Trading Signal Analysis with Pairs

Trading Strategy in the Stock Exchange

of Thailand

Natnarong Namwong, Woraphon Yamaka, and Roengchai Tansuchat(B)

Faculty of Economics, Chiang Mai University, Chiang Mai, Thailand

[email protected], [email protected], [email protected]

Abstract. Pair trading is a popular strategy. The concept of pair trad-

ing consists of two processes. First, find two stocks whose prices tend to
move together. Second, short the overvalued stock and buy the underval-
ued stock simultaneously when the spread between the prices diverges.
A profit between the prices will be made if the prices converge again.
The return from the strategy is often uncorrelated with market return
because the market risk can be eliminated by shorting the overvalued
stock and buying the undervalued stock simultaneously. Therefore, the
results of the strategy must have a lower risk than trading each stock
in the market. The purpose of this study is to find the trading signal
for the pair-trading strategy in Stock Exchange of Thailand (SET). The
proposed trading signal for pair-trading strategy can generate a higher
return when compared with trading in individual stocks. The investors
can employ it to gain the better profit in their portfolios. We propose
Markov-switching GARCH model for pair-trading strategy. The model is
considered with two regimes which have different variance in each regime.
The average of the two variances is used as trading signal. We apply the
pair-trading strategy to SET100 index based on the most 100 liquid
stocks. The pairs formation period is over 1 January 2016 to 12 Decem-
ber 2016 and the trading period is over 2 January 2017 to 29 December
2017. The empirical results show that the trading signals of the Markov-
switching GARCH model generate positive return for all selected pairs
and provide the highest return of up to 14.267%. The return from pair-
trading strategy is higher than the return from trading in individual
stocks.

Keywords: Pair-trading strategy · Markov switching model

GARCH model · Trading signal · SET100 index

1 Introduction

Pairs trading is a market-neutral strategy which trades stock in pairs with long
position in one stock and short position in the other. The strategy was invented
by Nunzio’s quantitative analysis group in 1980s. The concept of pair trading
c Springer Nature Switzerland AG 2019
V. Kreinovich and S. Sriboonchitta (Eds.): TES 2019, SCI 808, pp. 378–388, 2019.
https://doi.org/10.1007/978-3-030-04263-9_29
 Trading Signal Analysis with Pairs Trading Strategy 379

consists of two processes. First, find two stocks whose prices move together. Sec-
ond, short the overvalued stock and buy the undervalued stock simultaneously
when the prices diverge. A profit between the prices will be made if the prices
converge again. There are many methods for pair-trading strategy. Gatev, Goet-
zmann, and Rouwenhorst [11] used distance approach to find the stock prices
that move together in U.S. stocks during 1962–2002. The distance approach uses
the sum of squared differences between normalized stock prices and the mini-
mum values offer a corresponding pair. They used two standard deviations as
trading signal. Trade is opened when the normalized prices diverge more than
two historical standard deviations, and a profit will be made at the next crossing
of the normalized prices. This trading signal provides annualized excess return
of up to 11%. Do and Faff [7] used the same pair-trading strategy as Gatev et al.
[11] in U.S. stocks. They extended the data to 2006. The empirical results indi-
cated that the return from the trading signal based on two standard deviations
does not provide highly excess return as shown in the past. The dramatically
declining returns are caused by the increased magnitude of divergence rather
than an increase in the market efficiency.
Vidyamurthy [15] proposed cointegration method for pair-trading strategy.
The cointegration method applies cointegration test of Engle and Granger [10]
to identify the comoving of stocks, and then uses a standard deviation as trading
signal. However, Chen et al. [3] indicated that the cointegration method cannot
find enough cointegration pairs in the sample. Elliott, van der Hoek, and Malcolm
[8] modeled the spread between the prices but they did not show how to select
the pairs. Trade is opened when the predicted spread from the model is different
from the subsequent spread at significant level. Do et al. [6] criticized the model
proposed by Elliott et al. [8] and they proposed the stochastic residual spread
model for pair-trading strategy.
In some situations, the spread exhibits switching behavior which has changing
mean and variance. This causes the trading signal in the previous study to be
wrong because the spread is assumed to be a single regime process. Bock and
Mestel [1] applied Markov regime-switching model to the price ratio of a pair
and created a trading signal for pair-trading strategy. The pair-trading signal
with regime-switching model can generate positive return. Zhu, Yamaka, and
Sriboonchitta [17] proposed Markov Switching Regression GARCH model for
pair-trading signal and indicated the structural change in the pair return. They
applied the proposed pair-trading signal to SET50 index based on the most
50 liquid stocks. The empirical results showed that the pair-trading strategy is
more effective for investment management than individual stock method. Yang
et al. [16] used the distance approach to find the pairs in the S&P 500 stocks.
They created the spread model which combines the Markov regime-switching
model with mean reversion model and compared it to single regime models.
The empirical results indicated that the trading signal of their model provides
a better performance than other models in short-term trading.
To capture some facts of financial market, the autoregressive conditional
heteroscedastic (ARCH) model was proposed by Engle [9] and the generalized
380 N. Namwong et al.

