Statistical Arbitrage in the U.S.

Equities Market
Marco Avellaneda

and Jeong-Hyun Lee

First draft: July 11, 2008

This version: June 15, 2009
Abstract
We study model-driven statistical arbitrage in U.S. equities. The trading
signals are generated in two ways: using Principal Component Analysis
and using sector ETFs. In both cases, we consider the residuals, or idio-
syncratic components of stock returns, and model them as mean-reverting
processes. This leads naturally to contrarian trading signals.
The main contribution of the paper is the construction, back-testing
and comparison of market-neutral PCA- and ETF- based strategies ap-
plied to the broad universe of U.S. stocks. Back-testing shows that, af-
ter accounting for transaction costs, PCA-based strategies have an av-
erage annual Sharpe ratio of 1.44 over the period 1997 to 2007, with
much stronger performances prior to 2003. During 2003-2007, the aver-
age Sharpe ratio of PCA-based strategies was only 0.9. Strategies based
on ETFs achieved a Sharpe ratio of 1.1 from 1997 to 2007, experiencing
a similar degradation after 2002.
We also introduce a method to account for daily trading volume infor-
mation in the signals (which is akin to using trading time as opposed to
calendar time), and observe signicant improvement in performance in the
case of ETF-based signals. ETF strategies which use volume information
achieve a Sharpe ratio of 1.51 from 2003 to 2007.
The paper also relates the performance of mean-reversion statistical
arbitrage strategies with the stock market cycle. In particular, we study
in detail the performance of the strategies during the liquidity crisis of the
summer of 2007. We obtain results which are consistent with Khandani
and Lo (2007) and validate their unwinding theory for the quant fund
drawdown of August 2007.

Courant Institute of Mathematical Sciences, 251 Mercer Street, New York, N.Y. 10012
USA

Finance Concepts, 49-51 Avenue Victor-Hugo, 75116 Paris, France.

1
1 Introduction
The term statistical arbitrage encompasses a variety of strategies and investment
programs. Their common features are: (i) trading signals are systematic, or
rules-based, as opposed to driven by fundamentals, (ii) the trading book is
market-neutral, in the sense that it has zero beta with the market, and (iii) the
mechanism for generating excess returns is statistical. The idea is to make many
bets with positive expected returns, taking advantage of diversication across
stocks, to produce a low-volatility investment strategy which is uncorrelated
with the market. Holding periods range from a few seconds to days, weeks or
even longer.
Pairs-trading is widely assumed to be the ancestor of statistical arbitrage.
If stocks P and Q are in the same industry or have similar characteristics (e.g.
Exxon Mobile and Conoco Phillips), one expects the returns of the two stocks
to track each other after controlling for beta. Accordingly, if P
t
and Q
t
denote
the corresponding price time series, then we can model the system as
ln(P
t
/P
t0
) = (t t
0
) +ln(Q
t
/Q
t0
) + X
t
(1)
or, in its dierential version,
dP
t
P
t
= dt +
dQ
t
Q
t
+dX
t
, (2)
where X
t
is a stationary, or mean-reverting, process. This process will be re-
ferred to as the cointegration residual, or residual, for short, in the rest of the
paper. In many cases of interest, the drift is small compared to the uctua-
tions of X
t
and can therefore be neglected. This means that, after controlling for
beta, the long-short portfolio oscillates near some statistical equilibrium. The
model suggests a contrarian investment strategy in which we go long 1 dollar of
stock P and short dollars of stock Q if X
t
is small and, conversely, go short P
and long Q if X
t
is large. The portfolio is expected to produce a positive return
as valuations converge (see Pole (2007) for a comprehensive review on statistical
arbitrage and co-integration). The mean-reversion paradigm is typically asso-
ciated with market over-reaction: assets are temporarily under- or over-priced
with respect to one or several reference securities (Lo and MacKinley (1990)).
Another possibility is to consider scenarios in which one of the stocks is
expected to out-perform the other over a signicant period of time. In this
case the co-integration residual should not be stationary. This paper will be
principally concerned with mean-reversion, so we dont consider such scenarios.
Generalized pairs-trading, or trading groups of stocks against other groups
of stocks, is a natural extension of pairs-trading. To explain the idea, we con-
sider the sector of biotechnology stocks. We perform a regression/cointegration
analysis, following (1) or (2), for each stock in the sector with respect to a
benchmark sector index, e.g. the Biotechnology HOLDR (BBH). The role of
the stock Q would be played by BBH and P would an arbitrary stock in the
biotechnology sector. The analysis of the residuals, based of the magnitude of
2
X
t
, suggests typically that some stocks are cheap with respect to the sector,
others expensive and others fairly priced. A generalized pairs trading book, or
statistical arbitrage book, consists of a collection of pair trades of stocks rel-
ative to the ETF (or, more generally, factors that explain the systematic stock
returns). In some cases, an individual stock may be held long against a short
position in ETF, and in others we would short the stock and go long the ETF.
Due to netting of long and short positions, we expect that the net position in
ETFs will represent a small fraction of the total holdings. The trading book
will look therefore like a long/short portfolio of single stocks. This paper is
concerned with the design and performance-evaluation of such strategies.
The analysis of residuals is our starting point. Signals will be based on
relative-value pricing within a sector or a group of peers, by decomposing stock
returns into systematic and idiosyncratic components and statistically modeling
the idiosyncratic part. The general decomposition may look like
dP
t
P
t
= dt +
n

