The Distribution of Cross Sectional Momentum

Returns

Oh Kang Kwon †,? · Stephen Satchell †,‡

Current revision: August 14, 2017

Abstract. The cross sectional momentum (CSM) strategy, considered in Jegadeesh

and Titman (1993), is a trading strategy in which assets with highest past returns
are equally weighted and held long while assets with lowest returns are also equally
weighted but held short. Although anomalous returns from such momentum strate-
gies, indicating persistence or reversal in the relative performance of underlying as-
sets, have been the subject of numerous empirical studies, very little appears to be
known about the distributional properties of the CSM returns. Most of the known re-
sults in this regard, such as those obtained in Lo and MacKinlay (1990), Jegadeesh
and Titman (1993), Lewellen (2002), and Moskowitz et al. (2012), are limited to
expected values and the first order autocorrelations in the returns of alternative
momentum strategies in which the assets are weighted not equally, but instead in
proportion to their returns over the ranking period. In this paper, we derive the
density of CSM returns in analytic form, along with moments of all orders, under a
reasonably general set of assumptions on the underlying asset returns. In particular,
if the asset returns over the ranking and holding periods are independent then the
density of the CSM returns is shown to be a mixture of univariate normals. We also
obtain the density and the moments of equally weighted long only portfolios consist-
ing of assets from an arbitrary return percentile band that we refer to as quantile
portfolios.

1. Introduction

Momentum based trading strategies rely on the persistence or the reversal in the
relative performances of assets over successive time periods called the ranking and
? Corresponding author
†Discipline of Finance, Codrington Building (H69), The University of Sydney, NSW 2006,
Australia · ‡ Trinity College, University of Cambridge, Cambridge CB2 1TQ, UK
E-mail: ohkang.kwon@sydney.edu.au · ses999gb@yahoo.co.uk

Electronic copy available at: https://ssrn.com/abstract=3018366

2

holding periods respectively. For example, a cross sectional momentum (CSM) strat-
egy, considered for example in Jegadeesh and Titman (1993), sorts the returns from
n assets over a ranking period, and constructs a portfolio over the holding period
consisting of an equally weighted long position in the m+ best performing assets and
an equally weighted short position in the m− worst performing assets. Such strategies
have become popular in practice, due mainly to the fact that they tend to exhibit
positive expected returns and since momentum has emerged as a key component in
the factor based approach to investment attributed to Carhart (1997).
Since anomalous returns from momentum based trading strategies would indi-
cate a violation of the assumption of market efficiency, the returns generated by these
strategies have been the subject of considerable empirical research spanning extensive
asset classes, jurisdictions, and investment periods. For example, Jegadeesh and Tit-
man (1993) and Asness (1994) found that momentum strategies are profitable in the
US equities markets over the short to medium horizons from 1965 to 1989, Jegadeesh
and Titman (2001) found similar results for the period 1990 to 1998, and Israel and
Moskowitz (2013) extended momentum evidence to two periods from 1927 to 1965
and from 1990 to 2012. Analogous results were found for country equity indices in
Richards (1997), Asness et al. (1997), Chan et al. (2000), and Hameed and Yuanto
(2002), for emerging markets in Rouwenhorst (1998), for exchange rate markets in
Okunev and White (2003) and Menkhoff et al. (2012), for commodities in Erb and
Harvey (2006), for futures contracts in Moskowitz et al. (2012), and in industries in
Sefton and Scowcroft (2004). Anomalous momentum returns were also found in As-
ness et al. (2013) and Daniel and Moskowitz (2016) for markets in European Union,
Japan, United Kingdom and United States, and across asset classes including fixed
income, commodities, foreign exchange and equity from 1972 through 2013.
In contrast to the extensive literature on the empirical properties of momentum
based returns, there are relatively few that consider the distributional properties of
these returns from a theoretical viewpoint. Most of the known theoretical results,
obtained for example in Lo and MacKinlay (1990), Jegadeesh and Titman (1993),
Lewellen (2002), and Moskowitz et al. (2012), are concerned only with the expected
values and the first order autocorrelations of the returns from the so-called weighted
relative strength strategy in which the portfolio over the holding period is constructed
from all underlying assets weighted, essentially, in proportion to their returns over the
ranking period. In this paper, we address this gap by deriving, under the assumption
that asset returns are multivariate normal, the probability density of CSM returns
along with the moments of these returns. We also derive analogous results for a class
of strategies that consist only of the long part of CSM strategies that we call quantile
portfolios. These portfolios were considered by Grinblatt and Moskowitz (2004) for
the US equities market, Rey and Schmid (2007) for the Swiss market, and by Devito
et al. (2012) taking into account the effect of currency hedging. Quantile portfolios
of small sizes in the US market, suitable for private investors, were studied in Foltice
and Langer (2015), while Fisher et al. (2016) considered US quantile portfolios in
the context of factor combinations.
Now, let lr ∈ N be the length of the ranking period and lh ∈ N the length of
the holding period. Then the construction of a cross sectional momentum strategy
can be described in greater detail as follows. Firstly, in month t, all underlying
assets are ranked and sorted into quantiles, for example quintiles or deciles, based
on their returns over the past lr -month ranking period from month t − lf − 1 to t − 1.

Electronic copy available at: https://ssrn.com/abstract=3018366

 3

One month is omitted to avoid short-term reversals and similar phenomena. At this
time, the top quantile portfolio, consisting of “winners”, is purchased and the worst
performing quantile, consisting of “losers”, is shorted for the lh -month holding period
from month t to t + lh . The two portfolios may either be equally weighted or value
weighted. By assuming that underlying asset returns are Gaussian, we derive in this
paper the distribution and the moments of CSM returns in the general case, and in a
number of special cases under which resulting expressions simplify significantly. We
also derive analogous results for the long only quantile portfolios that allow us to
examine their distributional properties as we move from one quantile to the next in
a manner similar to the one followed in Daniel and Moskowitz (2016). Anticipating
our results, the densities obtained involve truncated normal distributions, which is
a result partially discussed in Grundy and Martin (2001) Appendix A. The results
generalize naturally to arbitrary fixed weight portfolios, albeit with added complexity
in notation.
The problem of determining the distributional properties of the CSM returns
is regarded as a 2-period problem. At any time t, the first period is the ranking
period from t − lr − 1 to t − 1 over which the asset returns are described by the
vector r t , and the second period is the holding period from t + 1 to t + lh over
which the corresponding asset returns are r t+1 . The vector (r 0t , r 0t+1 ) is assumed to
be Gaussian, and since we do not make any assumptions on stationarity the joint
distribution of r t and r t+1 is a 2n-dimensional vector of normal returns with means
and variances depending on t and t + 1 as in Assumption 1. This flexibility means
that the exclusion of the period t − 1 or the assumption lr 6= lh , for example, do not
pose any difficulties.
However, when we do assume stationarity so that E[r t ] = E[r t+1 ] with corre-
sponding restrictions on covariances, the condition lr = lh needs to be imposed.
Fortunately, many momentum strategies are of this form where the length of the
ranking and holding periods are equal. In this case, the exclusion of a period is still
consistent with strict stationarity, and since it is assumed that the vector return
process is Gaussian this is equivalent to weak stationarity.
At this point, potential computational difficulties associated with practical im-
plementation of the expressions derived in this paper require some discussion. For
example, the expression for the density of CSM returns obtained in Theorem 3 con-
sists of
 
n n!
=
n − m− − m+ , m − , m + (n − m− − m+ )!m− !m+ !

