Trading Strategies Used by Hedge Funds

Trading Strategies
Used by Hedge Funds

Submission for Award of Doctor of Philosophy
Timofei Bogomolov
Associate Professor John van der Hoek

(Principal Supervisor)
Professor Robert Elliott

(Associate Supervisor)
Professor Petko Kalev

(Associate Supervisor)
School of Mathematics and Statistics
Division of Information Technology,

Engineering and the Environment
University of South Australia
2012
Contents
List of Figures iv
List of Tables vi
Table of Abbreviations viii
Abstract ix
Declaration xi
Acknowledgements xii
1 Thesis Overview 1
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.1 Hedge funds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.2 Pairs trading . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Thesis Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2.1 Study 1. Performance evaluation of the pairs trading strategies . 7
1.2.2 Study 2. Pairs trading based on statistical variability of the
spread process . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.2.3 Study 3. Arbitrage strategies between listed Asia-Pacific stocks
and their NYSE ADRs when there is no overlap of trading . . . 9
1.3 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2 Research Background 12
2.1 Cointegration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.1.1 Cointegration and the CAPM . . . . . . . . . . . . . . . . . . . 17
2.1.2 Tests for cointegration . . . . . . . . . . . . . . . . . . . . . . . 19
2.1.3 Practical application of cointegration . . . . . . . . . . . . . . . 20
2.2 Mean-reverting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.3 Market Data and Software Used . . . . . . . . . . . . . . . . . . . . . . 25
3 Study 1. Performance Evaluation of the Pairs Trading Strategies 27

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.2.1 Methods used . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.2.2 Distance method . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.2.3 Cointegration approach . . . . . . . . . . . . . . . . . . . . . . . 32
3.2.4 Stochastic spread process method . . . . . . . . . . . . . . . . . 33
3.3 Computation of returns . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
i
3.3.1 Excess returns computation . . . . . . . . . . . . . . . . . . . . 36
3.3.2 Transaction costs . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.3.3 Data sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.3.4 Testing periods and benchmarks . . . . . . . . . . . . . . . . . . 42
3.3.5 Data snooping bias . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.4 Trading rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.4.1 Distance method . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.4.2 Cointegration method . . . . . . . . . . . . . . . . . . . . . . . 46
3.4.3 Stochastic spread method . . . . . . . . . . . . . . . . . . . . . 47
3.5 Empirical results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
4 Study 2. Pairs Trading Based on Statistical Variability of the Spread

Process 66
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4.2 Method of renko and kagi constructions . . . . . . . . . . . . . . . . . 70
4.2.1 Renko construction . . . . . . . . . . . . . . . . . . . . . . . . . 70
4.2.2 Kagi construction . . . . . . . . . . . . . . . . . . . . . . . . . . 72
4.2.3 Some properties of renko and kagi constructions . . . . . . . . . 74
4.2.4 Trading strategies . . . . . . . . . . . . . . . . . . . . . . . . . . 75
4.3 More on the renko and kagi constructions . . . . . . . . . . . . . . . . 78
4.3.1 Properties of H -constructions on the Ornstein–Uhlenbeck process 78
4.3.2 H -construction on the discrete process . . . . . . . . . . . . . . 80
4.4 Pairs trading by the contrarian H -strategy . . . . . . . . . . . . . . . . 84
4.4.1 Data sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
4.4.2 Stocks pre-selection . . . . . . . . . . . . . . . . . . . . . . . . . 85
4.4.3 Pairs formation . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
4.4.4 Trading rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
4.4.5 Excess return calculation and transaction costs . . . . . . . . . 89
4.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
4.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
5 Study 3. Arbitrage Strategies Between Listed Asian-Pacific Stocks

and Their NYSE ADRs When There Is No Overlap of Trading 101
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
5.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
5.2.1 Review of arbitrage for cross-listings . . . . . . . . . . . . . . . 103
5.2.2 Review of arbitrage for non-overlapping
traded cross-listings . . . . . . . . . . . . . . . . . . . . . . . . . 105
5.3 Research questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
5.4 Research Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
5.4.1 Sample and data description . . . . . . . . . . . . . . . . . . . . 113
5.4.2 Trading strategy . . . . . . . . . . . . . . . . . . . . . . . . . . 113
5.4.3 Excess returns calculation . . . . . . . . . . . . . . . . . . . . . 115
5.5 Empirical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
5.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
ii
6 Conclusion 125
6.1 Contributions to theory and knowledge . . . . . . . . . . . . . . . . . . 125
6.2 Research findings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
6.3 Practical implications . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
6.4 Limitations and future research . . . . . . . . . . . . . . . . . . . . . . 130
A Crossing Time for the Ornstein–Uhlenbeck Process 131

A.1 Most likely time to hit the mean for the Ornstein–Uhlenbeck process . 131
A.2 Estimation for optimal trigger level to start trading for the Ornstein–
Uhlenbeck process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
A.3 Numerical simulation for the first passage time of AR(1) . . . . . . . . 136
B Proofs of the Theorem and Axillary Lemmas 138
Bibliography 156
iii
List of Figures
3.1 Historical performance of the distance method of pairs trading on the