j=1

j
F
(j)
t
+ dX
t
, (3)
where the terms F
(j)
t
, j = 1, ..., n represent returns of risk-factors associated with
the market under consideration. This leads to the interesting question of how
to derive equation (3) in practice. The question also arises in classical portfolio
theory, but in a slightly dierent way: there we ask what constitutes a good
set of risk-factors from a risk-management point of view. Here, the emphasis
is instead on the residual that remains after the decomposition is done. The
main contribution of our paper will be to study how dierent sets of risk-factors
lead to dierent residuals and hence to dierent prot-loss (PNL) for statistical
arbitrage strategies.
Previous studies on mean-reversion and contrarian strategies include Lehmann
(1990), Lo and MacKinlay (1990) and Poterba and Summers (1988). In a recent
paper, Khandani and Lo (2007) discuss the performance of the Lo-MacKinlay
contrarian strategies in the context of the liquidity crisis of 2007 (see also refer-
ences therein). The latter strategies have several common features with the ones
developed in this paper. Khandani and Lo (2007) market-neutrality is enforced
by ranking stock returns by quantiles and trading winners-versus-losers, in a
dollar-neutral fashion. Here, we use risk-factors to extract trading signals, i.e.
to detect over- and under-performers. Our trading frequency is variable whereas
Khandani-Lo trade at xed time-intervals. On the parametric side, Poterba and
Summers (1988) study mean-reversion using auto-regressive models in the con-
text of international equity markets. The models of this paper dier from the
latter mostly in that we immunize stocks against market factors, i.e. we consider
mean-reversion of residuals (relative prices) and not of the prices themselves.
The paper is organized as follows: in Section 2, we study market-neutrality
using two dierent approaches. The rst method consists in extracting risk-
factors using Principal Component Analysis (Jollie (2002)). The second method
uses industry-sector ETFs as proxies for risk factors. Following other authors,
we show that PCA of the correlation matrix for the broad equity market in
3
the U.S. gives rise to risk-factors that have economic signicance because they
can be interpreted as long-short portfolios of industry sectors. However, the
stocks that contribute the most to a particular factor are not necessarily the
largest capitalization stocks in a given sector. This suggests that PCA-based
risk factors may not be as biased towards large-capitalization stocks as ETFs,
as the latter are generally capitalization-weighted. We also observe that the
variance explained by a xed number of PCA eigenvectors varies signicantly
across time, which leads us to conjecture that the number of explanatory factors
needed to describe stock returns (to separate systematic returns from residuals)
is variable and that this variability is linked with the investment cycle, or the
changes in the risk-premium for investing in the equity market.
1
This might
explain some of the dierences that we found in performance between the PCA
and ETF methods.
In Section 3 and 4, we construct the trading signals. This involves the statis-
tical estimation of the residual process for each stock at the close of each trading
day, using 60 days of historical data prior to that date. Estimation is always
done looking back 60 days from the trade date, thus simulating decisions which
might take place in real trading. The trading signals correspond to signicant
deviations of the residual process from its estimated mean. Using daily end-of-
day (EOD) data, we perform a calculation of daily trading signals, going back
to 1996 in the case of PCA strategies and to 2002 in the case of ETF strategies,
across the universe of stocks with market-capitalization of more than 1 billion
USD at the trade date. The condition that the company must have a given
capitalization at the trade date (as opposed to at the time when the paper was
written), avoids survivorship bias.
Estimation and trading rules are kept simple to avoid data-mining. For each
stock, the estimation of the residual process is done using a 60-day trailing
window because this corresponds roughly to one earnings cycle. The length
of the window is not changed from one stock to another. We select as entry
point for trading any residual that deviates by 1.25 standard deviations from
equilibrium, and we exit trades if the residual is less than 0.5 standard deviations
from equilibrium, uniformly across all stocks.
In Section 5 we back-test several strategies which use dierent sets of fac-
tors to generate residuals, namely: (i) synthetic ETFs based on capitalization-
weighted indices
2
, (ii) actual ETFs, (iii) a xed number of factors generated by
PCA, (iv) a variable number of factors generated by PCA. Due to the mecha-
nism described above used to generate trading signals, the simulation is always
out-of-sample, in the sense that the estimation of the residual process at time t
uses information available only for the 60 days prior to this time. In all trades,
we assume a slippage/transaction cost of 0.05% or 5 basis points per trade (a
round-trip transaction cost of 10 basis points).
1
See Scherer and Avellaneda (2002) for similar observations for Latin American debt se-
curities in the 1990s.
2
Synthetic ETFs are capitalization-weighted sector indexes formed with the stocks of each
industry that are present in the trading universe at the time the signal in calculated. We used
synthetic ETFs because most sector ETFs where launched only after 2002.
4
In Section 6, we consider a modication of the strategy in which signals
are estimated in trading time as opposed to calendar time. In the statistical
analysis, using trading time on EOD signals is eectively equivalent to multi-
plying daily returns by a factor which is inversely proportional to the trading
volume for the past day. This modication accentuates (i.e. tends to favor)
contrarian price signals taking place on low volume and mitigates (i.e. tends
not to favor) contrarian price signals which take place on high volume. It is
as if we believe more a print that occurs on high volume and are less ready
to bet against it. Back-testing the statistical arbitrage strategies using trading-
time signals leads to improvements in most strategies, suggesting that volume
information, is valuable in the context of mean-reversion strategies , even at the
EOD sampling frequency and not just only for intra-day trading.
In Section 7, we discuss the performance of statistical arbitrage in 2007,
and particularly around the inception of the liquidity crisis of August 2007. We
compare the performances of the mean-reversion strategies with the ones studied
in the recent work of Khandani and Lo (2007). Conclusions are presented in
Section 8.
2 A quantitative view of risk-factors and market-
neutrality
Let us denote by R
i

N
i=1
the returns of the dierent stocks in the trading
universe over an arbitrary one-day period (from close to close). Let F represent
the return of the market portfolio over the same period, (e.g. the return on
a capitalization-weighted index, such as the S&P 500). We can write, for each
stock in the universe,
R
i
=
i
F +

R
i
, (4)
which is a simple regression model decomposing stock returns into a systematic
component
i
F and an (uncorrelated) idiosyncratic component

R
i
. Alterna-
tively, we consider multi-factor models of the form
R
i
=
m

j=1

ij
F
j
+

R
i
. (5)
Here there are m factors, which can be thought of as the returns of benchmark
portfolios representing systematic factors. A trading portfolio is said to be
market-neutral if the dollar amounts Q
i

N
i=1
invested in each of the stocks are
such that

j
=
N

i=1

ij
Q
i
= 0, j = 1, 2, ..., m. (6)
The coecients
j
correspond to the portfolio betas, or projections of the port-
folio returns on the dierent factors. A market-neutral portfolio has vanishing
5
portfolio betas; it is uncorrelated with the market portfolio or factors that drive
the market returns. It follows that the portfolio returns satisfy
N

i=1
Q
i
R
i
=
N

i=1
Q
i

j=1

ij
F
j

+
N

i=1
Q
i

R
i
=
m

j=1

i=1

ij
Q
i

F
j
+
N

i=1
Q
i

R
i
=
N

i=1
Q
i

R
i
(7)
Thus, a market-neutral portfolio is aected only by idiosyncratic returns. We
shall see below that, in G8 economies, stock returns are explained by approxi-
mately m=15 factors (or between 10 and 20 factors), and that the systematic
component of stock returns explains approximately 50% of the variance (see
Plerou et al. (2002) and Laloux et al. (2000)). The question is how to dene
factors.
2.1 The PCA approach
A rst approach for extracting factors from data is to use Principal Components
Analysis (Jollie (2002)). This approach uses historical share-price data on a
cross-section of N stocks going back M days in history. For simplicity of expo-
sition, the cross-section is assumed to be identical to the investment universe,
although this need not be the case in practice.
3
Let us represent the stocks
return data, on any given date t
0
, going back M + 1 days as a matrix
R
ik
=
S
i(t0(k1)t)
S
i(t0kt)
S
i(t0kt)
, k = 1, ..., M, i = 1, ..., N,
where S
it
is the price of stock i at time t adjusted for dividends and t = 1/252.
Since some stocks are more volatile than others, it is convenient to work with
standardized returns,
Y
ik
=
R
ik
R
i

i
where
R
i
=
1
M
M

k=1
R
ik
and

2
i
=
1
M 1
M

k=1
(R
ik
R
i
)
2
3
For instance, the analysis can be restricted to the members of the S&P500 index in the
US, the Eurostoxx 350 in Europe, etc.
6
The empirical correlation matrix of the data is dened by

ij
=
1
M 1
M

k=1
Y
ik
Y
jk
, (8)
which is symmetric and non-negative denite. Notice that, for any index i, we
have

ii
=
1
M 1
M

k=1
(Y
ik
)
2
=
1
M 1
M

k=1
(R
ik
R
i
)
2

2
i
= 1.
The dimensions of are typically 500 by 500, or 1000 by 1000, but the data
is small relative to the number of parameters that need to be estimated. In
fact, if we consider daily returns, we are faced with the problem that very long
estimation windows M N dont make sense because they take into account
the distant past which is economically irrelevant. On the other hand, if we just
consider the behavior of the market over the past year, for example, then we are
faced with the fact that there are considerably more entries in the correlation
matrix than data points. In this paper, we always use an estimation window for
the correlation matrix of 1-year (252 trading days) prior to the trading date.
The commonly used solution to extract meaningful information from the
data is to model the correlation matrix
4
. We consider the eigenvectors and
eigenvalues of the empirical correlation matrix and rank the eigenvalues in de-
creasing order:
N
1
>
2