distinct terms in the general case. For the CSM strategy that takes long position in
the top decile and short position in the bottom decile from the S&P500 index, this
corresponds 500!/(400!(50!)2 ) distinct terms which is very large. However, in many
of the non-equity applications, the number of terms is much smaller. For example,
the empirical study of the US industry return data in Sefton and Scowcroft (2004)
corresponds to n = 10, m+ = 2, and m− = 2 and results in 1, 260 distinct terms in
the expression for the CSM return density, and the study in Foltice and Langer (2015)
corresponds to m+ = 1 and m− = 1 in a universe of n stocks giving n(n − 1) distinct
terms. For quantile portfolios, many authors consider choosing the best stock out of
n stocks, thereby leading to n distinct terms. In any case, it is shown that the returns
generated by Monte Carlo simulation provide an accurate approximation to the CSM
4

return densities, and is computationally feasible for practical applications, while

analytic expressions enable theoretical investigation of the distributional properties
of CSM returns.
In the special case where the markets are efficient, the returns in the ranking
period are independent of the returns in the holding period, and our results simplify
to give the CSM return as a mixture of univariate normal distributions. Finally,
although we have assumed that asset returns are Gaussian, our arguments will gen-
eralize to other multivariate distributions if we can characterize the density of the
holding periods returns conditional on the ranking period returns. The details of this
generalization is left for future research.
The remainder of this paper is organized as follows: Section 2 introduces the no-
tation and the key results on multivariate normal distributions, Section 3 considers
the distributional properties of CSM returns, and Section 4 examines the properties
of quantile portfolios. Numerical examples illustrating the different CSM return dis-
tributions that can be generated under the assumption of Gaussian asset returns are
given in Section 5, and the paper concludes with Section 6.

2. Notation and Preliminaries

For any x ∈ Rn , we will write xi for the i-th coordinate of x, and given y ∈ Rn write
x ≺ y if and only if xi < yi for all 1 ≤ i ≤ n. Similarly, given a matrix M ∈ Rm×k ,
we will write Mi,j for the (i, j)-th entry of M , and the transpose of a vector or a
matrix will be denoted by the superscript 0 . The density of an n-dimensional normal
distribution, with mean µ and covariance Σ, at x ∈ Rn will be denoted φn (x; µ, Σ),
and
Z
Φn [a; µ, Σ] = φn (x; µ, Σ) dx (1)
x≺ a

for any a ∈ Rn , where dx = dx1 · · · dxn . In general, given random variables

x1 , . . . , xn , their joint probability density function will be denoted fx1 ,··· ,xn .

Theorem 1. Let n1 , n2 ∈ N and suppose x ∼ Nn1 +n2 (µ, Σ), where

     
x µ1 Σ1,1 Σ1,2
x= 1 , µ= , Σ= , (2)
x2 µ2 Σ2,1 Σ2,2

with xi , µi ∈ Rni and Σi,j ∈ Rni ×nj for 1 ≤ i, j ≤ 2, and Σ positive definite. Then
the conditional distribution of x1 given x2 is normal with mean and covariance
−1
µx1 |x2 = µ1 + Σ1,2 Σ2,2 (x2 − µ2 ), (3)
−1
Σx1 |x2 = Σ1,1 − Σ1,2 Σ2,2 , Σ2,1 , (4)

respectively, and φn1 +n2 (x; µ, Σ) decomposes as

φn1 +n2 (x; µ, Σ) = φn1 (x̃1 ; µx1 |x2 , Σx1 |x2 )φn2 (x2 ; µ2 , Σ2,2 ). (5)

Proof. Refer to Muirhead (1982) Theorem 1.2.11. 

 5

For any n ∈ N, let Sn be the group of permutations of the set {1, 2, . . . , n}, and given
any τ ∈ Sn , denote by τi = τ (i) the image of 1 ≤ i ≤ n under τ . Recall that Sn acts
naturally on the set of polynomials, R[x1 , . . . , xn ] by the rule

τ p(x1 , . . . , xn ) = p(xτ1 , . . . , xτn )

for any polynomial p ∈ R[x1 , . . . , xn ] and τ ∈ Sn . Now, let n1 , n2 ∈ N be fixed, and

let pn1 ,2n2 ∈ R[x1 , . . . , xn1 +2n2 ] be the polynomial

n1
! n2
!
Y Y
pn1 ,2n2 (x1 , . . . , xn1 +2n2 ) = xi (xn1 +2i − xn1 +2i−1 )2 . (6)
i=1 i=1

Denote by Z(n1 , 2n2 ) the stabilizer of pn1 ,2n2 under the action of Sn1 +2n2 so that

Z(n1 , 2n2 ) = {τ ∈ Sn1 +2n2 | τ pn1 ,2n2 = pn1 ,2n2 }, (7)

and let Q(n1 , 2n2 ) = Sn1 +2n2 /Z(n1 , 2n2 ) be the quotient group,1 with elements of
Q(n1 , 2n2 ) identified with their coset representatives τ ∈ Sn1 +2n2 .

Theorem 2. Let x ∼ Nn (µ, Σ), where n ∈ N, µ ∈ Rn and Σ ∈ Rn×n is positive

definite. Given κ = (κ1 , . . . , κm ) ∈ {1, 2, . . . , n}m , if the κ-th moment of x is defined
by
" m
#
Y
µκ (x) = E x κi , (8)
i=1

then
k
! l
!
X X Y Y
µκ (x) = µ κτ i Σκτk+2i−1 ,κτk+2i . (9)
k,l∈N τ ∈Q(k,2l) i=1 i=1
k+2l=m

Proof. Refer to Withers (1985) Theorem 1.1. 

Corollary 1. Let x ∼ N1 (µ, σ 2 ). Then for any m ∈ N, the m-th moment of x is

given by
m  
X m
µm (x) = (i − 1)!! σ i µm−i , (10)
i
i=0
i even

Qb k2 c
where k!! = i=0 (k − 2i) is the double factorial of k ∈ N.

