Gold

ARBITRAGE IN GOLD SPOT (XAU/USD) & GOLD FUTURE
(COMEX) BY USING PAIR TRADING STRATEGY FROM

THRESHOLD CO-INTEGRATION MODEL
BY
MS. KRITTIKA THAMPHONGSRI
AN INDEPENDENT STUDY SUBMITTED IN PARTIAL

FULFILLMENT OF THE REQUIREMENTS FOR
THE DEGREE OF MASTER OF SCIENCE
PROGRAM IN FINANCE (INTERNATIONAL PROGRAM)
FACULTY OF COMMERCE AND ACCOUNTANCY
THAMMASAT UNIVERSITY
ACADEMIC YEAR 2018
COPYRIGHT OF THAMMASAT UNIVERSITY
Ref. code: 25616002042064LGM

ARBITRAGE IN GOLD SPOT (XAU/USD) & GOLD FUTURE
(COMEX) BY USING PAIR TRADING STRATEGY FROM
THRESHOLD CO-INTEGRATION MODEL
BY
MS. KRITTIKA THAMPHONGSRI
AN INDEPENDENT STUDY SUBMITTED IN PARTIAL

FULFILLMENT OF THE REQUIREMENTS FOR
THE DEGREE OF MASTER OF SCIENCE
PROGRAM IN FINANCE (INTERNATIONAL PROGRAM)
FACULTY OF COMMERCE AND ACCOUNTANCY
THAMMASAT UNIVERSITY
ACADEMIC YEAR 2018
COPYRIGHT OF THAMMASAT UNIVERSITY
Ref. code: 25616002042064LGM

(1)
Independent study title ARBITRAGE IN GOLD SPOT (XAU/USD) &

GOLD FUTURE (COMEX) BY USING PAIR
TRADING STRATEGY FROM THRESHOLD
CO-INTEGRATION MODEL
Author Ms. Krittika Thamphongsri
Degree Master of Science (Finance)
Major field/Faculty/University Master of Science Program in Finance
(International Program)
Faculty of Commerce and Accountancy
Thammasat University
Independent study advisor Associate Professor Tatre Jantarakolica, Ph.D.
Academic year 2018
ABSTRACT
Since the original concept of pair trading strategy that is choosing pair of
asset base on similar fundamental, could not be applied on this day because company
structural changes caused by M&A. Nowadays, pair trading strategy are used for pair
of future and spot that they have exactly same fundamental. This study applies pair
trading strategy with pair of gold spot and gold future in COMEX market that is global
equity traded by international investors. TVECM method applied base on the
assumption of asymmetric behavior of mispricing between two markets. In order to find
the performance of pair trading strategy by using TVECM as a trading signal, 1 minute,
5 minute and 30-minute frequency data during 1 August 2018 to 30 October 2018 are
collected and transaction cost would be included. Result shows that asymmetric
behavior occurs in every frequency series. The trading simulation result shows the best
arbitrage profit is 5-Minute Series. For the trading signal method, TVECM performs
better than traditional strategy (Two-Standard-deviation Rule).
Keywords: Threshold Co-Integration Model, Pair Trading, Two-Standard-Deviation

Rule, Asymmetric Behavior
Ref. code: 25616002042064LGM

(2)
ACKNOWLEDGEMENTS
Completion of my independent study would not have been possible by

without assistance of my advisor, Associate Professor Tatre Jantarakolica, Ph.D. who
kindly gave me a good recommendation and encouragement. I wish to express my
sincerely appreciation to my advisor for providing me an opportunity to do this topic
and for all his contribution. I also would like to thank to my committee, Assistant
Professor Chaiyuth Padungsaksawasdi, Ph.D., for many recommendations that is very
useful to improve my research.
This research idea is inspired by this paper “Thailand is Statistical arbitrage
in SET and TFEX: Pair Trading Strategy from Threshold Co-Integration Model” by
Mr. Surasak Choedpasuporn. That is one of the literature give me some inspiration to
study furthermore in this area. This research would not have been completed without
helping from Mr. Surasak Choedpasuporn, I would like to express my deeply
appreciation to him for his support about computer programming and many useful
recommendation during my research.
Finally, I also thank my friends and family for their help and understand
throughout my study at MIF.
MS. Krittika Thamphongsri
Ref. code: 25616002042064LGM

(3)
TABLE OF CONTENTS
Page
ABSTRACT (1)
ACKNOWLEDGEMENTS (2)
LIST OF TABLES (5)
LIST OF FIGURES (7)
LIST OF ABBREVIATIONS (8)
CHAPTER 1 INTRODUCTION 1
1.1 Statement of problems and motivations 1

1.2 Objectives of study and research questions 3
1.3 Research contribution 3
CHAPTER 2 REVIEW OF LITERATURE 4
2.1 Review of literature 4

2.2 Theoretical framework 5
2.2.1 Efficient Market Hypothesis 5
2.2.2 Cost of Carry 6
2.2.3 Threshold Co-Integration model 7
CHAPTER 3 RESEARCH METHODOLOGY 8
3.1 Data selection 8

