The profitability of pairs trading strategies: distance, cointegration, and

copula methodsI

Hossein Rada,, Rand Kwong Yew Lowa , Robert Faffa

a
UQ Business School, University of Queensland, Brisbane, 4072, Australia

Abstract

We examine and compare the performance of three different pairs trading strategies the distance
cointegration, and copula methods on the US equity market from 1962 to 2014 using a time-
varying series of trading costs. Using various performance measures, we conclude that cointegration
strategy performs as well as the distance method. However, the copula method shows relatively
poor performance. Particularly, the distance, cointegration, and copula methods show a mean
monthly excess return of 36, 33, and 5 bps after transaction costs and 88, 83, and 43 bps before
transaction costs. In recent years, the distance and cointegration methods have presented less
trading opportunities whereas this frequency remains stable for the copula method. While liquidity
factor is negatively correlated to all strategies returns, we find no evidence of their correlation to
market excess returns. All strategies show positive and significant alphas after accounting for
various risk-factors.
Keywords: pairs trading, copula, cointegration, quantitative strategies, statistical arbitrage
JEL classification: G11, G12, G14

I
We acknowledge the role of University of Queenslands Research Computing Centre and Barrine High Perfor-
mance Computing cluster in conducting this research.

Principal corresponding author
Email address: h.rad@business.uq.edu.au (Hossein Rad)
1. Introduction

Gatev et al. (2006) show that a simple pairs trading strategy (PTS), namely the Distance Method
(DM), generates profit over a long period. However, Do and Faff (2010) find that the profitability
of the strategy is declining. They associate this decrease to a reduction in arbitrage opportunities
during recent years as measured by the increase in the proportion of pairs that diverge but never
converge. Do and Faff (2012) show that the DM is largely unprofitable after 2002 when trading costs
are taken into account. Nonetheless, there are other mathematical concepts such as cointegration
and copulas that can be used to implement statistical arbitrage trading strategies. Although such
concepts are briefly introduced in the pairs trading literature, their performance has not been
robustly evaluated. Therefore, we evaluate the performance of two sophisticated PTSs, namely
copula and cointegration methods, using a long-term and comprehensive data set. We ascertain
if there is also a decline in pairs trading profitability from these more sophisticated methods and
determine the arbitrage risk-factors that may influence the profitability of PTSs.
PTSs comprise of two stages: first, the method applied to form pairs, and second, the criteria
for opening and closing positions. In the DM, securities whose prices are closely correlated are
grouped in pairs and traded when their prices diverge more than a specified amount. Nonetheless,
this method is the only strategy that has been tested thoroughly using extensive data sets, wide
variety of securities, and different financial markets (Gatev et al., 2006; Do and Faff, 2010, 2012;
Andrade et al., 2005; Perlin, 2009; Broussard and Vaihekoski, 2012). In the application of copulas
in pairs trading, Xie and Wu (2013) propose a PTS using copulas and Wu (2013) evaluates its
performance using three pre-selected pairs. Xie et al. (2014) explore 89 US stocks in the utility
industry over a sample period of less than a decade. Cointegration is another concept that can
be employed in a pairs trading framework (Vidyamurthy, 2004; Lin et al., 2006). Although Lin
et al. (2006) implement a cointegration PTS, their empirical analysis only examines two Australian
shares over a short sample period of one year. We examine the performance of a copula-based and
a cointegration-based PTS, using the CRSP data set from 1962 to 2014. Therefore, by using a
comprehensive data set spanning over 5 decades and containing all US stocks, our study is a robust
examination of different PTSs. By evaluating the performance of copula and cointegration based
trading strategies against the widely used DM, we establish with certainty whether these complex
methods yield better performance in the long-term.
Understanding the difference in performance outcomes between copula and cointegration PTSs,
versus the benchmark, the DM, will provide valuable insight into the source of pairs trading prof-
itability and the reasons behind the observed decline in the profitability of the DM. Has the market
become more efficient and the availability of arbitrage opportunities have diminished? Or are
methodologies more sophisticated than the simple DM required to exploit market inefficiencies?
For example, the simplicity of the DM may induce increased arbitrage activity, therefore resulting
in fewer arbitrage opportunities to exploit and a drop in profitability. The answer to these ques-

2
tions sheds light on the direction of future research by academics and practitioners in order to build
better performing strategies which will in turn further improve market conditions and efficiency.
Furthermore, we study how increasing the sophistication of methods by which pairs are selected
and traded can affect the quality and precision of the captured relationship within the pair and,
ultimately, the performance of PTSs. In theory, the presence of a cointegration relation between
two assets means that there is a long-term relationship between them. Exploiting this relationship
should allow us to accurately model the co-movement of the pair and use that to implement a
high-performance PTS. Using copulas for modeling the dependence structure between two assets,
instead of the elliptical dependence structure of covariance matrix, would also possibly lead to a
superior PTS by allowing for more flexibility in capturing asymmetries in the dependence structure
within pairs. Nevertheless, more complex models can also result in inferior performance, especially
out of sample, by introducing issues such as over-fitting. Moreover, the intensity of the comput-
ing power required to execute these complex algorithms may outweigh their relative performance
improvement over simpler strategies. This might result in loss of motivation in adapting such
strategies in practice.
We find that the cointegration method performs as well as the DM in most performance mea-
sures. Using lower partial moment and drawdown measures, we conclude that the cointegration
method performs better than other strategies before transaction costs are taken into account,
whereas after costs the DM is slightly superior. We find the copula methods performance not
to be as good as the other two methods. Nonetheless, the copula method is the only strategy in
which the frequency of trades have not fallen in recent years. The copula method has the highest
proportion of unconverged trades which contributes to its lower performance. We show that the
liquidity factor is negatively correlated to the strategies returns. Although, no such correlation
can be found regarding the market returns, which further demonstrates the market neutrality of
these strategies. Strategies alphas remain large and significant even after risk-factors are taken
into account.
The novel contributions of this paper to the relevant literature are twofold. First, we measure
the performance of two alternative PTSs, i.e. pairs trading using copulas and pairs trading using
cointegration, against a comprehensive data set consisting of all the shares in the US market from
1962 to 2014. Statistical arbitrage strategies more sophisticated than the DM, have rarely been
examined conclusively, and therefore their long-term performance remains unknown. Due to the
broad and long data sample used, this longitudinal study presents the first robust examination of
the performance of two relatively new PTSs, using cointegration and copula methods. Second, with
various performance metrics, we compare the three strategies and show if the increased complexity
in the pairs selection and trading criteria improve performance. In addition, with respect to studies
finding a recent decline in the performance of the DM, this comparison will lead to understand-
ing that if arbitrage opportunities are still available in the market, but perhaps due to increased

3
arbitrage activity, more complex methods are needed to take advantage of them.
The remainder of this paper is as follows. In section 2 we review some of the relevant literature
on pairs trading, copulas and cointegration. The data that we have used is discussed in section 3.
Section 4 covers the methodology of this research in detail. And finally, the results and conclusion
are presented in sections 5 and 6 respectively.

2. Literature review

Research on PTSs fall under the general finance literature on statistical arbitrage. Statistical
arbitrage refers to strategies that employ some statistical model or method to take advantage
of what appears to be mispricing between assets while maintaining a level of market neutrality.
Several statistical arbitrage strategies other than pairs trading have been developed. For instance,
Lo and MacKinlay (1990) propose a strategy that sells winner stocks and buys loser stocks in
an successful attempt to take advantage of the fact that some stock returns lead or lag those of
others. Khandani and Lo (2007) use the strategys performance in the early days of US sub-prime
mortgage crisis of 2007 to explain how unwinding positions triggered the initial loss experienced by
long-short hedge funds. Avellaneda and Lee (2010) implement two market-neutral mean-reverting
statistical arbitrage strategies using Principal Component Analysis (PCA) and regressing stock
returns on sector Exchange Traded Funds. In addition, studies have been done on statistical
arbitrage strategies in the commodities market. Specifically in commodities futures market, Erb
and Harvey (2006) and Miffre and Rallis (2007) use term structure information and momentum
signals to implement profitable trading strategies. In addition, Fuertes et al. (2010) propose trading
strategies on commodities futures market using term structure, momentum, and a combination of
the two (double-sort strategy). They show while all three strategies are profitable, the double-sort
strategy generates abnormally high returns.
Gatev et al. (2006) is one of the most comprehensive studies on pairs trading. They test the
most commonly used and simplest method of pairs trading, i.e. the DM, against the CRSP stocks
from 1962 to 2002. Their strategy yields a monthly excess return1 of 1.3%, before transaction costs,
for the DMs top 5 unrestricted2 pairs, and 1.4% for its top 20. In addition, after restricting the
formation of pairs to same-industry stocks, Gatev et al. (2006) report monthly excess returns of
1.1%, 0.6%, 0.8%, and 0.6% on top 20 pairs in utilities, transportation, financial, and industrial
sectors respectively. This study is an unbiased indication of the strategys performance as they
interpret the simplest method that practitioners employ as PTS. To avoid any criticisms that a
profitable trading rule is data-mined and applied in their study, the authors reevaluate the original
strategy after 4 years and show that it remains profitable.

