See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/273959237

Pairs trading and selection methods: Is cointegration superior?

Article in Applied Economics · February 2015

DOI: 10.1080/00036846.2014.975417

CITATIONS READS

60 8,283

2 authors:

Nicolas Huck Komivi Afawubo

ICN Business School University of Lorraine
17 PUBLICATIONS 818 CITATIONS 13 PUBLICATIONS 146 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Komivi Afawubo on 04 October 2018.

The user has requested enhancement of the downloaded file.

 Applied Economics

ISSN: 0003-6846 (Print) 1466-4283 (Online) Journal homepage: http://www.tandfonline.com/loi/raec20

Pairs trading and selection methods: is

cointegration superior?

Nicolas Huck & Komivi Afawubo

To cite this article: Nicolas Huck & Komivi Afawubo (2015) Pairs trading and selection methods: is
cointegration superior?, Applied Economics, 47:6, 599-613, DOI: 10.1080/00036846.2014.975417

To link to this article: https://doi.org/10.1080/00036846.2014.975417

Published online: 12 Nov 2014.

Submit your article to this journal

Article views: 1004

View Crossmark data

Citing articles: 16 View citing articles

Full Terms & Conditions of access and use can be found at

http://www.tandfonline.com/action/journalInformation?journalCode=raec20
Applied Economics, 2015
Vol. 47, No. 6, 599–613, http://dx.doi.org/10.1080/00036846.2014.975417

Pairs trading and selection

methods: is cointegration superior?
Nicolas Hucka,* and Komivi Afawubob,c
a
ICN Business School – CEREFIGE, Nancy-Metz, France
b
University of Lorraine – CEREFIGE, Nancy, France
c
University of Franche-Comté, CRESE, Besançon, France

Pairs trading is a popular dollar-neutral trading strategy. This article, using

the components of the S&P 500 index, explores the performance of a pairs
trading system based on various pairs selection methods. Whereas large
empirical applications in the literature focus on the distance method, this
article also deals with well-known statistical and econometric techniques
such as stationarity and cointegration which make the trading system
much more demanding from a computational point of view. Trades are
initiated when stocks deviate from their equilibrium. Our results confirm,
after controlling for risk and transaction costs, that the distance method
generates insignificant excess returns. While a pairs selection following
the stationarity criterion leads to a weak performance, this article reveals
that cointegration provides a high, stable and robust return.

Keywords: pairs trading; trading rules; distance; cointegration;

stationarity
JEL Classification: G11

I. Introduction By going long on the relatively undervalued stock

and short on the relatively overvalued stock, a profit
In its most common form, pairs trading involves may be made by unwinding the position upon ‘con-
forming a portfolio of two related stocks whose rela- vergence’ of the spread. The success of pairs trading,
tive pricing departs from its ‘equilibrium’. It is linked especially statistical arbitrages, depends heavily on
to cointegration (Bossaerts, 1988; Bossaerts and the modelling and forecasting of the spread time
Green, 1989) and correlation in stock prices, mean series. The ability to anticipate the ‘direction’ of
reversion, overreaction (De Bondt and Thaler, 1985; this spread is a key point. Whilst the strategy appears
Lo and Mackinlay, 1990), contrarian strategies simple and has, in fact, been widely implemented by
(Jegadeesh and Titman, 1993) and also to the law of traders and hedge funds, owing to the proprietary
one price. Pairs trading is one way to select stocks nature of the area there has been a limited amount
and to build a long/short dollar-neutral portfolio. In of published research until a recent burst of interest
reality, even such strategies require some outlay, if in the last few years. This includes several books on
only to meet margin calls and brokerage fees. the subject (Vidyamurthy, 2004; Whistler, 2004;

*Corresponding author. E-mail: [email protected]

