σ) + φ J × (x1 , x2 , w)0 ; 0, Σ

= φ (x1 , x2 , w ; 0, Σ) + φ (x1 , x2 , w ; 0, Ω) ,

IA – 5

Electronic copy available at: https://ssrn.com/abstract=3150525

where we have used that J = J −1 and that det Ω = det Σ. By the marginalization property of the
multivariate normal distribution, (X2 , W ) has density

Z ∞    
fX2 ,W (x2 , w) = fX1 ,X2 ,W (x1 , x2 , w) dx1 = φ x2 , w ; 0, Σ[2:3,2:3] + φ x2 , w ; 0, Ω[2:3,2:3] .
−∞

Define the conditional densities

φ(x1 , x2 , w ; 0, Σ) φ(x1 , x2 , w ; 0, Ω)
φ(x1 | x2 , w ; 0, Σ) =
and φ(x1 | x2 , w ; 0, Ω) =
.
φ x2 , w ; 0, Σ[2:3,2:3] φ x2 , w ; 0, Ω[2:3,2:3]

Then the conditional expectation can be written as

Z ∞
   fX1 ,X2 ,W (x1 , x2 , w)
E X1  {X2 , W } = {x2 , w} = x1 dx1
−∞ fX2 ,W (x2 , w)

φ x2 , w ; 0, Σ[2:3,2:3] (β [1] x2 + β [2] w) + φ x2 , w ; 0, Ω[2:3,2:3] (β [1] x2 − β [2] w)

=


φ x2 , w ; 0, Σ[2:3,2:3] + φ x2 , w ; 0, Ω[2:3,2:3]

φ x2 , w ; 0, Σ[2:3,2:3] − φ x2 , w ; 0, Ω[2:3,2:3] [2]

[1]
= β x2 +

β w,
φ x2 , w ; 0, Σ[2:3,2:3] + φ x2 , w ; 0, Ω[2:3,2:3]

where β is defined in the lemma and where we have used that

 −1  0
Ω[1,2:3] Ω[2:3,2:3] = J [2:3,2:3] β = β [1] , −β [2] .

Using the definition of Γ, we can write

1 [1,1] x2 +Γ[2,2] w 2 +2Γ[1,2] wx


[2:3,2:3]
 e − 2 (Γ 2 2 )
φ x2 , w ; 0, Σ = 1 ,
(2π det Σ[2:3,2:3] ) 2

1 [1,1] x2 +Γ[2,2] w 2 −2Γ[1,2] wx


[2:3,2:3]
 e − 2 (Γ 2 2 )
φ x2 , w ; 0, Ω = 1 ,
(2π det Σ[2:3,2:3] ) 2

where we have used that det Ω[2:3,2:3] = det Σ[2:3,2:3] . Hence,

[1,2] [1,2]
  [1] e−Γ wx2 − eΓ wx2 [2]
E X1 {X2 , W } = {x2 , w} = β x2 +

[1,2] [1,2]
β w
e−Γ wx2 + eΓ wx2
 
= β [1] x2 − tanh Γ[1,2] wx2 β [2] w,

which completes the proof.

IA – 6

Electronic copy available at: https://ssrn.com/abstract=3150525

IA.1.4. Proof of Proposition 1

Proof. Part 1 is an application of Lemma 1 with (X1 , X2 , X3 )0 = (r2 , r1 , 2ιxI )0 . Note that this
vector consists of jointly normally distributed variables with mean zero. The joint normality follows
because the variables are linear functions of {d , d,
e s , δ , z}, which are normally distributed and

mutually independent. Hence, if Σ is the vector’s covariance matrix, we immediately have

 
E [ r2 | {r1 , v1 }] = β [1] r1 − tanh Γ[1,2] v1 r1 β [2] v1 , (IA.12)

−1
where β = Σ[1,2:3] Γ and Γ = Σ[2:3,2:3] . It thus suffices to determine Σ, β, and Γ[1,2] .
To simplify the expressions for the elements of Σ, we also write xI in terms of the independent
shocks. Using Eqs. (IA.1) and (IA.9) and the price coefficients {A0 , Ad , As , Aδ , Az }, we can write

xI = Bd d + Bs s + Bδ δ + Bz z, (IA.13)

where the coefficients {Bd , Bs , Bδ , Bz } are given by

 
Bd = C −1 VbdI Vs−1 + Vδ−1 − Ad > 0,


(IA.14)
 
Bs = C −1 VbdI Vs−1 − As < 0, (IA.15)
 
Bδ = C −1 VbdI Vδ−1 − Aδ > 0, (IA.16)

Bz = −C −1 (Az + C) < 0, (IA.17)

with C = α(Vde + VbdI ) > 0, and where we have used that

(1 − χ)VbdU,s,p Vs−1 + Vp−1 + χVbdU,s Vs−1 < VbdU,s,p Vs−1 + Vp−1 < VbdI Vs−1 + Vδ−1 .