3.2 Details of COMEX Gold future contract specification and 9
transaction cost
Ref. code: 25616002042064LGM

(4)
Page
3.3 Research process 9
3.3.1 Co-Integration method 9
3.3.2 Threshold Vector Error Correction Model 11
3.3.3 Determine the parameter (Training period) 12
3.3.4 Trading rule (Execute period) 12
3.3.5 Performance measurement 13
CHAPTER 4 EMPIRICAL RESULT 14
4.1 Unit root test and long-run relationship estimation 14

4.2 Co-Integration estimation 14
4.3 Short run dynamic estimation 15
4.4 Threshold Vector Error Correction Model Estimation 18
4.5 Time rolling and portfolio performance 23
4.5.1 Hansen-Seo Test 25
4.5.2 Threshold Vector Error Correction Model Estimation 26
4.5.3 Trading rule performance measurement 28
CHAPTER 5 CONCLUSION 29
REFERENCES 31
BIOGRAPHY 32
Ref. code: 25616002042064LGM

(5)
LIST OF TABLES
Tables Page
3.1 COMEX Gold future contract specification 9
3.2 Transaction cost 9
4.1 ADF test of gold spot and gold future of each frequency 14
4.2 Johansen tests for Co-Integration of gold spot and gold future of each 15
frequency
4.3 SBIC Criteria for gold spot and gold future in 30-minute series 15
4.4 Result of VECM of lags 4 for gold spot and gold future in 30-minute 16
series
series
series
4.10 Result from TVECM estimation for gold spot and gold future in 30- 18
minute series
minute series
minute series of observation 1 to 13,000
minute series of observation 13,001 to 26,000
Ref. code: 25616002042064LGM

(6)
Tables Page
4.19 Performance of TVECM strategy with different training and execute 24
period for 1-Minute Series
4.22 Percentage number of time roll for threshold behavior and non- 26
threshold behavior
4.23 HS Test result 26
4.24 Compare performance of each trading rule 28
Ref. code: 25616002042064LGM

(7)
LIST OF FIGURES
Figures Page
1.1 Concept of pair trading 2
3.1 Gold spot price and Gold future price of 1-minute series during 1 8
August 2018 to 30 October 2018
3.2 Demonstration of three regimes 12
3.3 Demonstration of pair trading rule 13
Ref. code: 25616002042064LGM

(8)
LIST OF ABBREVIATIONS
Symbols/Abbreviations Terms
TVECM Threshold Vector Error Correction Model
VECM Vector Error Correction Model
ADF Augmented Dickey–Fuller
Ref. code: 25616002042064LGM

1
CHAPTER 1
INTRODUCTION
1.1 Statement of problems and motivations
In the globalization of a financial market, it is a direct consequence of free

international capital movement among the country. That provides the benefit for Thai
investors to invest in the foreign assets to allocate their portfolio. At the same time, this
event leads to increase the risk and volatility in global equity investing. Following this
situation, hedging strategy would be taken an interest, to reduce their risk with the
optimum return. In this study is about pair trading strategy, which it can make a profit
with any market condition and avoid some form of market risk.
Since 1980, pair trading has been popular in terms of statistical arbitrage
finding the profit situation from pricing inefficiencies between two assets for short
periods of time. It is usually used in the hedge funds and investment bank. This strategy
is due to two key of successes. The first is a pair of asset forming and the second is
method to identify the trading signal. The concept of pair trading strategy is following
the law of one price, which the price of same asset in difference market should be equal.
Once the mispricing occur, arbitrageur can make a profit from this situation. In this
work, we focus on the future and underlying assets that the gap between spot and future
should be equal to cost of carry. If the market is inefficiency, the gap would not be
constant then, we can get the profit from buying a cheaper asset and selling an expensive
one. After mispricing occurs in short term, it will revert to the mean in long term (mean
reversion), it has mean that two assets are co-integrated and have a historical long-term
relationship.
This study uses the pair of gold spot and gold future (COMEX) that is
global equity traded by international investor. Gold price depends on many factors such
as gold’s demand and supply, interest rate, U.S dollar, inflation, policies and etc. That
is not only global economic factor, but the investor behavior in different country also
affect volatility of gold price. This is a chance for proprietor to get higher arbitrage
profit from this higher volatility and liquidity.
Ref. code: 25616002042064LGM

2
The trading is simulated by taking short position in outperforming one and

taking long position in underperforming one and then close both of position when it
reverts to the mean.
Figure 1.1 Concept of pair trading from Choedpauporn S., (2014)
The second key of success is about the criteria to open or close the position.
The common method is Minimum distance method, and in the past many researchers
have developed a new method to get a better performance of pair trading for example
Co-integration method, Augmented Dickey fuller test and granger causality test. The
research in Thailand, Choedpauporn S. (2014) and Songyoo K. (2013) presented
TVECM to use as a trading signal for pair of SET and TFEX because they found that
the relationship between these two assets is not linear. From this asymmetric
behavioral, threshold co-integration should be applied to divide the behavior of
mispricing into 3 regimes for each speed of adjustment. The result shows that arbitrage
opportunities exist by using TVECM superior to traditional method.
The main focusing is a TVECM method applied in the global asset (gold
spot and gold future) base on the assumption of asymmetry behavior of mispricing of
two markets. The aim of this research is to find a relationship of spot and future of
global commodity asset and study a combination of this pair with a TVECM method to
assess its performance. This paper expects that this combination can generate more
profit from the higher volatility of this global asset and TVECM as an appropriate
trading signal.
Ref. code: 25616002042064LGM