© 2014 Taylor & Francis 599

600 N. Huck and K. Afawubo
Ehrman, 2006). These strategies exhibit, at least position is open. As underlined by the authors, both
according to the first studies including an application the 12- and 6-month periods are chosen arbitrarily.
over several decades, positive and significant ris- The distance approach merely exploits the statistical
k-adjusted returns which are not due exclusively to relationship of a pair at a price level. As the approach
a short term reversion phenomenon. Pairs trading is normative and economic free, it has the advantage
questions market efficiency. of not being exposed to model mis-specification and
The academic literature can be divided into three mis-estimation. On the other hand, this strategy lacks
main categories according to the methodology con- forecasting ability: if a ‘divergence’ is observed, the
sidered to select and trade pairs: assumption is that prices should converge in the
future because of the law of one price. When equili-
● The minimum distance approach, brium is reached or at the end of the 6-month trading
● Combined forecasts and Multi-Criteria period, the positions are closed out.
Decision Methods (MCDM), The second group of papers (Huck, 2009, 2010)
● The modelling of mean reversion (stationarity, combines neural networks for the purpose of fore-
cointegration, etc.) casting and multi-criteria decision aids (Outranking
methods, Electre III (Figueira et al., 2005)) for the
The most popular reference in the first category is selection of pairs. This approach is rather different
Gatev et al. (1999, 2006). It also includes Papadakis from the others because it is developed without refer-
and Wisocky (2008) and Engelberg et al. (2009), Do ence to any equilibrium model. At the moment, it has
and Faff (2010, 2012), Jacobs and Weber (2011), not been built for large indexes. This pairs trading
Mori and Ziobrowski (2011), Broussard and system will not be discussed in this work.
Vaihekoski (2012) and Huck (2013). These papers Several methodologies, coming from an econo-
explore different dimensions and implications of metric standpoint, have been used to model the
pairs trading strategies: accounting information, expected mean reversion phenomenon:
news, liquidity, sensitivity, transaction costs, etc.
Gatev et al. (2006), like many traders, envision a ● The cointegration concept (Engle and Granger,
simple algorithm for choosing pairs. The rule follows 1987; Johansen, 1988) can be an attempt to
the general outline of first ‘find stocks that move parameterize pairs trading. Generally speak-
together’ then ‘take a long short position when they ing, the framework is as follows: first, choose
diverge’ using a distance measure. They include a two cointegrated stock price series, then open a
trading system and the management of a portfolio long/short position when stocks deviate from
and consider a very large number of stocks using the their long-term equilibrium and finally, close
Center for Research in Security Prices (CRSP) data- the position after convergence or at the end of
base (about 2300 securities). This paper shows that the trading period. Mathematical aspects of
pairs trading after costs can be profitable. As a pos- pairs trading and cointegration are discussed
sible explanation, they indicate that pairs traders in Chiu and Wong (2011).
could be the disciplined investors taking advantage ● Testing stationarity of the price ratio between
of the undisciplined over-reaction displayed by indi- the two stocks is also an alternative (Baronyan
vidual investors. This point is in line with Jegadeesh et al., 2010).
and Titman (1995). The first version has been known ● A stochastic approach is used by Elliott et al.
for about 10 years and Do and Faff (2010), replicat- (2005) and Do et al. (2006). In a continuous
ing the Gatev et al. (2006) methodology with more setting, the first of these articles models the
recent data, report that the results of this strategy are difference between the two stock prices using
declining. a mean reverting Gaussian Markov chain
Pairs trading requires the selection and the trading model. The second studies the behaviour of
steps to be parameterized in some way. Gatev et al. the series at the return level.
(2006) use a simple SD strategy to select and trade
stocks. With daily data, they form pairs over a 12- Most of the time, this third group of works suffers
month period and trade them over the next 6 months. from the financial and practical point of view, and
Among the candidates chosen during the first stage, if from the fact that the real data application, if there is
prices diverge by more than two SDs, a long/short one, considers only a very limited number of stocks.
Pairs trading and selection methods 601
The articles based on a distance criterion or on com- may be highly time consuming. This is especially true
bined forecasts and MCDM propose fairly developed using techniques like stationarity or cointegration.
trading systems and the management of a portfolio Following Papadakis and Wisocky (2008), in order
over a period of many years. On the other hand, these to speed up computations, pairs selection is only per-
mathematical considerations can provide analytical formed using pairs that are likely to have high
results about the supposed speed of convergence of a comovement. If the return of two stocks, between
given series, the first time passage or the optimal the beginning and the end of the selection/
threshold for opening and closing positions. formation period, differs by more than 10%, the pair
The contribution of this article is empirical and is is automatically discarded from the whole process.
based on the S&P 500 index components. It aims at About 80% of the pairs will be affected in the
filling a lack of large and complete applications deal- applications.
ing with pairs trading performance and well-known
econometric methods such as stationarity and coin-
tegration. An article like Baronyan et al. (2010) per- The minimum distance method
forms a comparison between some methods using a
The distance method is now widely used in the
small index: the Dow Jones Industrial Average (30
academic literature since Gatev et al. (2006).
stocks). This leads to a piece of work which is quite
Details can also be found in Do and Faff (2010) or
demanding from a computational point of view com-
in Engelberg et al. (2009). The framework can be
pared to the distance method which is much faster.
divided into two stages (formation period and trad-
The empirical results confirm the weak perfor-
ing period).
mance of the distance approach in the last years
Initially, for each pair, we define a measure of
with US data. A pairs selection method based on
closeness. For each stock, we form pairs by finding
the stationarity of the price ratio does not provide
the partner that minimizes the sum of squared differ-
positive and significant returns after transaction
ences (SSD) in the normalized daily prices (inclusive
costs. Cointegration generates very high and stable
of dividends). Both prices are scaled to start at $1.
returns which are robust to transaction costs, risk
Among these pairs, for example, following Gatev
factors and data-snooping.
et al. (2006), the top 20 pairs with the lowest SSD
The rest of this article is organized as follows.
become candidates to be traded.
Section II reviews the different pairs selection
approaches. The data and the design of the applica-
tion are presented in Section III. Empirical results X
T  2
SSDi;j ¼ Pti  Ptj (1)
based on distance, stationarity and cointegration are t¼1
provided in Section IV. They cover dimensions like
transaction costs, sensitivity, stability, sector compo-
sition, risk exposure and data-snooping bias. Section with Pti and Ptj the normalized prices for stock i and
V concludes the article. stock j on day t, and T the number of trading days in
the formation period.
Prices are once again scaled to $1 at the begin-
ning of the trading period. Then, during the trading
II. Pairs Selection period, a long/short position (one dollar short in the
higher priced stock and one dollar long in the lower
This section discusses three selection methods that priced stock) is initiated in a pair whenever its
are suitable for pairs trading: normalized price difference, or spread, diverges by
more than a trigger, which is generally a multiple of
● the distance method, the SD from the historical spread computed over the
● the stationarity of the price ratio and formation period. Positions are unwound after con-
● the cointegration between stock prices. vergence or automatically at the end of the trading
period. This two-step sequence (formation/trading)
Exploring the full universe of pairs is, whatever the selection method, repeated every


( N ðN2 1Þ
¼ 124750 with S&P 500 index components, month. The formation period has a fixed length:
N being the number of stocks under consideration) one or two years for example in this article. The
602 N. Huck and K. Afawubo
data of the formation period are mobile windows historical reversal in the price spread used in Do and
with a 21-day lag. Faff (2010) working with the distance criterion. This
additional metric improves the performance of their
Stationarity and ADF test pairs distance trading system.
In order to generate profits in a pair-trade, the price Cointegration incorporates the idea of mean rever-
ratio between the two stocks needs to have a constant sion between stock prices. If two stocks are cointe-
mean and a constant volatility over time. A deviation grated, it means they share a long-term equilibrium
of the price ratio from this equilibrium state can thus relationship. Pairs trading, whatever the selection
be interpreted as a trading opportunity. In this article, method, will try to exploit deviations from an equili-
the selection method is based on the unit root brium asset-pricing framework with nonstationary
test developed by Dickey and Fuller (1979): this common factors (Jagannathan and Viswanathan,
econometric method is simple and popular. 1988; Bossaerts and Green, 1989; Chen and Knez,
Alternative tests to the Augmented Dickey–Fuller 1995). The application of the cointegration concept
(ADF) approach include for example Kwiatkowski to stock price analysis is that a system of nonstation-
et al. (1992). The ADF test, without trend, for a unit ary stock prices in level form can share common
root assesses the null hypothesis of a unit root using stochastic trends (Stock and Watson, 1988). As
the model: developed in Gatev et al. (2006), if the long and
short components fluctuate with common nonstation-
ary factors, then the prices of the component portfo-
yt ¼ c þ ϕ yt1 þ β1 Δyt1 þ   þ βp Δytp lios would be cointegrated and the pairs trading
þ εt (2) strategy would be expected to work.
The most familiar cointegration test has been
developed by Engle and Granger (1987). This is a
where Δ is the differencing operator, such that
two-step approach. Consider P1; t and P2; t are the
Δyt ¼ yt  yt1 . The number of lagged terms, p, is
prices of stocks 1 and 2 at time t and are Ið1Þ pro-
determined empirically (up to 10 lags) so that the
cesses (in order to avoid spurious regression). The
mean zero error term εt in the tested equation is
first step requires the regression of P1;t against P2;t
serially uncorrelated. The null hypothesis of a unit
root is:
P1; t  βP2; t ¼ μ þ εt (4)
Ho : ϕ ¼ 1 (3)