Australian market for top 5 pairs — accumulated returns without rein-
vestment. Axes are time and returns. . . . . . . . . . . . . . . . . . . . 53
Australian market for top 20 pairs — accumulated returns without rein-
vestment. Axes are time and returns. . . . . . . . . . . . . . . . . . . . 54
S&P 500 companies for top 5 pairs — accumulated returns without
reinvestment. Axes are time and returns. . . . . . . . . . . . . . . . . . 55
S&P 500 companies for top 20 pairs — accumulated returns without
3.5 Historical performance of the cointegration method of pairs trading on
the Australian market for top 5 pairs — accumulated returns without
the Australian market for top 20 pairs — accumulated returns without
the S&P 500 companies for top 5 pairs — accumulated returns without
the S&P 500 companies for top 20 pairs — accumulated returns without
3.9 Historical performance of the stochastic spread process method of pairs
trading on the Australian market for top 5 pairs — accumulated returns
without reinvestment. Axes are time and returns. . . . . . . . . . . . . 61
trading on the Australian market for top 20 pairs — accumulated returns
trading on the S&P 500 companies for top 5 pairs — accumulated returns
trading on the S&P 500 companies for top 20 pairs — accumulated
returns without reinvestment. Axes are time and returns. . . . . . . . . 64
4.1 Log prices spread process between two major Australian banks — Com-
monwealth (CBA) and Westpac (WBC). . . . . . . . . . . . . . . . . . 69
iv
4.2 Renko chart . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.3 Renko construction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
4.4 Kagi construction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.5 Strategy historical performance on the Australian market data set for
top 5 pairs portfolio before and after transaction cost — accumulated
4.6 Strategy historical performance on the Australian market data set for
4.7 Strategy historical performance on the S&P 500 data set for top 5 pairs
portfolio before and after transaction cost — accumulated returns with-
out reinvestment. Axes are time and returns. . . . . . . . . . . . . . . . 98
4.8 Strategy historical performance on the S&P 500 data set for top 20
pairs portfolio before and after transaction cost — accumulated returns
4.9 Strategy historical performance on the S&P 400 MidCap data set for
4.10 Strategy historical performance on the S&P 400 MidCap data set for
4.11 Strategy historical performance on the S&P 600 SmallCap data set for
4.12 Strategy historical performance on the S&P 600 SmallCap data set for
5.1 Example of a Group 1 company . . . . . . . . . . . . . . . . . . . . . . 122

5.4 Quick evolution of ICICI Bank Ltd from Group 1 to Group 3. . . . . . 123
A.1 Starting point to trade . . . . . . . . . . . . . . . . . . . . . . . . . . . 135

A.2 Empirical PDF for the first passage time . . . . . . . . . . . . . . . . . 137
A.3 Empirical CDF for the first passage time . . . . . . . . . . . . . . . . . 137
v
List of Tables
3.1 Statistics on data sets used in the Australian and US markets . . . . . 41

3.2 Historical performance of the benchmark indexes . . . . . . . . . . . . 50
3.3 Monthly excess returns statistics and risk measures for trading top 5
pairs by the distance method of pairs trading on the Australian market
before and after transaction cost. . . . . . . . . . . . . . . . . . . . . . 53
pairs by the distance method of pairs trading on the Australian market
pairs by the distance method of pairs trading on the S&P 500 companies
pairs by the distance method of pairs trading on the S&P 500 companies
pairs by the cointegration method of pairs trading on the Australian
market before and after transaction cost. . . . . . . . . . . . . . . . . . 57
pairs by the cointegration method of pairs trading on the Australian
market before and after transaction cost. . . . . . . . . . . . . . . . . 58
pairs by the cointegration method of pairs trading on the S&P 500 com-
panies before and after transaction cost. . . . . . . . . . . . . . . . . . 59
pairs by the cointegration method of pairs trading on the S&P 500 com-
panies before and after transaction cost. . . . . . . . . . . . . . . . . . 60
pairs by the stochastic spread process method of pairs trading on the
Australian market before and after transaction cost. . . . . . . . . . . 61
Australian market before and after transaction cost. . . . . . . . . . . 62
S&P 500 companies before and after transaction cost. . . . . . . . . . 63
S&P 500 companies before and after transaction cost. . . . . . . . . . 64
vi
3.15 Individual trade statistics based on trading top 5 and top 20 pairs for
the distance, cointegration and stochastic spread process methods . . . 65
4.1 Statistics on data sets used in the Australian and US markets . . . . . 84

4.2 Historical performance of the benchmark indexes . . . . . . . . . . . . 92
4.3 Monthly excess returns of the kagi pairs trading strategy with and with-
out transaction costs (0.10% per one trade per stock) . . . . . . . . . . 95
4.4 Risk measures based on the monthly excess returns before transaction
costs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
5.1 Descriptive statistics of the visible return spreads s1 (t), coefficient of