ARCH (GARCH) model was proposed by Bollerslev [2]. Both models have been
employed to describe the dynamic features of volatility. Chen et al. [3] used the
distance approach to find stock pairs from 30 companies in DJIA. They modeled
the return spread as the threshold autoregressive model with GARCH effects.
The threshold value of the model is used as trading signal. Chodchuangnirun,
Zhu, and Yamaka [5] used the distance approach to select the pairs from 36 com-
panies in DJIA, NYSE, and NASDAQ. They proposed three models consisting
of kink autoregressive GARCH model, threshold autoregressive GARCH model
and Markov switching autoregressive GARCH model to find the most appropri-
ate model for the selected pairs. These models are considered with two regimes
which are used as trading signal. The empirical results indicated that the Markov
switching autoregressive GARCH model is the best model for all return spreads
and the trading signal of the model can generate positive returns. In addition,
Chen et al. [4] used smooth transition heteroskedastic models with a second-
order logistic function to generate trading signal and suggested two pair-trading
strategies. The pairs were selected by using minimum square distance method.
They applied the pair-trading strategies to 36 stocks from U.S. stock market.
Their proposed strategies provide average annualized returns of at least 35.5%
without a transaction cost and at least 18.4% with a transaction cost.
In our previous study, the model captured with the volatility dynamics of the
return spread was employed to create a trading signal for pair-trading strategy.
We propose Markov-switching GARCH model to capture the volatility dynamics
of the return spread. The model keeps attention only on the conditional variance
structure. We consider two regimes in the model which has different variances in
each regime. The average of two variance is used as trading signal. The objective
of this study is to find the trading signal of the pair-trading strategy in Stock
Exchange of Thailand (SET). The proposed trading signal for pair-trading strat-
egy can generate higher return when compared with trading in individual stocks.
This is important for the investors because the investors must encounter both
market risk and specific risk when they trade each stock in the market. The
market risk may be eliminated by using the pair-trading strategy because risk
will be offset when short the overvalued stock and buy the undervalued stock
simultaneously. Therefore, the investors can employ the proposed trading signal
for pair-trading strategy to provide a lower risk than trading in individual stocks
and gain a better profit in their portfolios.
The remaining section is organized as follows: Sect. 2 presents the methodol-
ogy including Markov-switching GARCH model, the maximum likelihood esti-
mation, and the pair trading procedure. Section 3 presents the empirical results.
Finally, Section 4 contains the conclusion.

2 Methodology
2.1 Markov-Switching GARCH Models
The GARCH (1, 1) model of Bollerslev [2] is defined as
y t = μ + εt (1)
 Trading Signal Analysis with Pairs Trading Strategy 381