where μ denotes an intercept. Potential cointegration

under the alternative hypothesis, ϕ < 1. The exis- between the two stocks is examined via the analysis
tence of a unit root in the price ratio indicates pairs of the order of integration of the residuals εt using
trading conditions are met. In the stationarity and Dickey and Fuller (1979) test. Stocks are cointe-
cointegration (see below) based approaches, the grated if the residuals of the regression are stationary.
tests of the selected pairs are always significant at Instead of using the Engle and Granger (1987)
a 1% rate (or less). After the selection of a pair, if methodology, the Johansen (1988) approach is con-
the price ratio diverges, from its historical mean yt , sidered in this article. It takes the form of a likelihood
by more than a predefined threshold, a long/short ratio test and avoids the asymmetry problem in
position is open. The opening threshold is based treating variables. This approach tests the hypothesis
on the SD of the error term. Each month, the of r unrestricted cointegrating relationships in the
eligible pairs for the next 6 months are the pairs unrestricted Vector Autoregressive (VAR) model.
with the lowest ADF t-statistics. The null hypothesis of both the trace and maximum
tests is that there is no cointegration and the alter-
Cointegration native is that there is cointegration. The critical
The concept of cointegration has been used in the values may be found in Johansen (1996). As a pre-
pairs trading context by Vidyamurthy (2004), Lin liminary step before the Johansen (1988) test, a like-
et al. (2006), Bogomolov (2010) and Galenko et al. lihood ratio test (longer lags versus shorter lag
(2012). It shares connection with the frequency of lengths) is used to determine the optimal lag length
Pairs trading and selection methods 603
(up to 10 lags): this is a common approach in speci- method (distance, stationarity, cointegration), four
fying VAR models. parameterizations will be performed.
Cointegrated pairs with the highest trace statistics Evaluating the stability of trading rules matters
will be kept as eligible pairs for the trading step. (Falbo and Pelizzari, 2011). As underlined in Huck
During this last step, a deviation of the relation P1; t  (2013), the results of pairs trading are known to be
βP2; t from its historical mean, μ, will be interpreted highly sensitive to these key parameters. The aim of
as a trading opportunity. this article is of course not to find an ‘optimal’
parameterization of pairs trading strategies. The
objective is to establish whether or not, from a
III. Data and Design quite general point of view, one of the three selec-
tion methods (distance, stationarity, cointegration)
This section presents in details the pairs trading sys- may be considered as generating a significant and
tems based on the selection methods introduced in positive excess return. Testing multiple strategies/
the previous section. trading rules on the same data set leads to the well-
known problem of data-snooping. This issue will be
Basics controlled using Hansen (2005) test for Superior
The data used in this application are the prices of the Predictive Ability.
S&P 500 stocks1 (inclusive of dividends). These
stocks are among the most liquid in the world. As a Return computation, transaction costs and
consequence, transaction costs will be relatively low. benchmarks
In a pairs trading context, they will be estimated Since there are potentially several openings and clos-
using Do and Faff (2012). A selection procedure ings during the 6-month eligibility period for a given
starts every 21 trading days (about 1 month) and the pair, portfolio return computation is not a trivial
trading/eligibility period of a pair lasts 126 trading issue. Following Gatev et al. (2006) and others,
days (about 6 months, 6 × 21 days). The length of the excess returns are computed in order to evaluate the
trading period is chosen so that the selection process performance of the strategies. Two measures of
is recent and round-trips have time to occur using a excess return have been proposed: the return on
reasonable opening trigger. If the eligibility/trading committed capital and the fully invested return. The
period of each pair lasts for 6 months, the ‘entire flexibility of hedge funds’ funding leads to choose
portfolio’ is the sum of six overlapping ‘sub-portfo- the second one because it seems more realistic.
lios’ staggered by 1 month. The trading results pre- Opening and closing could happen on the same day
sented in the next section cover the period from as the trigger is reached: a popular alternative is a 1-
August 2000 to September 2011 (134 months). In day delay. The way returns are computed in this
fact, computation/trading starts and ends 5 months article differs slightly from the rest of the literature:
before and after so that the results always refer to a
period with a ‘complete portfolio’. ● The excess return of the portfolio is first com-
If the selection of parameters like the length of the puted on a daily basis as the mean excess return
formation and trading periods or the opening and among all pairs (equal weighted portfolio, at
closing triggers is done in a coherent way according least ten) opened a given day in the entire port-
to the trading system, it remains an arbitrary choice. folio (which could be considered as the sum of
This study considers two different lengths for the for- six portfolios that start 1 month apart). That
mation period (1 year (252 trading days) or 2 years way, the day-by-day composition of the portfo-
(504 trading days)) and 2 or 3 SDs for the opening lio is easily known (number of pairs, sectors).
trigger. Needless to say the greater the trigger, the ● Diversification is a crucial matter in portfolio
lower the number of openings and trades during the and risk management. Some selection methods
trading period. If most authors like Gatev et al. (2006) and opening triggers (especially with 3 SD),
consider the 2-SD rule, a more selective scheme is also from time to time, could lead the whole portfo-
examined. As a consequence, for each selection lio to be composed of less than 10 activated
1
The sample really considers 500 stocks only at the end of the trading period due to initial public offerings/newcomers in
the index: Google in 2004 for example.
604 N. Huck and K. Afawubo
pairs. In that case, the ‘missing’ positions will ● Does pairs trading generate profits without
be filled by a long position in the market index. transaction costs?
This switch from pairs trading positions to long ● Are pairs trading profits robust to transaction
exposure to the market occurs during a weak costs?
number of days as mentioned in Tables 3 and 4. ● Does one selection method dominates the
● The daily return (excluding transaction costs) of others?
the portfolio at time t, RPortfolio;t is computed as: ● Are all methods (stationarity, cointegration) as
parameter sensitive as the distance method?
● Is pairs trading still profitable after accounting
Pt
Pairs
Ri; t for traditional risk factors?
RPortfolio; t ¼
i¼1
(5) ● Does pairs trading outperform the S&P 500
Pairst index/equity premium?
● Which conclusions can be drawn about the
activity sectors dimension (in-sample/ forma-
where Pairst is the number of open pairs on
tion, out-of-sample/ trading)?
day t and Ri;t the return of the price ratio of the
ith pair open on day t. If Pairst < 10; Ri;t ¼
Xt ; "i 2 ½Pairst þ 1; 10 where Xt is the excess
return of the market on day t. Formation period and sector composition
● The estimation of pairs trading transaction
Tables 1 and 2 focus on two different aspects of
costs in this article is based on Do and Faff
the pairs selection, of which there are 6 ways
(2012). They consider these costs have three
((distance, stationarity, cointegration) × (1- or 2-
components: commissions, market impact
year formation period)). Table 1 reports the per-
(average one-way cost (commission + market
impact) of 30 bps (10 + 20) in our sample) and centage of identical eligible pairs between strate-
gies. Whatever the combination, the proportion of
short-selling constraints (constant loan fee of
1% per annum payable over the life of each pairs which are chosen the same month as eligible
pairs by two different selection methods is weak.
trade). Using the average number of trades per
Proportions are always below 13% indicating
6-month trading period, monthly estimations
strategies are clearly based on very different
of transaction costs can be computed2 for
pools of eligible pairs. The highest percentages
each strategy. These quite conservative values
of common eligible pairs are obtained with:
are reported in Tables 3 and 4.
● distance 1-year and distance 2-year,
IV. Empirical Results ● stationarity 2-year and cointegration 2-year.