3
1.2 Objectives of study and research questions
Research objective
To study a pair of global commodity asset by using a pair trading strategy
with Threshold Co-Integration Model. Focus on finding an arbitrage opportunity and
performance of trading simulation from pair of gold spot and gold future.
Research question
1. Find Gold future and gold spot return relationship in long and short term.
2. Find Arbitrage opportunities exist in the gold market by using the pair trading
strategy with TVECM as a trading signal.
1.3 Research contribution
To find the arbitrage opportunity reflected in inefficiency of the two

markets that are gold spot and gold future (COMEX) by using pair trading strategy.
This study uses the pair of gold spot and gold future (COMEX) that is global equity
traded by international investors. Therefore, these two assets have high daily volume
that will provide more liquidity. It is worldwide trading that provides more asymmetric
of information that lead to react in a different way among international investors. For
the method to identify trading signal in this study is TVECM as an appropriate method
to describe an asymmetric of return from negative and positive information effect.
This study has contributions to proprietary trader, institutional, hedge fund
traders and arbitrager to apply this combination of this asset pair with a TVECM
method to assess its performance. This study expects this combination can generate
more profit from the higher volatility of this global asset and TVECM as an appropriate
trading signal.
Ref. code: 25616002042064LGM

4
CHAPTER 2
REVIEW OF LITERATURE
2.1 Review of literature
The theories of pair trading strategy are Efficient Market Hypothesis,

arbitrage opportunity from mispricing and the law of one price. The efficient market
hypothesis (EMH) by Fama(1965) explains about three forms of market which are weak
form, semi-strong form and strong form. If the market efficient, the current price will
reflect by the stock information. If this concept is applied with future pricing, the future
price will be equal to spot price plus cost of carry in case of efficiency market that there
provides no arbitrage opportunity.
For imperfect market, Liu, Wong, An and Zhang (2014) this paper shows
asymmetric characteristic of return in Chinese and US commodity by using the
threshold framework. The result shows that asymmetric Threshold stochastic volatility
(THSV) model is better fit when compare to the symmetric model to forecast the future
price and volatility. This Threshold model can examine good and bad news asymmetry
on mean by divided into two regimes for good news and bad news. They also found
that future return has more response in negative news more than positive news. At last,
this asymmetry will revert into long-term equilibrium.
Therefore, in an imperfect market with asymmetric characteristic provides
the arbitrage opportunity between future market and spot market. In previous studies,
most studies in the Thailand focus on pair of SET and SET50 Future. Songyoo (2013)
and Choedpasuporn (2014), this both paper examine the performance of the TVECM
pair trading strategy by using pairs in Thailand’s SET and TFEX Markets. The result
shows that TVECM as a trading signal is superior to traditional pair trading strategy
because the application of threshold is suitable for asymmetric market condition.
Chunhachinda and Fuangkasem(2014), this literatures provides the
evidence on arbitrage opportunities in gold futures market investigate cross-market
arbitrage opportunities by using 5 minutes intraday data. Three major gold futures
markets: COMEX, MCX and TOCOM, and gold spot are examined. They investigated
Ref. code: 25616002042064LGM

5
two type of arbitrage opportunity; 1. futures-spot and 2. futures-futures by using cost

of carry model in order to obtain the upper and lower bounds of gold future price. The
result shows that they found the arbitrage profit in both type of arbitrage opportunity.
However, the cost of carry model is initially under assumption of perfect market and
non-arbitrage argument, which is might not completely fit the imperfect market case.
2.2 Theoretical framework
2.2.1 Efficient Market Hypothesis

The efficient market hypothesis (EMH) by Fama(1965) explains
about three forms of market which are weak form, semi-strong form and strong form.
If the market efficient, the current price will reflect by the stock information. If the gap
between forward and spot price is not constant, it means that the market is inefficient
and there have an arbitrage opportunity.
Fully Efficient Market
If the market is fully efficient, it is mean that the information arrives

symmetric among the market. Once the investor in two markets receive the new
information symmetric, the asset price will be reflected by the same in two markets.
Ft,T = Et(St) = St(the cost of carrying asset over time)
Base on fully efficient market, spot price of a commodity should

equal the futures price on the delivery date (coat of carry is vanished at the delivery
date). Before the delivery date, the prices of futures are determined by the spot price
and cost of carry following the equation.
Future Price = Spot Price + Cost of Carry
Ref. code: 25616002042064LGM

6
Therefore, if the gap between spot and future should be constant over
the time that means there has no an arbitrage opportunity in fully efficient market
condition.
Not Fully Efficient Market
If the market is not fully efficient, it is mean that the information arrives
asymmetric among the market. Therefore, the investors in two markets consider
different information, and then the trader will act in different way. It creates the
mispricing in short term, leading to exist the arbitrage opportunity.
Short run arbitrage

ft,T-ST(1+bt,T)>0
Arbitrager will take a ‘Short’ position in future asset (ft,T) and ‘Long’
position in underlying asset(ST).
ft,T-ST(1+bt,T)<0
Arbitrager will take a ‘Short’ position in underlying asset (ST) and ‘Long’
position in future asset (ft,T).
The arbitrager will take the riskless profit like this until the different of two
price is equal to zero or it’s revert to the mean then the arbitrager will close the position.
2.2.2 Cost of Carry