εt = ht ηt , ηt ∼ i.i.d.(0, 1) (2)
ht = α0 + α1 ε2t−1 + β1 ht−1 (3)
where yt are observations at time t. μ is conditional mean. The standard residual
ηt is i.i.d. with zero mean and unit variance. ht is conditional variance and the
parameters in the conditional variance are restricted by α1 > 0, β1 > 0 and
α1 + β1 < 1.
The study assumes the conditional mean to be zero and yt to be not serially
correlated, that is E(yt ) = 0 and E(yt yt−i ) = 0 for i = 0 and t > 0. The
standard residual ηt is normally distributed with zero mean and unit variance.
The conditional variance ht is allowed for regime switching in the parameters.
Therefore, the regime-switching GARCH is defined as process

yt = ht ηt , ηt ∼ N ID(0, 1) (4)
2
ht = α0,St + α1,St yt−1 + β1,St ht−1 (5)
where St is state variable.
The study follows the work of Haas et al. [12]. St is assumed as an unobserved
first-order ergodic homogeneous Markov chain with transition probability matrix
P . In this study, we consider the model with two regimes St ∈ {1, 2}. The
transition probabilities matrix P is written as
 
p11 p12
P = (6)
p21 p22

where pij = P (St = i|St−1 = j) for ∀i, j ∈ {1, 2} is the probability of state
changes moving from state i at time t − 1 to state j at time t. The  probability of
2
state changes pij for ∀i, j ∈ {1, 2} are restricted by 0 < pij < 1 and j=1 pij = 1.
From Eq. 5, ht is variance of observations yt conditional on the realization of
St ∈ {1, 2} and the conditional variance ht for St ∈ {1, 2} are assumed to follow
2 separate GARCH processes which evolve in parallel.

2.2 The Maximum Likelihood Estimation

The maximum likelihood estimation is used to estimate the model. The like-
lihood function is written as

n
L(Ψ |yt ) = f (yt |Ψ, yt−k ) (7)
t=1

where Ψ = (α0,St , α1,St , β1,St , pij ) is the vector of model parameters for ∀i, j ∈
{1, 2} and St ∈ {1, 2}. f (yt |Ψ, yt−k ) is defined as the density of yt given past
observations yt−k for k > 0 and model parameters Ψ . The conditional density of
yt is
 2  2
f (yt |Ψ, yt−i ) = pij zi,t−1 fD (yt |St = j, Ψ, yt−k ) (8)
i=1 j=1
382 N. Namwong et al.

where zi,t−1 is the filtered probability of state i at time t − 1 which was proposed
by Hamilton [13] and Hamilton [14] for details. fD (yt |St = j, Ψ, yt−k ) is density
of yt in state St = j given past observations yt−k for k > 0 and model parameters
Ψ.

2.3 Pair Trading Procedure

The process of pair trading strategy consists of two periods, which are forma-
tion period and trading period. In the formation period, the study selects pairs
by using the distance approach proposed by Gatev et al. [11]. The distance app-
roach is the sum of squared differences (SSD) between normalized price series as
follows:
n
B 2
SSD = t − xt )
(xA (9)
t=1

where xjt
is the normalized price of stock j at time t during the formation period.
The normalized price of stock is calculated by

Xtj − X̄ j
xjt = (10)
σj

where Xtj is the daily closing price of stock j at time t during the pairs formation
period, X̄ j is the average of X j and σ j is the standard deviation of X j .
The study selects the first five smallest SSD values. Then, the first five pairs
are used to calculate the return spread, which is the difference between the
two series of returns. Suppose, we have stocks A and B, the return spread is
calculated by
yt = rtA − rtB (11)
where rtA and rtB are the return series of normalized stock prices A and B,
respectively, at time t during the pairs formation period. The return series of
stock are calculated by
rtj = ln(Xtj /Xt−1
j
) (12)
In trading period, the study fits the Markov-switching GARCH model to the
five return spreads and creates the trading signal for each return spread. The
pair-trading signal was firstly proposed by Gatev et al. [11]. They used the two
standard deviations as the trading signal. Therefore, the model consists of two
regimes which have different variance in each regime. The square root of the
average of the two variances is used as trading signal. Our trading signal is

sell A buy B, if yt ≥ +δ h̃t
 (13)
buy A sell B, if yt ≤ −δ h̃t

where yt is observed return spread at time t during trading period. The critical
value δ is set at 1.645 (90% confidence). h̃t is the average of two variances
 Trading Signal Analysis with Pairs Trading Strategy 383

which were estimated in the formation period. The average of two variances is
calculated by

h̃t = E(ht,1 )P r(St = 1) + E(ht,2 )P r(St = 2) (14)

where E(ht,St ) is expectation of conditional variance of yt on the realization of