3
...
N
0.
We denote the corresponding eigenvectors by
v
(j)
=

v
(j)
1
, ...., v
(j)
N

, j = 1, ..., N.
A cursory analysis of the eigenvalues shows that the spectrum contains a few
large eigenvalues which are detached from the rest of the spectrum (see Figure
1). We can also look at the density of states
D(x, y) =
#of eigenvalues between x and y
N
(see Figure 2). For intervals (x, y) near zero, the function D(x, y) corresponds
to the bulk spectrum or noise spectrum of the correlation matrix. The
eigenvalues at the top of the spectrum which are isolated from the bulk spectrum
are obviously signicant. The problem that is immediately evident by looking
at Figures 1 and 2 is that there are fewer detached eigenvalues than industry
sectors. Therefore, we expect that the boundary between signicant and
4
We refer the reader to Laloux et al. (2000), and Plerou et al. (2002) who studied the
correlation matrix of the top 500 stocks in the US in this context.
7
Figure 1: Top 50 eigenvalues of the correlation matrix of market returns com-
puted on May 1 2007 estimated using a 1-year window and a universe of 1417
stocks (see also Table 3). (Eigenvalues are measured as percentage of explained
variance.)
8
Figure 2: The density of states for May 1-2007 estimated using a 1-year window,
corresponding to the same data used to generate Figure 1. Notice that there are
some detached eigenvalues, and a bulk spectrum. The relevant eigenvalues
includes the detached eigenvalues as well as a eigenvalues in the edge of the bulk
spectrum.
9
noise eigenvalues to be somewhat blurred and to correspond to be at the
edge of the bulk spectrum. This leads to two possibilities: (a) we take into
account a xed number of eigenvalues to extract the factors (assuming a number
close to the number of industry sectors) or (b) we take a variable number of
eigenvectors, depending on the estimation date, in such a way that a sum of the
retained eigenvalues exceeds a given percentage of the trace of the correlation
matrix. The latter condition is equivalent to saying that the truncation explains
a given percentage of the total variance of the system.
Let
1
, ...,
m
, m < N be the signicant eigenvalues in the above sense. For
each index j, we consider a the corresponding eigenportfolio, which is such
that the respective amounts invested in each of the stocks is dened as
Q
(j)
i
=
v
(j)
i

i
.
The eigenportfolio returns are therefore
F
jk
=
N

i=1
v
(j)
i

i
R
ik
j = 1, 2, ..., m. (9)
It is easy for the reader to check that the eigenportfolio returns are uncorrelated
in the sense that the empirical correlation of F
j
and F
j
vanishes for j = j

. The
factors in the PCA approach are the eigenportfolio returns.
Figure 3: Comparative evolution of the principal eigenportfolio and the
capitalization-weighted portfolio from May 2006 to April 2007. Both portfo-
lios exhibit similar behavior.
Each stock return in the investment universe can be decomposed into its
projection on the m factors and a residual, as in equation (4). Thus, the PCA
10
approach delivers a natural set of risk-factors that can be used to decompose our
returns. It is not dicult to verify that this approach corresponds to modeling
the correlation matrix of stock returns as a sum of a rank-m matrix correspond-
ing to the signicant spectrum and a diagonal matrix of full rank,

ij
=
m

k=0

k
v
(k)
i
v
(k)
j
+
2
ii

ij
,
where
ij
is the Kronecker delta and
2
ii
is given by

2
ii
= 1
m

k=0

k
v
(k)
i
v
(k)
i
so that
ii
= 1. This means that we keep only the signicant eigenvalues/eigenvectors
of the correlation matrix and add a diagonal noise matrix for the purposes of
conserving the total variance of the system.
2.2 Interpretation of the eigenvectors/eigenportfolios
As pointed out by several authors (see for instance, Laloux et al.(2000)), the
dominant eigenvector is associated with the market portfolio, in the sense
that all the coecients v
(1)
i
, i = 1, 2.., N are positive. Thus, the eigenport-
folio has positive weights Q
(1)
i
=
v
(1)
i
i
. We notice that these weights are in-
versely proportional to the stocks volatility. This weighting is consistent with
the capitalization-weighting, since larger capitalization companies tend to have
smaller volatilities. The two portfolios are not identical but are good proxies
for each other,
5
as shown in Figure 3.
To interpret the other eigenvectors, we observe that (i) the remaining eigen-
vectors must have components that are negative, in order to be orthogonal
to v
(1)
; (ii) given that there is no natural order in the stock universe, the
shape analysis that is used to interpret the PCA of interest-rate curves (Litter-
man and Scheinkman (1991) or equity volatility surfaces (Cont and Da Fonseca
(2002)) does not apply here. The analysis that we use here is inspired by Scherer
and Avellaneda (2002), who analyzed the correlation of sovereign bond yields
across dierent Latin American issuers (see also Plerou et. al.(2002) who made
similar observations). We rank the coecients of the eigenvectors in decreasing
order:
v
(2)
n1
v
(2)
n2
... v
(2)
n
N
,
the sequence n
i
representing a re-labeling of the companies. In this new order-
ing, we notice that the neighbors of a particular company tend to be in the
5
The positivity of the coecients of the rst eigenvector of the correlation matrix in the
case when all assets have non-negative correlation follows from Kreins Theorem. In practice,
the presence of commodity stocks and mining companies implies that there are always a few
negatively correlated stock pairs. In particular, this explains why there are a few negative
weights in the principal eigenportfolio in Figure 4.
11
same industry group. This property, which we call coherence, holds true for v
(2)
and for other high-ranking eigenvectors. As we descend in the spectrum towards
the noise eigenvectors, the property that nearby coecients correspond to rms
in the same industry is less true and coherence will not hold for eigenvectors of
the noise spectrum (almost by denition!). The eigenportfolios can therefore be
interpreted as long-short portfolios at the level of industries or sectors.
Figure 4: First eigenvector sorted by coecient size. The x-axis shows the ETF
corresponding to the industry sector of each stock.
2.3 The ETF approach: using the industries
Another method for extracting residuals consists in using the returns of sector
ETFs as factors. Table 3 shows a sample of industry sectors number of stocks
of companies with capitalization of more than 1 billion USD at the beginning
of January 2007, classied by sectors. It gives an idea of the dimensions of the
trading universe and the distribution of stocks corresponding to each industry
sector. We also include, for each industry, the ETF that can be used as a
risk-factor for the stocks in the sector for the simplied model (11).
Unlike the case of eigenportfolios, which are uncorrelated by construction,
ETF returns are correlated. This can lead to redundancies in the factor decom-
position: strongly correlated ETFs sometimes give rise to large factor loadings
with opposing signs for stocks that belong to or are strongly correlated with dif-
ferent ETFs. There are several approaches that can be used to remedy this: one
is a robust version of multiple regression aiming at sparse representations. For
example, the matching pursuit algorithm (Davis, Mallat & Avellaneda (1997))
12
Figure 5: Second eigenvector sorted by coecient size. Labels as in Figure 4.
Figure 6: Third eigenvector sorted by coecient size. Labels as in Figure 4.
13
Top 10 Stocks Bottom 10 Stocks
Energy, oil and gas Real estate, nancials, airlines
Suncor Energy Inc. American Airlines
Quicksilver Res. United Airlines
XTO Energy Marshall & Isley
Unit Corp. Fifth Third Bancorp
Range Resources BBT Corp.
Apache Corp. Continental Airlines
Schlumberger M & T Bank
Denbury Resources Inc. Colgate-Palmolive Company
Marathon Oil Corp. Target Corporation
Cabot Oil & Gas Corporation Alaska Air Group, Inc.
Table 1: The top 10 stocks and bottom 10 stocks in second eigenvector.
Top 10 Stocks Bottom 10 Stocks
Utility Semiconductor
Energy Corp. Arkansas Best Corp.
FPL Group, Inc. National Semiconductor Corp.
DTE Energy Company Lam Research Corp.
Pinnacle West Capital Corp. Cymer, Inc.
The Southern Company Intersil Corp.
Consolidated Edison, Inc. KLA-Tencor Corp.
Allegheny Energy, Inc. Fairchild Semiconductor International
Progress Energy, Inc. Broadcom Corp.
PG&E Corporation Cellcom Israel Ltd.
FirstEnergy Corp. Leggett & Platt, Inc.
Table 2: The top 10 stocks and bottom 10 stocks in third eigenvector.
14
which favors sparse representations is preferable to a full multiple regression.
Another class of regression methods known as ridge regression would achieve
the similar goal (see, for instance Jollie (2002)).
In this paper we use a simple approach. We associate to each stock a single
sector ETF (following the partition of the market shown in Table 3) and perform
a regression of the stock returns on the corresponding ETF returns, i.e.
R
i
= R
ETFi
+