The pairs trading results/performances are exam- Table 2 provides information about the sector compo-
ined through different dimensions/questions: sition, based on the GICS taxonomy, of eligible/

Table 1. Percentage of identical eligible pairs between strategies

1 year 2 years
Distance Stationarity Cointegration Cointegration Stationarity
2 years Distance 12.4 0.8 0.5 3.1 9.1
Stationarity 1.8 1.9 0.6 12.8
Cointegration 0.6 1.0 0.9
1 year Cointegration 1.8 9.8
Stationarity 6.2

2
As an example, see Table 3, transaction costs for the 1-year formation period, distance method, 2 SDs are 0.38%.




0:38  ð0:30 4 122 1:5Þþ1 .
Pairs trading and selection methods 605
Table 2. Sector composition (%) of eligible pairs between strategies
1 year 2 years
S&P 500 GICS code Distance Stationarity Cointegration Distance Stationarity Cointegration
8.2 10 Energy 10.8 7.8 8.5 9.3 8.5 8.3
5.8 15 Materials 7.2 6.0 5.4 5.9 5.3 5.9
12.4 20 Industrials 11.8 12.5 12.7 11.8 12.6 13.0
16.4 25 Consumer 8.6 16.6 16.5 10.9 16.3 16.0
discretionary
8.2 30 Consumer staples 17.2 8.7 8.6 20.5 9.4 9.4
10.2 35 Health care 10.5 11.7 10.4 9.3 12.3 11.4
16.2 40 Financials 18.7 15.9 15.9 17.6 15.5 15.7
14.0 45 Information 8.9 13.0 13.8 8.5 11.1 11.9
technology
1.6 50 Telecommunication 1.1 1.5 1.7 1.2 1.2 1.1
services
7.0 55 Utilities 5.2 6.4 6.5 5.1 7.9 7.3

candidate pairs and of the S&P 500 index. Pairs selec- during the period we considered (August 2000
tion based on stationarity or cointegration leads to the to September 2011).
same distribution as the index (eligible or traded pairs). ● The significance of the individual monthly
This point has been evaluated via a χ 2 test in Tables 3 excess returns is tested using t-statistics based
and 4. Differences are observed when comparing the on Newey–West SEs with six lags and the
pairs selection based on distance against the composi- consistent p-values of the bootstrap approach
tion of the S&P 500 index. The three main points are developed by Hansen (2005). On the one side,
an over-weighting of the Consumer Staples sector with t-statistics, 10 of 12 parameterizations
(GICS 30) and an under-weighting of the Consumer have positive and significant excess returns at
Discretionary (GICS 25) and Information Technology 5% level. The two failures occur with a 1-year
(GICS 45) sectors. This result is not surprising. The formation period associated with distance or
distance method, during the selection process, favours stationarity and a three SEs opening trigger.
low volatile stocks. The sector distribution of traded On the other side, the Hansen (2005) metho-
pairs comes close to the repartition of the market index. dology reports all trading strategies have sig-
Pairs trading sometimes focus on pairs with nificant positive excess returns (ignoring
stocks within the same sector. With the S&P 500 transaction costs).
index, this proportion is about 11.5%. Whatever ● As mentioned in Tables 3 and 4, the monthly
the selection method, for eligible or activated pairs, estimations of transaction costs go from 0.19
the proportion observed in our empirical results bps for the strategy with the lowest average
does not differ from the real proportion computed number of trades (2-year formation period,
with the index. cointegration method, 3 SEs) to 0.38 bps
(1-year formation period, distance method, 2
SEs). Including transaction costs, robust stra-
Performance and portfolio composition tegies at 5% level are the following:
The main information provided in Tables 3 and 4 is as – Cointegration whatever the length of the
follows: formation period or the opening trigger.
Returns are especially high with monthly
● Whatever the selection method, the opening excess returns, including transaction costs,
trigger or the length of the formation period, greater than 1.38% (1.68–0.30) and going up
the trading strategies, without transaction to about 5% over a period of more than
costs, have generated positive excess returns 10 years. These are clearly the main and
606 N. Huck and K. Afawubo
Table 3. Results: 1-year formation period
Design of the strategy (1-year formation period)
Method Distance Stationarity Cointegration
Opening trigger (nb of σ) 2 3 2 3 2 3
Excess returns (in $, per 100$, per month)
Without transaction costs 0.33 0.27 0.48 0.36 2.08 5.86
σ 1.51 1.75 2.89 3.25 2.59 3.83
t-Statistics (Newey–West) 2.91 1.83 2.20 1.58 7.92 15.70
Consistent p-values 0.00 0.03 0.02 0.01 0.00 0.00
(Hansen)
Trend (p-value) −0.003 (0.44) −0.002 (0.68) 0.003 (0.66) −0.002 (0.79) −0.012 (0.04) −0.010 (0.25)
Median 0.28 0.23 0.44 0.68 2.03 5.49
Skewness 0.17 0.21 0.46 0.37 0.28 0.44
Kurtosis 4.80 4.29 6.16 6.91 4.55 3.42
Min −4.54 −4.73 −8.79 −10.95 −6.49 −4.23
Max 6.25 6.60 12.67 14.51 11.19 17.82
Monthly returns > 0 (%) 61.19 59.70 58.21 58.96 80.60 97.01
Daily returns > 0 (%) 52.03 51.42 51.53 50.82 57.54 65.73
Sharpe Ratio 0.22 0.16 0.17 0.11 0.80 1.53
Monthly serial correlation 0.09 0.11 0.07 0.05 0.39 0.67
Monthly transaction costs 0.38 0.28 0.40 0.30 0.33 0.20
(estimation)
Consistent p-values 1.00 0.47 0.31 0.26 0.00 0.00
(Hansen) with T.C
Break/Stability (August 2000–September 2008 versus October 2008–September 2011)
Mean (SD) August 2000 to 0.36 (1.5) 0.27 (1.7) 0.31 (2.9) 0.28 (3.3) 2.02 (2.6) 5.68 (3.8)
September 2008
Mean (SD) October 2008 to 0.29 (1.6) 0.33 (1.8) 1.09 (2.7) 0.75 (3.0) 2.34 (2.5) 6.62 (3.8)
September 2011
t-Test 0.22 −0.17 −1.38 −0.74 −0.60 −1.28
Chow test (F distribution) 0.61 1.17 3.00 1.72 8.52* 8.36*
*Significant at the 5% level
Break date (Bai and Perron, 2003) July 2008 July 2008
Trading statistics and portfolio composition (per pair, per 6-month period)
Non traded pairs (%) 4.93 20.75 5.07 14.93 3.81 43.21
Non convergent pairs (%) 45.49 52.95 41.68 53.40 60.00 44.66
Single round trip pairs (%) 33.28 21.38 34.18 26.16 28.28 11.87
Multiple openings pairs (%) 16.31 4.93 19.07 5.52 7.91 0.26
Profitable trades (%) 62.59 58.31 61.90 57.45 66.82 76.00
Non Convergent profitable 27.93 37.85 26.11 35.41 47.76 69.79
trades (%)
Number of openings (σ) 1.5 (0.9) 1.0 (0.7) 1.6 (1.0) 1.1 (0.7) 1.2 (0.6) 0.6 (0.5)
Average time open (%) 57.0 41.4 60.4 47.0 45.2 13.6
Average number of open 70.5 (12.0) 50.9 (14.1) 77.9 (11.9) 58.0 (14.3) 55.3 (9.3) 17.2 (6.6)
pairs per day (σ)
Days with less than 10 0.00 0.00 0.00 0.00 0.00 13.59
actives pairs (%)
Days with less than 5 actives 0.00 0.00 0.00 0.00 0.00 1.03
pairs (%)
Average time (days) of a 28.3 (23.4) 33.5 (23.5) 29.0 (25.2) 36.0 (26.0) 34.5 (24.0) 35.5 (20.1)
convergent trade (σ)
σ of price ratio returns 1.27 1.27 2.44 2.44 2.52 2.52
(formation period)
Correlation between returns 0.57 0.57 0.33 0.33 0.33 0.33
(formation period,
eligible pairs)
(continued )
Pairs trading and selection methods 607
Table 3. Continued
Design of the strategy (1-year formation period)
Method Distance Stationarity Cointegration
Opening trigger (nb of σ) 2 3 2 3 2 3
Sectors: concordance/contingency χ 2 test p-value
S&P 500 versus candidate 0.00 0.00 0.99 0.99 1.00 1.00
pairs
S&P 500 versus traded pairs 0.00 0.00 0.99 1.00 1.00 1.00
traded pairs versus candidate 0.21 0.00 0.94 0.92 0.97 0.90
pairs
Factor model (t-stat)
Intercept 0.31 (3.2) 0.25 (2.1) 0.60 (3.2) 0.49 (2.2) 2.19 (11.1) 6.17 (18.9)
Market 0.00 (−0.1) 0.02 (0.5) 0.09 (1.1) 0.11 (1.2) 0.02 (0.5) −0.12 (−1.4)
SMB 0.07 (1.6) 0.04 (0.7) −0.01 (−0.1) 0.00 (0.0) −0.04 (−0.5) −0.23 (−1.7)
HML −0.02 (−0.4) 0.01 (0.2) −0.18 (−2.1) −0.20 (−2.0) −0.21 (−2.3) −0.41 (−2.9)
MOM −0.12 (−5.1) −0.15 (−5.9) −0.24 (−5.7) −0.26 (−5.3) −0.21 (−5.9) −0.30 (−5.6)
REV 0.12 (2.1) 0.17 (2.5) 0.20 (2.0) 0.23 (1.7) 0.33 (4.5) 0.46 (3.8)
r2 0.32 0.36 0.39 0.38 0.34 0.26
Long and short positions
Mean monthly excess 0.55 (3.56) 0.55 (3.68) 0.42 (5.53) 0.31 (5.64) 1.27 (5.42) 2.90 (5.30)
returns: long (σ)
Mean monthly excess 0.23 (3.23) 0.28 (3.24) −0.06 (4.51) −0.05 (4.36) −0.81 (4.66) −2.96 (4.95)
returns: short (σ)
Intercept (Factor model): 0.44 (3.0) 0.44 (2.6) 0.40 (2.4) 0.37 (1.6) 1.18 (6.8) 2.96 (9.7)
long (t-stat)
Intercept (Factor model): 0.14 (1.2) 0.20 (1.4) −0.19 (−1.1) −0.13 (−0.8) −1.00 (−5.8) −3.21 (−11.0)
short (t-stat)