1
Return on employed capital
2
Unrestricted pairs are the pairs that have been not been formed based on specific criteria such as belonging to
the same industry

4
 Do and Faff (2010) take the DM, introduced by Gatev et al. (2006), and closely examine the
strategy in order to investigate the source of its profits using CRSP data from 1962 to 2009. They
divide the data into three different periods, i.e. 1962-1988, 1989-2002, and 2003-2009 and report
strategys performance in each of these periods, in addition to the full sample, for the top 20 pairs.
They find that the strategy is profitable over a long period. Nevertheless, the DMs performance,
after experiencing its peak in 1970s and 1980s, begins to decline in the 1990s. Moreover, the findings
show that there are two exceptions to the plummeting performance of the strategy, that both occur
during bear markets. During these two periods, the DM shows solid performance, which is not
in line with the declining trend in the strategys profitability in recent years. By further studying
the properties of occurring trades, they attribute the decline in strategys profitability to, mainly,
worsening arbitrage risk, rather than to increasing market efficiency. They do so by dividing pairs
into 4 distinct groups and measuring the change in the proportion of trades that belong to each
group. Since the group that contains pairs that diverge but never converge (trades that open and
never close) represents the arbitrage risk, authors conclude that the observed increase in this groups
figure shows an increase in arbitrage risk. In other words, increased arbitrage risk is responsible
for 70 percent of the cut in pairs tradings profits, while improved market efficiency is responsible
for the remainder 30 percent. Moreover, they show that increased performance during the first
bear market, i.e. 2000-2002, is due to higher profits of pairs that complete more than one round-
trip trade rather than an increase in their number. By contrast, in the second bear market, i.e.
2007-2009, the increase in the number of these trades is the driver of strategys strong profitability.
Finally, the authors introduce two additional metrics to improve the pairs selection phase of the
DM: industry homogeneity and historical reversal frequency in pairs price spread. They point
out that the use of Fama and French (1997) 48-industry classification, as a proxy for industry
homogeneity, and number of zero crossings, as a measure for spread reversal frequency, results in
plausible improvements in profitability.
To improve the robustness of the previous study, Do and Faff (2012), using the same Gatev
et al. (2006) pairs trading framework, study the effects of trading costs on the profitability of
DM on US equity market from 1962 to 2009. They decompose direct trading costs into three
different components: commissions, market impact, and short selling costs and subtract them
from trade returns. Pointing out some drawbacks of the portfolio used in Gatev et al. (2006), they
construct and analyze a total of 28 alternative portfolios in addition to the baseline portfolio. These
alternative portfolios are constructed based on industry homogeneity and price reversal frequency,
introduced in Do and Faff (2010), along with two other portfolios that are the utility-only and
bank-only portfolios. The authors construct an estimate for commissions paid by all investors for
each year, based on the work of Jones (2002), data from Investment Technology Group (ITG),
and interpolation. The commission for one round-trip trade comes to an average of 34 bps over
the full sample period. Moreover, they conclude that investors bear an average extra cost of 26

5
bps for market impact. In summary, the calculated one-way transaction costs, which consist of
commissions and market impact for two subperiod and the full sample period is: 81 bps for 1963-
1988, 33 bps for 1989-2009, and 60 bps for the whole sample data, i.e. 1963-2009. Finally, a fixed
loan fee of 1% per annum for the life of each trade is also considered as short selling costs. After
taking the costs into account, Do and Faff (2012) conclude that DM on average is not profitable.
Despite that, for the duration of the sample period, the top 4 out of the 29 portfolios that they
constructed show moderate monthly profits of average 28 bps or 3.37% per annum. They also show
that for the period of 1989 to 2009 DM is profitable albeit the majority of its profit occurring in
the bear market of 2000-2002.
Several studies explore pairs trading using the DM in different international markets, sample
periods, and asset classes. (Andrade et al., 2005; Perlin, 2009; Broussard and Vaihekoski, 2012).
Elliott et al. (2005) provide an analytical framework for pairs trading by employing a Gaussian
Markov chain model for the pairs spread.
PTSs can be implemented using cointegration. Vidyamurthy (2004) discuss the theoretical
framework of pairs trading using cointegration based upon the error correction model representa-
tion of cointegrated series by Engle and Granger (1987). Caldeira and Moura (2013) implement
and test this framework on Sao Paolo exchange. They find that the strategy generates a 16.38%
excess return per annum with a sharpe ratio of 1.34 from 2005 to 2012. Lin et al. (2006) also
motivate the use of cointegration as a model that can capture the long-term equilibrium of the
price spread, while addressing the deficiencies of simpler statistical techniques used in pairs trading
such as correlation and regression analysis. By using cointegration coefficient weighting3 , Lin et al.
(2006) implement a theoretical framework that ensures some minimum nominal profit per trade
(MNPPT). They proceed to introduce a five step set of trading rules to use the framework in pairs
trading. Finally, their empirical analysis applies a small data set of 20 months sample period for
two Australian bank stocks and concludes that the MNPPT does not put excessive constraints
on trading if adapted along with commonly-used values for trading parameters such as open and
close trade triggers. In addition, Galenko et al. (2012) implements a PTS based on cointegration
and examines its performance with four exchange traded funds. Nevertheless, these studies share
a common shortcoming, where the empirical evidence provided to support the cointegration-based
PTSs is either non-present or gravely limited. For instance, the strategy proposed in Vidyamurthy
(2004) is not analyzed on real data and Caldeira and Moura (2013) use data from the Sao Paulo
stock exchange for a period of less than 7 years.
Copulas are used to model the asymmetric dependence of multivariate asset returns extensively
(Fortin and Kuzmics, 2002; Ang and Chen, 2002). With powerful properties including the free-
dom to choose marginal distributions and flexibility to model joint distributions, copulas have also

3
cointegration strategies mainly use cointegration coefficients to calculate the weights of the long and short
positions in order to maintain market neutrality of the combined long-short position

6
been frequently used in portfolio construction and asset allocation (Patton, 2004; Low et al., 2013).
However literature is limited on the application of copulas in quantitative trading strategies such
as PTS. Xie and Wu (2013), Wu (2013), and Xie et al. (2014) attempt to address this limitation.
Wu (2013) points out that the main drawbacks of distance and cointegration PTSs lie in the linear-
ity restriction and symmetry that correlation and cointegration enforce on the pairs dependence
structure. The application of copulas would be beneficial in relaxing these restrictions. By using
copulas to measure the relative undervaluation or overvaluation of one stock against the other, they
implement a PTS. Their study tests the DM, the cointegration method, and the proposed copula
method but only against 3 same-industry pairs. Also, the sample period is limited to 36 months.
The study does not enter the pairs selection phase of the strategy and only uses pre-nominated
pairs, which are selected from stocks with the same SIC code. Moreover, as regression analysis is
used to form the spread and the exact method of verifying the cointegration relationship is not
stated, spurious regression4 is also a possibility. The results show that the copula approach yields
higher returns than the other two approaches. The copula approach also presents more trading
opportunities than the distance and cointegration methods. Xie et al. (2014) employs a similar
methodology but use a broader data set to test it: the utility sotcks available (a total of 89 stock
after additional filtration) in CRSP from 2003 to 2012. Similarly, they show that the performance
of the copula strategy is superior to that of the DM used in Gatev et al. (2006). They also ob-
serve a fewer trades with negative returns for the copula strategy compared to the DM. Similar
to cointegration method, the main deficiency in these copula-based PTS studies, is the restricted
empirical evidence to robustly measure the performance of the strategies.