Cost of carry is the model to explain how to price index futures. According
to this model, theoretical future price depends on spot price of commodity (S0) and cost
of carrying the asset from now until the futures contract matures (C). The cost is
associate with four categories, which are transportation cost, insurance cost, storage
cost, and financing cost. In case of gold, the storage, transportation, and insurance costs
are very small. The largest cost is the financing cost to purchase the asset.
Ref. code: 25616002042064LGM

7
F0,T = 𝑆0 (1 + 𝐶)
Following Cost of Carry Model for gold is based on Edward M. Riley III (2014).
𝑇
F0, T = 𝑆0 (1 + 𝑟)365
F0, T is expected future price, 𝑆0 is the current value of spot price, r is the
risk-free interest rate on an investment. Theoretical pricing of gold futures depends on
three parameters. The first is the spot price, which is influenced by many factors such
as gold’s demand and supply, interest rate, U.S dollar, inflation, policies and etc.
Second parameter is about the risk free rate from the investment, which is referred to
Treasury bond rate. The last parameter is about the number of day until maturity (T).
However, Treasury rates(r) is important to determine the theoretical gold
price but it is not fluctuating much in short term. Therefore, in order to apply cost of
carry model to find a mispricing for intraday trade is not appropriate.
2.2.3 Threshold Co-Integration model

The short-term behavior of trader following the inefficiency market
evidence shows the non-linear lead-lag relationship Songyoo K., (2013). This non-
linear adjustment between two markets is according to the different new information
arrival and then the cost will approach to this information such as transaction cost.
Threshold co-integration model is applied to analyze nonlinear behavior.
Ref. code: 25616002042064LGM

8
CHAPTER 3
RESEARCH METHODOLOGY
3.1 Data selection
This study uses the pair of gold spot and gold future (COMEX). Pair trading
strategy will be used to perform arbitrage opportunity by using 1, 5 and 30 minutes
intraday price between pairs of assets (gold spot and gold future). Gold Future (GC) is
from Thomson Reuters Eikon and Gold Spot (XAU/USD) data is from Meta trader.
Gold future series in this study is GCZ8 and study period is three months during 1
August 2018 to 30 October 2018. GCZ8 has expired on 27 December 2018.
Figure 3.1 Gold spot price and Gold future price of 1-minute series during 1 August
2018 to 30 October 2018
Ref. code: 25616002042064LGM

9
3.2 Details of COMEX Gold future contract specification and transaction cost
Table 3.1 COMEX Gold future contract specification
Underlying asset Fine gold bar

Contract size 100 troy oz.
Quality specification 0.995 purity
Price quote USD per troy oz.
Tick size USD 0.1 per troy oz.
Settlement type Physical delivery
Table 3.2 Transaction cost
Asset type Transaction cost Reference

Gold spot 0.01 % XM
COMEX Gold future 12 $/Contract Phillip Securities
3.3 Research process
3.3.1 Co-Integration method

Augmented Dickey-Fuller (ADF), Test: Unit Root Test
The collected data is time series data. To analyze the time series data, it
must has a stationary property. It assumes that mean, variance and covariance have no
change depend on time.
To test the stationary property, Unit root test is conducted by using
Augmented Dickey-Fuller (ADF) following this equation.
∆xt= µ1+ γ xt-1+ µ2t + ∑∞

𝑖=1 𝛽𝑖 ∆𝑥𝑡−𝑗 + ϵt
Ref. code: 25616002042064LGM

10
xt represents the series of log future price or log spot price

µ, γ and 𝛽 are the coefficient
t is time
ϵt is error term
Hypothesis
H0: γ=0 (xt is non stationary)
Ha: γ≠0 (xt is stationary)
If the result is fail reject null hypothesis (γ=0), it means that xt has unit root
which is non-stationary. In this case, we need to test at first difference | (1) further. If
null hypothesis is rejected γ=0, xt is stationary (has no unit root) or | (0).
Co-Integration test
This is the technique to test the long-term relationship between two

variables. If two variables have a Co Integration property, it means that it has a long-
term relationship. To test Co-Integration, it uses the method of Engel and Granger.
Error Correction Model is following formula.
∆xt= α0+ αxZt-1+ Σ α1j∆xt-j + Σ α2j∆yt-j + ϵxt

∆yt= α0+ byZt-1+ Σ b1j∆yt-j + Σ b2j∆yt-j + ϵyt
Zt-1= xt-1- βyt-1
∆xt = xt - xt-1
∆yt = yt - yt-1
Zt-1 is the error term
β is Co-Integration coefficient
Ref. code: 25616002042064LGM

11
Hypothesis
H0: rank (∏) =0
Ha: rank (∏) ≠0
∏ is the Co-Integration matrix
If the result is fail reject null hypothesis, there is no Co-Integration. If null

hypothesis is rejected, there is Co-Integration.
3.3.2 Threshold Vector Error Correction Model