most impressive empirical finding of this Do and Faff, 2010), in the recent years, of the
article. Even if the frameworks are very dif- most documented parameterization (1-year for-
ferent, these results are in line with the posi- mation period with a 2 SD trigger) is confirmed
tive contribution of the frequency of in our study: including transaction costs, a
historical reversal in the price spread used small negative monthly excess return
to augment a distance approach in Do and (−0.05 = 0.33 − 0.38) is observed. This article
Faff (2010). In some way, cointegration gen- also shows that pairs trading returns are sensi-
eralizes this metric. tive, whatever the selection technique, to key
– Strategies with a 2-year formation period and parameters like the length of the formation
a high level for the opening trigger. period or the opening trigger.
– The use of several parameterizations (12) leads ● Figure 1 compares the cumulative excess
to take into account a data-snooping bias. In returns (including transaction costs) of the dif-
order to provide a proper adjustment, the ferent strategies against the equity premium.
Hansen (2005) methodology is performed. The indexes based on pairs trading are
With or without transaction costs, the main smoother than the one referring to the equity
conclusions are highly robust indicating that premium. Cointegration performed well what-
at least one trading strategy (cointegration ever the market conditions.
trading strategies in fact) is profitable. ● The parameterization with the best monthly
● These results are in line with the empirical excess returns (Cointegration, 1-year forma-
existing literature dealing with pairs trading tion period, 3 SEs) reports a 97% proportion
and the distance method. The weak and of months (134) with positive excess returns
declining performance (Gatev et al., 2006; (without transaction costs). If this number
Table 4. Results: 2-year formation period
608

Design of the strategy (2-year formation period)

Method Distance Stationarity Cointegration
Opening trigger (nb of σ) 2 3 2 3 2 3
Excess returns (in $, per 100$, per month)
Without transaction costs 0.44 0.47 0.58 0.64 1.68 3.77
σ 2.00 2.54 2.59 3.21 2.59 4.09
t-Statistics (Newey–West) 3.18 3.00 2.54 3.37 7.29 10.22
Consistent p-values (Hansen) 0.00 0.00 0.00 0.01 0.00 0.00
Trend (p-value) −0.002 (0.60) 0.001 (0.84) −0.003 (0.60) −0.002 (0.69) −0.003 (0.65) −0.005 (0.59)
Median 0.28 0.58 0.58 0.52 1.60 3.32
Skewness −0.14 −0.65 0.42 0.62 0.31 0.74
Kurtosis 3.21 5.23 4.95 4.87 3.06 5.83
Min −5.61 −10.37 −6.93 −7.78 −4.06 −8.88
Max 5.21 7.15 11.44 12.01 8.89 22.59
Monthly returns > 0 (%) 58.21 60.45 58.96 58.21 73.13 83.58
Daily returns > 0 (%) 51.46 51.17 51.89 51.14 56.19 59.00
Sharpe ratio 0.22 0.19 0.23 0.20 0.65 0.92
Monthly serial correlation 0.04 0.00 0.06 −0.02 0.29 0.41
Monthly transaction costs (estimation) 0.29 0.20 0.31 0.23 0.30 0.19
Consistent p-values (Hansen) with T.C 0.11 0.02 0.09 0.05 0.00 0.00
Break/Stability (August 2000–September 2008 versus October 2008–September 2011)
Mean (SD) August 2000 to September 2008 0.46 (2.0) 0.41 (2.6) 0.48 (2.5) 0.50 (3.1) 1.61 (2.6) 3.55 (4.2)
Mean (SD) October 2008 to September 2011 0.46 (1.9) 0.77 (2.4) 0.98 (2.7) 1.20 (3.5) 1.97 (2.4) 4.66 (3.8)
t-Test −0.01 −0.73 −0.98 −1.12 −0.72 −1.39
Chow test (F distribution) 0.30 0.39 2.79 2.91 3.17* 4.92*
*Significant at the 5% level
Break date (Bai and Perron, 2003) July 2008