R
i
.
where ETF
i
is associated with stock i.
3 A relative-value model for equity valuation
We propose a quantitative approach to stock valuation based on its relative
performance within industry sector ETFs or, alternatively, with respect to the
constructed PCA factors. In Section 4, we present a modication of this ap-
proach which take into account the trading volume in the stocks, within a similar
framework. Our investment model is purely based on price and volume data,
although in principle it could be extended to include fundamental factors, such
changes in analysts recommendations, earnings momentum, and other quan-
tiable factors.
We shall use continuous-time notation and denote stock prices by S
i
(t), ...., S
N
(t),
where t is time measured in years from some arbitrary starting date. Based on
the multi-factor models introduced in the previous section, we assume that stock
returns satisfy the system of stochastic dierential equations
dS
i
(t)
S
i
(t)
=
i
dt +
N

j=1

ij
dI
j
(t)
I
j
(t)
+dX
i
(t), (10)
where the term
N

j=1

ij
dI
j
(t)
I
j
(t)
represents the systematic component of returns (driven by the returns of the
eigenportfolios or the ETFs). The coecients
ij
are the corresponding factor
loadings.
In the case of ETF factors, we work with the model
dS
i
(t)
S
i
(t)
=
i
dt +
i
dI(t)
I(t)
+dX
i
(t), (11)
where I(t) is the ETF corresponding to the stock under consideration.
6
In both cases, the idiosyncratic component of the return is
d

X
i
(t) =
i
dt +dX
i
(t).
6
In other words, we analyze a pair-trade between each stock and its assigned ETF.
15
Table 3: Trading universe on January 1, 2007: breakdown by sectors. The
market is therefore partitioned into 15 sectors and each stock is associated with
an ETF.
16
Here,
i
represents the drift of the idiosyncratic component, i.e.
i
dt is the
excess rate of return of the stock in relation to market or industry sector over the
relevant period. The term dX
i
(t) is assumed to be the increment of a stationary
stochastic process which models price uctuations which are not reected the
factors.
Based on these assumptions, we introduce a simple parametric model for
X
i
(t) which can be estimated easily, namely, the Ornstein-Uhlenbeck process:
dX
i
(t) =
i
(m
i
X
i
(t)) dt +
i
dW
i
(t),
i
> 0. (12)
This process is stationary and auto-regressive with lag 1 (AR-1 model).
7
In
particular, the increment dX
i
(t) has unconditional mean equal to zero and con-
ditional mean equal to
E dX
i
(t)[X
i
(s), s t =
i
(m
i
X
i
(t)) dt .
The conditional mean, or forecast of expected daily returns of the residual
process, is positive or negative according to the sign of m
i
X
i
(t).
The parameters of the stochastic dierential equation,
i
,
i
, m
i
and
i
,are
specic to each stock. They are assumed de facto to vary slowly in relation
to the Brownian motion increments dW
i
(t), in the time-window of interest. In
the simulations, we estimate the residual processes for each stock on a window
of length 60 days, assuming implicitly that the parameters are constant over
the window. We accept this hypothesis on stocks for which the speed of mean-
reversion (the estimate of ) is suciently high and reject it for stocks having a
slow speed of mean-reversion. Details on how to estimate the model are given
in the next section and in the Appendix.
It follows from (12) that
X
i
(t
0
+t) = e
it
X
i
(t
0
) +

1 e
it

m
i
+
i
t0+t

t0
e
i(t0+ts)
dW
i
(s) .
(13)
Letting t tend to innity, we see that equilibrium probability distribution for
the process X
i
(t) is normal with
E X
i
(t) = m
i
and V ar X
i
(t) =

2
i
2
i
. (14)
According to Equation (10), an investment in a market-neutral long-short port-
folio in which the agent is long $1 in the stock and short
ij
dollars in the j
th
factor (or the ETF, in the case of the ETF framework) has an expected 1-day
return
7
Among the myriad of mean-reverting processes, we choose the simplest one to model
residuals. We encourage the reader interested in practical implementations to experiment
with other models, as we have in the course of this research.
17

i
dt +
i
(m
i
X
i
(t)) dt .
The second term corresponds to the models prediction for the return based on
the position of the stationary process X
i
(t): it forecasts a negative return if
X
i
(t) is suciently high and a positive return if X
i
(t) is suciently low.
The parameter
i
is called the speed of mean-reversion and

i
=
1

i
represents the characteristic time-scale for mean reversion. If 1 the stock
reverts quickly to its mean and the eect of the drift is negligible. In our
strategies, and to be consistent with the estimation procedure that uses constant
parameters, we are interested in stocks with fast mean-reversion, i.e. such that

i
< T
1
.
4 Signal generation
Based on this simple model, we dened several trading signals. We considered
an estimation window of 60 business days i.e. T
1
= 60/252. We selected stocks
with mean-reversion times less than 1/2 period ( > 252/30 = 8.4). Typical
descriptive statistics for signal estimation are presented in Table 5. For the
details of the estimation of the O-U process and more statistical details on
signal generation see the Appendix.
4.1 Pure mean-reversion
We focus only on the process X
i
(t), neglecting the drift
i
. We know that the
equilibrium variance is