This paper Songyoo K., (2013) mentioned that the dynamic
adjustment between two markets is non-linear process, it will occur when the prices are
deviate from the equilibrium described by asymmetric information, which TVECM is
developed to capture this asymmetric market behavior. If the Time-series data has a co-
Integration property. Next step is testing a threshold or nonlinear property of time
series. Therefore, Grid search is used to find all possible pair of β (Co-Integrating
vector) and γ (Threshold value) within constraint, and then test the pair to find the
optimum β and γ by using SBIC selection criteria.
The three regimes base on different of adjustment process are
constructed by Threshold Vector Error Correction technique. Three regimes are
following model.
Regime1, when (xt-1- β yt-1) ≤ γa

∆xt= α0+ α1xZt-1+ Σ α11j∆xt-j + Σ α12j∆yt-j + ϵxt
∆yt= α0+ b1yZt-1+ Σ b11j∆yt-j + Σ b12j∆yt-j + ϵyt
Regime2, when γa ≤ (xt-1- β yt-1) ≤ γb

Regime1, when γb < (xt-1- β yt-1)

Ref. code: 25616002042064LGM

12
3.3.3 Determine the parameter (Training period)

This step is the first step of time rolling procedure, which is to find
the parameters needed to construct the regime. After TVECM and grid search method,
provide the optimum β and γ. It will be used to construct three regimes of investor
decision to use as trading signal. In different regime, speed of adjustment might be
different. Under no adjustment regime is no arbitrage band or regime 2. The arbitrager
will invest only the mispricing occur which is under regime 1 and 3.
Figure 3.2 Demonstration of three regimes from Choedpauporn S., (2014)
3.3.4 Trading rule (Execute period)

This step is the second step of time rolling procedure, which will
execute the trading rule by using the parameter from training period.
Trading rule is based on the reference paper (Choedpasuporn. S,
(2014). The open the position is taken when the graph go into the area of arbitrage band
(1st and 3rd regime) and Close the position is taken When the graph reverts to the 2nd
regime (no arbitrage band).
Ref. code: 25616002042064LGM

13
Figure 3.3 Demonstration of pair trading rule proposed from Choedpasuporn. S., (2014)
3.3.5 Performance measurement

The return of open position is price at short position minus price at
open position. The return from close position equal to price at short position minus
price at open position. The time rolling is conducted for performance measurement as
mention above which divided by three periods. First is training period, second is
execute trading rule by using parameters from training period, third is moving training
period forward in the same length of execution period. Do this procedure until the end
of sample period.
Ref. code: 25616002042064LGM

14
CHAPTER 4
EMPIRICAL RESULT
4.1 Unit root test and long-run relationship estimation
To analyze the time series data, stationary property would be checked by

using Augmented Dickey-Fuller (ADF). The result in every series is reject null
hypothesis which means every series have no unit root or stationary. Both log of gold
spot series and log of gold future series are I (0).
Table 4.1 ADF test of gold spot and gold future of each frequency
Frequency Series ADF statistics Critical value Conclusion
1 Minute ln Gold spot -171.706 -3.410 Stationary
ln Gold future -174.436 -2.860 Stationary
4.2 Co-Integration estimation
Base on fully efficient hypothesis, spot and future price will revert to the
long term mean. To test the long term relationship between two assets, we applied
Johansen test. If two assets have a Co-Integration property, it means that it has a long-
term relationship. The result in Table 4.2, shows that null hypothesis of every pair series
is rejected, there is Co-Integration property.
Ref. code: 25616002042064LGM

15
Table 4.2 Johansen tests for Co-Integration of gold spot and gold future of each
frequency
Frequency Price series Johansen tests Critical value Conclusion

Statistic
1 Minute ln Gold spot 591.1582 15.41 Co-Integration
ln Gold future
ln Gold future
ln Gold future
4.3 Short run dynamic estimation
VECM will be used to estimate the short run dynamic. First, SBIC criteria
will be used to find the optimal lags by choosing the lowest number of SBIC. That
optimal number of lags will be used to estimate TVECM later.
Table 4.3 SBIC Criteria for gold spot and gold future in 30-minute series
No.of Lags SBIC Criteria
1 -89976.29
2 -89953.79
3 -89972.93
4 -89997.92(Optimal Lag)
5 -89964.16
Ref. code: 25616002042064LGM

16
Table 4.4 Result of VECM of lags 4 for gold spot and gold future in 30-minute series
Gold spot Coefficient Standard error
Correction Term -0.3121 0.0195
Gold future Coefficient Standard error
Correction Term 0.0173 0.1092
1 -492576.2
2 -492673.3
3 -492882.1
4 -493126.4
5 -493406(Optimal Lag)
Ref. code: 25616002042064LGM

17
1 -1914963
2 -1914969(Optimal Lag)
3 -1914917
4 -1914842
5 -1914792
Following VECM model, the most important is a sign of speed of

adjustment or coefficient of correction term that tells responsible for long-term
equilibrium.
The VECM estimation result. For gold spot equation, the coefficient is high
magnitude of negative sign represents downward adjustment. For gold future equation,
the coefficient is not differences from zero, it represents no adjustment. Therefore, it is
the convergence process by downward of spot adjustment and no future adjustment.
For gold commodity system, it implied that when both price deviates from,
spot price will move downward direction to converge to future price. For variance, spot
price has lower variance than future price that means spot price has a stronger trend
than the future price.
Ref. code: 25616002042064LGM