3. Data

Our data set consists of daily data of the stocks in CRSP5 from July 1st, 1962 to December 31st,
2014. The CRSP database includes stocks in all major US equity markets, i.e. NYSE, NYSE MKT,
and NASDAQ. The data set sample period is 13216 days (630 months) and includes a total of 23616
stocks. Note that, data is not available for all stocks throughout the entire period. In accordance
with Do and Faff (2010) and Do and Faff (2012), we restrict our sample to ordinary shares, which
are identified by share codes 10 and 11 in the CRSP database. In order to avoid relatively high
trading costs and complications, we have further restricted our sample to liquid stocks. This is
done by removing the bottom decile stocks, in terms of market cap, in each formation period. For
the same reason, stocks with prices less than $1 in the formation period are also not considered.
To increase the robustness of our results and to replicate practical trading environments as closely
as possible, we use trading volume to filter out stocks that have at least one day without trading

4
Spurious regression is the case when not-cointegrated I(1) series are regressed against each other, and based on
regression statistics a relationship is falsely verified. (Sims et al., 1990; Srensen, 2005)
5
Center of Research for Security Prices

7
in any formation period in the respective trading period. In summary, out data set is consistent
with that of Do and Faff (2012).

4. Research Method

Pairs trading is a mean-reverting or contrarian investment strategy. It assumes a certain price

relationship between two, and in some cases more than two, securities. Since pairs trading is a
long-short strategy, modeling this relationship would allow us to take advantage of any short-term
deviations from it by buying the undervalued and selling short the overvalued security simultane-
ously. Assuming this relationship holds, we would close, or reverse, the two opened positions once
the relationship is restored and pocket the profit. We examine the performance of three different
PTSs using CRSP database consisting of US stocks from 1962 to 2014.
First, in DM (detailed in section 4.1), potential security pairs are sorted based on the sum of
squared differences (SSD) in their normalized prices during the formation period. After the pairs
are formed, their spread is monitored throughout the trading period and any deviations beyond
a certain threshold in that spread would trigger the opening of two simultaneous long and short
positions. We use this strategy as our main benchmark and evaluate other strategies by comparing
their performance to that of this benchmark.
Second, we apply cointegration to pairs trading (detailed in section 4.2). By definition, coin-
tegrated time series maintain a long-term equilibrium and any deviation from this equilibrium is
caused by white noise and will be corrected as the series evolve through time (Vidyamurthy, 2004)..
Using the statistical model of cointegration, we are able to incorporate the mean-reverting attribute
of this statistical property in PTS. Our work is similar to the strategy outlined in (Vidyamurthy,
2004). After selecting nominated cointegrated pairs using the two-step Engle-Granger method (En-
gle and Granger, 1987), we extract their stationary spread. Any deviation from this spread is by
definition temporary and thus can be used to open long and short positions.
Third, copulas are used (detailed in section 4.3) to model the relationship between stocks of a
pair and to detect pairs deviations from their most probable relative pricing (Xie et al., 2014). We
define two mispriced indices, which represent the relative under or overvaluation of stocks of a pair,
and use them as criteria to open long-short positions when stocks move away from their relative
fair prices.
In contrast to other works, our study gathers comprehensive empirical evidence on evaluating
the performance of one well known and two relatively new PTSs, both independently and compared
with each other.

8
4.1. The distance method
In DM, we calculate the spread between the normalized prices of all possible combinations of stock
pairs during the formation period. The formation period is chosen to be 12 months6 . The nor-
malized price is defined as the cumulative return index, adjusted for dividends and other corporate
actions, and scaled to $1 at the beginning of the formation period. We then select 20 of those
combinations that have the least sum of squared spreads, or sum of squared differences (SSD), to
form the nominated pairs to trade in the following trading period, which is chosen to be 6 months.
The standard deviation of the spread during the formation period is also recorded to be used as
the trading criterion. A specific stock can participate in forming more than one pair as long as the
other stock of the pair varies. Thus, the same stock might be re-selected in different pairs.
At the beginning of the trading period, prices are once again rescaled to $1 and the spread is
recalculated and monitored. When the spread diverges by two or more from its historical standard
deviation (calculated in the formation period), we simultaneously open a long and a short position in
the pair depending on the direction of the divergence. The two positions are closed (reversed) once
the spread converges to zero again. The pair is then monitored for another potential divergence and
therefore can complete multiple round-trip trades during the trading period. We run the strategy
each month, without waiting 6 months for the current trading period to complete. As a result,
we have 6 overlapping portfolios, with each portfolio associated with a trading period that has
started in a different month. This implementation of the DM is in accordance with Gatev et al.
(2006) , Do and Faff (2010) and Do and Faff (2012).

4.2. The Cointegration Method

4.2.1 Framework
A non-stationary time series Xt is called I(1) if its first difference forms a stationary process, i.e.
I(0) (Lin et al., 2006). Now, consider X1,t and X2,t to be two I(1) time series. If we can find a linear
combination of the two time series that is stationary, X1,t and X2,t are said to be cointegrated. In
other words, X1,t and X2,t are cointegrated if there exists a non-zero real number so that:

X2,t X1,t = ut (1)

where is the cointegration coefficient and ut is a stationary series, i.e. I(0), called cointegration
errors. By using the Grangers theorem (Engle and Granger, 1987), the cointegration relationship
can be equivalently shown in a Error Correction Model framework (ECM) (Vidyamurthy, 2004).
Based on ECM, the cointegrated series have long-term equilibrium and, while short-term deviations
from this equilibrium occur, they will be corrected, through time, by the error term in the ECM.

6
month refers to calender month unless stated otherwise.

9
The ECM representation of the cointegration relationship between time series X1,t and X2,t is:

X2,t X2,t1 = X2 (X2,t1 X1,t1 ) + X2,t (2)

X1,t X1,t1 = X1 (X2,t1 X1,t1 ) + X1,t

Equation (2) shows that the evolution of a time series, for example X2,t , consists of a white noise,
X2,t , and an error correction term, X2 (X2,t1 X1,t1 ), which pulls the time series towards
its long-term equilibrium as the series evolves through time. This mean-reverting property of
cointegrated series can be used in implementing PTSs. We can further split the error correction
term into X2 , the rate of correction, and (X2,t1 X1,t1 ), which we recall from equation (1) to
be the cointegration relation. The cointegration relation shows the deviation of the process from
its long-term equilibrium. We define the spread series as the scaled difference in the price of two
stocks:
spreadt = X2,t X1,t (3)

We now assume that we buy one share of stock 2 and sell short share of stock 1 at time t 1.
X1,t and X2,t represent the price series of stocks 1 and 2 respectively. The profit of this trade at
time t, is given by:
(X2,t X2,t1 ) (X1,t X1,t1 ) (4)

By rearranging the above equation we get:

(X2,t X1,t ) (X2,t1 X1,t1 ) = spreadt spreadt1 (5)

Thus, the profit of buying one share of stock 2 and selling share of stock 1 for the period t is
given by the change in the spread for that period. As we also know from equation (1), the spread is
stationary by definition and has therefore mean-reverting properties. We can use this to construct a
quantitative trading strategy that uses deviations from the long-term equilibrium of a cointegrated
pair to open long and short positions. Positions are unwinded once the equilibrium is restored, as
a consequence of being stationary .

4.2.2 Trading Strategy

The trading strategy we implement is in line with the trading strategy discussed in Vidyamurthy
(2004).We use two criteria to implement the pairs selection phase of the cointegration PTSs, while
the first criterion is the same as the procedure used in the DM. We sort all possible combination
of pairs based on their SSD in normalized price during the formation period, which is set to 12
months to allow for consistency with the DM. Second, we test each of the pairs with the least SSD
for cointegration, using their cumulative return series in the formation period. We eliminate those
that are not cointegrated and estimate the cointegration coefficient of those that are. We continue

10
until the top 20 pairs with minimum SSDs, which are also cointegrated, are selected to be traded
in the following trading period.
Due to simplicity, we use the two-step Engle-Granger approach (Engle and Granger, 1987)
to test for the existence of cointegration between nominated pairs and to estimate cointegration
coefficient . In this procedure, the cointegration regression is estimated using OLS in the first step
and the Error Correction Model (ECM) is estimated in the second step. For each nominated pairs,
we then form the spread defined in equation (3) and calculate the spreads mean e and standard
deviation e , with the data of the formation period. These parameters are later used in the trading
period as trades open and close triggers. From equations (1) and (3) we can rewrite the spread
as: spreadt = et where et I(0), with mean e and standard deviation e . We now define the
normalized cointegration spread as:
spread e
(6)
e
During the trading period, we compute and monitor the normalized spread, using , e , and e
that we inferred in the previous step. Similar to what we did in the DM, we simultaneously open
and close long and short positions when the spread diverges beyond 2. However, the values of long
and short positions vary from those in the DM. By construction, if the spread drops below -2, we
buy 1 dollar worth of stock 2 and sell short dollar worth of stock 1. Equivalently, if the spread
moves above +2, we should sell short 1 dollar worth of stock 2 and buy dollar worth of stock 1.
Instead, as multiplying the trade proportions by the same number does not affect the cointegration
relation, we sell short 1/ dollar worth of stock 2 and buy 1 dollar worth of stock 1 when the
spread moves above the +2 threshold. By always buying $1, this transformation in position values
allows for consistency in calculating trade returns. We close both positions once the spread returns
to zero, which translates into the pair returning to their long-term equilibrium. The pair is again
monitored for other potential round-trip trades for the remainder of the trading period. Similar to
the implementation of the DM, we rerun the strategy each month and do not wait for the expiry
of the trading period and therefore have 6 overlapping portfolios each month.