lag one auto-regression of the spread s2 (t) and coefficient of predictive
power of s1 (t) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
5.2 Monthly and individual trades excess returns of the strategy with con-
version between ADR and local shares . . . . . . . . . . . . . . . . . . 118
5.3 Monthly and individual trades excess returns of the pairs trading style
strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
vii
Table of Abbreviations
ADF Augmented Dickey–Fuller test
ADR American Depositary Receipt
APT Arbitrage Pricing Theory
AR Autoregressive process
ASIC Australian Securities & Investments Commission
ASX Australian Stock Exchange
AUM Assets Under Management
CAPM Capital Asset Pricing Model
DF Dickey–Fuller test
FX Foreign Exchange
GARCH Generalized Autoregressive Conditional Heteroskedasticity model
GFC Global Financial Crisis
LOP Law of One Price
MA Moving Averages
NASDAQ National Association of Securities Dealers Automated Quotations
NYSE New York Stock Exchange
OLS Ordinary Least Squares
OU Ornstein–Uhlenbeck process
PCA Principal Component Analysis
RAP Risk Adjusted Performance
ROI Return On Investment
SIRCA Securities Industry Research Centre of Asia-Pacific
viii
Abstract
Hedge funds are an important part of the modern financial system. As it follows from
the name, hedge funds try to apply market neutral trading strategies, such as pairs
trading and arbitrage, and generate stable income, which should be independent of
general market fluctuations. This research uses a quantitative approach to study sev-
eral different market neutral strategies.
This thesis consists of three self-contained studies. Each study has been presented at
conferences and submitted for journal publication as independent papers.
In the first study is an empirical research, where I review three methods of pairs
trading cited in academic literature (Gatev et al., 2006; Vidyamurthy, 2004; Elliott
et al., 2005). The trading rules have been defined for each method as well as a testing
framework which allows one to compare different methods. The performance of each
strategy is assessed using Australian and US stock exchanges daily data covering the
last 15 years including the time of the recent Global Financial Crisis. While all three
methods demonstrate a highly desirable property of having a low correlation with mar-
ket benchmarks, returns on these three strategies are low and not consistent over time.
This suggests the need for improvements or alternative method, which is developed in
the second study.
In the second study, I review a probabilistic–statistical approach for time series anal-
ysis introduced by Pastukhov (2005) for Brownian motion and extend it to the case
ix
of Ornstein–Uhlenbeck and AR(1) processes. I give a theorem about H-volatility of
the Ornstein–Uhlenbeck process. I then build on this theory, and propose a new non-
parametric method of pairs trading based on the volatility of the spread process. The
method does not try to find a mean of the spread process and trade towards it like
other methods of pairs trading. It constantly stays in the market alternating ’short’
and ’long’ on the spread with trading signals defined by the analysis of variability of
the spread. The empirical tests on the US and Australian market data demonstrate the
superiority of this new method in terms of returns and returns volatilities compared
to the existing methods of pairs trading examined in the first study. In addition, it
provides groundwork for a discussion of the effect of different levels of market efficiency
on these trading strategies.
In the third study, I investigate the problem of an arbitrage style trading between
the markets without overlap in trading hours. These markets exist in the US and
Asia-Pacific countries — Australia, Japan, New Zealand, China (Hong Kong), In-
donesia, India, Philippines, South Korea, Taiwan. I suggest a method to segregate
these markets into three groups with different levels of market efficiency and segmenta-
tion/integration. I then test two approaches for trading between markets without time
overlap on real market data: (a) trading with conversion between ADRs and shares,
and (b) pairs trading style without conversion. Results show that the proposed method
allows the identification of moderately segmented markets and generates economically
significant profit at a reasonably low level of risk from trading on those markets.
This thesis extends the theory of auto-regressive and mean-reverting processes and
introduces a new property of the Ornstein–Uhlenbeck process. In financial terms the
overall contribution of this thesis is in providing support for the current financial the-
ories (CAPM, APT). While the thesis explores the possibility of generating abnormal
returns through trading based on the analysis of historical prices, that is clearly demon-
strated that abnormal returns are the result of exploiting market inefficiencies.
x
Declaration
I declare that:
• this thesis presents work carried out by myself and does not incorporate without
acknowledgment any material previously submitted for a degree or diploma in
any university;
• to the best of my knowledge it does not contain any materials previously published
or written by another person except where due reference is made in the text; and
all substantive contributions by others to the work presented, including jointly
authored publications, is clearly acknowledged.
Timofei Bogomolov
xi
Acknowledgements
I would like to thank my supervisors Associate Professor John van der Hoek, Professor
Robert Elliott and Professor Petko Kalev for their leadership and guidance.
I would like to acknowledge the support of the Centre for Applied Financial Studies at
UniSA and Securities Industry Research Centre of Asia-Pacific in providing the data
for this thesis.
I also thank the organisers and participants of the 2nd Finance and Corporate Gov-
ernance Conference in Melbourne (April 2011) and Quantitative Methods in Finance
Conference in Sydney (December 2011) for the opportunity to present my research and
gain valuable feedback provided on my presentation.
Final thank you goes to my wife Svetlana for her patience, love and support.
xii
xiii
Chapter 1
Thesis Overview
1.1 Introduction
Most investment funds and individual investors use popular financial market theo-
ries like the Capital Asset Pricing Model (CAPM) (Sharpe, 1964; Treynor, 1962) and
Arbitrage Pricing Theory (APT) (Fama and French, 1993) to construct investment
portfolios. As a result, the investors passively depend on general market movements
and expected returns using these models. The recent Global Financial Crisis (GFC)
has demonstrated a high level of dependance from markets.
For example, many investment and superannuation funds lost up to 40% of their value.
Alternative approaches to investment, that are unaffected by general market move-
ments, attract constant interest from investment professionals. Such approaches allow
the diversification of risks of the market portfolio, and generate profits regardless of
the market up or down movements.
1.1.1 Hedge funds
Hedge Funds (or Absolute Return Funds) offer such alternatives and often declare
themselves as neutral to market movements. Nowadays, the hedge funds industry is
an important part of the modern financial system. Even the most conservative institu-
1
tional investors, such as superannuation funds, allocate about 10% of the total capital
to hedge funds.
Alfred Jones is known as a founder of the first hedge fund in 1949. He kept a bal-
anced portfolio of long and short positions to eliminate the effect of general market
movements and gained from the relative pricing of the assets in his portfolio (Ineichen
and Silberstein, 2008). His fund was really hedged, but was not named a ‘hedge fund’
— that name was introduced in 1966 by Carol Loomis in his article about Alfred
Jones (Loomis, 1966).
It is impossible to know what types of strategies are employed by each individual hedge
fund due to the proprietary character of any information related to the hedge funds
industry. However, 138 out of 278 Australian hedge funds included in the report of
Fund Monitors.com self-proclaim using some kind of market neutral strategies such as
pairs trading or statistical arbitrage.
While not all hedge funds deliver what they promise (Brown and Goetzmann, 2003;
Patton, 2009), the general interest in the hedge fund industry is strong. Fifteen years
ago, total assets under management (AUM) for the hedge funds industry was around
US$450 billion. The most recent estimation of AUM is between US$1.8 and US$4
trillion (Ineichen and Silberstein, 2008).
1.1.2 Pairs trading
One popular trading strategy used by hedge funds is called pairs trading. It is based
on a simple idea of simultaneously taking short and long positions in different assets.
Many authors attribute its development to Nunzio Tartaglia in the middle of the 1980s.
The trading strategy used by Alfred Jones was similar to pairs trading but without a
strong quantitative approach to stock selection.
2
Pairs trading provides two main benefits:
1. The strategy is market-neutral. This means that it is immune to market fluctu-