Proof. Follows
from theorem 2, since the inner sum in (9), corresponding to 2l = i,
consists of mi (i − 1)!! identical terms all equal to σ i µm−i . 

1 Refer to Rotman (1995) Chapter 2 for the details on quotient groups.

6

3. Cross Sectional Momentum Returns

We begin with a general framework under which to investigate the distribution of

cross sectional momentum (CSM) returns, and show how well-known models for
asset returns are special cases of this framework.
Fix 0 < m+ , m− , n ∈ N such that m+ + m− ≤ n, and for each 1 ≤ i ≤ n denote
by ri,t the return on asset i at time t. Let

r t = (r1,t , . . . , rn,t )0 , (11)

and for any τ ∈ Sn define

m+ m−
1 X 1 X
rτ,m± ,t = rτi ,t − rτn−m− +i ,t , (12)
m+ m−
i=1 i=1
xτi ,t = rτi+1 ,t − rτi ,t , 1 ≤ i ≤ n − 1, (13)
0
xτ,t = (xτ1 ,t , . . . , xτn−1 ,t ) , (14)

0
z τ,t = rτ,m± ,t+1 , x0τ,t . (15)

Assumption 1. The vector (r 0t+1 , r 0t )0 has a multivariate normal distribution for

all t ∈ N.

The next proposition shows that returns following a vector autoregressive moving
average process, VARMA(p, q), satisfies the above assumption.

Proposition 1. Suppose r t ∼ VARMA(p, q), so that

p q
X X
rt = α + Ai r t−i + εt + Mi εt−i , (16)
i=1 i=1

where α ∈ Rn , Ai ∈ Rn×n for 1 ≤ i ≤ p, Mi ∈ Rn×n for 1 ≤ i ≤ q, εt ∼ N (0n , Σ)

with Σ ∈ Rn×n positive definite for all t ∈ N, E[εs ε0t ] = δs,t Σ for all s, t ∈ N, and
δs,t denotes the Kronecker delta so that
(
1, s = t,
δs,t =
0, s 6= t.

If (r 00 , . . . , r 0p−1 )0 is multivariate normal and independent of εt , or the VARMA(p, q)

process satisfies conditions for weak stationarity, then (r 0t+1 , r 0t ) satisfies assump-
tion 1.

Proof. This follows from the well-known properties of the VARMA(p, q) process. 

Processes that play a key role in the investigation of CSM returns are z τ,t defined in
(15), and the next result provides the distribution of these processes under assump-
tion 1.
 7

Proposition 2. If the return process r t satisfies assumption 1, then z τ,t follows a

multivariate normal distribution for all τ ∈ Sn and t ∈ N. In particular, z τ,t is
multivariate normal if r t ∼ VARMA(p, q), and either (r 00 , . . . , r 0p−1 )0 is multivariate
normal and independent of εt , or the VARMA(p, q) process satisfies conditions for
weak stationarity.
Proof. For any τ ∈ Sn , let Pτ be the n ×
0 n permutation matrix corresponding to τ ,
0 0
let ιm± = m+ 1m+ , 0, . . . , 0, − m− 1m− ∈ Rn where 1m± = (1, 1, . . . , 1)0 ∈ Rm± ,
1 1

and define
−1 1 0 0 · · · 0 0 0
 
 0 −1 1 0 · · · 0 0 0
Dn−1 = ∈ R(n−1)×n . (17)
. . . . . . . . . . . . . . . . . . . . .
0 0 0 0 · · · 0 −1 1
Then xτ,t = Dn−1 Pτ r t and rτ,m± ,t+1 = ι0m± Pτ r t+1 , and so z τ,t can be written as
   0  
rτ,m± ,t+1 ιm± Pτ O1×n r t+1
z τ,t = = . (18)
xτ,t O(n−1)×n Dn−1 Pτ rt

0
Hence, z τ,t is a linear transformation of r 0t+1 , r 0t , which is assumed to be multi-
variate normal, and so z τ,t is itself multivariate normal. 
Assuming (r 0t+1 , r 0t )0 satisfies assumption 1, we may write
     
r t+1 µt+1 Λ Λ
∼N , t+1,t+1 t+1,t , (19)
rt µt Λt,t+1 Λt,t

with µu ∈ Rn , Λu,v ∈ Rn×n , and Λu,u are positive definite for u, v ∈ {t, t + 1}. As
discussed in the introduction, the 2-period model with returns specified by (19) is
quite general, since models with longer ranking and holding periods can be trans-
formed to a model of this type. The decomposition of z τ,t in (18) then establishes
the next result.
Corollary 2. Let (r 0t+1 , r 0t )0 be given by (19). Then for any τ ∈ Sn and t ∈ N,

ι0m± Pτ µt+1 ι0 P Λ P0ι ι0 P Λ P 0 D0

   
z τ,t ∼ N , m± τ t+1,t+1 0τ m± m± τ t+1,t 0τ n−1 0 . (20)
Dn−1 Pτ µt Dn−1 Pτ Λt,t+1 Pτ ιm± Dn−1 Pτ Λt,t Pτ Dn−1
Definition 1. The (m+ , m− )-cross sectional momentum return, rm± ,t+1 , at time
t + 1 is defined by
X
rm± ,t+1 = I{xτ,t ≺0n−1 } rτ,m± ,t+1 , (21)
τ ∈Sn

where IA , for any subset A ⊂ Rn , denotes the indicator function on the set A.
The following theorem expresses the density of CSM returns based on m+ equally
weighted long positions and m− equally weighted short positions from a universe
of n assets as the sum over the permutations in Sn of terms that are essentially
the product of a univariate normal density function and an (n − 1)-dimensional
cumulative normal distribution function. This could easily be generalized to two
arbitrary fixed-weight portfolios each of whose weights add up to one.
8

Theorem 3. Let frm±,t+1 (r) be the probability density of the (m+ , m− )-cross sec-
tional momentum return rm±,t+1 . If (r 0t+1 , r 0t )0 satisfies assumption 1, then we have
X
2


frm±,t+1 (r) = φ1 r; r̄τ,m± ,t+1 , ςr,τ,t+1
τ ∈Sn (22)
−2


Φn−1 0; x̄τ,t + ςr,τ,t+1 r − r̄τ,m± ,t+1 λτ,t , Λx|r,τ,t ,

where r̄m± ,τ,t+1 = ι0m± Pτ µt+1 , x̄τ,t = Dn−1 Pτ µt ,

2
ςr,τ,t+1 = ι0m± Pτ Λt+1,t+1 Pτ0 ιm± ,
λτ,t = Dn−1 Pτ Λt,t+1 Pτ0 ιm± ,
Λx|r,τ,t = Dn−1 Pτ Λt,t Pτ0 Dn−1
0 −2
− ςr,τ,t+1 λτ,t λ0τ,t .