Using Eqs. (IA.9) and (IA.13), Σ has diagonal elements

Σ[1,1] = V[r2 ] = Vde + (1 − Ad )2 Vd + A2s Vs + A2δ Vδ + A2z Vz ,

Σ[2,2] = V[r1 ] = A2d Vd + A2s Vs + A2δ Vδ + A2z Vz ,

Σ[3,3] = V[2ιxI ] = 4ι2 Bd

2 2 2
Vδ + Bz2 Vz ,


Vd + Bs Vs + Bδ

IA – 7

Electronic copy available at: https://ssrn.com/abstract=3150525

and off-diagonal elements

Σ[1,2] = COV[r2 , r1 ] = (1 − Ad )Ad Vd − A2s Vs + A2δ Vδ + A2z Vz ,

Σ[1,3] = COV[r2 , 2ιxI ] = 2ι (1 − Ad )Bd Vd − (As Bs Vs + Aδ Bδ Vδ + Az Bz Vz ) ,

Σ[2,3] = COV[r1 , 2ιxI ] = 2ι Ad Bd Vd + As Bs Vs + Aδ Bδ Vδ + Az Bz Vz .

We can now determine β and Γ[1,2] . First, note that we can write

1  0
[3,3] [1,2] [2,3] [1,3] [2,2] [1,3] [2,3] [1,2]
β= Σ Σ − Σ Σ , Σ Σ − Σ Σ ,
Σ[2,2] Σ[3,3] − (Σ[2,3] )2

where Σ[2,2] Σ[3,3] − (Σ[2,3] )2 = det Γ > 0. Direct computation gives that

Σ[3,3] Σ[1,2] − Σ[2,3] Σ[1,3] = 0, (IA.18)

which means that r2 and r1 are conditionally independent given xI . Hence, β [1] = 0, which readily
Σ[1,3] 1 COV[r2 ,xI ]
implies β [2] = Σ[3,3]
= 2ι V[xI ] . Furthermore, we have

−Σ[2,3] 1 −COV[r1 , xI ]
Γ[1,2] = = < 0,
[2,2]
Σ Σ [3,3] [2,3]
− (Σ ) 2 2ι V[r1 ]V[xI ] − COV[r1 , xI ]2

where the negativity follows by noting that we can write

COV[r1 , xI ] = Aδ (Bd Vd + Bδ Vδ ) + Az Bz Vz > 0, (IA.19)

since Bd Vd + Bs Vs = 0. Hence, part 1 follows by setting βv = β [2] , γvr = −Γ[1,2] > 0, and using
that − tanh(x) = tanh(−x) for any x.
For part 2, we have

γvr βv v12 r1 + O (v1 r1 )3

E [ r2 | {r1 , v1 }] = tanh(γvr v1 r1 )βv v1 =
 
γvr βv v 21 + 2γvr βv v 1 (v1 − v 1 ) r1 + O (v1 − v 1 )2 r1 + O (v1 r1 )3



=

−γvr βv v 21 r1 + (2γvr βv v 1 ) v1 r1 + O (v1 − v 1 )2 r1 ,

=

q q
2Σ[3,3] 8V[xI ]
where v 1 = E[v1 ] = π = ι π is expected volume. Hence, part 2 follows by setting
ψr = −γvr βv v1 2 and ψvr = 2γvr βv v1 .

IA – 8

Electronic copy available at: https://ssrn.com/abstract=3150525

 For part 3, note that ψvr and ψr have opposite signs, and that the sign of ψvr equals that of
COV[r2 ,2ιxI ]
βv = V[2ιxI ] . Using Eqs. (IA.18), we have

Σ[3,3] Σ[1,2] V[r1 ]COV[r2 , r1 ]

COV[r2 , 2ιxI ] = [2,3]
= .
Σ COV[r2 , 2ιxI ]

Since COV[r2 , 2ιxI ] > 0 by Eq. (IA.19), the sign of βv in turn equals that of the autocovariance in
returns, COV[r2 , r1 ]. Appendix IA.1.2 therefore implies that (i) under REE, it holds that βv ≤ 0
for any Vz ≥ 0 with equality for Vz = 0, and (ii) for χ > 0, it holds that βv > 0 given sufficiently
low Vz . This proves part 3 and completes the proof.

IA.1.5. Proof of Proposition 2

0
Proof. We apply Lemma 1 with (X1 , X2 , X3 )0 = d − d, r1 , 2ιxI , which consists of jointly normally
distributed variables with mean zero. Let Σ
e be this vector’s covariance matrix. Furthermore, let

Σ be the covariance matrix of (r2 , r1 , 2ιxI )0 , as in the proof of Proposition 1. Note that we have
e [2:3,2:3] = Σ[2:3,2:3] . We can therefore write
Σ

E d − d  {r1 , v1 } = βe[1] r1 − tanh (−γvr v1 r1 ) βe[2] v1 ,

  
(IA.20)

−1
e [1,2:3] Γ with Γ = Σ[2:3,2:3]
where βe = Σ and where γvr = −Γ[1,2] > 0 as in Proposition 1 and its
proof. The remaining elements of Σ
e are

e [1,1] = V[d − d] = V e + Vd ,

Σ d

e [1,2] = COV[d − d, r1 ] = Ad Vd ,

Σ
e [1,3] = COV[d − d, 2ιxI ] = 2ιBd Vd .
Σ

1 COV[d,r1 ]V[xI ]−COV[r1 ,xI ]COV[d,xI ] 1 COV[d,xI ]V[r1 ]−COV[r1 ,xI ]COV[d,r1 ]
It follows that βe[1] = 2ι V[r1 ]V[xI ]−COV[r1 ,xI ]2
and βe[2] = 2ι V[r1 ]V[xI ]−COV[r1 ,xI ]2
.
Defining βer = βe[1] and βev = βe[2] , we have

 
E d − d  {r1 , v1 } = βer r1 + tanh (γvr v1 r1 ) βev v1 = βer + γvr βev v12 r1 + O (v1 r1 )3
  

   
βer − γvr βev v 21 r1 + 2γvr βev v 1 v1 r1 + O (v1 − v 1 )2 r1 ,


=

where v 1 = E[v1 ]. Hence, the proposition follows with ψer = βer − γvr βev v1 2 and ψevr = 2γvr βev v1 .