ations.
2. The strategy is self-financing (dollar-neutral ). This means that the investors do

not need to invest their own money.
The last statement is not exactly true as one still needs to make some initial invest-
ments to open and maintain a margin account and pay transaction costs. However,
the sums involved are relatively small for individual traders and even smaller for insti-
tutional investors. At this stage we can ignore these expenses and assume the strategy
to be a zero investment, that is dollar-neutral.
Market neutrality is appealing for superannuation funds, insurance companies and

other risk averse investors, who take capital protection as their first priority. On the
other hand, dollar neutrality makes pairs trading a naturally leveraged product that is
desirable by investors who target high returns even at the price of higher risk through
the use of leverage.
The clear disadvantage of the pairs trading is the higher transaction costs, as the
investor trades a number of different assets each time the investor opens or closes po-
sitions on the pair.
The general description of a pairs trading strategy is simple: find two stocks that have
historically moved together, open positions when they deviate too far from each other
and close them out when stocks converge together (Gatev et al., 2006). To create a
pairs trading strategy investors should answer three questions:
1. What does it mean to ‘move together’ ? These assets should form the long-short
portfolio.
3
2. How ‘far’ should these assets deviate before applying the strategy? This will
determine when to open positions.
3. What does it mean that the assets ‘converge together’ and what does one do if
this never happens? This will determine when to close positions.
Do et al. (2006) review different approaches to pairs trading described in academic

literature, and separate them into three groups:
1. the distance method used by Gatev et al. (2006);
2. the cointegration method described by Vidyamurthy (2004); and
3. the stochastic spread method proposed by Elliott et al. (2005).
The distance method is the only method of pairs trading which has been widely tested
on the real market data. The first objective of this thesis is testing and evaluating all
these strategies with the careful consideration of the transaction costs.
All of the above mentioned approaches can be considered as special cases of one general
trading strategy in which the spread between different real assets is regarded as a sin-
gle synthetic asset. The theory of cointegration (Engle and Granger, 1987; Alexander,
2001) provides an explanation to the nature of a possible relationship between different
assets and ways to construct the spread (Alexander and Dimitriu, 2005). The same
statistical methods used for the analysis of real asset time series can also be applied to
the spread process.
Existing methods of pairs trading exploit the fundamental idea about price equilibrium
or the spread mean-reversion. Positions should be taken if there is a deviation from an
equilibrium in the hope that a spread (or prices) would return to their long-run his-
torical equilibrium. Positions can be unwound when the spread returns to equilibrium
values or prices cross each other.
4
However, in practice it is difficult to estimate the true parameters of the spread pro-
cess — such as a mean, variance and a speed of mean-reversion. Even if the assets are
cointegrated, a very long history of data may be required to make good estimations.
The required time series might be longer than the available data, so any estimation
made from an existing data sample may not be reliable.
The second objective of this thesis is the development and testing of an alternative
method of pairs trading — trading of the spread process using trading strategies gen-
erally applied to ‘normal’ assets. This approach might have a number of advantages
compared with the trading of normal assets:
• the investor can expect the spread process to be more stable than the prices of
individual assets involved due to the strong relations between cointegrated assets;
• the spread process is invisible for all other investors, so it cannot be affected by
behavioural factors affecting the pricing — for example, support/resistance levels
on rounding numbers or overreaction;
• the natural benefits of pairs trading being market-neutral and dollar-neutral.
A trading style closely related to pairs trading is arbitrage — simultaneously buying

and selling the same asset on different markets. An example is provided by the shares
of the cross-listed companies traded in different countries. Profit from classical arbi-
trage is generated from the differences of the asset prices on different markets. It is
guaranteed and risk free. Numerous studies have investigated the existence and prof-
itability of arbitrage. However, the outcome of the strategy becomes uncertain if it is
impossible to execute both trades at the same time. On one hand, this is a perfect pair
to trade, which makes trading quite appealing. On other hand, there is a potentially
unlimited risk of holding an unhedged position during the time gap between opening
positions on different markets.
The third objective of this thesis is to investigate to what extent arbitrage and pairs
5
trading strategies can be applied to the markets without a time overlap in trading
hours. Is it possible to control risks? What market conditions can facilitate this kind
of trading?
6
1.2 Thesis Structure
In this thesis I have the following research objectives:
1. carry out wide scale empirical testing of the pairs trading strategies described in
the existing academic literature on the real market data;
2. provide improvements and theoretical justification for the existing methods of