St for St ∈ {1, 2}. P r(St ) for St ∈ {1, 2} is unconditional probabilities in each
regime St .
The study will short stock A and buy stock B when the subsequent
 return
spread is greater than or equivalent to the upper boundary (+δ h̃t ). A  profit
will be made when the return spread crosses over the lower boundary (−δ h̃t ).
Conversely, buy stock A and short stock B when the  subsequent return spread
is less than or equivalent to the lower boundary (−δ h̃t ). A profit will be made
when the subsequent spread crosses over the predicted value of the upper bound-
ary (+δ h̃t ). Finally, the study calculates the average trading return on shorting
stock A and buying stock B as follows

A B
1 Xsold Xsold
R1 = − ln A + ln B (15)
D Xbought Xbought

where D is the number of holding days. The average trading return on buying
stock A and shorting stock B is as follows

A B
1 Xsold Xsold
R2 = ln A − ln B (16)
D Xbought Xbought

3 Empirical Results
3.1 Data Description

In this study, we use the daily closing prices of stocks in SET100 index based
on the most 100 liquid stocks. The data is obtained from Thomson Reuters data
stream over a 2-year period from 1 January 2016 to 29 December 2017. There
are 94 stock prices in the SET100 index to be used because the data of the
remaining 6 stock prices is fewer than the others. The data are separated into
two periods for pair trading procedure. The pairs formation period is over 1
January 2016 to 12 December 2016 and the trading period is over 2 January
2017 to 29 December 2017.

3.2 Pairs Selection

In the pairs formation period, the number of possible pairs is 4,371 (C294 ). We
calculate the sum of squared differences (SSD) between the normalized prices for
4,371 pairs. Table 1 shows the first five pairs ranked with smallest SSD values.
384 N. Namwong et al.

Table 1. The first five pairs with the smallest SSD value

No. Stock A Stock B SSD values

1 ADVANCE INTOUCH 11.828
2 CBG PTG 12.021
3 PTG TKN 12.230
4 CBG COM7 13.460
5 BIG GLOBAL 14.857

3.3 The Estimation Results of Markov-Switching GARCH Model

Before the return spread of the five pairs will be estimated with the
Markov-switching GARCH model during the pair formation period, the return
spreads must be tested for unit root. The study employs the Augmented Dickey
Fuller (ADF) test. The test results show ADF test rejects the null hypothesis
for all pairs that the return spreads have a unit root in level. Therefore, the
return spreads for the five pairs are stationary. Table 2 shows the parameters for
the return spreads estimated with Markov-switching GARCH model when the
residual term in the model is assumed as a normal distribution. The model is
focused on two states which have different variance. For the first state (St = 1),
the estimated parameters in the conditional variance (ht ) are α0,St =1 , α1,St =1 ,
and β1,St =1 . For the second state (St = 2), the estimated parameters in the con-
ditional variance (ht ) are α0,St =2 , α1,St =2 , and β1,St =2 . The sum of α1,St + β1,St
for each regime is defined as the volatility persistence which shows the time is
needed for shocks in volatility to disappear. From Table 3, The volatility per-
sistence of ADVANCE-INTOUCH, CBG-COM7, and BIG-GLOBAL pairs for
the second regime is higher than the first regime. This indicates that the second
regime takes longer time for shocks in volatility to disappear. For CBG-PTG
and PTG-TKN pairs, the first regime uses longer time for shocks in volatility
to disappear because the volatility persistence for the first regime is higher than
the second regime. Therefore, the investor must encounter different volatility in

Table 2. Parameter estimates for the return spread of the five pairs

PARAMETER ADVANCE-INTOUCH CBG-PTG PTG-TKN CBG-COM7 BIG-GLOBAL

α0,St =1 3.46E–05 4.74E–07 2.49E–05 0.0004 3.02E–06
α1,St =1 0.1487 0.2542 0.0059 2.42E–05 0.0012
β1,St =1 0.0001 0.0247 0.9183 0.0006 0.1328
α0,St =2 0.0002 0.0007 0.0003 0.0001 0.0008
α1,St =2 0.0013 0.1066 0.0006 0.0001 0.0132
β1,St =2 0.6607 0.0339 0.8307 0.9427 0.3924
p11 0.9385 0.0003 0.6804 0.918 0.2063
p22 0.4907 0.6868 0.4874 0.9359 0.9323
 Trading Signal Analysis with Pairs Trading Strategy 385

each period of time. Moreover, the estimated transition probabilities for CBG-
PTG, CBG-COM7, and BIG-GLOBAL pairs indicate that the second regime
is more persistent than the first regime. For ADVANCE-INTOUCH and PTG-
TKN pairs, the estimated transition probabilities indicate the first regime is
more persistent than the second regime.