IA – 9

Electronic copy available at: https://ssrn.com/abstract=3150525

 IA.2. Additional results

Table IA.1: Portfolio characteristics for double sorts on one-month returns and turnover.
This table shows time-series averages of portfolio characteristics for the double sorts on previous month’s
return (r1,0 ) and turnover (TO 1,0 ) in Table I. Portfolio r1,0 is the time-series average of each portfolios’
monthly value-weighted r1,0 , and similarly for portfolio TO 1,0 . Average market capitalization is the time-
series average of each portfolio’s monthly equal-weighted average market capitalization in millions of dollars.
Average number of firms is the time-series average of each portfolios’ monthly number of firms. The sample
excludes financial firms. Data are monthly and cover July 1963 to December 2018.

r1,0 deciles

TO 1,0 deciles Low 2 3 4 5 6 7 8 9 High

Portfolio r1,0
Low −0.14 −0.07 −0.04 −0.02 0.00 0.02 0.04 0.06 0.09 0.19
2 −0.14 −0.07 −0.04 −0.02 0.00 0.02 0.04 0.06 0.09 0.18
3 −0.14 −0.07 −0.04 −0.02 0.00 0.02 0.04 0.06 0.09 0.17
4 −0.13 −0.07 −0.04 −0.02 0.00 0.02 0.04 0.06 0.09 0.18
5 −0.14 −0.07 −0.04 −0.02 0.00 0.02 0.04 0.06 0.09 0.18
6 −0.14 −0.07 −0.04 −0.02 0.00 0.02 0.04 0.06 0.09 0.18
7 −0.14 −0.07 −0.04 −0.02 0.00 0.02 0.04 0.06 0.10 0.19
8 −0.14 −0.08 −0.04 −0.02 0.00 0.02 0.04 0.06 0.10 0.20
9 −0.16 −0.08 −0.04 −0.02 0.00 0.02 0.04 0.06 0.10 0.22
High −0.18 −0.08 −0.04 −0.02 0.00 0.02 0.04 0.06 0.10 0.28

Portfolio TO 1,0
Low 0.02 0.02 0.02 0.02 0.02 0.02 0.02 0.02 0.02 0.02
2 0.04 0.03 0.03 0.03 0.03 0.03 0.03 0.04 0.04 0.05
3 0.06 0.05 0.04 0.04 0.04 0.04 0.04 0.04 0.05 0.06
4 0.07 0.06 0.05 0.05 0.05 0.05 0.05 0.05 0.06 0.08
5 0.09 0.07 0.06 0.06 0.06 0.06 0.06 0.06 0.07 0.10
6 0.11 0.08 0.07 0.07 0.07 0.07 0.07 0.07 0.08 0.12
7 0.14 0.10 0.09 0.08 0.08 0.08 0.08 0.09 0.10 0.14
8 0.17 0.12 0.10 0.10 0.09 0.09 0.10 0.11 0.12 0.18
9 0.23 0.15 0.13 0.12 0.12 0.12 0.13 0.13 0.16 0.24
High 0.43 0.27 0.24 0.22 0.21 0.21 0.22 0.24 0.28 0.47

Average market capitalization ($ mio.)

Low 176 414 803 1,003 985 1,174 1182 1,227 952 364
2 530 1,880 2,957 3,642 4,806 4,167 4,212 3,866 2,789 994
3 711 2,169 3,140 4,219 4,481 4,498 4,439 4,117 3,340 1,290
4 930 2,275 3,262 3,676 3,753 4,009 3,945 3,875 3,116 1,265
5 905 2,376 2,815 3,445 3,753 3,566 3,842 3,550 3,016 1,396
6 934 2,485 2,829 3,485 3,499 3,335 3,391 3,411 2,654 1,284
7 923 2,142 2,488 3,010 3,259 3,347 3,167 2,990 2,389 1,206
8 925 1,886 2,198 2,660 2,811 2,805 2,858 2,565 2,154 1,158
9 855 1,675 1,984 2,079 2,180 2,496 2,364 2,151 1,919 1,132
High 722 1,403 1,589 1,697 1,706 1,756 1,667 1,779 1,511 898

Average number of firms

Low 177 95 83 74 68 66 63 66 72 124
2 78 41 34 30 28 27 27 29 33 59
3 58 32 27 24 23 23 22 24 27 47
4 49 29 24 22 21 21 21 22 24 42
5 44 26 22 21 20 20 20 21 23 39
6 39 24 21 20 19 19 20 20 22 36
7 37 24 21 20 20 19 19 20 22 36
8 36 24 22 20 20 20 20 21 23 37
9 35 24 22 21 21 21 21 22 24 37
High 41 29 27 26 26 26 26 27 29 47

IA – 10

Electronic copy available at: https://ssrn.com/abstract=3150525

Table IA.2: Short-term momentum’s factor exposures and abnormal returns relative to the
q-factors. This table shows time-series regression results for the STMOM and STREV* strategies. The
explanatory variable are the factors from Hou, Xue, and Zhang’s (2015) q-factor model. Test-statistics (in
parentheses) are adjusted for heteroscedasticity and autocorrelation. Data are monthly and cover January
1967 to December 2018, where the start date is determined by the availability of the q-factors.