Trading statistics and portfolio composition (per pair, per 6-month period)
Non traded pairs (%) 19.03 45.86 13.43 34.89 12.09 47.16
Non convergent pairs (%) 50.60 41.64 51.42 49.40 56.53 43.21
Single round trip pairs (%) 24.25 10.82 27.54 13.62 25.97 9.14
Multiple openings pairs (%) 6.12 1.68 7.61 2.09 5.41 0.49
Profitable trades (%) 61.04 58.79 58.89 59.14 64.00 71.11
Non Convergent profitable trades (%) 38.40 45.78 35.17 46.00 45.94 64.89
Number of openings (σ) 1.0 (0.8) 0.6 (0.6) 1.1 (0.7) 0.7 (0.6) 1.1 (0.6) 0.5 (0.5)
Average time open (%) 42.0 24.9 49.9 32.9 43.2 15.6
Average number of open pairs per day (σ) 51.3 (14.3) 30.4 (14.0) 63.7 (14.3) 40.4 (15.9) 53.2 (12.4) 19.8 (9.6)
N. Huck and K. Afawubo
Days with less than 10 actives pairs (%) 0.00 2.63 0.00 0.00 0.00 18.79
Days with less than 5 actives pairs (%) 0.00 0.00 0.00 0.00 0.00 2.56
Average time (days) of a convergent trade (σ) 32.2 (25.1) 36.4 (24.5) 33.7 (25.6) 38.1 (25.7) 36.2 (24.9) 37.5 (19.8)
σ of price ratio returns (formation period) 1.45 1.45 2.47 2.47 2.54 2.54
Correlation between returns (formation period, eligible pairs) 0.54 0.54 0.33 0.33 0.34 0.34
Pairs trading and selection methods

Sectors: concordance/contingency χ 2 test p-value

S&P 500 versus candidate pairs 0.00 0 00 0.62 0.62 0.89 0.89
S&P 500 versus traded pairs 0.00 0 00 0.84 0.91 0.92 0.80
traded pairs versus candidate pairs 0.00 0 00 0.63 0.57 0.96 0.66

Factor model (t-stat)

Intercept 0.48 (3.9) 0.53 (3.2) 0.67 (4.0) 0.75 (3.4) 1.84 (10.3) 4.08 (14.5)
Market −0.02 (−0.4) 0.02 (0.3) 0.06 (1.6) 0.00 (0.0) 0.02 (0.4) −0.01 (−0.1)
SMB −0.04 (−0.3) −0.01 (−0.1) −0.06 (−0.5) 0.00 (0.0) 0.00 (0.0) −0.14 (−1.2)
HML −0.06 (−0.9) −0.08 (−1.0) −0.08 (−1.3) −0.16 (−2.0) −0.33 (−2.9) −0.48 (−2.7)
MOM −0.17 (−6.6) −0.21 (−5.3) −0.20 (−6.7) −0.26 (−6.3) −0.18 (−3.9) −0.32 (−3.7)
REV 0.18 (2.0) 0.20 (2.0) 0.16 (2.3) 0.21 (2.3) 0.27 (3.2) 0.53 (4.0)
r2 0.26 0.28 0.27 0.28 0.33 0.32

Long and short positions

Mean monthly excess returns: long (σ) 0.55 (4.10) 0.65 (4.39) 0.65 (5.65) 0.69 (5.93) 0.98 (5.32) 1.88 (5.22)
Mean monthly excess returns: short (σ) 0.11 (3.63) 0.18 (3.68) 0.09 (4.80) 0.07 (5.07) −0.70 (4.57) −1.89 (4.47)
Intercept (Factor model): long (t-stat) 0.40 (2.3) 0.54 (2.5) 0.56 (3.4) 0.63 (3.1) 0.92 (5.8) 1.91 (9.1)
Intercept (Factor model): short (t-stat) −0.07 (−0.4) 0.02 (0.1) −0.10 (−0.6) −0.12 (−0.6) −0.92 (−6.3) −2.16 (−10.9)
609
610 N. Huck and K. Afawubo
(1) Pairs that never trade throughout the trading
10 Coint, 1 y
period.
(2) Pairs that do open but never converge in time
during the 6-month trading period. This part
5
Coint, 2 y represents the behaviour of about 50% of
activated pairs. All nonconvergent trades
2 can be divided into two groups: winners and
Stat, 1 y Dist, 2 y Stat, 2 y
losers. This sub-split is very informative.
1 (3) Pairs that have one round-trip (profitable by
Dist, 1 y
definition) trade. The higher the opening
Equity premium
trigger, the higher the ratio between the
0.5
proportion of pairs that do open but never
2001 2003 2005 2007 2009 2011
converge in time and the proportion of
Fig. 1. Cumulative excess returns (log scale) includ- pairs that have one round-trip trade.
ing transaction costs pairs trading strategies with a (4) Pairs that have multiple openings (at least 2),
2-SD opening trigger versus equity premium possibly a final nonconvergent trade. At most
19% of eligible pairs have two or more
openings.
appears very high, it has to be compared to the
85% obtained by Gatev et al. (2006) with their Even if all strategies have a proportion of profit-
top-20 strategy over a period of about 40 years able trades which is high, some have weak and
(474 months). insignificant excess returns. This point is explained
by a large number of unprofitable trades when
The S&P 500 index is known in the literature to focusing on nonconvergent trades (rates of success
have been subject to excess comovement (Shiller, can fall below 40%) and the absence of stop-loss:
1989; Vijh, 1994; Barberis et al., 2005). This the possibility of trades leading to very important
effect could question our results. Kasch and losses whereas profits are (more or less) limited3 to
Sarkar (2014) improve the methodology dealing the amplitude of the opening trigger. This confirms
with the analysis of this empirical finding. The that a large part of pairs trading profits comes from
period they consider covers the one studied in this the first days after trade initiation and that the
article. They show that the permanent changes in introduction of a stop loss mechanism or a limita-
market value and return comovement, previously tion of the open period of nonconvergent trades
attributed to S&P 500 index additions, reflect could be useful. Engelberg et al. (2009) indicate
well-established regularities in asset returns inde- that the profitability of the distance-based strategy
pendent of index membership. decreases exponentially over time. The most suc-
The proportion of profitable trades, a good cessful parameterizations (cointegration combined
directional forecasting ability, is a key point for with a restrictive opening trigger, 3 SD) are the
the success of a trading strategy. Whatever the para- only cases where the proportion of profitable non-
meterization/method, as we could imagine with convergent trades stays (well) above 50%: this
monthly positive excess returns, the proportions of approach reduces nonconvergent risk.
profitable trades (at least 57%), ignoring transaction Some comparisons between the two most interest-
costs, are significantly greater than 50% which is a ing selection methods, distance and cointegration,
benchmark to consider with dollar-neutral position. can be done:
For cointegration-based strategies, more than 64% of
the trades are profitable. Using a classification ● The cointegration approach chooses eligible
slightly different from Do and Faff (2010), eligible pairs with a greater level of volatility of the
pairs are split into four categories: price ratio (about 2.5 versus 1.3) and a weaker