Proof. For each permutation τ ∈ Sn , define a subset Rτ ⊂ Rn by

Rτ = (x1 , . . . , xn ) ∈ Rn | xτi+1 − xτi < 0 .

(23)
Then Rτ ∩ Rυ = ∅ for τ 6= υ ∈ Sn , and Rn = ∪τ ∈Sn Rτ up to a set of measure zero,
and so
Z
frm± ,t+1 (r) = frm± ,t+1 ,rt (r, r) dr
r∈Rn
XZ XZ
= frm± ,t+1 ,rt (r, r) dr = frτ,m,t+1 ,rt (r, r) dr,
τ ∈Sn r∈Rτ τ ∈Sn r∈Rτ

where the final equality follows from the fact that the summands of rm± ,t+1 in (21),
other than I{xτ,t ≺0n−1 } rτ,m± ,t+1 , vanish on Rτ . Now, for each τ ∈ Sn consider the
integral
Z
Iτ (r) = frτ,m± ,t+1 ,rt (r, r) dr,
r∈Rτ

and let gτ : Rn → Rn be defined by

   
1 O1×(n−1) r
gτ (r) = Pτ r = τ1 ,
O(n−1)×1 Dn−1 xτ

0
where xτ = xτ1 , . . . , xτn−1 and xτi = rτi+1 − rτi . Then gτ is invertible, and the
Jacobian of gτ is 1. So making the change of variables from r to (rτ1 , x0τ )0 , we obtain
Z
Iτ (r) = frτ,m± ,t+1 ,rτ1 ,xτ,t (r, s, xτ ) ds dxτ ,
n−1
(rτ1 ,x0τ )0 ∈R×R−

since xτi < 0 on Rτ for 1 ≤ i ≤ n−1. Applying the properties of conditional densities
gives
Z Z ∞
Iτ (r) = frτ1 ,xτ,t |rτ,m± ,t+1 (s, xτ | r) frτ,m± ,t+1 (r) ds dxτ
xτ ∈Rn−1
−
−∞
Z
= frτ,m± ,t+1 (r) Jτ (xτ , r)fxτ,t | rτ,m± ,t+1 (xτ | r) dxτ ,
xτ ∈Rn−1
−
 9

where
Z ∞
Jτ (xτ , r) = frτ1 |xτ,t ,rτ,m± ,t+1 (s | xτ , r) ds.
−∞

But the conditional density of rτ1 given xτ,t , rτ,m± ,t+1 is univariate normal by
theorem 1, and so Jτ (xτ , r) = 1 for all (xτ , r) ∈ Rn−1
− × R and
Z
Iτ (r) = frτ,m± ,t+1 (r) fxτ,t | rτ,m± ,t+1 (xτ | r) dxτ . (24)
xτ ∈Rn−1
−

Now, by corollary 2, we have

ςr,τ,t+1 λ0τ,t
     2 
rτ,m± ,t+1 r̄τ,m± ,t+1
z τ,t = ∼N , , (25)
xτ,t x̄τ,t λτ,t Λx|r,τ,t

and so setting x1 = xτ,t and x2 = rτ,m± ,t+1 in theorem 1 gives

−2



fxτ,t | rτ,m± ,t+1 (xτ | r) = φn−1 x; x̄τ,t + ςr,τ,t+1 r − r̄τ,m± ,t+1 λτ,t , Λx|r,τ,t . (26)

Computing the integral of fxτ,t | rτ,m± ,t+1 (xτ | r) over xτ ∈ Rn−1

− gives

−2


Iτ (r) = frτ,m± ,t+1 (r) Φn−1 0; x̄τ,t + ςr,τ,t+1 r − r̄τ,m± ,t+1 λτ,t , Λx|r,τ,t ,

2


and since rτ,m± ,t+1 ∼ N r̄τ,m± ,t+1 , ςr,τ,t+1 , we have

2


Iτ (r) = φ1 r; r̄τ,m± ,t+1 , ςr,τ,t+1
−2


Φn−1 0; x̄τ,t + ςr,τ,t+1 r − r̄τ,m± ,t+1 λτ,t , Λx|r,τ,t .

Summing Iτ (r) over τ ∈ Sn gives (22). 

Using an alternative decomposition of frτ,m± ,t+1 ,xτ,t (r, x), by conditioning rm± ,t+1
on xτ,t gives the following equivalent expression for frm± ,t+1 (r).

Corollary 3. If the assumption in theorem 3 is satisfied, then

frτ,m±,t+1 (r)
XZ  
(27)
2
= φ1 r; r̄τ,m±,t+1|x , ςr|x,τ,t+1 φn−1 (x; x̄τ,t , Λx,τ,t ) dx,
τ ∈Sn x∈Rn−1
−

where Λx,τ,t = Dn−1 Pτ Λt,t Pτ0 Dn−1

0 2
, ςr|x,τ,t+1 2
= ςr,τ,t+1 − λ0τ,t Λ−1
x,τ,t λτ,t , and

r̄τ,m± ,t+1|x = r̄τ,m± ,t+1 + λ0τ,t Λ−1

x,τ,t (x − x̄τ,t )
10

Proof. It follows from the definition of conditional densities that

fxτ,t | rτ,m± ,t+1 (xτ | r)frτ,m± ,t+1 (r) = frτ,m± ,t+1 | xτ,t (r | xτ )fxτ,t (xτ ),

and substituting the right-hand side into the expression for Iτ (r) in (24) gives
Z
Iτ (r) = frτ,m± ,t+1 | xτ,t (r | xτ )fxτ,t (xτ ) dxτ .
n−1
xτ ∈R−

Setting x1 = rτ,m± ,t+1 and x2 = xτ,t in theorem 1 implies

 
2
frτ,m± ,t+1 | xτ,t (r, xτ ) = φ1 r; r̄τ,m±,t+1|x , ςr|x,τ,t+1 ,

and since fxτ,t (xτ ) = φn−1 (xτ ; x̄τ,t , Λx,τ,t ), we have

Z  
2
Iτ (r) = φ1 r; r̄τ,m±,t+1|x , ςr|x,τ,t+1 φn−1 (xτ ; x̄τ,t , Λx,τ,t ) dxτ
xτ ∈Rn−1
−

and summing over τ ∈ Sn gives (27). 

In the special case n = 2 and m± = 1, the density, frm±,t+1 , can be made more
explicit.

Corollary 4. Let n = 2, m± = 1, and suppose (r 0t+1 , r 0t )0 satisfies assumption 1.