4.3. Copula Method

4.3.1 Framework
A copula is a function that links marginal distribution functions to their joint distribution function.
It captures the dependence structure between the marginal distributions. A copula function is
defined as joint multivariate distribution function with uniform univariate marginal distributions:

C(u1 , u2 , , un ) = P (U1 u1 , U2 u2, , Un un ) (7)

Where ui [0, 1], i = 1, 2, , n. Now, suppose X1 , X2 , , Xn are n random variables with con-
tinuous distribution functions F1 (x1 ), F2 (x2 ), , Fn (xn ). Since a random variable with arbitrary

11
distribution can be transformed to a uniform random variable by feeding it into its distrubution
function, i.e. Ui = F (Xi ) where Ui U nif orm(0, 1), we can define the copula function of random
variables X1 , X2 , , Xn as:

F (x1 , x2 , , xn ) = C(F1 (x1 ), F2 (x2 ), , Fn (xn )) (8)

If Fi and C are differentiable 1 and n times respectively, we can write the joint probability density
function (pdf) f as the product of marginal density functions fi (xi ) and the copula density function
c:
f (x1 , x2 , , xn ) = f1 (x1 ) f2 (x2 ) fn (xn ) c(F1 (x1 ), F2 (x2 ), , Fn (xn )) (9)

where the copula density function c is given by diffrentiating the copula function, C, n times with
respect to each marginal:

n C(u1 , u2 , , un )
c(u1 , u2 , , un ) = (10)
u1 u2 un

Equation (9) allows us to decompose a multivariate distribution into two parts. One, the indi-
vidual marginal probability density functions and two, the copula density function. Consequently,
since all the characteristics of marginal distributions are captured in their pdfs and all the char-
acteristics of the joint distribution are represented by the joint pdf, the copula density function
should contain all the dependence characteristics of the marginal distributions.
Therefore, copulas allow for higher flexibility in modeling multivariate distributions. They allow
the marginal distributions to be modeled independently from each other, and no assumption on the
joint behavior of the marginals is required. Moreover, the choice of copula is also not dependent
on the marginal distributions. Thus, by using copulas, the linearity restriction that applies to
the dependence structure of multivariate random variables in a traditional dependence setting is
relaxed. Thus, depending on the chosen copulas, different dependence structures can be modeled
to allow for any asymmetries.
Now, let X1 and X2 be two random variables with probability functions F1 (x1 ) and F2 (x2 ) and
joint bivariate distribution function F (X1 , X2 ). We have U1 = F1 (X1 ) and U2 = F1 (X2 ) where
U1 , U2 U nif orm(0, 1) and their copula function C(u1 , u2 ) = P (U1 u1 , U2 u2 ). By definition,
the partial derivative of the copula function gives the conditional distribution function (Aas et al.,
2009):
C(u1 , u2 )
h1 (u1 |u2 ) = P (U1 u1 |U2 = u2 ) = (11)
u2
C(u1 , u2 )
h2 (u2 |u1 ) = P (U2 u2 |U1 = u1 ) =
u1
Using functions h1 and h2 , we can estimate the probability of outcomes where one random variable
is less than a certain value, given the other random variable has a specific value. This is practically

12
useful in a PTS where we can estimate the probability of one stock of the pair moving higher or
lower than its current price given the price of the other stock.

4.3.2 Trading Strategy

Similar to the DM, we first sort all possible pairs based on sum of squared differences (SSD) in their
normalized price during the formation period and nominate 20 pairs with the least SSDs to trade
in the trading period. Normalized price are calculated using the cumulative return index, inclusive
of dividends and adjusted for other corporate actions, and are scaled to $1 at the beginning of
trading period. The formation period is kept at 6 months to be consistent with the distance and
cointegration methods.
Then, we fit nominated pairs to copulas in 2 separate steps. (Hatherley and Alcock, 2007) First,
for each pair, we fit the daily returns of the formation period to marginal distributions and find the
two distributions that best fit the each stock. Marginal distributions are selected from Extreme
Value, Generalized Extreme Value, Logistic, and Normal distribution, and independently fitted for
each stock of the pair. In the second step, with the estimated parameters from the previous step,
we nominate the copula that best fits the uniform marginals and estimate its parameter. Copulas
that are tested in this step are Clayton, Rotated Clayton, Gumbel, and Rotated Gumbel. The best
fitting copula is the copula that describes the dependence structure between the pair returns better
than others. In quantifiable terms, the best copula is chosen by maximizing the log likelihood of
each copula density function and calculating the corresponding AIC and BIC. The copula associated
with the highest AIC and BIC is then selected as the best fitting copula.
In each day during the trading period, using the daily realizations of random variables U1
and U2 , which represent the daily returns of two stock of a pair, we calculate the conditional
probabilities, h1 and h2 functions defined in Equation (11), for each nominated pair. A value of
0.5 for h1 is interpreted as 50% chance for the random variable U1 , which is the price of stock 1, to
be below its current realization, which is its Todays price, given the current price of stock 2. The
same interpretation is valid for h2 , which demonstrates the same conditional probability for stock
2. Accordingly, conditional probability values above 0.5 show that chances for the stock price to fall
below its current realization is higher than they are for it to rise, while values below 0.5 predict an
increase in the stock price compared to its current value is more probable than a decrease. Similar
to Xie et al. (2014), we define two mispriced indices:

m1,t = h1 (u1 |u2 ) 0.5 = P (U1 u1 |U2 = u2 ) 0.5 (12)

m2,t = h2 (u2 |u1 ) 0.5 = P (U2 u2 |U1 = u1 ) 0.5

Each day, we calculate the cumulative mispriced indicies M1 and M2 , which are set to zero at the
beginning of the trading period:
M1,t = M1,t1 + m1,t (13)

13
 M2,t = M2,t1 + m2,t

Positive M1 and negative M2 is interpreted as stock 1 being overvalued relative to stock 2.

Negative M1 and positive M2 is interpreted as the opposite. We have arbitrarily set the strategy
to open a long short position once one of the cumulative mispriced indices is above 0.4 and the
other one is below -0.4 at the same time. The positions are then unwound when both cumulative
mispriced indices return to zero. The pair is then monitored for other possible trades throughout
the remainder of the trading period. Like in the two previous methods, formation and trading
periods are set to 12 and 6 months respectively and the strategy is run every month so we have 6
overlapping portfolios at each month.

4.4. Transaction Costs

Costs play a vital role in the profitability of PTSs. Each pairs trading complete trade consists of
two roundtrip trades. In addition to that, an implicit market impact and short selling costs are also
applicable. Since, the sum of these costs can be large, they can degrade the profitability of PTSs
once taken into account. Do and Faff (2012) have studied the subject of transaction costs on pairs
trading comprehensively. We use a time-varying data set of transaction costs inline with Do and
Faff (2012). The motivation behind this is that commissions are the first element of transaction
costs to be considered. And, since commissions have changed considerably over the last 50 years
that we are using as our time span, a flat commission system distorts the accuracy of our study. To
that extent, we use the institutional commissions that Do and Faff (2012) calculated which starts
from 70 bps in 1962 and gradually declines to 9 bps for recent years. Again, similar to that study,
we have divided our time period into 2 subperiods and used a different market impact estimate for
each subperiod: 30 bps for 1962-1988 and 20 bps for 1989 onward7 . As we screen out stocks that
have low dollar value and low market cap, we assume the remainder domain is relatively cheap to
short sell and therefore do not take into account short selling costs. It is worth noting that we
double these costs to cover two roundtrip trades.