eq,i
=

i

2
i
=
i

i
2
Accordingly, we dene the dimensionless variable
s
i
=
X
i
(t) m
i

eq,i
. (15)
We call this variable the s-score.
8
See Figure 7 for a graph showing the evolution
of the s-score for residuals of JPM against the Financial SPDR, XLF. The s-
score measures the distance to equilibrium of the cointegrated residual in units
standard deviations, i.e. how far away a given stock is from the theoretical
equilibrium value associated with our model.
8
See the Appendix for practical details on estimating the s-score.
18
Our basic trading signal based on mean-reversion is
buy to open if s
i
< s
bo
sell to open if s
i
> +s
so
close short position if s
i
< +s
bc
close long position s
i
> s
sc
(16)
where the cuto values are determined empirically. Entering a trade, e.g. buy to
open, means buying one dollar of the corresponding stock and selling
i
dollars
of its sector ETF or, in the case of using multiple factors,
i1
dollars of ETF
#1,
i2
dollars of ETF #2, ...,
im
dollars of ETF #m. Similarly, closing a long
position means selling stock and buying ETFs.
Since we expressed all quantities in dimensionless variables, we expect the
cutos s
bo
, s
bo
, s
bc
, s
sc
to be valid across the dierent stocks. We selected the
cutos empirically, based on simulating strategies from 2000 to 2004 in the case
of ETF factors. Based on this analysis, we found that a good choice of cutos
is
s
bo
= s
so
= 1.25
s
bc
= 0.75 and s
sc
= 0.50
Thus, we enter opening trades when the s-score exceeds 1.25 in absolute value.
We close long trades when the s-score reaches -0.50. Closing short trades sooner,
at s=0.75, gives slightly better results than 0.50 in the training period of 2000-
2002, so we use this slightly asymmetric rule in backtesting.
The rationale for opening trades only when the s-score s
i
is far from equilib-
rium is to trade only when we think that we detected an anomalous excursion
of the co-integration residual. We then need to consider when we close trades.
Closing trades when the s-score is near zero also makes sense, since we expect
most stocks to be near equilibrium most of the time. Thus, our trading rule
detects stocks with large excursions and trades assuming these excursions will
revert to the mean in a period of the order of the mean-reversion time
i
.
4.2 Mean-reversion with drift
In the previous section, the presence of the drift was ignored. That is, we
assumed that the drift was statistically insignicant in comparison with the am-
plitude of excursions
eq
= /

2. In this section, we show how to incorporate

the drift. This leads to a modied s-score.
We consider the conditional expectation of the residual return over a period
of time dt, namely,
19
Figure 7: Evolution of the s-score of JPM ( vs. XLF ) from January 2006 to
December 2007.

i
dt +
i
(m
i
X
i
(t)) dt =
i

i
+m
i
X
i
(t)

dt
=
i

eq,i
s
i

dt.
This suggests that the dimensionless decision variable is the modied s-score
s
mod,i
= s
i


i

eq,i
= s
i


i

eq,i
. (17)
To make contact with the analysis of the pure mean-reversion strategy, con-
sider for example the case of shorting stock. In the previous framework, we
short stock if the s-score is large enough. The modied s-score is larger if
i
is
negative, and smaller if
i
is positive. Therefore, it will be harder to generate
a short signal if we think that the residual has an upward drift and easier to
short if we think that the residual has a downward drift. If the s-score is zero,
the signal reduces to buying when the drift is high enough and selling when the
drift is low. Since the drift can be interpreted as the slope of a 60-day moving
average, we have therefore a built-in momentum strategy in this second sig-
nal. A calibration exercise using the training period 2000-2004 showed that the
cutos dened in the previous strategy are also acceptable for this one.
20
We observed empirically, through the calculation of signals, that the drift
has values of the order of 15 basis points, the average expected reversion
time is 7 days, and the equilibrium volatility of residuals is of the order of
300 bps. The expected average shift for the modied s-score is of the order of
0.15 7/300 0.3. In backtesting simulations, the eect of incorporating a
drift in these time-scales of a few days is minor. Thus, for the sake of brevity,
we do not present back-testing results with the modied s-scores.
In essence, we claim that the residual process has no signicant drift, or at
least that trading based on the simple constant drift estimated from the model
does not improve trading performance over assuming that = 0. We could sat
that, in aggregate, stock returns have negligible momentum after controlling for
industry/size factors, on the trading scale of interest.
5 Back-testing results
The back-testing experiments consist in running the signals through historical
data, simulating trades in all the stocks in the universe according to the signals
in (16). Estimation of parameters (betas, residuals) and signal evaluations are
performed daily. Estimation of the parameters sometimes leads to values of

i
< 8.4. When
i
crosses this threshold, we reject the model and (i) do not
open trades or (ii) close open trades.
We assume that all trades are done at the closing price of that day. As
mentioned previously, we assume a round-trip transaction cost per trade of 10
basis points, to incorporate an estimate of price slippage and other costs as a
single friction coecient.
Let E
t
represent the portfolio equity at time t. The basic PNL equation for
the strategy has the following form:
E
t+t
= E
t
+E
t
r t +
N

i=1
Q
it
R
it

i=1
Q
it

r t
+
N

i=1
Q
it
D
it
/S
it

N

i=1
[Q
i (t+t)
Q
it
[ ,
Q
it
= E
t

t
,
where Q
it
represents the investment in stock i at time t, R
it
is the stock return
from corresponding to the period (t, t+t), r represents the interest rate (assum-
ing, for simplicity, no spread between lending and borrowing rates), t = 1/252,
D
it
is the dividend payable to holders of stock i over the period (t, t +t)(when
t=ex-dividend date), S
it
is the price of stock i at time t, and = 0.0005 is the
slippage term alluded to above. The last line in the equation states that the
money invested in stock i is proportional to the total equity in the portfolio. The
proportionality factor,
t
, is stock-independent and chosen so that the portfolio
has a desired level of leverage on average.
21
For example, if we expect to have 100 stocks long and 100 short and we wish
to have a 2+2 leverage, then
t
= 2/100.
9
Another way to think of
t
is as
the maximum fraction of the equity that can be invested in any given stock.
10
The choice of 2+2 leverage was made to target a volatility of approximately 10%
by backtesting in the period 2002-2004. Of course, the choice of leverage does
not aect the Sharpe-ratio and other choices would give comparable results after
standardization ( Khandani and Lo (2007) run a 1/2+1/2 contrarian strategy
over a dierent, broader, stock universe).
Given the discrete nature of the signals, the investment strategy that we
propose is bang-bang: there is no continuous trading. Instead, the full amount
is invested on the stock once the signal is active (buy-to-open, short-to-open)
and the position is unwound when the s-score indicates a closing signal. This
all-or-nothing strategy, which might seem inecient at rst glance, turns out
to outperform making continuous portfolio adjustments, probably due to model
mis-specication.
5.1 Synthetic ETFs as factors
The rst set of experiments were done using 15 synthetic capitalization-weighted
industry-sector indices as risk-factors (see Table 3). The reason for using syn-
thetic ETFs was that most sector ETFs were launched after 2000. In order
to be able to back-test strategies going back to 1996, when most ETFs did
not exist and to compare the strategies with PCA, we decided to construct
capitalization-weighted sector indices and to construct residuals based on these
indices.
A series of daily returns for a synthetic index is calculated for each sector
and recorded for the 60 days preceding the estimation date. We then perform a
regression of stock returns on the its sector index and extract the corresponding
residual series and trading signals.
To ensure market-neutrality, we added to the portfolio an S&P 500 index
hedge using the SPY, which was adjusted daily and kept the overall portfolio
beta-neutral. In other words, due to the fact that the synthetic ETFs are not
traded instruments, we trade the stocks according to the signal and buy or sell
the SP 500 Index ETF in the appropriate amount so as to be Beta-neutral.
Since we expect that, on average, stocks are correctly priced, we experi-
mented with adjusting the mean of the OU processes exogenously. We intro-
duced the adjusted means for the residuals
m
i
= m
i