Denote by id ∈ S2 the identity permutation, and define
√
µu = x̄id,u , ςu,v = cov(xid,u , xid,v ), ςu = ςu,u

for u, v ∈ {t, t + 1}, and %t,t+1 = ςt,t+1 /(ςt ςt+1 ). Then

fr1± ,t+1 (r)

 
2 %t,t+1 ςt
(r + µt+1 ) , ςt2 1 − %2t,t+1



= φ1 r; −µt+1 , ςt+1 Φ1 0; µt −
ςt+1 (28)
 
2 %t,t+1 ςt
(r − µt+1 ) , ςt2 1 − %2t,t+1 .



+ φ1 r; µt+1 , ςt+1 Φ1 0; −µt −
ςt+1

Proof. Let (1 2) ∈ S2 be the non-trivial permutation. Then since xid,u = r1,u − r2,u
and x(1 2),u = r2,u − r1,u for u ∈ {t, t + 1}, in the notation introduced in theorem 3
2 2
ςr,τ,t+1 = ςt+1 , λτ,t = −%t,t+1 ςt ςt+1 ,
2
ςt2 , 2
1 − %2t,t+1


ςr,τ,t = Λx|r,τ,t = ςt+1

for all τ ∈ S2 , and

r̄id,1± ,t+1 = −µt+1 , r̄(1 2),1± ,t+1 = µt+1 ,

x̄id,1± ,t = µt , x̄(1 2),1± ,t = −µt ,

and substituting these terms into (22) gives the density in (28). 
 11

The density, frm±,t+1 (r), simplifies significantly when the asset returns have the
same mean and pairwise covariances. The next result is of some interest since it
gives conditions under which the CSM return density is normal.
Corollary 5. If (r 0t+1 , r 0t )0 satisfies assumption 1, and the means and covariances
in (19) take the degenerate form
µu = µu 1n , Λu,v = δu,v σu σv (1 − ρu,v )In×n + ρu,v σu σv 1n 10n , (29)
where µu ∈ R, σu ∈ R+ , ρu,v ∈ (−1, 1), and u, v ∈ {t, t + 1}, then in the notation of
theorem 3 we have
2


frm±,t+1 (r) = φ1 r; 0, ςr,id,t+1 ,
where id ∈ Sn is the identity element.
Proof. If µu and Λu,v are as given in (29), then r̄m±,τ,t+1 = 0, x̄τ,t = 0, and
λτ,t = 0 for all τ ∈ Sn , and the remaining terms that appear in (22) are independent
of the permutation τ ∈ Sn . So the summands are identical, and since there are n!
summands, we have
2

 
frm±,t+1 (r) =n! φ1 r; 0, ςr,id,t+1 Φn−1 0; 0, Λx|r,id,t .
Integrating both sides over r ∈ R, and noting that the cumulative distribution
 term
on the right-hand side is independent of r, gives Φn−1 0; 0, Λx|r,id,t = 1/n! and so
the result follows. 
Another special case in which frm±,t+1 (r) simplifies considerably is when r t and r t+1
are independent.
Theorem 4. If (r 0t+1 , r 0t )0 satisfies assumption 1, and r t and r t+1 are independent,
then
X
2


frm±,t+1 (r) = φ1 r; r̄τ,m± ,t+1 , ςr,τ,t+1 Φn−1 [0; x̄τ,t , Λx,τ,t ] , (30)
τ ∈Sn
P  
and τ ∈Sn Φn−1 0; x̄τ,t , Λx|r,τ,t = 1.
Proof. Since λτ,t = 0n−1 in this case, (30) follows immediately from (22), and the
final statement follows from integrating both sides of (30) over r ∈ R. 
The decomposition in (30) is of particular interest, since it expresses frm±,t+1 (r) as
a mixture of univariate normal densities with weights summing to unity.
For any p ∈ N, denote by µp (rm±,t+1 ) the p-th moment of rm±,t+1 . Then
µp (rm±,t+1 ) can be obtained in terms of the moments of truncated multivariate
normal distributions.
Proposition 3. If (r 0t+1 , r 0t )0 satisfies assumption 1, then
p   p−k  
X X p k
X p − k p−k−l
µp (rm±,t+1 ) = (k − 1)!! ςr|x,τ,t+1 r̄τ,m±,t+1|x
k l
τ ∈Sn k=0, l=0
k even (31)
Z

l
λ0τ,t Λ−1
x,τ,t (x − x̄τ,t ) φn−1 (x; x̄τ,t , Λx,τ,t ) dx,
x∈Rn−1
−

2
where Λx,τ,t , r̄τ,m±,t+1|x , and ςr|x,t+1,τ are as defined in corollary 3.
12

Proof. Using the density, frm± ,t+1 (r), from (27) allows µp (rm±,t+1 ) to be expressed
as the sum
XZ Z  
φn−1 (x; x̄τ,t , Λx,τ,t ) rp φ1 r; r̄τ,m±,t+1|x , ςr|x,τ,t+1
2
dr dx.
τ ∈Sn x∈Rn−1
−
r∈R

Now, applying the expression for the univariate moments from (10) and the definition
of r̄τ,m±,t+1|x from corollary 3, inner integrals can be computed as follows:
Z  
rp φ1 r; r̄τ,m±,t+1|x , ςr|x,τ,t+1
2
dr
r∈R
p  
X p
p−k
= k
(k − 1)!! ςr|x,τ,t+1 r̄τ,m±,t+1|x + λ0τ,t Λ−1
x,τ,t (x − xτ,t )
k
k=0
k even
p   p−k  
X p X p − k p−k−l
l
= k
(k − 1)!! ςr|x,τ,t+1 r̄τ,m±,t+1|x λ0τ,t Λ−1
x,τ,t (x − xτ,t ) .
k l
k=0 l=0
k even

Substituting back into the expression for µp (rm±,t+1 ) gives (31). 

In the special case where r t and r t+1 are independent, a more explicit expression for
µp (rm±,t+1 ) can be obtained.
Corollary 6. If (r 0t+1 , r 0t )0 satisfies assumption 1, and r t and r t+1 are independent,
then
p  
X X p k p−k
µp (rm±,t+1 ) = Φn−1 [0; x̄τ,t , Λx,τ,t ] (k − 1)!! ςr,τ,t+1 r̄τ,m ± ,t+1
. (32)
k
τ ∈Sn k=0,
k even

Proof. In this case, we have λτ,t = 0n−1 , and the only non-zero term in the innermost
sum in (31) is where l = 0. The corresponding integral reduces to
Z
φn−1 (x; x̄τ,t , Λx,τ,t ) dx = Φn−1 [0; x̄τ,t , Λx,τ,t ] ,
x∈Rn−1
−

and substituting back into (31) gives (32). 