4.5. Performance measurement

The performance of the three PTSs are recorded and compared based on various performance
measures including returns. In accordance with Gatev et al. (2006) and Do and Faff (2010), two
types of returns are calculated: return on committed capital and return on employed capital.
Return on employed capital for one month is calculated as the sum of marked-to-market returns
on that months traded pairs divided by the number of pairs that have traded during that month.
Return on committed capital for one month however, is calculated as the sum of marked-to-market
returns on traded pairs divided by the number of pairs that were nominated to trade in that month,

7
See Do and Faff (2012) Section 3 for full details on commissions and market impact estimations.

14
regardless of whether they actually traded or not. This type of return, which is a more conservative
return measure as compared to the former, more realistically mimics what a hedge fund might use
to report returns, as it takes into account the opportunity cost of the capital that has been allocated
for trading.
As mentioned, we run the strategy each month and do not wait for a trading period to be
complete and therefore have 6 overlapping portfolios each month. The monthly excess return of
a strategy is calculated as the equally weighted average return on these 6 portfolios. Note that,
as the trades neither necessarily start at the beginning of the trading period nor complete exactly
at the end of the trading period, the full capital is not always locked in a trade. Also, there are
months where no trading occurs. Since we do not allocate any interest to the capital when it is not
involved in a trade, the performance is underestimated.
Positions that are opened in the distance and copula methods are $1 long-short positions. $1
long-short positions are defined as opening a long positions worth $1 and a short positions worth
$1 simultaneously. Since the money raised from shorting a stock can be used to buy the other
stock, these positions are self-financing and do not require any capital to trade. However, for the
sake of calculating returns, we adapt the widely used concept of using $1 as the total value of each
long-short position. In the cointegration method, by definition, long and short positions are not
valued equally. However, since we have designed the methodology to ensure a $1 long position for
every trade, we assume an average $1 value for each long-short position.

5. Results

Table 1 reports monthly excess return distribution for each of the three strategies from 1962-2014
both before and after transaction costs in two sections. Section 1 shows the Return on Employed
Capital, while section 2 reports return on committed capital (see section 4.5 for details on calcu-
lations). As both return measures achieve similar results and rankings for the strategies and to
avoid unnecessary complications, we use return on employed capital, hereafter simply referred to
as return, to report results for the remainder of this paper, unless stated otherwise. The average
monthly excess return of the DM before transaction costs is 0.88 percent which is in line with the
0.90 percent reported in Do and Faff (2010). Results presented in section 1 show that, while the
DM and cointegration method are both showing statistically and economically significant and very
similar average monthly excess returns (before and after transaction costs), the copula method
after-cost excess return is relatively small at 5 bps, albeit being statistically significant at 10%.
However, before transaction costs, the copula method is producing a significant 43 bps average
excess return. Moreover, all three strategies show small standard deviations, with the lowest be-
longing to the copula method with 0.0064 after costs and the highest to DM the highest with 0.0108
before costs. With the highest sharpe ratio and lowest value at risk measures, the cointegration
method is the superior strategy in risk-adjusted terms among the three strategies, albeit being only

15
slightly better the distance method. Also, while neither of the strategies show normally distributed
returns, the cointegration method is the only strategy whose returns are positively skewed after
transaction costs. In addition, the return on committed capital measure, presented in section 2,
produces very similar reuslts. It is worth noting that despite this return measure being more con-
servative regarding positive returns, it has the tendency to underestimate losses since the whole
number of pairs is always used in its denominator. This is seen in the copula methods statistics
where its after-costs committed capital return is 1 bps higher than its employed capital. Due to
the same fact, smaller VaR figures are calculated when committed capital return is used. However,
as stated above using either of the return measures will result in same strategy ranking and very
similar results.
Table 1: Pairs Trading Strategies Monthly Excess Return
This table reports key distribution statistics for the monthly excess return time series of three different PTSs i.e.
DM, cointegration, and copula methods for July 1962 to December 2014. Formation and trading period for all
strategies are set to 12 and 6 months respectively. The column labled JB Test tests the null hypothesis of normality
of the series using the Jarque-Bera Test. Null hypothesis is rejected for all strategies.

VaR CVaR JB Test

Strategy Mean t-stat Std. Dev. Sharpe Ratio Skewness Kurtosis
(95%) (95%) p.value

Section 1: Return on Employed Capital

Panel A: after transaction costs
Distance 0.0036 5.2666 0.0108 0.3343 -0.4334 13.7338 -0.0112 -0.0192 0
Cointegration 0.0033 5.0287 0.0097 0.3354 0.3257 8.8402 -0.0112 -0.0172 0
Copula 0.0005 1.7025 0.0064 0.0828 -0.4886 6.5801 -0.0097 -0.0144 0

Panel B: before transaction costs

Distance 0.0088 7.4201 0.0119 0.7423 0.1756 9.2750 -0.0081 -0.0151 0
Cointegration 0.0083 7.3171 0.0109 0.7619 0.8068 7.4610 -0.0067 -0.0126 0
Copula 0.0043 7.1482 0.0069 0.6212 -0.2717 6.1143 -0.0064 -0.0108 0
Section 2: Return on Committed Capital
Panel A: after transaction costs
Distance 0.0030 5.3827 0.0087 0.3413 -0.5295 25.4432 -0.0077 -0.0140 0
Cointegration 0.0028 5.0842 0.0084 0.3339 0.6164 12.7261 -0.0093 -0.0144 0
Copula 0.0006 2.2729 0.0055 0.1105 -0.3761 6.0309 -0.0087 -0.0119 0

Panel B: before transaction costs

Distance 0.0066 7.0261 0.0099 0.6644 0.7377 15.1934 -0.0048 -0.0105 0
Cointegration 0.0065 7.0230 0.0095 0.6884 1.3880 10.9995 -0.0055 -0.0104 0
Copula 0.0034 7.1282 0.0058 0.5799 -0.2130 5.8125 -0.0060 -0.0093 0
*** significant at 1% level
** significant at 5% level
* significant at 10% level

To better realize the relative performance of the strategies, Figure 1 compares the cumulative
excess return for each of the three strategies from 1963-2014. It can be seen how the DM and coin-
tegration methods performance are almost identical to each other with the cointegration method
slightly underperforming compared to the DM. In contrast, the copula method performs poorly in
terms of cumulative excess return. However, the gap between the copula method and the other two

16
strategies is narrower on a risk-adjusted basis, as shown in figure 2.
Figure 1: Cumulative Excess Return
This figure shows how an investment of $1 evolves through time in each of the strategies. Excess return over S&P500
value-weighted return index(including distributions) is used as a benchmark.
11
Distance
Cointegration
Copula
9
Cumulative Excess Return ($)

1965 1970 1975 1980 1985 1990 1995 2000 2005 2010 2015
Time

The 2-year moving average sharpe ratio also confirms the nearly identical performance of the
DM and the cointegration method. In addition, it shows the risk-adjusted performance of all three
strategies fluctuate greatly, but maintain a upward trend until around 1992, where they experience
their peak and the trend reverses. More importantly, the reversal in this trend happens in the
copula and cointegration methods as well as the DM and so, neither of the strategies avoid this
decline. However, after the downward trend begins, the gap between the risk-adjusted performance
of the copula method and the other two strategies becomes smaller than before. It appears that
as we move closer to recent years, the three startegies show a very close risk-adjusted performance
and there is no clear winner among them.
As the return series of neither of the strategies are normally distributed, sharpe ratio, as the
classic risk-adjusted measure, has the potential to underestimate risk and so, overestimate the
risk-adjusted performance (Eling, 2008). In order to avoid such issue, we further analyze the risk
profile of the three PTSs, using performance measures that replace the standard deviation in favor
of a more suitable measure. Table 2 reports various risk-adjusted metrics for each of the strategies,
before and after transaction costs. The measures are divided into two main groups: lower partial
moment measures and drawdown measures. Lower partial moment measures take into account only
the negative deviations of returns from a specified minimum acceptable return. As a result, they
more appropriately account for risk as opposed to the sharpe ratio which considers positive and
negative deviations equally. Omega is defined as the ratio of returns above a threshold to returns
below that threshold (in our case the threshold is 0% and returns are excess returns). Sortino
ratio is the ratio of average excess return (return minus a threshold) to the absolute value of lower

17
Figure 2: 2-Year Moving Average Sharpe Ratio
This figure shows 24-month moving average sharpe ratio for the three strategies.
1.5 Distance
Cointegration
Copula