1
N
N

j=1
m
j
, i = 1, 2, ..., N. (18)
9
2+2 leverage means 2 dollars long and 2 dollars short per dollar of equity in the portfolio.
In practice, t is adjusted only for new positions, so as not to incur transaction costs for stock
which are already held in the portfolio.
10
Other renements that can be made have to do with using dierent leverage according to
the companys market capitalization or choosing a sector-dependent leverage that is inversely
proportional to the average volatility of the sector.
22
This modication has the eect of removing model bias; it expresses that we
expect that on average the mean of the residual is zero. We obtained consis-
tently better results in back-testing when using m
i
instead of m
i
. Therefore, we
adopted this centering approach in all the simulations (synthetic ETFs, ETFs,
and PCA).
The results of back-testing with synthetic ETFs are shown in Figure 8, Figure
9 and Table 4.
Figure 8: Historical PNL for the strategy using synthetic ETFs as factors from
1996-2007. The strategy does not produce signicant returns after 2004.
5.2 Actual ETFs
Back-testing with actual ETFs was possible only from 2002 onward, due to the
fact that many ETFs did not exist before. We backtested the strategy going
back to 2002, using regression on the ETF assigned to each stock to generate
residuals. The results are displayed on Figure 9 and Table 5.
The simulations suggest that using actual ETFs improves performance con-
siderably in relation to synthetic ETFs. A possible explanation for this improve-
ment is that ETFs are traded instruments, whereas the synthetic ETFs are not,
thus providing better price information. Another possible explanation is hedg-
ing: in the case of actual ETFs we were able to neutralize the portfolio across
23
Table 4: Sharpe ratios for the strategy using synthetic ETFs as factors : 1996-
2007. The Sharpe Ratio is dened as (r)/, where , r, are the annualized
return, interest rate and standard deviation of the PNL. The best years are
1996-2002. Sector Sharpe ratios assume beta-neutrality with respect to the SP
500 index.
24
Figure 9: Historical PNL for the strategy using actual ETFs as factors, com-
pared with the one using synthetic ETFs : 2002-2007. Notice the strong out-
performance by the strategy which uses actual ETFs.
25
Table 5: Sharpe ratios for actual 15 ETFs as factors : 2002-2007. Industry
Sharpe ratios assume beta-neutrality with respect to the corresponding ETF.
We observe, for the purpose of comparison, that the average Sharpe ratio from
2003 to 2007 was 0.6. Sharpe Ratios above 1.0 where obtained in 2002 and 2004.
26
each industry using ETFs as hedges, whereas the back-testing with synthetic
ETFs is beta-neutral with respect to the SPY .
5.3 PCA with 15 eigenportfolios
The back-testing results for signals generated with 15 PCA factors are shown
in Figures 10 and 11 and Table 6. In the case of PCA factors, we trade the
signals as explained in the previous section. To ensure market neutrality, we
hedge daily with the SP 500 tracker as in the case of synthetic ETFS.
The rst set of results, displayed in Figure 10, shows the progression of the
equity in the two portfolios corresponding to synthetic ETFs and PCA with 15
eigenportfolios.
The second set of results, displayed in Figure 11, compare the performances
of PCA with 15 eigenportfolios with the strategy with actual ETFs. We observe
that the 15-PCA strategy out-performs the actual ETF strategy since 2002. It
is noteworthy that this is the case even if the actual ETF strategy is hedged
sector-by-sector but the 15-PCA strategy is hedged only at the level of SPY.
Figure 10: PNL corresponding 15 PCA factors, compared with synthetic ETFs
from 1997-2007. The reason for starting in 1997 is that we need one year of data
to compute the initial correlation matrix. The PCA strategy produces superior
results, particularly after 2002.
27
Figure 11: Comparison of strategies with 15 PCA factors and the using actual
ETFs in the period 2002-2007. 15-PCA outperforms signicantly the strategy
with actual ETFs.
28
Table 6: Sharpe ratios for 15 PCA factors : 1997-2007. We start in 1997 due to
the fact that we need 1 years worth of data to compute the initial correlation
matrix. The best performances, with Sharpe ratios above 2.0 were 200,2001,2002
and 2004. Sectors are beta- neutral with respect to the S&P 500 index.
29
5.4 Using a variable number of PCA factors
We also back-tested strategies based on a variable number of factors, with the
number of factors chosen so as to explain a given level of variance. In this
approach, we retain a certain number of eigen-portfolios (factors) such that the
sum of the corresponding eigenvectors is equal to a set percentage of the trace
of the correlation matrix.
The number of eigenvalues (or eigenvectors) which are needed to explain
55% of the total variance varies in time. This variability is displayed in Figure
12 and Figure 13. We also looked at other cutos and report similar results in
Figure 15. The periods over which the number of eigenvectors needed to explain
a given level of variance is low appear to be those when the risk-premium for
equities is relatively high. For instance, the latter parts of 2002 and 2007,
which correspond to the aftermath of the internet bubble and the bursting of
the subprime bubble, are periods for which the variance is concentrated on
a few top eigenvectors/eigenvalues. In contrast, 2004-2006 is a period where
the variance is distributed across a much larger set of modes. The equity risk
premium, as represented by the VIX, reaches a historical low in early 2007.
Figure 12: Number of signicant eigenvectors needed to explain the variance
of the correlation matrix at the 55% level, from 2002 to February 2008. The
estimation window for the correlation matrix is 252 days. The boundary of the
shaded region represents the VIX CBOT Volatility Index (measured in percent-
age points).
Back-testing the strategy with 55% explained variance shows that the strat-
30
Figure 13: Percentage of variance explained by the top 15 eigenvectors: 2002-
February 2008. Notice the increase in the Summer of 2007.
Figure 14: Comparison of the PNLs for the xed explained variance (55%) of
PCA and the 15 PCA strategy: 2002-2007. The performance of the 15 PCA
strategy is slightly superior.
31
Table 7: Sharpe ratios for the xed explained variance (55%) of PCA : 2003-
2007. Sectors are neutral with respect to the S&P 500 index.
32
egy is comparable but slightly inferior to taking 15 eigenvectors. The corre-
sponding results are presented in Figure 17 and Table 9. As before, we trade
daily according to the s-score levels and keep the portfolio beta-neutral with
respect to the S&P 500 index.
In the same vein, we studied the performances of other strategies with a
variable number of eigenportfolios explaining dierent levels of variance. In
Table 8 and Figure 16, we display the performances of strategies using 45%,
55% and 65% compared with the PCA strategies with 1 eigenportfolio and with
15 eigenportfolios.
The conclusion is that 55% PCA performs best among the three strategies
and underperforms slightly the 15 PCA strategy. We also observe that taking a
high cuto such as 75% of explained variance leads invariably to steady losses,
probably due to the fact that transaction costs dominate the small residual
variance which remains in the system after defactoring. Using too many factors
lead to noise trading!
On the opposite side of the spectrum, using only one eigenportfolio, as in
the Capital Asset Pricing Model, gives rise to lower speeds of mean-reversion,
higher residual volatilities and worst Sharpe ratios. (See Figures 25, 26 ).
Figure 15: Time-evolution of number of PCA factors for dierent levels of ex-
plained variance: 2002-2007. The decay in the number of factors is associated
with the onset of the subprime crisis in the summer of 2007.
33
Figure 16: PNL for dierent variance truncation level:2002-2007
Table 8: Sharpe ratios for variable PCA strategies: 2002- 2007
34
6 Taking trading volume into account
In this section, we incorporate volume information to the mean-reversion signals.
Let V
t
represent the cumulative share volume transacted until time t starting
from an arbitrary reference time t
0
(say, the date at which the stock was rst
issued). This is an increasing function which can be viewed as a sum of daily
trading volumes and approximated as an integral:
V
t
=

V
k

t

t0

V
s
ds.
Historical prices can be viewed on a uniform time grid or on a uniform volume
grid (i.e. the price evolution each time one share is traded). If we denote the
latter prices by P
V
, we have
S
t+t
S
t
= P
V (t+t)
P
V (t)
=
P
V (t+t)
P
V (t)
V (t + t) V (t)
(V (t + t) V (t)) . (19)
Thus, the price change per share traded over the period of interest is
P
V (t+t)
P
V (t)
V (t + t) V (t)
=
S
t+t
S
t
V (t + t) V (t)
.
This suggests that, instead of the classical daily returns, we use the modied
returns
R
t
=
S
t+t
S
t
S
t
'V `
V (t + t) V (t)
= R
t