pairs trading;
3. develop an alternative method for pairs trading, evaluate its performance on the
market data and compare the results with other existing methods;
4. propose and evaluate an arbitrage-style (or pairs trading style) trading strategy
between markets without overlap in trading hours.
To address the above research objectives, I present three studies independently analysing
different approaches and aspects of the phenomena of market neutral trading and, in
particular, of pairs trading.
1.2.1 Study 1. Performance evaluation of the pairs trading
strategies
The first study is empirical research. It reviews the three most cited methods of pairs
trading: the distance (Gatev et al., 2006), cointegration (Vidyamurthy, 2004) and
stochastic spread method (Elliott et al., 2005). The last two methods have never been
tested on real market data.
The trading rules have been defined for each method as well as a testing framework
which allows one to compare different methods. The performance of each strategy is
empirically assessed using daily data covering the last 15 years of their history on the
Australian and US stock exchanges. The testing period includes the GFC. This allows
to evaluate the true degree of market neutrality of all methods of pairs trading.
7
The stochastic spread method is the only parametric method of pairs trading. Using
the distribution of the first crossing time for the Ornstein–Uhlenbeck process, I provide
a theoretical justification for the level of deviation from the mean which should be used
as a trigger to start trading (Appendix A). Previous research takes that level arbitrarily.
The discussion on the methods of excess returns calculation sheds light on the problem
of leverage in pairs trading which has not been clearly addressed before. Pairs trading
is a naturally leveraged product, so to be able to compare performance and the risk
level of different trading strategies, they need to have the same level of leverage. In
most cases, this means using de-leveraged returns and risk measures. This information
is also useful in comparing potential returns with transaction costs.
All three methods demonstrate statistically significant average excess returns from 5%
to 12% per year in most of the tests for the US and Australian markets. However, those
returns are not consistent over time. Transaction costs have a very negative effect on
the performance of the strategies — after accounting for the transaction costs, the
profitability reduces by 15–40%. The testing results demonstrate limited practical
value of these strategies in their current form, suggesting the need for substantial
improvements.
1.2.2 Study 2. Pairs trading based on statistical variability of
the spread process
This study proposes a new nonparametric approach to pairs trading based on renko
and kagi constructions which originated from Japanese charting indicators and were
introduced to academic studies by Pastukhov (2005).
This research considers a spread process between log prices of two assets but it does
not try to find a long-run mean of the process and trade towards it like other methods
of pairs trading. Instead, it exploits statistical information about the variability of the
8
tradable process. The only assumption needed is that the statistical properties of the
spread process volatility remain reasonably constant. That is a much milder assump-
tion than an assumption about constant mean of the spread process widely used in the
pairs trading literature.
The study provides a detailed review of new techniques to measure the process volatil-
ity proposed by Pastukhov (2005), and an application of these new measures to trading.
I then extend Pastukhov’s research by considering its application to different types of
continuous and discrete processes. Next, I formulate and prove a theorem about renko
and kagi constructions for the case of the Ornstein–Uhlenbeck type spread process,
which provides a theoretical justification of the profitability of the proposed method of
pairs trading for such processes.
The tests of the method on the daily market data of US and Australian stock exchanges,
show statistically significant average excess returns ranging from 1.4% to 3.6% per
month. The proposed strategy has almost zero correlation with market. However, it
is not absolutely market neutral as the level of profitability depends on the level of
market volatility. The study concludes with a discussion about market efficiency as
a possible reason for the different levels of profitability on different markets and in
different market conditions.
1.2.3 Study 3. Arbitrage strategies between listed Asia-Pacific
stocks and their NYSE ADRs when there is no overlap
of trading
A large amount of academic literature is dedicated to research of arbitrage between

cross-listed assets, in particular shares listed on the stock exchange of the country of
company’s origin and American Depositary Receipts (ADRs) on the same company in
the New York stock exchange (NYSE). These two securities are almost perfect substi-
tutes for each other and, as a result, make an ideal pair for arbitrage or pairs trading.
9
So, for the case of shares and corresponding ADRs, the stage of pairs selection is ig-
nored as a natural pair is formed.
One of the most important requirements of pairs trading and arbitrage strategies execu-
tion is a simultaneous trading on both legs of the pair. This prompts the development
of high speed execution and high frequency trading. However, there is a large class of
assets which are naturally perfect pairs but cannot be traded simultaneously. They are
the shares of Asia-Pacific companies and their ADRs on NYSE.
Asia-Pacific markets (Australia, Hong Kong, India, Indonesia, Japan, New Zealand,
Philippines, Singapore, South Korea, Taiwan) do not have a time overlap in trading
hours with NYSE. The investor bears a significant risk while holding just one leg of
a pair during the time gap between markets. A high degree of uncertainty involved
in the trading between Asia-Pacific and New York markets was the reason why this
problem did not receive more attention from academia.
In this research, I test two methods of trading between markets without time overlap
and demonstrate that this trading can be profitable for markets with a moderate level
of segmentation. I propose a method for spread process analysis to segregate pairs
share-ADR (or countries) into three groups with different levels of market segmenta-
tion/integration based on the parameters estimated from their historical prices.
Group 1 (Australia, Japan) and Group 3 (India, South Korea, Taiwan) should be
avoided for trading for different reasons, which are discussed later in this study. The
decision not to engage in trading can be made in advance based on results of the anal-
ysis of historical prices proposed in this research.
Group 2 (New Zealand, Hong Kong, Indonesia, Philippines) provides consistent and
economically significant monthly return in excess of 2% after transaction costs with
10
relatively low level of risk. Annualised Sharpe ratios for most pairs are above 2. Corre-
lation with market indexes, which are used as a proxy for general market performance,
is negligibly small.
The remainder of this thesis proceeds as follows: Chapter 2 gives an introduction to the
theoretical background for all methods of pairs trading. Chapters 3, 4 and 5 present
three independent studies of different aspects of market neutral trading strategies.
Chapter 6 provides conclusions, discusses research limitations, and provides directions
for future research. Theoretical results that are to be used in Chapters 3 and 4 have
been placed in Appendices A and B.
1.3 Contributions
This research focuses on the performance evaluation of the existing methods of pairs
trading, on the construction of the new approaches to identify assets for pairs forma-
tion, and on developing new rules for pairs trading.
This research provides a bridge between financial mathematics, statistical theory and
the practical applications of this theory to the financial markets. It will help the in-
vestment industry to obtain a better understanding of the benefits and shortfalls of
different pairs trading strategies and the risks involved. It also demonstrates that prof-
its generated from pairs trading are a result of exploiting market inefficiencies. While
exploring opportunities for absolute returns, this study does not violate the major the-
ories of financial markets.
This research also contributes to the theory of auto-regressive processes, mean-reverting