1
Sharpe Ratio

0.5

-0.5

-1
1970 1975 1980 1985 1990 1995 2000 2005 2010 2015
Time

partial moment. Kappa 3, is the ratio of average excess return to the lower partial moment8 . One
the other hand, drawdown metrics measure the magnitude of losses of a portfolio over a period
of time. Maximum drawdown is defined as the maximum possible loss that could have occured
during a time period. Calmar ratio is defined as the ratio of average excess return to the maximum
drawdown. Streling ratio is the ratio of average excess return to the average of n most significant
continuous drawdowns, thus reducing the sensitivity of the measure to outliers. Slightly different
to maximum drawdown, continuous drawdown is defined as the maximum incurred loss that is not
interrupted by positive returns. Finally, Burke ratio is the ratio of average excess return to the
square root of sum of the squared n most significant drawdowns9 .
Cointegration method shows the best before-cost risk-adjusted performance with the highest
figures for all but the maximum drawdown measure. However, its after-cost performance falls
slightly below that of the DM with the aid of measures such as Kappa 3, maximum drawdown,
and Calmar ratio. The copula method is the poorest PTS among the three, except for showing
the least before-cost maximum drawdown. The major contributor to the low performance of the
copula method is the insignificance of its mean returns. This can be further verified by the fact
that the strategy devolves from having the best maximum drawdown to the worst when transaction
costs are taken into account. Although not with that magnitude, the other two strategies also suffer
considerably when transaction costs are taken into account. We can see the after-costs Omega ratio
decreasing by 70 percent or more for all strategies. The same decrease is observed in the Sortino
ratio of DM and the cointegration method, while the for copula method this figure rises to above

8
For detailed explanation and calculation of lower partial moment measures, refer to Eling and Schuhmacher
(2007)
9
For detailed explanation and calculation of drawdown measures, refer to Schuhmacher and Eling (2011)

18
90 percent.
Table 2: Pairs Trading Strategies Risk-adjusted Performance
This table reports key risk-adjusted performance measures for the monthly excess return time series of three different
PTSs i.e. DM, cointegration, and copula methods for July 1962 to December 2014. Formation and trading period
for all strategies are set to 12 and 6 months respectively.

lower partial moments measures drawdown measures

Omega Sortino Ratio Kappa 3 Max Drawdown Calmar Ratio Sterling Ratio Burke Ratio

Panel A: after transaction costs

Distance 2.6230 0.5983 0.3148 0.1152 0.0313 0.3246 0.0180
Cointegration 2.5691 0.6376 0.3860 0.1704 0.0191 0.3238 0.0168
Copula 1.2531 0.1211 0.0788 0.2012 0.0027 0.0562 0.0032

Panel B: before transaction costs

Distance 8.8186 1.9332 0.9252 0.0737 0.1198 0.9804 0.0687
Cointegration 9.1848 2.3843 1.3090 0.0449 0.1842 1.0004 0.0782
Copula 5.4008 1.3859 0.7631 0.0399 0.1068 0.6636 0.0461

The distribution of trades allows us to further investigate the dynamics of PTSs profitability.
Figure 3 illustrates the trade distributions after transaction costs for each strategy. As we can
see all the strategies have fatter left tail than right which makes extreme negative returns more
likely than extreme positives. This is due to the fact that for all strategies, we use some criteria to
close a trade once it has converged, but we allow the unconverged trades to remain open for the
duration of the trading period. In other words, for the DM and cointegration method, whereby the
criteria for opening and closing positions are directly related to the prices of stocks, the magnitude
of profit for each trade is by definition bound to the scale of divergence that is used in the opening
criteria. Therefore, the profits per each trade is generally limited and so we do not observe fat right
tails. Although rarely, higher profits do occur when a pairs spread suddenly diverges overnight by
much more than just the triggering amount. Similar scenario can also happen on the convergence
of a pair. Contrary to profit-making trades, trades that do not converge can accumulate a big loss
before being forced to close by the strategy at the end of their trading period, which results in
fat left tails. In the copula method, the opening and closing criteria are based on the probability
of relative mispricing within pairs, rather than being directly related to the spread. Therefore,
the observed fatter right tail in its distribution is not out of ordinary. In practice, an additional
risk-limiting criteria for closing a trade, i.e. a stop-loss measure, can potentially limit the extreme
losses. In that case, strategies such as the copula method have the potential to perform well as
their profits are not bound to some specific amount.
Table 3 reports further statistics on the converged and unconverged trades. As expected, all
strategies show positive average return for their converged and negative average return for their
unconverged trades. Moreover, the cointegration method has the highest average return (4.8 per-
cent) and sharpe ratio (1.7) of converged trades among the three strategies. On the other hand, the
DM has the highest percentage of converged trades at 61.62 percent while cointegration method is

19
Figure 3: Distribution of trade returns
This figure shows the distribution of trade returns after transaction costs for each of the three strategies.
(a) Distance (b) Cointegration (c) Copula
12000 12000 9000

8000
10000 10000
7000

8000 8000 6000

Frequency

Frequency

Frequency
5000
6000 6000
4000

4000 4000 3000

2000
2000 2000
1000

0 0 0
-0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 -1 -0.5 0 0.5 1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6
Trade Returns Trade Returns Trade Returns

second with 60.69 percent. Interestingly, the copula method also shows an average return of 3.2
percent and a sharpe ratio of 0.81 for its converged trades. The copula method also shows the
highest sharpe ratio for the unconverged trades among the three strategies. Nonetheless, the con-
siderably high proportion of unconverged trades for the copula method degrades its performance
to only 0.1 percent average return for all trades. In fact, the copula method is the only strategy
that has less converged trades than unconverged. In addition, the DM, with around 21 days, has
the lowest average number of days that takes for its converging trades to converge, whereas for
the copula method this figure is a little less than 25 days. This means that the pairs in the DM
are less vulnerable to risks such as the fundamental risk as they are open for a shorter period.
In addition, more than 96 percent of the converged trades in the cointegration method generate
positve return, higher than the DM with 93 and copula method with 85 percents. This mean
that if the cointegration relation remains present during the trading period, i.e. pair converges,
the cointegration relation is more capable of generating a profitable trade than any other criteria
used in strategies. If the unconverged trades are also considered, the cointegration strategy still
generates the most profitable trades as demonstrated in figure 4. However, the smaller average
returns of this strategy diminishes this advantage and results in an almost equivalent performance
of the DM and the cointegration method.
In order to explain the source of pairs trading profitability and understand whether its perfor-
mance is a compensation for risk, we regress the strategies monthly excess return series against
several risk factors. First, we examine the strategies against the Fama and French (1993) 3 factors
plus the momentum and a liquidity factors. We added the liquidity factor since the it is argued
that liquidity shock is a source of pairs trading profitability (Engelberg et al., 2009). Similar to
Do and Faff (2012), we use Pastor-Stambaughs Innovations in Aggregate Liquidity series from
Wharton Research Data Services (WRDS) (see Pastor and Stambaugh (2003) for the detail on the
liquidity factor).
Results from Table 4 show that, the profits of PTSs are not fully explained by the the risk factors.
In fact, large and significant alphas are observed for all the strategies. It is worth mentioning that

20
Figure 4: Proportion of Positive and Negative Trades
This figure shows the proportion of trades with positive and negative returns after transaction costs for each of the
three strategies.
(a) Distance (b) Cointegration (c) Copula
positive return positive return positive return
negative return negative return negative return

30%
33%

44%

56%

67%
70%

alphas are very close to their corresponding strategys mean of monthly excess return, implying
that after being adjusted for these risk-factors, strategies profits remain unaffected. Moreover, the
momentum factor is negatively correlated with strategies profits, with significant t-statistics both
before and after transaction costs. Similarly, liquidity has also a considerable negative effect on
profits, but only for the DM and the cointegration method. Interestingly, liquidity can not explain
the profits of the copula method neither before nor after costs, as opposed to researches such as
Engelberg et al. (2009) which suggest otherwise. This further demonstrates the different nature
of the copula method in comparison to the other PTSs. Surprisingly, no other factor, including
the market excess return, is correlated to the strategies returns, which is more evidence for the
market-neutrality of these strategies. From an investors perspective, this market neutrality can
have diversifying effects on portfolios and so decrease the risk of portfolios, especially the ones that
are somehow correlated with the market.
Next, we regress the monthly excess returns of the three strategies against the recent Fama and
French (2014) 5 factor model. This model is an attempt by the authors to improve their well-known
3 factor model by introducing 2 additional factors: profitability and investment factors. Robust
Minus Weak (RMW), a measure of profitability, is defined as as the difference between returns
on robust and weak profitability portfolios, while Conservative Minus Aggressive (CMA), which
represents the investment factor, is defined as the difference between the returns of conservative
and aggressive portfolios. Results, reported in Table 5, show that there is a negative correlation
between before-cost pairs profit and RMW for two of the strategies. This relationship is statistically
significant at 1 and 5 percent for the DM and cointegration method respectively, albeit the size of
the regression coefficients pointing out the smallness of these correlations economical significance.
Nonetheless, RMW only affects the after-cost profitability of the DM, which is again economically
insignificant. Notably, the size factor has a relation to some PTSs return, although at a economi-
cally small level. Similar to prior results, alphas remain large and statistically significant at 1% for
all strategies, with the exception of after-costs copula method, which shows that the risk-factors

21
are unable to account for the profits generated by the strategies.