3.4 Pair Trading Strategy

In the trading period, the estimated parameters of two regimes in conditional

variance ht are used to calculate the expectation of conditional variance E(ht,St )
for each regime St ∈ {1, 2}. Then, the estimated transition probabilities, p11 and
p22 , are used to calculate the unconditional probabilities P r(St ) for St ∈ {1, 2}.
Both E(ht,St ) and P r(St ) for St ∈ {1, 2} are used to calculate the average of two
variances in Eq. 14. We will use the average of the two variances to determine
trading signal in Eq. 13.
Figures 1, 2, 3, 4 and 5 show the return spreads of five pairs during the trading
period, 2 January 2017 to 29 December 2017. The blue line is upper boundary
and the orange line is lower boundary. The two lines are indicative of trading
entry and exit signals. According to Eqs. 15 and 16, the study will calculate a

0.05
0.04
0.03
0.02
0.01
0
-0.01
-0.02
-0.03

Upper boundary Lower boundary Return spread

Fig. 1. The return spread of ADVANCE-INTOUCH pair

0.2
0.15
0.1
0.05
0
-0.05
-0.1
-0.15

Upper boundary Lower boundary Return spread

Fig. 2. The return spread of CBG-PTG pair

386 N. Namwong et al.

0.15
0.1
0.05
0
-0.05
-0.1
-0.15
-0.2

Upper boundary Lower boundary Return spread

Fig. 3. The return spread of PTG-TKN pair

0.1

0.05

-0.05

-0.1

-0.15

Upper boundary Lower boundary Return spread

Fig. 4. The return spread of CBG-COM7 pair

0.15

0.1

0.05

-0.05

-0.1

Upper boundary Lower boundary Return spread

Fig. 5. The return spread of BIG-GLOBAL pair

 Trading Signal Analysis with Pairs Trading Strategy 387

profit when the return spread crosses both the upper boundary and the lower
boundary. If the return spread does not cross over the other one boundary before
the end of the trading period, the profit gains or losses are calculated at the last
day of trading period.

Table 3. The average return of individual stocks and the return of five pairs from 2
January 2017 to 29 December 2017 (one-year trading period)

Pairs Stock A Average Return Stock B Average Return Pair Return

1 ADVANCE 0.100% INTOUCH 0.047% 2.379%
2 CBG 0.022% PTG −0.156% 14.267%
3 PTG −0.156% TKN −0.112% 11.242%
4 CBG 0.022% COM7 0.084% 10.859%
5 BIG −0.205% GLOBAL −0.005% 0.075%

In Table 3, the average returns of each stock are calculated by buying a stock
on the first day and selling it on the last day during the trading period, and then a
profit gain or loss is divided by the holding day of the trading period. In the one-
year trading period, the five pairs provide positive return and the pair return for
each pair is higher than the average return of stocks in its pair. The CBG-PTG
pairs provide the highest return of up to 14.267%. Therefore, our trading signal
for pair-trading strategy can generate excess return when compared with the
individual stocks. This indicates that the pair-trading strategy is better than
trading in individual stocks. The investors can employ the proposed trading
signal for pair-trading strategy to gain a better profit in their portfolios.