Intercepts, slopes, and test-statistics (in parentheses)

from time-series regressions of the form yt = α + β 0 Xt + t

Short-term momentum Short-term reversal*

Independent (1) (2) (3) (4) (5) (6)

variable

Intercept 1.42 1.62 1.65 −1.38 −1.25 −1.43

(4.74) (5.50) (4.47) (−6.57) (−6.07) (−5.95)
MKT −0.39 −0.39 −0.26 −0.17
(−4.55) (−3.91) (−3.79) (−2.65)
ME 0.03 −0.21
(0.15) (−1.75)
ROE −0.19 0.32
(−0.88) (2.43)
I/A 0.17 0.06
(0.62) (0.30)
Adj. R2 4.2% 4.3% 4.0% 7.5%

IA – 11

Electronic copy available at: https://ssrn.com/abstract=3150525

 Table IA.3: Independent double sorts on one-month returns and turnover. This table shows portfolios from independent double sorts on the
previous month’s return (r1,0 ) and turnover (TO 1,0 ). The sorts are into deciles based on NYSE breakpoints and the resulting portfolios are value-weighted
and rebalanced at the end of each month. The table also shows the performance of long-short strategies across the deciles. Test-statistics (in parentheses)
are adjusted for heteroscedasticity and autocorrelation. Data are monthly and cover July 1969 to December 2018, where the start date is to ensure
non-empty portfolios as a result of the independent sorts.

r1,0 deciles r1,0 strategies

Low 2 3 4 5 6 7 8 9 High E[re ] αFF6 αq

TO 1,0 deciles
Portfolio excess return
Low 1.28 1.28 0.83 0.93 0.63 1.01 0.78 0.74 0.38 −0.14 −1.42 −1.53 −1.57
(−6.11) (−5.95) (−6.27)
2 1.71 1.06 1.02 0.95 1.05 0.88 0.75 0.73 0.64 −0.03 −1.74 −1.76 −1.73
(−5.41) (−4.61) (−4.51)
3 1.36 1.27 1.27 1.02 1.16 1.00 0.59 0.75 0.36 0.47 −0.89 −1.05 −1.15
(−2.54) (−2.75) (−3.04)
4 1.62 1.46 1.40 1.36 0.90 1.06 0.82 0.59 0.82 0.09 −1.53 −1.82 −1.73
(−5.15) (−5.88) (−4.90)

IA – 12
5 1.52 1.55 1.38 1.02 1.21 0.86 0.86 0.98 0.61 0.46 −1.06 −1.29 −1.37
(−3.22) (−3.00) (−3.37)
6 1.55 1.10 1.32 1.24 0.82 0.84 0.99 1.07 0.77 0.72 −0.83 −0.96 −0.90
(−2.59) (−2.63) (−2.36)
7 1.13 1.01 1.17 0.95 1.04 1.21 0.92 0.98 0.80 0.49 −0.64 −0.77 −0.67
(−3.10) (−2.73) (−2.39)
8 1.33 1.10 1.14 1.06 1.19 1.18 0.96 0.84 0.88 0.79 −0.55 −0.43 −0.40
(−2.07) (−1.39) (−1.18)
9 0.94 1.10 1.17 1.24 1.14 0.95 1.05 1.26 1.00 0.69 −0.25 −0.31 −0.23
(−0.91) (−1.07) (−0.56)
High 0.12 0.78 0.98 1.00 0.86 0.48 0.89 0.96 0.90 1.12 1.00 0.98 1.06
(4.02) (3.40) (3.31)
TO 1,0 strategies

Electronic copy available at: https://ssrn.com/abstract=3150525

E[re ] −1.16 −0.50 0.15 0.08 0.23 −0.53 0.10 0.23 0.52 1.26
(−4.63) (−1.93) (0.46) (0.23) (0.68) (−1.72) (0.35) (0.75) (1.72) (5.06)
αFF6 −1.15 −0.34 0.40 0.07 0.33 −0.35 0.28 0.34 0.74 1.36
(−4.47) (−1.21) (1.33) (0.23) (1.04) (−1.22) (1.01) (1.17) (2.73) (5.26)
αq −1.18 −0.16 0.38 0.04 0.31 −0.38 0.23 0.26 0.78 1.46
(−4.38) (−0.59) (1.14) (0.11) (0.87) (−1.18) (0.77) (0.74) (2.45) (5.19)
Table IA.4: Portfolio characteristics for independent double sorts on one-month returns and
turnover. This table shows time-series averages of portfolio characteristics for the independent double
sorts on previous month’s return (r1,0 ) and turnover (TO 1,0 ) in Table IA.3. Portfolio r1,0 is the time-
series average of each portfolios’ monthly value-weighted r1,0 , and similarly for portfolio TO 1,0 . Average
market capitalization is the time-series average of each portfolio’s monthly equal-weighted average market
capitalization in millions of dollars. Average number of firms is the time-series average of each portfolios’
monthly number of firms. Data are monthly and cover July 1969 to December 2018, where the start date is
to ensure non-empty portfolios as a result of the independent sorts.