22
 Table 3: Converged and Unconverged Trades Return Series
This table reports key distribution statistics for converged and unconverged trade return series after tranasction costs. The St.D and S.R columns
report standard deviation and sharpe ratio respectively. Two columns unde the Days Open title show the mean and median number of days that a
converged trade remains open. Positive Trades (%) shows the percentange of converged trades with positive returns. Note that all calculations are based
on after-cost returns.

Converged Trades Unconverged Trades

Days Open Positive

Strategy % of Trades Mean St.D. S.R. Skewness % of Trades Mean St.D. S.R. Skewness
Mean Median Trades (%)

Distance 61.62 0.0411 0.0264 1.5560 1.2806 21.3287 14 93.09 38.38 -0.0379 0.0722 -0.5250 -1.9868
Cointegration 60.69 0.0480 0.0282 1.7003 1.8603 22.7878 15 96.94 39.31 -0.0360 0.0763 -0.4720 -1.9716
Copula 43.96 0.0320 0.0394 0.8116 1.6438 24.7462 15 84.79 56.04 -0.0232 0.0775 -0.2999 -0.8497

Table 4: Monthly Risk Profile: Fama French 3 Factors + Momentum and Liquidity
This table shows results of regressing after costs (panel A) and before costs (panel B) monthly return series against Fama-French 3 factors plus momentum
and liquidity. The column labled Alpha reports the estimated regression intercept while Factor columns report the estimated coefficient for each
factor. Stambaugh (2003)s Innovations in Aggregate liquidity measure from WRDS is used as the liquidity factor. The t-stat shows the test statistic for
each regression coefficient, calculated using Newey-West standard errors with 6 lags.

23
Market Excess Return Small minus Big Low minus High Momentum Liquidity

Strategy Alpha t-stat Factor t-stat Factor t-stat Factor t-stat Factor t-stat Factor t-stat

Panel A: after transaction costs

Distance 0.0039 6.6702 0.0107 0.7291 -0.0167 -1.1679 -0.0072 -0.4157 -0.0316 -2.4950 -0.0396 -3.3246
Cointegration 0.0035 6.5505 0.0084 0.6850 -0.0163 -1.2404 -0.0033 -0.2506 -0.0299 -2.9125 -0.0377 -3.7144
Copula 0.0006 1.8116 0.0116 1.5249 0.0048 0.5557 0.0010 0.0978 -0.0163 -2.2354 -0.0030 -0.5344

Panel B: before transaction costs

Distance 0.0092 12.7340 0.0086 0.5680 -0.0096 -0.6008 -0.0014 -0.0682 -0.0318 -2.2437 -0.0479 -3.7234
Cointegration 0.0086 12.4213 0.0054 0.4259 -0.0084 -0.5401 0.0024 0.1516 -0.0304 -2.5983 -0.0456 -4.1276
Copula 0.0044 10.8845 0.0083 1.0358 0.0111 1.1828 0.0034 0.3182 -0.0139 -1.8743 -0.0051 -0.8544
 Table 5: Monthly Risk Profile: Fama French 5 Factors
This table shows results of regressing after costs (panel A) and before costs (panel B) monthly return series against Fama-French 5 factors (Fama and
French, 2014). The column labled Alpha reports the estimated regression intercept while Factor columns report the estimated coefficient for each
factor. The t-stat shows the test statistic for each regression coefficient, calculated using Newey-West standard errors with 6 lags.

Market Excess Return Small minus Big Low minus High Robust minus Weak Conservative minus Aggressive

Strategy Alpha t-stat Factor t-stat Factor t-stat Factor t-stat Factor t-stat Factor t-stat

Panel A: after transaction costs

Distance 0.0040 6.3153 -0.0000 -0.2156 -0.0004 -2.4217 0.0001 0.4438 -0.0007 -2.2580 -0.0001 -0.3367

24
Cointegration 0.0035 5.7479 -0.0001 -0.4690 -0.0003 -1.7851 -0.0001 -0.3819 -0.0004 -1.5643 0.0001 0.3724
Copula 0.0004 1.2726 0.0002 1.9367 0.0000 0.0051 -0.0001 -0.5078 -0.0001 -0.7569 0.0003 1.0620

Panel B: before transaction costs

Distance 0.0094 12.0278 -0.0001 -0.5591 -0.0004 -2.0923 0.0002 0.6308 -0.0009 -2.6212 -0.0002 -0.3985
Cointegration 0.0087 11.3561 -0.0002 -0.8597 -0.0003 -1.3988 -0.0000 -0.0293 -0.0006 -1.9923 0.0001 0.1970
Copula 0.0043 10.4685 0.0001 1.1250 0.0000 0.3893 -0.0000 -0.1252 -0.0002 -1.1920 0.0002 0.6577
* significant at 10% level
** significant at 5% level
*** significant at 1% level
 Next, we compare the performance of PTSs in different time periods. We divide the full data
set into 2 different periods: crisis and Normal. Crisis is defined as the period including the top 20%
of years in which the US equity market shows its worst performance (a total of 11 out of 53 years).
Normal refers to the period which is not included in the crisis. Figure 5 reports the risk-adjusted
performance of the strategies in these periods.
First and foremost, with the exception of the DMs sortino ratio, all strategies show better per-
formance in the crisis period compared to the normal period. Interestingly, while the cointegration
method shows a performance slightly inferior to that of the distance method in normal period, it
clearly outperforms all strategies in the crisis period. This superiority is most obvious in the sortino
ratio where the cointegration methods figure rises from 0.6 in the normal period to nearly 0.8 in
the crisis period as opposed to the DM that experiences a fall in its sortino ratio from above 0.6
to 0.6 for the same periods. These results and similar results from other studies (see e.g. Do and
Faff (2010)) motivate the use of such market neutral strategies in portfolios as protection against
turbulent market conditions. In times that other passive investment strategies tend to perform
poorly, an investment strategy that includes a PTS such as the cointegration method, could benefit
from an increase in the risk-adjusted performance measure.
Figure 5: Average Monthly Performance in Crisis and Normal Periods
This figure shows the performance of each strategy on two different crisis and normal subperiods. Crisis is defined
as the 20% annual subperiods in which the US stock market has shown its poorest performance (11 worst-performing
years out of 53) and Normal consists of the remaining annual subperiods. S&P500 value-weighted index is used as
a proxy for the US stock market.
(a) Sharpe Ratio (b) Sortino Ratio (c) Mean/CVaR(95%)
0.5 0.9 0.3
Distance Distance Distance
0.45 Cointegration 0.8 Cointegration Cointegration
Copula Copula 0.25 Copula
0.4 0.7
0.35
Mean/CVaR(95%)

0.6 0.2
Sharpe Ratio

Sortino Ratio

0.3
0.5
0.25 0.15
0.4
0.2
0.3 0.1
0.15

0.1 0.2
0.05
0.05 0.1

0 0 0
Crisis Normal Crisis Normal Crisis Normal

Figure 6 reports the each strategies mean monthly excess return over 5-year periods for the
duration of the study, both before and after transaction costs. The strategies best performance
period is 1982-1986 if transaction costs are taken into account (6a), but 1972-1976 if they are not
(6b). Interestingly, we can see a decrease in the trade count in the 1982-1986 period (6c). Since
the magnitude of trading costs are considerable, this decrease causes the after-cost performance
to rise. For the rest of the periods, as the average trade count remains relatively constant, the
decline in before-cost performance directly affects the after-cost performance. Notably, the DM
and cointegration method have experienced a 40 and a 35 percent drop in their average number
of trades from the 2007-2011 to the 2012-2014 period respectively, whereas the copula methods