3
The part of the initial deviation greater than the opening threshold plus the part exceeding the crossing with the
equilibrium after convergence.
Pairs trading and selection methods 611
correlation (about 0.33 versus 0.55) between cointegration-based strategies. The significance of
the returns of the components. these returns shows that the performance of pairs trad-
● Normalizing opening triggers in percentage, the ing with cointegration, even after accounting for data-
triggers are slightly lower for the distance method snooping bias, is robust to standard risk factors.
than for the cointegration approach. Even if coin- As a contrarian strategy, it is important to examine
tegration strategies have very high excess returns, the returns of the long and short components of pairs:
they still have convergence times greater than it provides a better understanding of the pairs trading
those of the distance method (about 30 days). profits. The monthly excess returns and the intercepts
These excess returns cannot be explained by the of the five-factor model are reported in Tables 3 and
selection/choice of pairs which are likely to con- 4. Dealing with cointegration-based strategies,
verge very quickly. The profits of cointegration results indicate that the abnormal returns of the port-
approaches come much more from a good ability folio are equally driven by the long and short posi-
to forecast direction than from the detection of tions. This symmetry shows, in our sample, that the
very short-term deviations. selection of pairs based on cointegration is a power-
ful tool to exploit mean reversion. Compared to
As motivated by Gatev et al. (2006) and Engelberg Gatev et al. (2006), this is a major difference. They
et al. (2009), a Fama and French (1993) three-factor observe, in a distance framework during a period of
regression4 augmented by two other factors is per- about forty years ending in 2002, that more than two
formed in order to analyse the risk exposure of pairs thirds of their abnormal returns come from the short
trading monthly returns. The two additional factors part of the portfolio.
control for the empirical findings of Jegadeesh and The period considered in this article involves 2008
Titman (1993), Jegadeesh (1990) and Lehmann financial crisis. Through different dimensions, the
(1990). The independent variables are standard factor stability of the performances around this event is
returns: examined. Between August 2000 and September
2011, for 11 of the 12 parameterizations, the returns
● the market excess return (Rm  Rf ) where Rf is do not exhibit a significant trend at 5% level. The
the 30-day Treasury bill return, exception concerns one strategy dealing with cointe-
● a size factor: the difference between small and gration (1-year formation period, 2 SDs opening
big stocks (SMB), trigger): a slight decline is observed. Splitting data
● a value factor: the difference between value around September and October 2008, the key obser-
and growth stocks (HML), vations are the following:
● a momentum factor: the difference between
portfolios of year-long winners minus year- ● Whatever the strategy, using a t-test, there is no
long losers (MOM), significant difference in returns between pre-
● a reversal factor: the difference between port- and post-crisis periods. A slight increase of the
folios of last month losers minus last month results can nevertheless be noticed for all coin-
winners (REV). tegration-based strategies.
● A Chow (1960) test indicates, at 5% level, the
In line with the literature, our pairs trading excess presence of a structural break around this crisis
returns are market neutral: the exposure to the market for 3 of the 4 cointegration-based methods.5
is insignificant. Furthermore, between the different stra- The linear regressions of the post-crisis data
tegies, the signs of this parameter vary. The exposures show an increase of the constants and signifi-
to momentum and reversals have the predicted signs cant negative trends. These structural breaks
and are statistically significant at 5% level in nearly all pointed out by the Chow test are in fact
cases. Risk adjusted alpha are statistically significantly explained by the very high performance of
positive for all strategies excluding transaction costs. the cointegration-based strategies during the
Including costs, alpha remains highly positive for 6-month period starting in October 2008.