processes and the study of the Ornstein–Uhlenbeck process by investigating new prop-
erties of those processes.
11
Chapter 2
Research Background
This chapter reviews the theory which unifies all three studies presented in this thesis.
I review the cointegration theory, its connection with the Capital Asset Pricing Model,
and the theory of mean-reverting processes. I then describe the data sets used for
testing in all three studies.
2.1 Cointegration
Cointegration is a property which two or more time series can possess. In brief, a
collection of non-stationary time series is cointegrated if a linear combination of these
time series is weakly stationary.
Definition. A time series {xt ; t = 0, 1, 2, . . . } is said to be weakly stationary (also

called covariance stationarity or wide stationarity) if its expectation and autocovariance
do not depend on time t (Alexander, 2001, pg. 317). More explicitly:
• E[xt ] is a constant
• Cov[xt , xt−s ] depends only on the lag s
• if we take s = 0, then the covariance is just the variance — written V[xt ], and it
has the same value for each t.
12
Definition. The first difference of a time series {xt } is the series of changes {∆xt }
where.
∆xt := xt − xt−1
Definition. A time series is integrated of order n, written yt ∼ I(n), if its stochastic

part is non-stationary but it becomes stationary after differencing a minimum of n
times (Alexander, 2001, pg. 322).
An I(0) process is already stationary. The first differencing of an I(1) process is sta-
tionary. An example of an I(1) process is a random walk model:
yt = yt−1 + εt ,
where {εt } is independent and identically distributed (iid) ∼ N (0, 1), or we could just
assume that {εt } be stationary.
The first difference of the process {yt } is
4yt = yt − yt−1 = εt ∼ I(0).
The process {yt } becomes stationary after the first differencing, so it is I(1).
It has been widely assumed in the efficient markets theory that log prices are random
walk, so the log price process is I(1) and then the returns process is I(0).
Normally, a linear combination of a number of I(1) time series is again I(1). However,
for some I(1) processes a linear combination of them could be I(0). This phenomena
was observed and described by Granger (1981), Granger and Weiss (1983) and Engle
and Granger (1987). They called this cointegration.
Definition. A set of I(1) time series is cointegrated if there is a linear combination

of these time series that is stationary (Alexander, 2001, pg. 349)).
13
So, if we have two time series {xt }, {yt } ∼ I(1) for which
yt − βxt = c + εt , (2.1)
where c and β are constants, and {εt } ∼ I(0), we see that {xt } and {yt } are cointegrated.
The proposed parameters c and β can be estimated by ordinary least square (OLS)
regression of {yt } on {xt }. They are the intercept and slope of the regression line of
{yt } on {xt }.
yt = c + βxt + εt . (2.2)
We can also write (2.1) in a more general form for n time series
β1 y1,t + β2 y2,t + . . . + βn yn,t = c + εt , (2.3)
where βi = {β1 , β2 , β3 , . . . , βn } is the cointegration vector of the set of time series

{yi,t }, i = 1, 2, ..., n.
A collection of n cointegrated time series may have up to (n − 1) cointegration vectors.

Cov(yt ,xt )
The cointegration vector for the two time series is (1, −β), where β = Var(xt )
is a
cointegration coefficient.
Real life examples of cointegration occur widely in economics. Fuel prices at petrol
stations in different locations are cointegrated, as are the relationships between con-
sumptions and income or short-term and long-term interest rates (Alexander, 2001;
Vidyamurthy, 2004).
Cointegrated time series have long-run relationships or common stochastic trends. Con-
sider time series data for petrol prices {xt } and {yt } at two different petrol retail outlets.
14
Obviously, the petrol price depends on the oil price process {ωt }
xt = ωt + εxt
yt = α ωt + εyt ,
where {εxt } and {εyt } could be iid noise processes ∼ I(0).
So a spread between {xt } and {yt } could be
αxt − yt = (αωt + α εxt ) − (α ωt + εyt ) = α εxt − εyt .
The common stochastic trend {ωt } is eliminated and the resulting linear combination
of two stochastic processes {xt } and {yt } is a stationary process.
With times series being cointegrated, it means that a spread process (or a linear com-
bination) is stationary. In real world scenarios, it can be expected that such a spread
process, while not stationary, may be at least relatively stable — that is, one where the
process parameters are changing slowly. One can trade the (portfolio) spread process
using an approach proposed by Elliott et al. (2005). An example of a trading strategy
based on cointegration has been provided by Vidyamurthy (2004).
We can consider cointegrated time series for prices, log prices or continuously com-
pounded (accumulated) returns. Vidyamurthy (2004) uses share log prices (this is
virtually the same as considering accumulated returns). Therefore several log price
time series being cointegrated is the same as the accumulated returns being cointe-
grated.
This can be illustrated as follows. Consider two time series of prices {Ptx } and {Pty }.
15
The price Pt (of either) at any point in time t is
Pt = Pt−1 ert ,
where rt is return over the period [t − 1, t];
t
!
X
Pt = P0 exp ri .
i=1
Hence the log price is