25
average trade count has dropped by only 15 percent in the same period. This shows that in
recent years, pairs whose prices closely follow each other diverge less frequently than they used
to before. However, as their complication increases, trading strategies are able to produce trading
opportunities more frequently. On the other hand, as the costs of trading have dramatically declined
over the years, it might be possible to reduce the threshold by which positions are opened, and still
be able to generate enough profit to cover for costs. This reduction in the threshold will allow the
strategies to capture smaller deviations from the equilibrium relation, which is seen in recent years,
and so by increasing the number of trades their performance could be enhanced.
Figure 6: Periodic Performance of Pairs Trading Strategies
This figure shows the mean of monthly excess return of each strategy in different periods 6a shows performance after
transaction cost, while 6b represents the same figure before transaction costs.
(a) After Transaction Costs
0.01

0.008
Distance
Cointegration
mean excess return

0.006 Copula

0.004

0.002

-0.002

-0.004
63-66 67-71 72-76 77-81 82-86 87-91 92-96 97-01 02-06 07-11 12-14
period

(b) Before Transaction Costs

0.016 Distance
Cointegration
mean excess return

0.014
Copula
0.012
0.01
0.008
0.006
0.004
0.002
0
63-66 67-71 72-76 77-81 82-86 87-91 92-96 97-01 02-06 07-11 12-14
period

(c) Trade Count

600
Average Yearly Trade Count

Distance
500 Cointegration
Copula
400

300

200

100

0
63-66 67-71 72-76 77-81 82-86 87-91 92-96 97-01 02-06 07-11 12-14
period

26
6. Conclusion

We examine and compare the performance of three pairs trading strategies using daily stock data
from July 1962 to December 2014: the distance, cointegration, and copula methods. We use a
time-varying series of trading costs for ascertaining the strategies performance with high accuracy
and robustness. We find that on a risk-adjusted performance level the cointegration strategy
performs as well as the DM. We also examine a relatively new PTS using copulas. Although at
this point we find its performance to be inferior to that of the DM and cointegration method,
certain attributes to it motivates further attention. First, in recent years that the simple strategies
suffer form a decline in trading opportunities, the copula method remains stable in presenting such
opportunities. Second, the copula method shows returns comparable to those of other methods
in its converged trades. However, its relatively high proportion of unconverged trades diminishes
such profits. Therefore, any attempt to increase the ratio of converged trades or limit their losses
(by for example implementing a stop-loss), would result in an enhanced performance. In the wake
of the mentioned recent decline in arbitrage opportunities that are captured by simple methods,
this potential improvement in the performance of a more complex method could be promising.
Moreover, we show that the markets excess return fails to account for the performance of the three
PTSs. Also that, the strategies perform better in crisis. Therefore, the use of such strategies can
be encouraged in investment portfolios for both diversifying risk and capturing alpha. We use the
SSD as at least one of the, if not the only, criteria in the pair formation phase of all three PTSs.
Nonetheless, there is a distinct possibility to form pairs based on cointegration and copula. Future
research could aim to examine the performance of such strategies.

27
7. References

Aas, K., Czado, C., Frigessi, A., Bakken, H., Apr. 2009. Pair-copula constructions of multiple
dependence. Insurance: Mathematics and Economics 44 (2), 182198.

Andrade, S., Di Pietro, V., Seasholes, M., 2005. Understanding the profitability of pairs trading.
Unpublished working paper, UC Berkeley, Northwestern University.

Ang, A., Chen, J., 2002. Asymmetric correlations of equity portfolios. Journal of Financial Eco-
nomics 63 (3), 443494.

Avellaneda, M., Lee, J.-H., 2010. Statistical arbitrage in the US equities market. Quantitative
Finance 10 (7), 761782.

Broussard, J. P., Vaihekoski, M., Dec. 2012. Profitability of pairs trading strategy in an illiquid
market with multiple share classes. Journal of International Financial Markets, Institutions and
Money 22 (5), 11881201.

Caldeira, J., Moura, G. V., 2013. Selection of a portfolio of pairs based on cointegration: A statis-
tical arbitrage strategy. Available at SSRN 2196391.

Do, B., Faff, R., Jul. 2010. Does Simple Pairs Trading Still Work? Financial Analysts Journal
66 (4), 8395.

Do, B., Faff, R., 2012. Are Pairs Trading Profits Robust to Trading Costs? Journal of Financial
Research 35 (2), 261287.

Eling, M., 2008. Does the measure matter in the mutual fund industry? Financial Analysts Journal
64 (3), 5466.

Eling, M., Schuhmacher, F., 2007. Does the choice of performance measure influence the evaluation
of hedge funds? Journal of Banking & Finance 31 (9), 26322647.

Elliott, R. J., Van Der Hoek*, J., Malcolm, W. P., 2005. Pairs trading. Quantitative Finance 5 (3),
271276.

Engelberg, J., Gao, P., Jagannathan, R., 2009. An anatomy of pairs trading: the role of idiosyn-
cratic news, common information and liquidity. In: Third Singapore International Conference on
Finance.

Engle, R. F., Granger, C. W., 1987. Co-integration and error correction: representation, estimation,
and testing. Econometrica: journal of the Econometric Society, 251276.

Erb, C. B., Harvey, C. R., 2006. The strategic and tactical value of commodity futures. Financial
Analysts Journal 62 (2), 6997.

28
Fama, E. F., French, K. R., 1993. Common risk factors in the returns on stocks and bonds. Journal
of financial economics 33 (1), 356.

Fama, E. F., French, K. R., 1997. Industry costs of equity. Journal of financial economics 43 (2),
153193.

Fama, E. F., French, K. R., 2014. A five-factor asset pricing model. Journal of Financial Economics.

Fortin, I., Kuzmics, C., 2002. Tail-dependence in stock-return pairs. Intelligent Systems in Account-
ing, Finance and Management 11 (2), 89107.

Fuertes, A.-M., Miffre, J., Rallis, G., 2010. Tactical allocation in commodity futures markets:
Combining momentum and term structure signals. Journal of Banking & Finance 34 (10), 2530
2548.

Galenko, A., Popova, E., Popova, I., 2012. Trading in the Presence of Cointegration. The Journal
of Alternative Investments 15 (1), 8597,6.

Gatev, E., Goetzmann, W. N., Rouwenhorst, K. G., 2006. Pairs trading: Performance of a relative-
value arbitrage rule. Review of Financial Studies 19 (3), 797827.

Hatherley, A., Alcock, J., 2007. Portfolio construction incorporating asymmetric dependence struc-
tures: a users guide. Accounting and Finance 47 (3), 447472.

Jones, C. M., 2002. A century of stock market liquidity and trading costs. Available at SSRN
313681.

Khandani, A. E., Lo, A. W., 2007. What happened to the quants in August 2007? Journal of
Investment Management 5 (4), 554.

Lin, Y.-X., McCRAE, M., Gulati, C., 2006. Loss protection in pairs trading through minimum
profit bounds: A cointegration approach. Advances in Decision Sciences 2006.

Lo, A. W., MacKinlay, A. C., 1990. When are contrarian profits due to stock market overreaction?
Review of Financial studies 3 (2), 175205.

Low, R. K. Y., Alcock, J., Faff, R., Brailsford, T., 2013. Canonical vine copulas in the context
of modern portfolio management: Are they worth it? Journal of Banking and Finance 37 (8),
30853099.

Miffre, J., Rallis, G., 2007. Momentum strategies in commodity futures markets. Journal of Banking
& Finance 31 (6), 18631886.

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

29
Patton, A. J., 2004. On the Out-of-Sample Importance of Skewness and Asymmetric Dependence
for Asset Allocation. Journal of Financial Econometrics 2 (1), 130168.

Perlin, M. S., 2009. Evaluation of pairs-trading strategy at the Brazilian financial market. Journal
of Derivatives & Hedge Funds 15 (2), 122136.

Schuhmacher, F., Eling, M., 2011. Sufficient conditions for expected utility to imply drawdown-
based performance rankings. Journal of Banking & Finance 35 (9), 23112318.

Sims, C. A., Stock, J. H., Watson, M. W., 1990. Inference in linear time series models with some
unit roots. Econometrica: Journal of the Econometric Society, 113144.

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

Srensen, B. E., 2005. 1 Cointegration.

Vidyamurthy, G., 2004. Pairs trading: quantitative methods and analysis. J. Wiley, Hoboken, N.J.

Wu, Y., 2013. Pairs trading: a copula approach. Journal of derivatives & hedge funds 19 (1), 12
30.

Xie, W., Liew, Q. R., Wu, Y., Zou, X., 2014. Pairs Trading with Copulas. Available at SSRN
2383185.

Xie, W., Wu, Y., 2013. Copula-Based Pairs Trading Strategy. In: Asian Finance Association (AsFA)
2013 Conference.

30