r1,0 deciles

TO 1,0 deciles Low 2 3 4 5 6 7 8 9 High

Portfolio r1,0
Low −0.15 −0.08 −0.05 −0.02 0.00 0.02 0.04 0.06 0.10 0.20
2 −0.14 −0.08 −0.05 −0.02 0.00 0.02 0.04 0.06 0.10 0.19
3 −0.14 −0.08 −0.05 −0.02 0.00 0.02 0.04 0.06 0.10 0.19
4 −0.14 −0.08 −0.05 −0.02 0.00 0.02 0.04 0.06 0.10 0.18
5 −0.14 −0.08 −0.05 −0.02 0.00 0.02 0.04 0.06 0.10 0.18
6 −0.14 −0.08 −0.05 −0.02 0.00 0.02 0.04 0.06 0.10 0.18
7 −0.14 −0.08 −0.05 −0.02 0.00 0.02 0.04 0.06 0.10 0.18
8 −0.14 −0.08 −0.05 −0.02 0.00 0.02 0.04 0.06 0.10 0.18
9 −0.15 −0.08 −0.05 −0.02 0.00 0.02 0.04 0.06 0.10 0.19
High −0.17 −0.08 −0.05 −0.02 0.00 0.02 0.04 0.06 0.10 0.24

Portfolio TO 1,0
Low 0.02 0.02 0.02 0.02 0.02 0.02 0.02 0.02 0.02 0.02
2 0.04 0.04 0.04 0.04 0.04 0.04 0.04 0.04 0.04 0.04
3 0.05 0.05 0.05 0.05 0.05 0.05 0.05 0.05 0.05 0.05
4 0.06 0.06 0.06 0.06 0.06 0.06 0.06 0.06 0.06 0.06
5 0.07 0.07 0.07 0.07 0.07 0.07 0.07 0.07 0.07 0.07
6 0.08 0.08 0.08 0.08 0.08 0.08 0.08 0.08 0.08 0.08
7 0.10 0.10 0.10 0.10 0.10 0.10 0.10 0.10 0.10 0.10
8 0.13 0.13 0.13 0.13 0.13 0.13 0.13 0.13 0.13 0.13
9 0.17 0.17 0.17 0.17 0.17 0.17 0.17 0.17 0.17 0.17
High 0.34 0.30 0.30 0.29 0.30 0.29 0.30 0.29 0.30 0.36

Average market capitalization ($ mio.)

Low 126 352 690 1,058 999 1,160 1,201 1,014 667 276
2 444 1,911 3,381 4,306 5,005 4,795 4,405 4,069 2,919 830
3 627 2,308 3,544 4,099 4,487 4,361 4,337 4,241 3,338 969
4 744 2,493 3,300 3,803 4,023 4,108 4,207 4,098 3,281 1,200
5 842 2,534 3,106 3,716 3,762 3,556 3,748 3,888 3,292 1,288
6 925 2,508 2,782 3,300 3,655 3,714 3,536 3,453 2,858 1,428
7 959 2,355 2,649 2,869 2,963 3,107 3,121 3,110 2,672 1,421
8 988 2,119 2,326 2,481 2,514 2,879 2,746 2,483 2,314 1,353
9 1,032 1,741 1,898 2,070 2,055 2,272 2,224 2,294 2,064 1,232
High 875 1,523 1,772 1,691 1,661 1,716 1,730 1,805 1,693 1,090

Average number of firms

Low 179 107 96 87 79 75 68 68 67 96
2 68 42 38 35 34 33 32 32 32 46
3 49 32 29 29 28 27 27 27 27 37
4 42 28 26 25 25 25 24 25 25 34
5 39 26 24 23 23 23 23 24 25 34
6 38 26 23 22 21 21 22 23 25 36
7 40 26 22 20 20 20 20 23 26 40
8 44 26 21 19 18 18 19 22 27 48
9 52 27 20 18 16 17 18 21 29 60
High 81 26 18 15 14 14 16 20 29 101

IA – 13

Electronic copy available at: https://ssrn.com/abstract=3150525

 A B
500
Short-term momentum Short-term momentum
Average cumulative sum of excess returns (%)

10 Short-term reversal* Short-term reversal*

Short-term reversal

5 Cumulative excess return (%)

0

-500
-5

-10 -1000

1 6 12 18 24 30 36 1995 2000 2005 2010 2015

Months after portfolio formation

Figure IA.1: International short-term momentum returns: Persistence and historical per-
formance. Panel A shows the average cumulative sums of post-formation excess returns to each of the
international short-term momentum (STMOM) and international short-term reversal* (STREV*) strategues
along with 95% confidence bands. Panel B shows a time-series plot of cumulative sums of excess returns
to the international STMOM and STREV* strategies as well as a conventional short-term reversal strategy.
Data are monthly and cover January 1993 to December 2018.

IA – 14

Electronic copy available at: https://ssrn.com/abstract=3150525

Table IA.5: Short-term momentum and end-of-month effects: International evidence. This
table shows average excess returns and abnormal returns to international zero-cost long-short strategies that
buy the previous month’s winners and sell the previous month’s losers among stocks with different turnover.
The underlying portfolios are from double sorts on the previous month’s return and turnover. We use
independent double sorts to form country-specific portfolios that are value-weighted and rebalanced at the
end of each month. We then weight each country’s portfolio by the country’s total market capitalization for
the previous month to form each international portfolio. All returns and market values are in U.S. dollars
and excess returns are above the monthly U.S. T-bill rate. Abnormal returns are relative to Fama and
French’s (2017) developed markets five-factor model including the momentum factor (DMFF6). In panel A,
the sorting variables exclude their end-of-month values (r1,0−EOM and TO 1,0−EOM ) measured at the month’s
last three trading days. In Panel B, the sorting sorting variables are just the end-of-month values (rEOM and
TO EOM ) measured at the month’s last three trading days. Test-statistics (in parentheses) are adjusted for
heteroscedasticity and autocorrelation. Data are monthly and cover January 1993 to December 2018.