18
4.4 Threshold Vector Error Correction Model Estimation
Since both of future and spot price series have Co-Integration property or
long-term relationship. Next step is to check threshold behavior by using Hansen-Seo.
For pair of gold spot and gold future in 1 minute, 5 minutes and 30 minutes
price series. SBIC criteria will be used to find the optimal lags by choosing the lowest
number of SBIC. That optimal number of lags will be used to estimate TVECM later.
1 -89784.81
2 -90235
3* -90412.93(Optimal Lag)
4 -90170.07
5 -90091.78
Table 4.10 Result from TVECM estimation for gold spot and gold future in 30-minute series
Parameters Value
Co-Integration Vector (1, - 0.9995884 )
Number of Lags 3
Threshold Values 9.866792e-05 , 0.001629885
Coefficient ECT gold spot (Upper regime) -0.3530
Coefficient ECT gold future (Upper regime) 0.1258
Coefficient ECT gold spot (middle regime) -0.4259
Coefficient ECT gold future (middle regime) 0.0542
Coefficient ECT gold spot (Lower regime) -0.0933
Coefficient ECT gold future (Lower regime) 0.2222
Percentage of Observations in each regime (Upper, Middle & lower) 80.8% 16.9% 2.3%
Sample size 2876
Ref. code: 25616002042064LGM

19
1 -492382.5
2 -492418.3
3 -492580.1
4 -492746.1
Table 4.12 Result from TVECM estimation for gold spot and gold future in 5-minute
series
Parameters Value
Number of Lags 5
Threshold Values -0.001168842, 0.002370551
Coefficient ECT gold future (Upper regime) -0.0088
Coefficient ECT gold future (Lower regime) -0.0144
Percentage of Observations in each regime (Upper, Middle & lower) 22.9%, 66.8%, 10.3%
Sample size 15624
Ref. code: 25616002042064LGM

20
2 -446133.1
3 -446017
4 -445886.8
5 -445765.7
For 1-minute series, total observation is 55,683. Due to too high frequency
data, the computer programming will be crashed if it is run at once. Therefore, we must
divide all sample (total 55,683 observation) into a small sample and test separately that
the result show in Table 18, Table 19, Table 20 and Table 21.
series of observation 1 to 13,000
Parameters Value
Number of Lags 1
Percentage of Observations in each regime (Upper, Middle & lower) 11.3%, 43.9%, 44.8%
Sample size 13000
Ref. code: 25616002042064LGM

21
series of observation 13,001 to 26,000
Parameters Value
Number of Lags 1
Coefficient ECT gold future (middle regime) -0.0047
Sample size 13000
Parameters Value
Number of Lags 1
Threshold Values -0.001376559, 8.772296e-05
Ref. code: 25616002042064LGM

22
Table 4.16 Continued
Parameters Value
Sample size 13000
Parameters Value
Number of Lags 1
Threshold Values -0.001085098, -0.0002108487
Coefficient ECT gold future (Upper regime) -0.0044
Sample size 13000
Ref. code: 25616002042064LGM

23
Parameters Value
Number of Lags 1
Threshold Values -0.003216164,-0.002234495
Coefficient ECT gold spot (middle regime) 0.0606
Percentage of Observations in each regime (Upper, Middle & lower) 5.4% 5% 89.6%
Sample size 3683
4.5 Time rolling and portfolio performance
When time series data is analyzed, the assumption of estimated parameter

is constant over the time. However, realistic market situation has changed and it is not
reasonable to assume that the parameter constants. To overcome this problem, time
rolling technique is applied to estimate the performance more realistic. Concept of time
rolling is dividing the sample into sub-samples to estimate the parameter and portfolio
returns would be calculated by moving the window forward.
First, in order to perform rolling regression, we have to know which size of
subsample is proper to use. Base on the reference paper (Choedpasuporn. S, (2014)
found that the training and execute period that generates the highest portfolio return are
5 trading days and 1 trading day respectively. It means that they used 5 trading days to
estimate parameters and repeated doing this when 1-day pass. This research is for
Thailand stock spot (SET) and future (TFEX).
Ref. code: 25616002042064LGM

24
For this study, we use global assets that are continuously traded 23 hours
per day. Therefore, we used trial and error method by shorten the duration of training
period and execute period. Result shows the best return for 1-minute series is 1-day
trading period and half-day execute period, and the best return for 5 and 30 minutes is
half day trading period and half day execute period. Then, we compare portfolio
performance with period from reference paper (5 days trading period and 1 day execute
period). For unite conversion is based on 21 trading hours per day, that is the trading
hours after cleaning data. We fix half day of execute period because it is practical to
use in the reality that there have night and day session, and vary only training period.
Table 4.19 Performance of TVECM strategy with different training and execute period
for 1-Minute Series
5 days : 1 day 1 day : half day half day : half

Training : Execute period day
6300:1260 1260:630 630:630

observations observations observations
No. of Time Roll 40 87 88

1-Minute Series
No. of 592 1,674 1,674
transactions
Gross Profit 20,642 60,100 59,791
Transaction cost 14,181 40,146 40,171
Net Profit 6,461 19,954 19,620
Ref. code: 25616002042064LGM

25
Table 4.20 Performance of TVECM strategy with different training and execute period
for 5-Minute Series
5 days : 1 day 1 day : half day half day : half day