We now consider the special case of 2 assets where r t+1 and r t are independent for
all t, and show that the first three moments from this model exhibit many of the
characteristics of empirical momentum portfolios.
Proposition 4. Assume that the conditions in corollary 4 are satisfied, and suppose


moreover that r t+1 and r t are independent for all t. Let πt = P x(12),1± ,t > 0 ,
h i
where (1, 2) denotes the non-trivial permutation in S2 , µ3,x,t = E x3(12),1± ,t , and
r̄t = r̄1,t − r̄t,2 . Then we have
E[r1± ,t+1 ] = (2π − 1)r̄t+1 , (33)
2 2 2


var[r1± ,t+1 ] = 1 − (2πt − 1) r̄t+1 + ςt+1 , (34)
2 3 2



skew[r1± ,t+1 ] = (2πt − 1) µ3,x,t+1 − 3r̄t+1 ςt+1 + r̄t+1 (2πt − 1) − 3 , (35)
2
where ςt+1 = var[x(12),1± ,t+1 ].
 13

Proof. Follows from direct computation using the independence of r t+1 and r t . 

Note that E[r1± ,t+1 ] > 0 if r̄1,t+1 > r̄2,t+1 , and this will hold if r1,t+1 − r2,t+1 has
a symmetric location-scale distribution around (r̄1,t+1 − r̄2,t+1 ). Next, the form of
2
var[r1± ,t+1 ] shows that variance of the CSM portfolio return is higher than ςt+1 .
2 3
Finally, and most interestingly, since µ3,x,t+1 = 3r̄t+1 ςt+1 + r̄t+1 in the case of nor-
mality, we have
3
(2πt − 1)2 − 2



skew[r1± ,t+1 ] = (2πt − 1) r̄t+1 .

So skew[r1± ,t+1 ] is always negative if 2πt − 1 is positive, and we have established

that, even when there is no autocorrelation, the CSM portfolio has positive expected
returns and negative skewness as long as we long, on average, the stock with the
highest expected return.

4. Quantile Portfolio Returns

It is common in the momentum literature to consider the properties of decile port-

folios of long momentum returns, while practitioners tend to use quintile portfolios.
In what follows, we will refer to an equally weighted long only portfolio consisting
of assets from an arbitrary percentile band as a quantile portfolio.
Let n ∈ N and let r t = (r1,t , . . . , rn,t )0 be the vector of asset returns as in the
previous section, and recall that a partition of n is a sequence χ = (n1 , . . . , nq ),
where ni , q ∈ N, q ≥ 1, and n = n1 + n2 + · · · + nq . Denote by Πn the set of all
partitions of n, and for any permutation τ ∈ Sn , define χ̄0 = 0,

χ̄i = χ̄i−1 + ni , 1≤i≤q (36)

χ̄i
1 X
rχ,τ,i,t = rτj ,t , 1≤i≤q (37)
ni
j=χ̄i−1 +1
0
r χ,τ,t = (rχ,τ,1,t , . . . , rχ,τ,q,t ) , (38)

0
z χ,τ,t = r 0χ,τ,t+1 , x0τ,t , (39)

where xτ,t is as defined in (14). An argument similar to those used in proposition 2

establishes the next result.
Proposition 5. Let Pτ be the permutation matrix corresponding to τ ∈ Sn , Dn−1
the (n − 1) × n matrix defined in (17), and for any χ = (χ1 , . . . , χq ) ∈ Πn let
1
0
ιχ,i = 00χ̄i−1 , 10ni , 00n−χ̄i ∈ Rn , 1≤i≤q (40)
ni
0
ιχ = (ιχ,1 , . . . , ιχ,q ) ∈ Rq×n . (41)

If r t satisfies assumption 1, then for any χ ∈ Πn , τ ∈ Sn , and t ∈ N we have

P 0 ι0 ιχ Pτ Λt+1,t Pτ0 Dn−1

0
   
ιχ Pτ r̄ t+1 ι P Λ
z χ,τ,t ∼ N , χ τ t+1,t+1 τ0 χ0 0 0 , (42)
Dn−1 Pτ r̄ t Dn−1 Pτ Λt,t+1 Pτ ιχ Dn−1 Pτ Λt+1,t Pτ Dn−1

where Λu,v are as defined in (19) for u, v ∈ {t, t + 1}.

14

Proof. For any χ = (χ1 , . . . , χq ) ∈ Πn and τ ∈ Sn , we have

    
rχ,τ,t+1 ιχ Pτ Oq×n r t+1
z χ,τ,t = = , (43)
xτ,t O(n−1)×n Dn−1 Pτ rt

0
and since r 0t+1 , r 0t is multivariate normal, and z χ,τ,t is a linear transformation of

0
r 0t+1 , r 0t ,
it follows that z χ,τ,t is also normally distributed with mean and covari-
ance as given in (42). 

Definition 2. For any χ ∈ Πn , the χ-quantile return, r χ,t+1 , at time t + 1 is defined

as
X
r χ,t+1 = I{xτ,t ≺0n−1 } r χ,τ,t+1 . (44)
τ ∈Sn

Theorem 5. Let frχ,t+1 (r) be the probability density of r χ,t+1 . If (r 0t+1 , r 0t )0 satisfies
assumption 1, then we have
X
frχ,t+1 (r) = φq (r; r̄ χ,τ,t+1 , Λr,χ,τ,t+1 )
τ ∈Sn (45)
Φn−1 0; x̄τ,t + Λr,x,χ,τ,t Λ−1
 
r,χ,τ,t+1 (r − r̄ χ,τ,t+1 ) , Λx|r,χ,τ,t ,

where r̄ χ,τ,t+1 = ιχ Pτ µt+1 , x̄τ,t = Dn−1 Pτ µt ,

Λr,χ,τ,t+1 = ιχ Pτ Λt+1,t+1 Pτ0 ι0χ ,

Λr,x,χ,τ,t = Dn−1 Pτ Λt,t+1 Pτ0 ι0χ ,
Λx|r,χ,τ,t = Dn−1 Pτ Λt+1,t Pτ0 Dn−1
0
− Λr,x,χ,τ,t Λ−1 0
r,χ,τ,t+1 Λr,x,χ,τ,t .

Proof. Let Rτ , for any τ ∈ Sn , be as defined in (23). Then applying arguments

similar to those used in the proof of theorem 3 show that

XZ
frχ,t+1 (r) = frχ,τ,t+1 ,xτ,t (r, x) dx.
n−1
τ ∈Sn x∈R−

Now, since z χ,τ,t = (r 0χ,τ,t+1 , x0τ,t )0 is multivariate normal by proposition 5, setting

x1 = xτ,t and x2 = r χ,τ,t+1 in theorem 1 gives

frχ,τ,t+1 ,xτ,t (r, x) = φq (r; r̄ χ,τ,t+1 , Λr,χ,τ,t+1 )

φn−1 x; x̄τ,t + Λr,x,χ,τ,t Λ−1


r,χ,τ,t+1 (r − r̄ χ,τ,t+1 ) , Λx|r,χ,τ,t .

Integrating over x ∈ Rn−1

− and summing over τ ∈ Sn completes the proof. 