Panel A Panel B

Performance of r1,0−EOM strategies Performance of rEOM strategies

within TO 1,0−EOM deciles within TO EOM deciles

TO 1,0−EOM decile E[re ] αDMFF6 TO EOM deciles E[re ] αDMFF6

Low −1.35 −0.98 Low −3.29 −3.46

(−2.82) (−1.88) (−7.97) (−7.06)
2 −0.91 0.16 2 −2.11 −2.16
(−1.88) (0.28) (−3.22) (−3.76)
3 −0.78 −0.61 3 −2.06 −2.13
(−1.50) (−1.06) (−4.83) (−5.03)
4 0.30 0.89 4 −2.47 −2.77
(0.81) (1.93) (−6.09) (−6.38)
5 −0.19 −0.06 5 −2.42 −2.32
(−0.49) (−0.12) (−4.72) (−4.94)
6 −0.04 −0.25 6 −2.09 −1.85
(−0.09) (−0.63) (−4.68) (−3.72)
7 0.62 0.56 7 −1.76 −2.00
(1.68) (1.48) (−4.79) (−4.67)
8 0.48 0.66 8 −1.13 −1.47
(1.23) (1.69) (−2.62) (−3.49)
9 0.98 1.08 9 −0.88 −0.97
(3.83) (3.43) (−2.46) (−2.32)
High 1.66 1.71 High −0.42 −0.25
(3.96) (3.78) (−1.06) (−0.71)

IA – 15

Electronic copy available at: https://ssrn.com/abstract=3150525

 Table IA.6: Double sorts on one-month returns and turnover excluding earnings announcers. This table shows portfolios double sorted
on previous month’s return (r1,0 ) and turnover (TO 1,0 ) excluding earnings announcers, i.e., excluding firms whose most recent earnings announcement
date (Compustat’s RDQ) fell in the previous month. We use conditional sorts into deciles based on NYSE breakpoints, first on r1,0 and then on TO 1,0 .
Portfolios are value-weighted and rebalanced at the end of each month. The table also shows the performance of long-short strategies across the deciles.
Test-statistics (in parentheses) are adjusted for heteroscedasticity and autocorrelation. Data are monthly and cover January 1972 to December 2018,
where the start date is determined by the availability of data on quarterly earnings announcement dates in Compustat.

r1,0 deciles r1,0 strategies

Low 2 3 4 5 6 7 8 9 High E[re ] αFF6 αq

TO 1,0 deciles
Portfolio excess return
Low 1.66 1.26 1.03 0.76 0.71 0.90 0.64 0.60 0.18 −0.11 −1.77 −1.83 −1.85
(−6.07) (−5.38) (−5.31)
2 1.48 1.68 1.14 1.08 1.16 1.07 0.82 0.84 0.51 −0.36 −1.84 −1.96 −1.99
(−5.21) (−5.22) (−5.37)
3 1.92 1.41 1.03 1.39 0.71 1.03 0.84 0.58 0.72 0.69 −1.23 −1.35 −1.26
(−3.56) (−3.44) (−3.33)
4 1.82 1.56 1.02 1.05 1.14 1.06 0.71 0.75 0.75 0.68 −1.15 −1.39 −1.26
(−3.38) (−4.00) (−2.99)

IA – 16
5 1.34 1.29 1.42 1.34 0.93 1.22 0.79 0.75 0.80 0.40 −0.94 −0.94 −0.84
(−2.69) (−2.32) (−2.17)
6 1.04 1.25 1.07 1.47 1.35 0.98 1.25 1.09 0.73 0.94 −0.10 −0.09 0.08
(−0.26) (−0.21) (0.16)
7 0.95 1.31 1.23 0.93 0.96 1.09 1.11 1.26 0.98 0.63 −0.32 −0.47 −0.32
(−0.94) (−1.26) (−0.63)
8 1.31 1.14 1.30 1.04 1.11 0.89 0.99 1.14 1.00 0.55 −0.76 −0.87 −0.88
(−2.30) (−2.23) (−2.03)
9 0.92 1.20 1.13 1.01 1.44 1.59 1.00 0.87 0.75 0.89 −0.03 −0.25 −0.02
(−0.07) (−0.59) (−0.03)
High 0.02 1.24 1.55 1.35 1.08 1.12 0.91 1.12 0.86 1.27 1.26 1.28 1.39
(2.65) (2.40) (2.92)

Electronic copy available at: https://ssrn.com/abstract=3150525

TO 1,0 strategies
E[re ] −1.64 −0.02 0.52 0.59 0.37 0.21 0.26 0.52 0.68 1.38
(−5.11) (−0.07) (1.87) (1.84) (1.09) (0.66) (0.79) (1.66) (2.20) (3.94)
αFF6 −1.68 0.17 0.60 0.51 0.54 0.29 0.35 0.61 0.73 1.44
(−4.75) (0.56) (2.38) (1.60) (1.70) (0.86) (1.08) (1.84) (2.46) (3.54)
αq −1.73 0.23 0.53 0.51 0.58 0.24 0.32 0.61 0.65 1.51
(−4.96) (0.71) (1.97) (1.42) (1.61) (0.70) (0.98) (1.55) (2.13) (3.64)
Table IA.7: Short-term momentum and industry momentum. This table shows time-series regres-
sion results for the short-term momentum (STMOM) strategy and for Moskowitz and Grinblatt’s (1999)
one-month industry momentum (IMOM) strategy. In Panel A, the additional controls are the FF6 factors.
In Panel B, the additional controls are the q-factors. Test-statistics (in parentheses) are adjusted for het-
eroscedasticity and autocorrelation. The sample excludes financial firms. Data are monthly and cover July
1963 (Panel A) or January 1967 (Panel B) to December 2018, where the start date in Panel B is determined
by the availability of the q-factors.