Training : Execute period
1260:252 252:126 126:126

5-Minute Series
No. of 592 762 1,070
transactions
Gross Profit 52,123 56,413 76,748
Net Profit 37,889 38,113 51,054
Table 4.21 Performance of TVECM strategy with different training and execute
period for 30-Minute Series.
5 days : 1 day 1 day : half day half day : half day

Training : Execute period
210:42 42:21 21:21

30-Minute Series
No. of 192 376 472
transactions
Gross Profit 16,199 32,672 41,446
Net Profit 11,599 23,630 30,105
4.5.1 Hansen-Seo Test

Before performing the trading simulation, we employ TVECM,
Hansen-Seo Test is used to check whether it is threshold behavior or nonlinear model.
Hansen-Seo Test will be tested for every frequency series that result show in Table 4.22.
Ref. code: 25616002042064LGM

26
Table 4.22 Percentage number of time roll for threshold behavior and non-threshold
behavior
1-Minute Series 5-Minute Series 30-Minute Series
% No. of time roll 48.6% 26.6% 1.1%

P-value <0.05
(Threshold behavior)
% No. of time roll 51.4% 73.4% 98.9%

P-value >0.05
The result shows that asymmetric behavior highly occurs in high

frequency series and decreases in low frequency series because it might be that
asymmetric information arrived among the market decreases as time passes.
Table 4.23 HS Test result
95% Critical Value P-Value Conclusion
1-Minute Series 22.47 0.03 Reject
5-Minute Series 21.57 0 Reject
30-Minute Series 19.55 0.025 Reject
The total sample is used to test threshold behavior, which the result
is reject null hypothesis. Strong evidence of threshold nonlinearity is found in every
frequency series.
4.5.2 Threshold Vector Error Correction Model Estimation

After Hansen-Seo is employed to test threshold or nonlinear property
of time series. Threshold Vector Error Correction Model (TVECM) is developed to
capture this asymmetric market behavior of the adjustment process and then separate it
into different regime with different speed of adjustment. This step, TVECM method
will be used to estimate adjustment parameter (coefficient of error correction term), β
(Co-Integrating vector) and γ (Threshold value).
Ref. code: 25616002042064LGM

27
The estimation process has three steps. First, SBIC selection criteria
is used to find the optimum lag term. We varied number of lag term and then select the
lowest value of SBIC. Second step is estimate β and threshold value. Then β is used to
calculate the error correction term (zt-1=st-1 - βf t-1). Threshold value is use to construct
a three regimes which are one of no arbitrage band and two of arbitrage profitability
band. Third step is simulating the portfolio when size of error correction term more or
less than threshold value ((xt-1- β yt-1) ≤ γa or γb < (xt-1- β yt-1)), we will open the position.
Close the position is taken when size of error correction term is in between two
threshold value (γa ≤ (xt-1- β yt-1) ≤ γb ) or graph reverts to the 2nd regime (no arbitrage
band).
Regime1, when (xt-1- β yt-1) ≤ γa

Regime2, when γa ≤ (xt-1- β yt-1) ≤ γb

Regime1, when γb < (xt-1- β yt-1)

Ref. code: 25616002042064LGM

28
4.5.3 Trading rule performance measurement
Table 4.24 Compare performance of each trading rule
Trading Rule 1-Minute Series 5-Minute Series 30-Minute

Series
1 day : half day half day : half half day : half

Training : Execute period day day
1260 / 630 126 / 126 21 / 21

TVECM No. of 1,674 1,070 472

transactions
Gross Profit 60,100 76,748 41,446
Net Profit($) 19,954 51,054 30,105
Traditional Pair No. of 516 432 80

Trading transactions
(Two-Standard-
deviation Rule) Gross Profit 25,518 40,108 12,211
Net Profit($) 13,146 29,732 10,286
For TVECM trading strategy, 5-Minute Series generates the best net
profit of $ 51,054, when compares with traditional pair trading.
Ref. code: 25616002042064LGM

29
CHAPTER 5
CONCLUSION
This study uses the pair of gold spot and gold future (COMEX) that is
global equity traded by international investors. The main focusing is a TVECM method
applied in the global asset (gold spot and gold future) base on assumption of asymmetry
behavior of mispricing of two markets. The aim of this research is to find relationship
of spot and future of global commodity asset and study a combination this pair with a
TVECM method to assess its performance.
This study uses the pair of gold spot and gold future (COMEX). Pair trading
strategy with TVECM as a trading signal would be used to perform arbitrage
opportunity by using 1, 5 and 30 minutes frequency.
The long run relationship, the result show that every price series have Co-
Integration property. Co-Integrating vector, estimated by Johansen MLE approach to
study characteristic when both assets converge to the long-term mean. For gold
commodity system, convergence pattern is downward of spot adjustment and no future
adjustment.
For threshold behavior, it is estimated by using Hansen-Seo Test to check
whether this is threshold behavior or nonlinear model. The result shows that asymmetric
behavior occurs this gold market.
The result of trading simulation, the arbitrage profit occurs by using this
pair of asset. TVECM method shows superior result than traditional method. The best
frequency series is 5-Minute Series that generates the best net profit of $ 51,054.
However, this study uses the pair of gold spot and gold future (COMEX)
that is global equity traded by international investors. Therefore, these two assets have
high daily volume that will provide more liquidity. It is worldwide trading that we
expect more asymmetric of information lead to react in a different way among
international investors. We expect that it can generate more arbitrage profit from the
theses reason of this global asset. To prove this, it must to compare global gold assets
with Thai gold assets. However, Thai gold spot is not given in intraday data, then it
made us lack of information to simulate trading by using pair Thai gold. Further research
Ref. code: 25616002042064LGM