'V `
V (t + t) V (t)

(20)
where 'V ` indicates the average, or typical, daily trading volume calculated
over a given trailing window. Measuring mean-reversion in trading time is
equivalent to rescaling stock returns as in (20).
The modied returns R
t
are equal to the classical returns if the daily trading
volume is typical. If the trading volume is low, the factor on the right-hand side
of the last equation is larger than unity and R
t
> R
t
. Conversely, if volume
is high then R
t
< R
t
. The concrete eect of the trading-time modication
is that mean-reversion strategies are sensitive to how much trading was done
immediately before the signal was triggered. If the stock rallies on high volume,
an open-to-short signal using classical returns may be triggered. However, if
the volume is suciently large, then the modied return is much smaller so the
residual will not necessarily indicate a shorting signal. Similarly, buying stocks
that drop on high volume is discouraged by the trading-time approach.
We backtested the ETF and PCA strategies using the trading-time approach.
The window for calculating the trading volume was taken to be ten trading days,
35
a choice dictated by the fact that the measured volume should be over a period
smaller than the estimation window for the residuals but long enough to average
spikes.
11
We found that the technique increases the PNL and the Sharpe ratios un-
equivocally for strategies with ETF-generated signals (Figure 17 and Table 9).
For PCA-based strategies, trading time framework does not seem to produce
a signicant improvement (Figure 18 and Table 10 ). Finally, we nd that
the ETF strategy using trading time is comparable in performance to the 15-
PCA/55% PCA strategies until 2006 and performs slightly better after that, if
we exclude the August 2007 drawdown (Figure 19 and Table 10 ).
Figure 17: Comparison of strategies corresponding to signals generated using
trading time vs. using calendar time, using actual ETFs as factors : 2002-2007.
7 A closer look at 2007
It is well-known that 2007 was a very challenging year for quantitative hedge
funds; see Khandani and Lo (2007), Barr (2007), the Associated Press (2007)
and Rusli (2007). After a mediocre performance in the rst half of the year, sta-
tistical arbitrage strategies experienced a large drawdown followed by a partial
11
This value was chosen as an educated rst guess. We did not optimize the trading-time
window length in this study.
36
Table 9: Sharpe ratios for signals in trading time using actual ETFs as factors :
2002-2007. Sector portfolios are beta-neutral with respect to the corresponding
ETF.
Figure 18: Comparison of signals in trading time vs. actual time using 15 PCAs
as factors : 2002-2007
37
Table 10: Sharpe ratios for signals in trading time using 15 PCAs as factors :
2002-2007
38
recovery in the second week of August 2007. Unfortunately for many man-
agers, the size of the drawdown was so large that many had to de-leverage their
portfolios and did not recover to pre-August levels (see the above references for
details). Our backtesting results are consistent with the real-world events of
2007 and also show a strong drawdown in August 2007 (see below). This draw-
down was rst reproduced in back-testing by Khandani and Lo(2007) using their
classical contrarian strategy.
We analyzed carefully the performance for all the strategies in this paper in
2007, with the exception of synthetic ETFs (which are known to underperform
the others). Figure 20 displays graphically the results.
First, we found that performance of all strategies was at or slightly negative
in the rst part of the year. In early August, we found that all the mean-
reversion strategies experienced a large, sudden drawdown followed by a recovery
in about 10 days. In certain cases, our strategies tracked almost identically the
Khandani-Lo (2007) simulation after adjusting for leverage (KL used 4+4 in
2007 leverage whereas we use 2+2 in this paper). PCA-based strategies showed
more resilience during the liquidity event, with a drawdown of 5% as opposed
to 10% for the ETF-based strategies (see Figure 20).
Khandani and Lo (2007) suggest that the events of 2007 could have been
due to a liquidity shock caused by funds unwinding their positions. As we have
seen, market-neutral statistical arbitrage strategies result in leveraged portfolios
with hundreds of long and short positions. While each position is small and has
probably small impact, the aggregate eect of exiting simultaneously hundreds
of positions may have produced the spike shown in Figure 21.
A closer look at the dierent sectors shows that the Technology and Con-
sumer Discretionary sectors were strongly aected by the shock and more so
in terms of price movements than Financials and Real Estate; see Figure 22,
which gives a breakdown of the performance of the dierent industry sectors
in August 2007. This apparently paradoxical result whereby sectors that are
uncorrelated with Financials experience large volatility is consistent with the
unwinding theory of Khandani and Lo. In fact, it shows a sudden divergence
of the mean-reversion strategy in sectors unrelated to nancials and real estate.
8 Conclusions
We presented a systematic approach for constructing market-neutral portfolio
strategies based on mean-reversion. The approach is based on decomposing
stock returns into systematic and idiosyncratic components. This extraction of
residuals is done using dierent denitions of risk-factors: (i) ETFs as proxies
for industry factors or (ii) a PCA-based approach where we extract factors from
eigenvectors of the empirical correlation matrix of returns.
We compared extensively the ETF and PCA methods. In the ETF method,
we used 15 liquid ETFs to representing the systematic price uctuations, and
we de-trended stock returns by regressing each of them on the associated ETF
39
Figure 19: Comparison of ETF and PCA strategies using trading time. The
ETF strategy in trading time clearly outperforms the PCA strategy.
40
Figure 20: Zoom on 2007. Performance of strategies with ETF factors with
trading time, ETF factor with calendar time and 15 PCA. All strategies exhibit
a similar drawdown in August 2007, like the contrarian strategy of Khandani
and Lo (2007).
41
Figure 21: Comparison of the ETF strategy in trading time with Khandani &
Lo during August 2007. The leverage is 2+2 for both strategies.
Figure 22: Breakdown of performance per sector in the ETF strategy, Aug 2007.
All sectors are beta-hedged with ETFs.
42
returns.
With this denition of systematic-idiosyncratic decomposition, the system-
atic component of stock returns explains between 40% and 60% of the variance.
This suggests, on the PCA side, that the number of factors needed to explain
stock returns should be equal to the number of eigenvalues needed to explain
approximately 50% of the variance of the empirical correlation matrix. In prac-
tice, we found that this number to vary across time, and lies somewhere between
10 and 30 factors. We also observe that this number varies inversely to the value
of the VIX Option volatility index, suggesting more factors are needed to ex-
plain stock returns when volatility is low, and less in times of crisis, or large
cross-sectional volatility.
A word on hedging. For the trading strategies based on actual ETFs, we
hedge the portfolio daily, on each sector, with ETFs. Hence, the resulting
portfolio is beta-neutral per sector. In the case of strategies generated using
synthetic ETFs or using the PCA factors, we did not enforce beta-neutrality
per sector. Instead we traded all the signals and hedged the overall portfolio
beta using SPY, the S&P 500 tracker. The reason for this is that we wanted
to evaluate the performance of the synthetic ETF and the PCA strategies be-
ginning in the mid 1990s, when there were few sector ETFs. ( Hedging with
ETFs was not possible then, with the exception of SPY, which was available
also during that period.)
12
We backtested the strategies using data from 1996 to 2007 for synthetic
ETFs and PCA. Backtesting with actual ETFs was only possible since 2002,
after the sector ETFs were created. We then compared all strategies over the
relevant time periods, and particularly after 2002, when ETFs were available.
The best performing strategies were the ETF strategy in trading time and the
15-PCA strategy. A PCA strategy with a variable number of factors explaining
55% of the variance also performed relatively well.
Estimation of signals in trading time is equivalent, for EOD trading, to
weighting returns inversely to the daily traded volume. This additional fea-
ture, which incorporates trading volume into our signals, appears to benet
particularly the ETF strategy and to make it competitive with PCA. Given its
simplicity, this means that the ETF-based in trading time strategy might merit
further attention and renements.
We also noted that, in general, the performance of mean-reversion strategies
appears to benet from market conditions in which the number of explanatory
factors is relatively small. That is, mean-reversion statistical arbitrage works
better when we can explain 50% of the variance with a relatively small number
of eigenvalues/eigenvectors. The reason for this is that if the true number of
factors is very large (> 25) then using 15 factors will not be enough to defactor
the returns, so residuals contain market information that the model is not
12
Of course, we could have hedged the PCA portfolios using sector ETFs after 2002, but
we preferred not to do this in order to keep ideas simple. Other more sophisticated risk-
management methods, such as trading with limits on each PCA factor exposure, would have
been possible. We leave these more sophisticated portfolio strategies as avenues for future
research.
43
able to detect. If, on the other hand, we use a large number of factors, the
corresponding residuals have small variance, and thus the opportunity of mak-
ing money, especially in the presence of transaction costs, is diminished. We
conjecture therefore that markets having many stocks (i.e. investment opportu-
nities), and yet are in conditions in which they are driven by a small number of
explanatory factors, may be the most propitious for this strategy. This conjec-
ture might be useful for performance evaluation of this class of strategies and
to determine when they may provide good investment opportunities, although
more studies, particularly after 2007, may be required to ascertain this.
Finally, we reproduced the results of Khandani and Lo (2007) and thus place
our strategies in the same broad universality class as the contrarian strategies
of their paper. Interestingly enough, an analysis of PNL at the sector level
shows that the spike of August 2007 was more pronounced in sectors such as
Technology and Consumer Discretionary than in Financials and Real Estate,
conrming the plausibility of the unwinding theory of Khandani and Lo.
Acknowledgement. We thank Raghu Sundaram, Jay Zhu, Stanley Zhang,
Sridevi Ramaswamy, Rich Tong, Sasha Stoikov, Sasha Stanton and many col-
laborators and students who contributed, directly or indirectly, to the ideas
presented here. We are also grateful to the referees for their comments and
suggestions. All errors and omissions are ours, of course.
9 Appendix: estimation of the residual process
We describe our approach for the estimating co-integration residuals as Ornstein-
Uhlenbeck processes and for the calculation of s-scores. We do not claim
that this is the most sophisticated or ecient method for estimating the price
processes, but simply one that can be readily used (and almost certainly im-
proved) by practitioners.
For simplicity, we describe the estimation of the OU parameters for the case
of ETF regressions, the case of PCA being similar. The rst step is to estimate
the regression
R
S
n
=
0
+R
I
n
+
n
, n = 1, 2, ..., 60.
relating stock returns to the corresponding ETF returns. Here we assume that
returns are chronologically ordered, and R
S
60
is the last observed return, based
on the variation of the closing prices from yesterday to today. Recalling the
model (10), we set
=
0
/t =
0
252.
Next, we dene auxiliary process
X
k
=
k