4 Conclusion
The objective of this study is to find trading signal for pair-trading strategy
in Stock Exchange of Thailand (SET) and the proposed pair-trading signal can
generate higher return when compared with trading in individual stocks. In pre-
vious study, the model captured with the volatility dynamics of the return spread
is employed to create a trading signal for pair-trading strategy. Therefore, we
propose the Markov-switching GARCH model to capture the volatility dynamics
of the return spread. The model keeps attention only on the conditional vari-
ance structure and focuses on two regimes which have changing variance in each
regime. The average of the two variances is used as trading signal for pair-trading
strategy. The study applies the pair-trading strategy to SET100 index based on
the most 100 liquid stocks. The pairs formation period is over 1 January 2016 to
12 December 2016 and the trading period is over 2 January 2017 to 29 December
2017.
The empirical results suggest the pair-trading signal of the Markov-switching
GARCH model generates positive return for all selected pairs and provides the
388 N. Namwong et al.

highest return of up to 14.267%. Moreover, the return from pair-trading strategy

is higher than the return from trading in individual stocks. The investors can
employ the proposed trading signal for pair-trading strategy to gain a better
profit from the SET100 index. For future study, the transaction costs and rolling
the pairs formation and the trading period should be included for analysis.

References
1. Bock, M., Mestel, R.: A regime-switching relative value arbitrage rule. In: Oper-
ations Research Proceedings 2008, pp. 9–14. Springer, Heidelberg (2009)
2. Bollerslev, T.: Generalized autoregressive conditional heteroskedasticity. J. Econ.
31(3), 307–327 (1986)
3. Chen, C.W., Chen, M., Chen, S.Y.: Pairs trading via three-regime threshold
autoregressive GARCH models. In: Modeling Dependence in Econometrics, pp.
127–140. Springer, Cham (2014)
4. Chen, C.W., Wang, Z., Sriboonchitta, S., Lee, S.: Pair trading based on quantile
forecasting of smooth transition GARCH models. N. Am. J. Econ. Financ. 39,
38–55 (2017)
5. Chodchuangnirun, B., Zhu, K., Yamaka, W.: Pairs trading via nonlinear autore-
gressive GARCH models. In: International Symposium on Integrated Uncertainty
in Knowledge Modelling and Decision Making, pp. 276–288. Springer, Cham,
March 2018
6. Do, B., Faff, R., Hamza, K.: A new approach to modeling and estimation for pairs
trading. In: Proceedings of 2006 Financial Management Association European
Conference, pp. 87–99, May 2006
7. Do, B., Faff, R.: Does simple pairs trading still work? Financ. Anal. J. 66(4),
83–95 (2010)
8. Elliott, R.J., Van Der Hoek, J., Malcolm, W.P.: Pairs trading. Quant. Financ.
5(3), 271–276 (2005)
9. Engle, R.F.: Autoregressive conditional heteroscedasticity with estimates of the
variance of United Kingdom inflation. Econ. J. Econ. Soc. 50, 987–1007 (1982)
10. Engle, R.F., Granger, C.W.: Co-integration and error correction: representation,
estimation, and testing. Econ. J. Econ. Soc. 55, 251–276 (1987)
11. Gatev, E., Goetzmann, W.N., Rouwenhorst, K.G.: Pairs trading: performance of
a relative-value arbitrage rule. Rev. Financ. Stud. 19(3), 797–827 (2006)
12. Haas, M., Mittnik, S., Paolella, M.S.: A new approach to Markov-switching
GARCH models. J. Financ. Econ. 2(4), 493–530 (2004)
13. Hamilton, J.D.: A new approach to the economic analysis of nonstationary time
series and the business cycle. Econ. J. Econ. Soc. 57, 357–384 (1989)
14. Hamilton, J.D.: Time Series Analysis, vol. 2, pp. 690–696. Princeton University
Press, Princeton (1994)
15. Vidyamurthy, G.: Pairs Trading: Quantitative Methods and Analysis, vol. 217.
John Wiley & Sons (2004)
16. Yang, J.W., Tsai, S.Y., Shyu, S.D., Chang, C.C.: Pairs trading: the performance
of a stochastic spread model with regime switching-evidence from the S&P 500.
Int. Rev. Econ. Financ. 43, 139–150 (2016)
17. Zhu, K., Yamaka, W., Sriboonchitta, S.: Pair trading rule with switching regres-
sion GARCH model. In: International Symposium on Integrated Uncertainty in
Knowledge Modelling and Decision Making, pp. 586–598. Springer, Cham, Novem-
ber 2016