t
X
log Pt = log P0 + ri .
i=1
Assuming cointegrated log prices, we have the following relation
log Pty = c + β log Ptx + εt ,
where {εt } ∼ I(0). Therefore
εt = log Pty − β log Ptx − c

t t
!
X X
= log P0y + riy − β log P0x + rix −c
i=1 i=1
" t t
#
X X
= [log P0y − β log P0x − c] + riy − β rix ,
i=1 i=1
where rix and riy are returns on stock x and y.
The first square bracket is a constant and does not effect the dynamics of the stochastic
process {εt }. One can re-write the above formula as a spread process between two
continuously compounded returns {xt } and {yt } of the assets {Ptx } and {Pty }.
εt = c + yt − βxt ,
Pt y Pt x
where yt = i=1 ri , xt = i=1 ri and {εt } ∼ I(0).
16
A more rigorous presentation for a general form of the above proposition is provided by
Galenko et al. (2007), who prove that i time series of log prices (i ≥ 2) are cointegrated
Pn i i
if the spread process of continuously compounded returns εt = i=1 β yt has the
following properties:
E[εt ] = 0,
∞
X
Var(εt ) = −2 Cov[εt , εt−p ] ,
p=1
∞
X
p Cov[εt , εt−p ] < ∞ .
p=1
The only assumption needed is that {yti ; i = 1, 2, . . . , n} are I(1). The process {εt } is
stationary with mean 0 and some variance σ 2 < ∞. While it need not be normal, we
will often consider the case that {εt } is iid Gaussian N (0, σ 2 ).
2.1.1 Cointegration and the CAPM
Cointegration has a direct link to the Capital Asset Pricing Model (CAPM) proposed by
Sharpe (1964) and Treynor (1962), based on the works of Harry Markowitz. According
to the CAPM, the excess rate of return r of an asset is given by
r = βrm + α, (2.4)
where rm is the excess return of the market portfolio, β is a systematic or market risk
and α is unsystematic or asset-related risk. The excess returns are taken as a difference
between the real returns and the risk-free interest rate. This will simplify my presen-
tation.
Consider two assets with excess returns rA and rB :
r A = β A r m + αA (2.5)
17
r B = β B r m + αB . (2.6)
Continuously compounded or accumulated returns for each asset are
t
X t
X t
X
riA = β A rim + αiA (2.7)
i=1 i=1 i=1
t
X t
X t
X
riB = β B rim + αiB . (2.8)
i=1 i=1 i=1
Using equation (2.7) one can obtain
t Pt A
− ti=1 αiA
P
i=1 ri
X
rim = (2.9)
i=1
βA
and insert this into equation (2.8)
t t t
! t
X βB X X X
riB = A riA − αiA + αiB . (2.10)
i=1
β i=1 i=1 i=1
Hence,
t t t t
X βB X A X B βB X A
riB − A r = αi − A α . (2.11)
i=1
β i=1 i i=1
β i=1 i
Taking the accumulated excess returns of assets A and B as RtA and RtB and coefficient
βB
γ= βA
we get the following equation:
t
X t
X
RtB − γRtA = αiB −γ αiA = εt , (2.12)
i=1 i=1
where {εt } represents a stochastic process derived from a relationship between two
series of assets-related risk factors. The process {εt } is independent of the market and
this makes it a subject of interest (or exploited) for market-neutral trading strategies.
If the process {εt } is stationary it can be loosely said that assets A and B are thus
cointegrated. This demonstrates that cointegration and pairs trading based on cointe-
gration do not violate the fundamental theories of asset pricing.
18
2.1.2 Tests for cointegration
If there are only two time series considered, the most used and recommended test for
cointegration is the so called Engle–Granger method. It uses a two-step approach:
1. Estimate OLS regression on the I(1) data: regress {xt } onto {yt } or {yt } onto
{xt }. In the case of two time series it does not matter which one is selected as
dependant. As a result, the following model is obtained for predicted time series
{yt∗ }
yt∗ = a + b xt ,
where a and b are constants.
2. Apply a stationarity test (unit-root test) to the residuals from that regression
εt = yt − yt∗ = yt − (a + b xt ).
These tests are the Dickey-Fuller test (DF), augmented Dickey-Fuller test (ADF),
Durbin-Hausmann test and the Schmidt-Phillips test.
If {εt } is stationary, then {xt } and {yt } are cointegrated.
If there are more than two time series, the Engle-Granger method is not the best ap-
proach. It is unable to find all possible cointegration vectors and results depend on
which variable is chosen to be the dependant one. In this case a more appropriate
method of testing is the so called Johansen test. It is based on the eigenvalues of a
stochastic matrix and is similar to principal component analysis (PCA) (Johansen and
Juselius, 1990). Procedures for panel cointegration can also be used (Banerjee, 1999).
19
2.1.3 Practical application of cointegration
If two or more stocks are cointegrated then they have a long-run relationship. This
can be exploited for pairs trading. In other words, the statistics of the spread process
{εt } of the long-short portfolio is known.
εt = c + yt − βxt , (2.13)
where {yt } represents the long part of the portfolio and {βxt } the short part. The
spread process {εt } is stationary with mean 0 and some finite variance σ 2 .
This spread process can be traded like a normal asset:
1. If the spread is higher than an upper selected threshold, one can sell the spread
— sell the long part {yt } and buy the short part {βxt } of the portfolio. This is
brought about by taking position (1, −β) in the stocks. This will only be dollar
neutral if β = 1.
If the spread is lower than a lower selected threshold, buy the spread — buy
the long part {yt } and sell the short part {βxt } of the portfolio, and proceed as
above.
As the spread process is stationary there is a strong likelihood that it will move
towards its mean rather than away from it. We can then proceed in various ways.
2. An open position can be closed when the spread crosses 0 for the first time.
3. Alternatively, an open position can be closed when the opposite trading signal is
generated. For example, if one has a short position on spread, one should wait
until the signal to buy the spread is generated and simultaneously close the short
position on spread and open the long one. In this case, one constantly stays in
the market and keeps open positions.
Assuming that the spread process is mean-reverting, the Elliott et al. (2005) approach
can be used to approximate parameters of a process and trade in a similar way as
20
before, but with consideration for holding time and spread deviation (selecting thresh-
olds) estimated based on the parameters of the spread process mean-reversion.
Most research on pairs trading takes the cointegration theory framework. Gatev et al.
(2006) do not do any tests for cointegration and do not use cointegration for pairs for-
mation but admitted that they worked in the spirit of the cointegrated prices literature.
Summarising,
• only cointegrated series should be used for pairs trading, and
• the cointegration vector provides weights for each time series inside the long-short
portfolio used for pairs trading.
Cointegration is usually attributed to the long-run econometric relationship. In prac-