Intercepts, slopes, and test-statistics (in parentheses)

from time-series regressions of the form yt = α + β 0 Xt + t

Short-term momentum Industry momentum

Independent (1) (2) (3) (4) (5) (6) (7)

variable

Panel A: Controls are the FF6 factors

Intercept 1.37 0.79 0.93 0.97 0.83 0.62 0.51

(4.74) (2.74) (3.08) (4.54) (3.42) (2.87) (2.28)
MKT −0.31 −0.07 0.01
(−3.87) (−1.03) (0.14)
SMB 0.07 −0.14 −0.14
(0.53) (−1.64) (−1.68)
HML −0.02 0.08 0.07
(−0.09) (0.81) (0.75)
RMW −0.30 −0.19 −0.09
(−1.50) (−1.31) (−0.63)
CMA 0.15 0.05 0.01
(0.63) (0.28) (0.04)
MOM 0.16 0.33 0.25
(1.08) (4.15) (3.53)
STMOM 0.26 0.23
(6.77) (5.53)
IMOM 0.59 0.53
(7.69) (7.73)
Adj. R2 15.1% 18.2% 6.7% 15.1% 18.2%

Panel B: Controls are the q-factors

Intercept 1.42 0.87 1.15 0.94 0.87 0.57 0.45

(4.74) (2.89) (3.69) (4.20) (3.38) (2.53) (1.62)
MKT −0.33 −0.10 0.00
(−3.78) (−1.28) (0.01)
ME 0.07 −0.06 −0.07
(0.39) (−0.50) (−0.92)
ROE −0.30 0.19 0.24
(−1.72) (1.24) (1.93)
I/A 0.13 0.08 0.04
(0.60) (0.37) (0.22)
STMOM 0.26 0.26
(6.49) (5.87)
IMOM 0.58 0.57
(7.35) (7.12)
Adj. R2 15.0% 18.2% 1.8% 15.0% 16.1%

Electronic copy available at: https://ssrn.com/abstract=3150525

Table IA.8: Short-term momentum and crash risk. This table shows downside- and crash risk
measures for the short-term momentum (STMOM) strategy. For comparison, it also shows the corresponding
measures for the market and/or the conventional momentum strategy.
Panel A shows, for each strategy, the skewness and kurtosis of log(1+rte +rf t ), where rte is the strategy’s
monthly excess return and rf t is the monthly risk-free rate. The corresponding test-statistics (in parentheses)
are for the null of normally distributed returns. Panel A also shows the two momentum strategies’ coskewness
with the market, computed as the slope coefficient from a univariate regression of εt (re ) on (rmt e 2
) , where
e
the latter is the squared excess market return and where εt (r ) is the residual from a univariate regression of
rte on rmt
e
. Finally, Panel A shows the two momentum strategies’ down-side β, computed as the coefficient
βD from the regression rte = α + βrmt e e
+ βD max{0, −rmt } + t .
Panel B shows results from Daniel and Moskowitz’s (2016) market-timing regressions, where the most
general form is rte = (α0 + αB IB,t−1 ) + (β0 + (βB + βB,U IU,t )IB,t−1 ) rmt
e
+ t . Here, IB,t−1 is an ex-ante “bear
market” indicator, which equals 1 if the cumulative market excess return over the previous 24 months is
negative and is otherwise zero, and IU,t is a contemporaneous “up-market” indicator, which equals 1 if the
market excess return is positive in month t and is otherwise zero. The intercept coefficients, α0 and αB , are
multiplied by 100 and are thus stated in % per month.
The sample excludes financial firms. Data are monthly and cover July 1963 to December 2018.

Panel A: Higher-order moments and downside beta

Statistic Market Short-term momentum Conventional momentum

Skew −0.78 −0.49 −1.55

(−7.44) (−4.92) (−12.24)
Kurtosis 5.85 6.64 12.51
(7.04) (7.88) (11.17)
Co-skewness −0.65 −1.34
(−0.77) (−1.98)
Downside β −0.14 −0.42
(−0.61) (−2.38)

Panel B: Market-timing regression results

Intercepts, slopes, and test-statistics (in parentheses)

from time-series regressions of strategy returns

Short-term momentum Conventional momentum

Coefficient Variable (1) (2) (3) (4) (5) (6)

α0 1 1.59 1.84 1.84 1.25 1.33 1.33

(4.91) (5.01) (5.03) (4.96) (4.63) (4.63)
αB 1B,t−1 −1.17 −3.07 −0.42 0.52
(−1.52) (−2.48) (−0.70) (0.53)
β0 e
rmt −0.38 −0.25 −0.25 −0.08 0.17 0.17
(−5.27) (−2.71) (−2.72) (−1.35) (2.43) (2.43)
βB e
1B,t−1 × rmt −0.36 −0.77 −0.69 −0.48
(−2.41) (−2.99) (−5.87) (−2.37)
βB,U e
1B,t−1 × 1U,t × rmt 0.81 −0.40
(1.96) (−1.24)
Adj. R2 4.0% 5.1% 5.5% 0.1% 5.2% 5.3%