j=1

j
k = 1, 2, ..., 60,
44
which can viewed as a discrete version of X(t), the OU process that we are
estimating. Notice that the regression forces the residuals to have mean zero,
so we have
X
60
= 0.
The vanishing of X
60
is an artifact of the regression, due to the fact that the
betas and the residuals are estimated using the same sample.
13
The estimation of the OU parameters is done by solving the 1-lag regression
model
X
n+1
= a + bX
n
+
n+1
, n = 1, ..., 59.
According to (13), we have
a = m

1 e
t

b = e
t
Variance() =
2
1 e
2t
2
whence
= log(b) 252
m =
a
1 b
=

Variance() 2
1 b
2

eq
=

Variance()
1 b
2
(21)
Fast mean-reversion (compared to the 60-day estimation window) requires that
> 252/30, which corresponds to mean-reversion times of the order of 1.5
months at most. In this case, 0 < b < 0.9672 and the above formulas make
sense. If b is too close to 1, the mean-reversion time is too long and the model
is rejected for the stock under consideration.
Notice that the s-score, which is dened theoretically as
s =
X(t) m

eq
becomes, since X(t) = X
60
= 0,
s =
m

eq
=
a

1 b
2
(1 b)

Variance()
.
13
This does not have to be the case. For instance, we can use 90 days to estimate the
regression and 60 days to estimate the process.
45
The last caveat is that we found that centered means work better, so we set
m =
a
1 b

a
1 b

where brackets denote averaging over dierent stocks. The s-score is therefore,
s =
m

eq
=
a

1 b
2
(1 b)

Variance()
+

a
1 b

1 b
2
Variance()
(22)
References
Associated Press, Quant funds endure August turmoil. The Motley Fool, De-
cember 6, 2007.
Barr, A., Quant quake shakes hedge-fund giants Goldman, Renaissance, AQR
see losses, but also sense opportunity, Marketwatch, August 13, 2007.
Cont, R., Da Fonseca, J., Dynamics of implied volatility surfaces. Quantita-
tive Finance, 2002, Vol 2, No 1, 45-60.
Davis, G., Mallat, S. and Avellaneda, M., Adaptive greedy approximations.
Constructive Approximations, 1997, Vol. 13, No. 1, 57-98.
Jollie, I. T., Principal Components Analysis, Springer Series in Statistics,
Springer-Verlag, Heidelberg, 2002.
Khandani, A. E. and Lo, A. W., What happened to the quants in August
2007? SSRN, 2007.
Laloux, L., Cizeau, P., Potters, M. and Bouchaud, J. P., Random matrix the-
ory and nancial correlations. International Journal of Theoretical and Applied
Finance, 2000, Vol. 3, No. 3, 391-397.
Lehmann, B., Fads, martingales, and market eciency. Quarterly Journal of
Economics, 1990, Vol. 105, No.1, 1-28.
Litterman, R. and Scheinkman, J. A., Common factors aecting bond returns.
Journal of Fixed Income, June 1991, 54-61.
Lo, A. W. and MacKinlay, A. C., When are contrarian prots due to stock
market overreaction? The Review of Financial Studies, 1990, Vol. 3, No. 2,
175-205.
Plerou, V., Gopikrishnan, P., Rosenow, B., Amaral, L. N., Guhr, T. and Stanley,
46
H. E., Random matrix approach to cross correlations in nancial data. Phys.
Rev., 2002, E 65, 066126.
Pole, A., Statistical arbitrage: Algorithmic trading insights and techniques, Wi-
ley Finance, 2007.
Poterba, J. M. and Summers, L. H., Mean reversion in stock prices: evidence
and implications. Journal of Financial Economics, 1988, Vol. 22, 27-59.
Potters, M., Bouchaud, J. P. and Laloux, L., Financial application of random
matrix theory: old laces and new pieces. Acta Physica Polonica B, 2005, Vol.
36, No. 9, 2767.
Rusli, E. M., Goldman Sachs Alpha to Fail?, Forbes.com, August 9, 2007.
Scherer, K. P. and Avellaneda, M., All for one and one for all? A principal
component analysis of Latin American brady bond debt from 1994 to 2000. In-
ternational Journal of Theoretical and Applied Finance, 2002, Vol. 5, No. 1,
79-106.
47