Arguments analogous to those used in corollary 3 establishes the following alternative

expression for frχ,t+1 (r).
 15

Corollary 7. If the assumption in theorem 5 is satisfied, then

frχ,t+1 (r)
XZ
(46)
= φq r; r̄ χ|x,τ,t+1 , Λr|x,χ,τ,t+1 φn−1 (x; x̄τ,t , Λx,τ,t ) dx,
τ ∈Sn x∈Rn−1
−

where Λx,τ,t = Dn−1 Pτ Λt+1,t Pτ0 Dn−1

0
,

Λr|x,χ,τ,t+1 = Λr,χ,τ,t+1 − Λ0r,x,χ,τ,t Λ−1

x,τ,t+1 Λr,x,χ,τ,t ,

r̄ χ|x,τ,t+1 = r̄ χ,τ,t+1 + Λ0r,x,χ,τ,t Λ−1

x,τ,t+1 (x − x̄τ,t ).

The distribution of r χ,t+1 in the case of independent returns and in the degenerate
case can be obtained more explicitly.

Corollary 8. If (r 0t+1 , r 0t )0 satisfies assumption 1, and r t and r t+1 are independent,

then
X
frχ,t+1 (r) = φq (r; r̄ χ,τ,t+1 , Λr,χ,τ,t+1 ) Φn−1 [0; x̄τ,t , Λx,τ,t ] , (47)
τ ∈Sn

P
and τ ∈Sn Φn−1 [0; x̄τ,t , Λx,τ,t ] = 1.

Proof. Follows immediately from (45) since Λr,x,χ,τ,t = O(n−1)×q in this case. 

Corollary 9. If the assumptions in corollary 5 are satisfied, then in the notation of

theorem 5 we have

frχ,t+1 (r) = n! φq (r; µt+1 1q , Λr,χ,id,t+1 )

(48)
Φn−1 0; Λr,x,χ,id,t Λ−1
 
r,χ,id,t+1 (r − r̄ χ,id,t+1 ) , Λx|r,χ,id,t .

Proof. Follows from an argument similar to those used in corollary 5, since in this
case, r̄ χ,τ,t+1 = µt+1 1q and x̄τ,t = 0 for all τ ∈ Sn , while the remaining terms in
(45) are independent of τ ∈ Sn . 

Recall that for any p ∈ N and κ ∈ {1, 2, . . . , q}p , the κ-th moment, µκ (r χ,t+1 ), of
r χ,t+1 is defined by
" p #
Y
µκ (r χ,t+1 ) = E rχ,κi ,t+1 , (49)
i=1

where r χ,t+1 = (rχ,1,t+1 , . . . , rχ,q,t+1 ). Note that p represents the order of the mo-
ment being considered so that, for example, moments corresponding to p = 1 are
subportfolio means of the form E [rχ,j,t+1 ]. As was the case with rm± ,t+1 , the mo-
ments of r χ,t+1 can be obtained in terms of the moments of truncated multivariate
normal distributions.
16

Proposition 6. If (r 0t+1 , r 0t )0 satisfies assumption 1, then for any κ ∈ {1, 2, . . . , q}p

l
!
X X X Y

µκ (r χ,t+1 ) = Λr|x,χ,τ,t+1 κυk+2i−1 ,κυk+2i
τ ∈Sn k,l∈N υ∈Q(k,2l) i=1
k+2l=p
Z k
!
Y
Λ0r,x,χ,τ,t


r̄ χ,τ,t+1 + (x − x̄τ,t ) κυi
φn−1 (x; x̄τ,t , Λx,τ,t ) dx,
x∈Rn−1
− i=1

where Λr|x,χ,τ,t+1 is as defined in corollary 7.

Proof. Follows from an argument similar to those in the proof of proposition 3. 

Corollary 10. If (r 0t+1 , r 0t )0 satisfies assumption 1, and r t and r t+1 are indepen-
dent, then for any κ ∈ {1, 2, . . . , q}p we have
X
µκ (r χ,t+1 ) = Φn−1 [0; x̄τ,t , Λx,τ,t ]
τ ∈Sn
k
! l
!
X X Y Y
r̄χ,τ,t+1,κυi (Λr,χ,τ,t+1 )κυk+2i−1 ,κυk+2i .
k,l∈N υ∈Q(k,2l) i=1 i=1
k+2l=p

Proof. In this case, Λr,x,χ,τ,t = Oq×(n−1) and Λr|x,χ,τ,t+1 = Λr,χ,τ,t+1 . 

5. Numerical Analysis

We begin by examining the different shapes of CSM return densities that can be
obtained by considering the impact of increasing holding period while keeping the
ranking period fixed. For this, we set the ranking period to 1 month and vary the
holding period lengths from 1 month to 24 months. Monthly returns on Microsoft
and Exxon Mobil from March 2015 to April 2017 were used to compute the means
and covariances for the 2 asset example, and the data for Ford Motor Company
was added for the 3 asset case. The means and the covariances of the returns for
the 3 companies are shown in table 1. The returns were assumed to be temporally

MSFT XOM F
Mean 0.023376 -0.000714 -0.012367
0.004861 0.000573 0.001128
Covariance 0.000573 0.001811 0.000703
0.001128 0.000703 0.002330

Table 1 Means and covariances of the monthly returns for Microsoft (MSFT), Exxon Mobil
(XOM), and Ford Motor Company (F) over the period March 2015 to April 2017.

independent, and the densities were computed using the analytic expression (22),
with bivariate normal distribution required for the 3 asset case computed using the
 17

16
1 month holding period
6 month holding period
12 12 month holding period
24 month holding period

r
−0.25 0.25

Fig. 1 Densities of 2-asset CSM returns with ranking period fixed at 1 month and holding
period ranging from 1 month to 24 months. Returns over all holding periods were normalized
to represent monthly returns.

16
1 month holding period
6 month holding period
12 12 month holding period
24 month holding period

r
−0.25 0.25

Fig. 2 Densities of 3-asset CSM returns with ranking period fixed at 1 month and holding
period ranging from 1 month to 24 months. Returns over all holding periods were normalized
to represent monthly returns.

approximation given in West (2004). The densities for the 2 and 3 asset cases are
shown in figures 1 and 2 respectively.
The figures show that the CSM return density is approximately normal for short
holding periods, but progressively becomes more skewed and bimodal with increasing
holding period. This is consistent with the discussion following proposition 4 since,
18

in particular, the difference in the maximum and minimum mean returns, viz. r̄t+1
in the notation of proposition 4, increases with holding period length.
Since the analytic expression (22) for CSM return density becomes computa-
tionally infeasible as the number of assets, n, becomes large due to the number of
summands, we demonstrate a computationally efficient method for computing the
densities using Monte Carlo simulation. A comparison of the densities, computed
analytically and by Monte Carlo simulation, in the case of the 3 asset example with
Microsoft, Exxon Mobil, and Ford Motor Company with 1 month ranking period and
24 month holding period is shown in figure 3. Monte Carlo simulation used 262,143
paths and the time taken was approximately 1.28 seconds on a PC with Core i5-
4750 CPU @ 3.20GHz and 8 GB of RAM. As the figure shows, the density obtained
from simulation closely mirrors the analytical counterpart, and could be used for
computing densities and moments in practical situations.