IA – 18

Electronic copy available at: https://ssrn.com/abstract=3150525

Table IA.9: Short-term momentum and volatility effects. This table shows the performance of zero-
cost long-short strategies that buy the previous month’s winners and sell the previous month’s losers among
stocks with different values for the previous month’s ‘residual turnover’ (RTO 1,0 ; Panel A) and volatility
(σ1,0 ; Panel B). Portfolios are from double sorts on the previous month’s return (r1,0 ) and either RTO 1,0 or
σ1,0 . We use conditional sorts into deciles based on NYSE breakpoints, first on r1,0 and then on either RTO 1,0
or σ1,0 . Portfolios are value-weighted and rebalanced at the end of each month. RTO 1,0 is the residual from
cross-sectional regression of the previous month’s share turnover on σ1,0 , estimated using WLS with market
equity as weights, and σ1,0 is the standard deviation of the previous month’s daily stock returns using a
minimum of 15 daily observations. Test-statistics (in parentheses) are adjusted for heteroscedasticity and
autocorrelation. The sample excludes financial firms. Data are monthly and cover July 1963 to December
2018, except when applying the q-factors, which are available from January 1967.

Panel A Panel B
Performance of r1,0 strategies Performance of r1,0 strategies
within RTO 1,0 deciles within σ1,0 deciles

RTO 1,0 decile E[re ] αFF6 αq σ1,0 decile E[re ] αFF6 αq

Low −1.54 −1.59 −1.57 Low −2.36 −2.45 −2.56

(−5.55) (−5.97) (−5.13) (−11.51) (−11.13) (−11.59)
2 −0.77 −0.97 −0.87 2 −1.72 −1.80 −1.81
(−3.07) (−3.31) (−2.21) (−7.12) (−6.61) (−6.55)
3 −0.91 −0.96 −1.02 3 −1.73 −1.90 −2.04
(−3.37) (−3.28) (−2.82) (−6.17) (−6.96) (−6.55)
4 −1.38 −1.36 −1.45 4 −1.66 −1.75 −1.77
(−5.13) (−5.06) (−4.80) (−6.06) (−5.60) (−5.64)
5 −0.63 −0.75 −0.64 5 −1.72 −1.99 −1.92
(−2.42) (−2.86) (−2.18) (−6.78) (−6.75) (−6.77)
6 −0.72 −0.89 −0.75 6 −1.86 −1.97 −2.06
(−3.00) (−3.09) (−2.24) (−7.45) (−7.00) (−7.56)
7 −0.49 −0.59 −0.44 7 −1.16 −1.17 −1.20
(−1.95) (−2.03) (−1.17) (−4.86) (−4.54) (−4.42)
8 −0.15 −0.32 −0.27 8 −1.67 −1.72 −1.79
(−0.54) (−0.89) (−0.69) (−5.60) (−5.01) (−4.58)
9 −0.06 −0.11 0.06 9 −1.16 −1.29 −1.29
(−0.26) (−0.46) (0.16) (−4.74) (−4.91) (−4.41)
High 1.06 0.96 1.21 High −0.65 −0.86 −0.87
(3.86) (3.07) (3.47) (−2.25) (−2.63) (−2.41)

IA – 19

Electronic copy available at: https://ssrn.com/abstract=3150525

Table IA.10: Short-term momentum and longer formation periods. This table shows time-series
regression results for alternative short-term momentum and short-term reversal* strategies constructed using
longer formation periods. The strategies are constructed similar to their one-month counterparts in Table I,
except that the sorting variables are cumulative return and average monthly share turnover for the previous
2, . . . , 6 months. The explanatory variables are the benchmark STMOM and STREV∗ strategies from Table I
based on a one-month formation period. Test-statistics (in parentheses) are adjusted for heteroscedasticity
and autocorrelation. The sample excludes financial firms. Data are monthly and cover January 1963 to
December 2018.

Intercepts, slopes, and test-statistics (in parentheses)

from time-series regressions of the form yt = α + β 0 Xt + t

2 months 3 months 4 months 5 months 6 months

formation formation formation formation formation

Independent (1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
variable

Panel A: Short-term momentum strategies with varying formation periods

Intercept 0.99 0.02 1.00 0.15 1.06 0.38 1.24 0.64 1.19 0.62
(2.95) (0.08) (3.11) (0.57) (3.16) (1.36) (3.78) (1.93) (3.25) (1.80)
STMOM 0.71 0.62 0.50 0.45 0.42
(15.12) (13.70) (13.38) (9.50) (7.35)
Adj. R2 46.5% 34.0% 21.5% 17.3% 14.9%

Panel B: Short-term reversal* strategies with varying formation periods

Intercept −1.07 −0.13 −0.88 −0.02 −0.49 0.33 −0.48 0.27 −0.49 0.24
(−4.70) (−0.60) (−3.78) (−0.07) (−2.05) (1.46) (−1.87) (1.10) (−1.84) (0.94)
STREV* 0.67 0.61 0.58 0.53 0.51
(12.74) (11.21) (10.02) (9.20) (6.86)
Adj. R2 36.1% 28.4% 23.9% 18.9% 16.1%

IA – 20

Electronic copy available at: https://ssrn.com/abstract=3150525