30
can use pair of Thai gold future with gold asset in other country to compare the trading
performance from gold asset in different market.
Ref. code: 25616002042064LGM

31
REFERENCES
1. Choedpauporn S. (2014). Thailand is Statistical arbitrage in SET and TFEX:

Pair Trading Strategy from Threshold Co-Integration Model. College of management,
Mahidol University
2. Songyoo K. (2013). Technical Trading Strategy in Spot and Future Markets:
Arbitrage Signaling. Faculty of Economics, Thammasat University
3. Martin, M., Kofman, P. and Vorst, C.F. (1998). A threshold error-correction
model for intraday future and index returns. Journal of applied econometrics, 13, 245-
263.
4. Edward M. Riley III (2014). The Cost-of-carry model and volatility: an analysis
of gold futures contracts pricing. University of Richmond
5. Qingfu L, Ieokhou W, Yunbi A, Jinqing Z (2014). Asymmetric Information and
Volatility Forecasting in Commodity Futures Markets. Pacific-Basin Finance Journal
26 (2014) 79–97
6. Upinder S. D, Dennis J. L, Taiji W(1997). Volatility, information, and double
versus walrasian auction pricing in US and Japanese futures markets. Journal of
Banking & Finance 21 (1997) 1045-1061
7. Pornchai C, Rapeesorn F(2014). Arbitrage Opportunities in Major Gold Futures
Markets: Evidence from High Frequency Data. Faculty of Commerce and
Accountancy, Thammasat University
Ref. code: 25616002042064LGM

32
BIOGRAPHY
Name Ms. Krittika Thamphongsri

Date of birth March27, 1992
Educational attainment 2018 Master in Finance, Thammasat University
Work position Credit Analyst
Bangkok Bank
Ref. code: 25616002042064LGM

Gold

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Gold

Uploaded by

Copyright:

Available Formats

ARBITRAGE IN GOLD SPOT (XAU/USD) & GOLD FUTURE

(COMEX) BY USING PAIR TRADING STRATEGY FROM

MS. KRITTIKA THAMPHONGSRI

AN INDEPENDENT STUDY SUBMITTED IN PARTIAL

Ref. code: 25616002042064LGM

MS. KRITTIKA THAMPHONGSRI

AN INDEPENDENT STUDY SUBMITTED IN PARTIAL

Ref. code: 25616002042064LGM

Independent study title ARBITRAGE IN GOLD SPOT (XAU/USD) &

Keywords: Threshold Co-Integration Model, Pair Trading, Two-Standard-Deviation

Ref. code: 25616002042064LGM

Completion of my independent study would not have been possible by

MS. Krittika Thamphongsri

Ref. code: 25616002042064LGM

LIST OF TABLES (5)

LIST OF FIGURES (7)

LIST OF ABBREVIATIONS (8)

1.1 Statement of problems and motivations 1

CHAPTER 2 REVIEW OF LITERATURE 4

2.1 Review of literature 4

CHAPTER 3 RESEARCH METHODOLOGY 8

3.1 Data selection 8

Ref. code: 25616002042064LGM

CHAPTER 4 EMPIRICAL RESULT 14

4.1 Unit root test and long-run relationship estimation 14

Ref. code: 25616002042064LGM

Ref. code: 25616002042064LGM

Ref. code: 25616002042064LGM

Ref. code: 25616002042064LGM

Ref. code: 25616002042064LGM

1.1 Statement of problems and motivations

In the globalization of a financial market, it is a direct consequence of free

Ref. code: 25616002042064LGM

The trading is simulated by taking short position in outperforming one and

Figure 1.1 Concept of pair trading from Choedpauporn S., (2014)

Ref. code: 25616002042064LGM

1.2 Objectives of study and research questions

1.3 Research contribution

To find the arbitrage opportunity reflected in inefficiency of the two

Ref. code: 25616002042064LGM

2.1 Review of literature

The theories of pair trading strategy are Efficient Market Hypothesis,

Ref. code: 25616002042064LGM

two type of arbitrage opportunity; 1. futures-spot and 2. futures-futures by using cost

2.2 Theoretical framework

2.2.1 Efficient Market Hypothesis

Fully Efficient Market

If the market is fully efficient, it is mean that the information arrives

Ft,T = Et(St) = St(the cost of carrying asset over time)

Base on fully efficient market, spot price of a commodity should

Future Price = Spot Price + Cost of Carry

Ref. code: 25616002042064LGM

Not Fully Efficient Market

Short run arbitrage

2.2.2 Cost of Carry

Ref. code: 25616002042064LGM

2.2.3 Threshold Co-Integration model

Ref. code: 25616002042064LGM

3.1 Data selection

Ref. code: 25616002042064LGM

Table 3.1 COMEX Gold future contract specification

Underlying asset Fine gold bar