16
◦
◦◦◦◦ ◦◦
◦ ◦◦
◦
◦ ◦
12 ◦ ◦
◦ ◦
◦ ◦
◦ ◦
◦
◦8 ◦
◦ ◦◦
◦ ◦
◦◦ ◦◦ ◦◦◦◦ ◦
◦◦ ◦
◦◦
◦◦ ◦
◦◦
◦
◦◦
4
◦
◦ ◦
◦◦
◦ ◦
◦
◦◦◦ ◦◦
◦◦◦ ◦◦◦
◦◦◦◦◦◦◦◦◦◦◦◦◦◦◦◦◦ ◦◦◦◦◦◦◦◦◦◦◦◦◦
r
−0.10 0.10

Fig. 3 Comparison of analytical and Monte Carlo densities for 3 assets. Solid curve is the
analytic density, and the curve traced out by ◦ is the density computed by Monte Carlo.

6. Conclusion

In this paper, analytic expressions for the density and the moments of cross sectional
momentum (CSM) returns were derived under the assumption that underlying as-
set returns are multivariate normal. In the special case where the asset returns in
the ranking and holding periods are independent, which we regard as an efficient
market situation, it was shown that the CSM return density reduces to a mixture of
univariate normals. Analogous results were also obtained for the so-called quantile
portfolios that are long only portfolios consisting of assets with returns in a given
quantile over the ranking period. Extension of the results to other return processes
is possible, and is the subject of future research.
 19

References

Asness, A. (1994), Variables that Explain Stock Returns, PhD Dissertation, University of
Chicago.
Asness, C. S., Liew, J. M. and Stevens, J. M. (1997), ‘Parallels between the cross-sectional
predictability of stock and country returns’, Journal of Portfolio Management 23, 79–87.
Asness, C. S., Moskowitz, T. J. and Pedersen, R. L. (2013), ‘Value and momentum everywhere’,
The Journal of Finance 58, 929–985.
Carhart, M. M. (1997), ‘On Persistence in Mutual Fund Performance’, Journal of Finance
52(1), 57–82.
Chan, K., Hameed, A. and Tong, W. (2000), ‘Profitability of Momentum Strategies in the
International Equity Markets’, Journal of Financial and Quantitative Analysis 35, 153–
172.
Daniel, K. and Moskowitz, T. J. (2016), ‘Momentum crashes’, Journal of Financial Economics
122(2), 221–247.
Devito, V., Kaiser, L., Menichetti, M. and Veress, A. (2012), Long-Only Momentum, Currency
Hedging and Transaction Costs: Implications for a Swiss Equity Investor. Available online
at: https://ssrn.com/abstract=2338506.
Erb, C. B. and Harvey, C. R. (2006), ‘The strategic and tactical value of commodity futures’,
Financial Analysts Journal 62, 69–97.
Fisher, G. S., Shah, R. and Titman, S. (2016), ‘Combining Value and Momentum’, Journal of
Investment Management 14(2), 33–48.
Foltice, B. and Langer, T. (2015), Profitable Momentum Trading Strategies for Individual
Investors, Scholarship and Professional Work – Business. Paper 264, Butler University.
Grinblatt, M. and Moskowitz, T. J. (2004), ‘Predicting stock price movements from past
returns: The role of consistency and tax-loss selling’, Journal of Financial Economics
71, 541–579.
Grundy, B. and Martin, J. S. (2001), ‘Understanding the nature of the risks and the source of
the rewards to momentum investing’, Review of Financial Studies 14, 29–78.
Hameed, A. and Yuanto, K. (2002), ‘Momentum Strategies: Evidence from the Pacific Basin
Stock Markets’, Journal of Financial Research 25(3), 383–397.
Israel, R. and Moskowitz, T. J. (2013), ‘The Role of Shorting, Firm Size, and Time on Market
Anomalies’, Journal of Financial Economics 108, 275–301.
Jegadeesh, N. and Titman, S. (1993), ‘Returns to buying winners and selling losers: Implica-
tions for stock market efficiency’, The Journal of Finance 48(1), 65–91.
Jegadeesh, N. and Titman, S. (2001), ‘Profitability and Momentum Strategies: An Evaluation
of Alternative Explanations’, Journal of Finance 56, 699–720.
Lewellen, J. (2002), ‘Momentum and autocorrelation in Stock Returns’, Review of Financial
Studies 15, 533–563.
Lo, A. and MacKinlay, C. (1990), ‘When are Contrarian Profits due to Stock Marjket Overre-
action?’, Review of Financial Studies 3(2), 175–205.
Menkhoff, L., Sarno, L., Schmeling, M. and Schrimpf, A. (2012), ‘Currency momentum strate-
gies’, Journal of Financial Economics 106, 660–684.
Moskowitz, T., Ooi, Y. H. and Pedersen, L. H. (2012), ‘Time Series Momentum’, Journal of
Financial Economics 104(2), 228–250.
Muirhead, R. J. (1982), Aspects of Multivariate Statistical Theory, John Wiley & Sons, New
Jersey.
Okunev, J. and White, D. (2003), ‘Do momentum-based strategies still work in foreign currency
markets?’, Journal of Financial and Quantitative Analysis 38, 425–447.
Rey, D. M. and Schmid, M. M. (2007), ‘Feasible Momentum Strategies: Evidence from the
Swiss Stock Market. Financial Markets and Portfolio Management’, Financial Markets
and Portfolio Management 21(3), 325–352.
Richards, A. (1997), ‘Winner-Loser Reversals in National Stock Market Indices: Can They be
Explained?’, Journal of Finance 52, 2129–2144.
Rotman, J. (1995), An Introduction to the Theory of Groups, Springer-Verlag, New York.
Rouwenhorst, K. G. (1998), ‘International Momentum Strategies’, Journal of Finance 53, 267–
284.
Sefton, J. and Scowcroft, A. (2004), A Decomposition of Portfolio Momentum Returns, Dis-
cussion Paper TBS/DP04/9, Tanaka Business School, Imperial College London.
West, G. (2004), Better Approximations to Cumulative Normal Functions, Working paper.
Withers, C. S. (1985), ‘The Moments of the Multivariate Normal’, Bulletin of Australian
Mathematics Society 32, 103–107.