4
Data and details on construction of these factors series can be found from Ken French’s website: http://mba.tuck.
dartmouth.edu/pages/faculty/ken.french/data_library.html
5
The fourth one, 2-year formation period, 2 SD, is significant at 10% level.
612 N. Huck and K. Afawubo
● In the Chow test, the date of the break is an input. Do and Faff (2010) regress pairs returns against
In order to be more general, the Bai and Perron market volatility over the same month but did not
(2003) test is also used. The breaks which have find any significant effect. A new direction could
been detected by this approach at 5% level are analyse the influence of conditioning openings to
reported in Tables 3 and 4. For each of the three a certain level of volatility/Vix.
strategies mentioned above, a single break in
July 2008 is found. With a 3-month difference,
this is in line with the Chow test and the hypoth-
esis (October 2008) that has been made. References
Bai, J. and Perron, P. (2003) Critical values for multiple
structural change tests, Econometrics Journal, 6,
72–8. doi:10.1111/1368-423X.00102
Barberis, N., Shleifer, A. and Wurgler, J. (2005)
Comovement, Journal of Financial Economics, 75,
V. Concluding Remarks 283–317. doi:10.1016/j.jfineco.2004.04.003
Baronyan, S., Boduroglu, I. and Sener, E. (2010)
This article extends the literature on pairs trading stra- Investigation of stochastic pairs trading strategies
tegies in providing new empirical results for real-world under different volatility regimes, Manchester
applications. The contribution of this article lies on a School, 78, 114–34.
rigorous evaluation of different methods to select and Bogomolov, T. (2010) Pairs Trading in the Land Down
Under, SSRN eLibrary. Available at http://dx.doi.
to trade pairs. While the most documented selection org/10.2139/ssrn.1717295
technique is based on a distance criteria, more Bossaerts, P. (1988) Common nonstationary components
advanced pairs selection techniques like stationarity of asset prices, Journal of Economic Dynamics &
and cointegration are considered here. Control, 12, 347–64. doi:10.1016/0165-1889(88)
Our results confirm the weak excess returns observed 90045-0
Bossaerts, P. and Green, R. (1989) A general equilibrium
in the recent publications for the minimum distance
model of changing risk premia: theory and tests,
method. The trades initiated following the stationarity Review of Financial Studies, 2, 467–93.
of the price ratio are not able to generate, after transac- Broussard, J. P. and Vaihekoski, M. (2012)
tion costs, large and significant excess returns. The key Profitability of pairs trading strategy in an illiquid
empirical fact revealed by this study concerns cointe- market with multiple share classes, Journal of
gration. After controlling for risk factors, transaction International Financial Markets, Institutions &
Money, 22, 1188–201. doi:10.1016/j.intfin.2012.
costs and data-snooping biases, cointegration-based 06.002
pairs trading exhibits high and robust positive alpha. Chen, Z. and Knez, P. (1995) Measurement of market
During a period of more than 10 years, even the least integration and arbitrage, Review of Financial
profitable parameterization dealing with cointegration Studies, 8, 287–325. doi:10.1093/rfs/8.2.287
delivers excess returns greater than 1.38% per month. Chiu, M. C. and Wong, H. Y. (2011) Mean–variance
portfolio selection of cointegrated assets, Journal of
Returns can rise up to 5% per month. Cointegration
Economic Dynamics and Control, 35, 1369–85.
reduces significantly nonconvergence risk. doi:10.1016/j.jedc.2011.04.003
Points to be explored on that topic stay numerous Chow, G. C. (1960) Tests of equality between sets of
and include: coefficients in two linear regressions,
Econometrica, 28, 591–605. doi:10.2307/1910133
● A comparison with international stock indices De Bondt, W. F. M. and Thaler, R. (1985) Does the stock
in other developed countries (e.g. EuroStoxx market overreact?, The Journal of Finance, 40,
793–805. doi:10.1111/j.1540-6261.1985.tb05004.x
600, Nikkei 225), which are less documented Dickey, D. A. and Fuller, W. A. (1979) Distribution of the
than the US market, to see whether the results estimators for autoregressive time series with a unit
of this article hold. root, Journal of the American Statistical Association,
● A deeper analysis of the sources of pairs trad- 74, 427–31.
ing profits especially with cointegration: Is the Do, B. and Faff, R. (2010) Does simple pairs trading still
birth of the divergence informative? What is work?, Financial Analysts Journal, 66, 83–95.
doi:10.2469/faj.v66.n4.1
the impact of market overreaction? Do, B. and Faff, R. (2012) Are pairs trading profits robust
● If pairs trading returns are not related to the equity to trading costs?, Journal of Financial Research, 35,
premium, does the volatility/Vix index matter? 261–87. doi:10.1111/j.1475-6803.2012.01317.x
 Pairs trading and selection methods 613
Do, B., Faff, R. and Hamza, K. (2006) A New Approach to Department of Finance, Northwestern University,
Modeling and Estimation for Pairs Trading, Monash Evanston, IL.
University, Caulfield. Jegadeesh, N. (1990) Evidence of predictable behavior of
Elliott, R. J., Van Der Hoek, J. and Malcolm, W. P. (2005) security returns, The Journal of Finance, 45, 881–98.
Pairs trading, Quantitative Finance, 5, 271–76. doi:10.1111/j.1540-6261.1990.tb05110.x
doi:10.1080/14697680500149370 Jegadeesh, N. and Titman, S. (1993) Returns to buying
Engelberg, J., Gao, P. and Jagannathan, R. (2009) An winners and selling losers: implications for stock
Anatomy of Pairs Trading: the Role of Idiosyncratic market efficiency, The Journal of Finance, 48,
News, Common Information and Liquidity, SSRN 65–91. doi:10.1111/j.1540-6261.1993.tb04702.x
eLibrary. Jegadeesh, N. and Titman, S. (1995) Overreaction, delayed
Engle, R. F. and Granger, C. W. J. (1987) Co-integration and reaction, and contrarian profits, Review of Financial
error correction: representation, estimation and testing, Studies, 8, 973–93. doi:10.1093/rfs/8.4.973
Econometrica, 55, 251–76. doi:10.2307/1913236 Johansen, S. (1988) Statistical analysis of cointegration
Ehrman, D. (2006) The Handbook of Pairs Trading: stra- vectors, Journal of Economic Dynamics & Control,
tegies Using Equities, Options and Futures, John 12, 231–54. doi:10.1016/0165-1889(88)90041-3
Wiley, Hoboken, NJ. Johansen, S. (1996) Likelihood-Based Inference in
Falbo, P. and Pelizzari, C. (2011) Stable classes of techni- Cointegrated Vector Autoregressive Models, Oxford
cal trading rules, Applied Economics, 43, 1769–85. University Press, Oxford.
doi:10.1080/09603100802676239 Kasch, M. and Sarkar, A. (2014) Is There an S&P 500
Fama, E. F. and French, K. R. (1993) Common risk factors Index Effect?, SSRN eLibrary. Available at http://dx.
in the returns on stocks and bonds, Journal of doi.org/10.2139/ssrn.2171235
Financial Economics, 33, 3–56. doi:10.1016/0304- Kwiatkowski, D., Phillips, P. C., Schmidt, P. et al. (1992)
405X(93)90023-5 Testing the null hypothesis of stationarity against the
Figueira, J., Greco, S. and Ehrgott, M. (2005) Multiple alternative of a unit root, Journal of Econometrics,
Criteria Decision Analysis: State of the Art Surveys, 54, 159–78. doi:10.1016/0304-4076(92)90104-Y
Springer Science Business Media, New York. Lehmann, B. N. (1990) Fads, martingales, and market
Galenko, A., Popova, E. and Popova, I. (2012) Trading in efficiency, The Quarterly Journal of Economics,
the presence of cointegration, Journal of Alternative 105, 1–28. doi:10.2307/2937816
Investments, 15, 85–97. Lin, Y.-X., Michael, M. and Chandra, G. (2006) Loss
Gatev, E., Goetzmann, W. N. and Rouwenhorst, K. G. protection in pairs trading through minimum profit
(1999) Pairs trading: performance of a relative-value bounds: a cointegration approach, Journal of Applied
arbitrage rule, NBER Working Paper 7032, National Mathematics and Decision Sciences, 2006, 1–14.
Bureau of Economic Research, Cambridge, MA. doi:10.1155/JAMDS/2006/73803
Gatev, E., Goetzmann, W. N. and Rouwenhorst, K. G. Lo, A. W. and Mackinlay, A. C. (1990) When are contra-
(2006) Pairs trading: performance of a relative- rian profits due to stock market overreaction?, Review
value arbitrage rule, Review of Financial Studies, of Financial Studies, 3, 175–205. doi:10.1093/rfs/
19, 797–827. doi:10.1093/rfs/hhj020 3.2.175
Hansen, P. R. (2005) A test for superior predictive ability, Mori, M. and Ziobrowski, A. J. (2011) Performance of
Journal of Business and Economic Statistics, 23, pairs trading strategy in the US REIT market, Real
365–80. doi:10.1198/073500105000000063 Estate Economics, 39, 409–28. doi:10.1111/j.1540-
Huck, N. (2009) Pairs selection and outranking: an appli- 6229.2010.00302.x
cation to the S&P 100 index, European Journal of Papadakis, G. and Wisocky, P. (2008) Pairs trading and
Operational Research, 196, 819–25. doi:10.1016/j. accounting information, Working Paper, MIT Sloan
ejor.2008.03.025 School of Management, Cambridge, MA.
Huck, N. (2010) Pairs trading and outranking: the multi- Shiller, R. J. (1989) Comovements in stock prices and
step-ahead forecasting case, European Journal of comovements in dividends, The Journal of Finance,
Operational Research, 207, 1702–16. doi:10.1016/ 44, 719–30. doi:10.1111/j.1540-6261.1989.tb04387.x
j.ejor.2010.06.043 Stock, J. H. and Watson, M. W. (1988) Variable
Huck, N. (2013) The high sensitivity of pairs trading trends in economic time series, Journal of
returns, Applied Economics Letters, 20, 1301–4. Economic Perspectives, 2, 147–74. doi:10.1257/
doi:10.1080/13504851.2013.802121 jep.2.3.147
Jacobs, H. and Weber, M. (2011) Losing Sight of the Trees Vidyamurthy, G. (2004) Pairs Trading, Quantitative
for the Forest? Pairs Trading and Attention Shifts, Methods and Analysis, John Wiley, Hoboken, NJ.
SSRN eLibrary. Available at http://papers.ssrn.com/ Vijh, A. M. (1994) S&P 500 trading strategies and stock
sol3/papers.cfm?abstract_id=2023539 (accessed 3 betas, Review of Financial Studies, 7, 215–51.
November 2014). doi:10.1093/rfs/7.1.215
Jagannathan, R. and Viswanathan, S. (1988) Linear Whistler, M. (2004) Trading Pairs – Capturing Profits
factor pricing, term structure of interest rates and and Hedging Risk with Statistical Arbitrage
the small firm anomaly, Working Paper 57, Strategies, John Wiley, Hoboken, NJ.

View publication stats