tice, there is the issue of how long is a long-run relationship? On one hand, considering
the full history knowledge of all stocks, it is unlikely that one could find any cointe-
grated stocks. On the other hand, if one uses too short a history, the estimation of the
spread process parameters could be just a result of randomness and far away from the
true values.
If we consider a finite investment horizon, it is reasonable to analyse only the last n

observations. Gatev et al. (2006) use the last 12 months of history for the 6 months
investment horizon. Elliott et al. (2005) take only n last observations and test if the
spread process has stable parameters over that time. If it has, then they assume that
those parameters could stay stable for some time. This is exactly the same assump-
tion that we take testing cointegration based on the last n observations. If the spread
process has constant mean and variance — being stationary on the history of the last
n observations, then one might assume that the parameters do not change too quickly.
This can be assumed because the cointegration vector is known to be ‘super-consistent’
(Alexander, 2001).
21
Classical dollar neutral pairs trading for two stocks only uses the cointegration vector
(1, −1). It might look like an ambiguous approach, but it does make sense from a prac-
tical viewpoint. If the weight of one asset in the pair is much larger than the other,
it means that we assume much higher volatility in the returns of the first asset and
reduce an exposure to that asset. In the real market data it is a very rare situation that
some stocks consistently have much higher volatility (several times) than the others.
Stock returns are quite homogenous. Increasing or decreasing the weight means the
investor becomes overexposed to one or another stock, which increases the market risk
and compromises one of the major benefits of pairs trading — market neutrality.
Preliminary empirical testing shows that the cointegration vector, as well as the pa-
rameters of the spread process, change over time. This means the spread process is
not stationary in terms of the strict classical definition. Nevertheless, these parameters
change slowly, so groups of approximately cointegrated shares could be suitable for
pairs trading.
2.2 Mean-reverting
A stochastic process is mean-reverting if it regularly returns to its long-run equilibrium

or mean. While the mean-reverting property is not very popular in the modelling of
real assets prices, it is natural to use it to model ‘relative value’ in the context of
strategies like pairs trading, where one assumes that the spread process moves around
some long-term equilibrium (Boguslavsky and Boguslavskaya, 2003).
A simple form of a mean-reverting process for discrete time-series is an autoregressive

process of order 1, termed AR(1),
yt = α yt−1 + εt , (2.14)
where {εt } is iid N (0, σ 2 ) and |α| < 1. The process (2.14) has a stationary mean-
22
reverting solution with
σ2
E[yt ] = 0 and Var[yt ] = .
1 − α2
Stationary time series are always mean-reverting. The terms of the series cannot move
too far from the mean due to the finite variance. The speed of mean-reversion α
is determined by autocovariance — it is fast if autocovariance is small and slow if
autocovariance is large (Alexander, 2001, p. 319). If α = 0 then the mean-reversion is
instantaneous and the process {yt } is a white noise.
σ2
E[yt yt−s ] = αs
1 − α2
As a result, the speed of mean-reversion α can be estimated as a ratio of the lag one
autocovariance to the variance of the process
E[yt yt−1 ]
α=
Var[yt ]
One of the most popular mean-reverting models is the Ornstein–Uhlenbeck process

named by the authors of the paper that first introduced it — Uhlenbeck and Ornstein
(1930).
If a stochastic process is stationary, Gaussian, Markovian, and continuous in proba-

bility, it is an Ornstein–Uhlenbeck process and it has the following equation (Finch,
2004):
dXt = θ(µ − Xt )dt + σdWt , (2.15)
where {Wt } is a standard Brownian motion and µ, θ > 0, σ > 0 are constants.
Expected value of Xt is
E[Xt ] = µ + (E[X0 ] − µ)e−θt = E[X0 ]e−θt + µ(1 − e−θt ). (2.16)
23
The variance of {Xt } is
−2θt σ2
Var[Xt ] = Var[X0 ]e + (1 − e−2θt ) (2.17)
2θ
as
Z t
−θt −θt
Xt = X0 e + µ(1 − e )+ σeθ(s−t) dWs . (2.18)
0
For stationarity, this Ornstein–Uhlenbeck process must be initialised with
σ2

X0 ∼ N µ, .
2θ
One can obtain a discretised form of equation (2.15), which takes the form of a first-
order autoregressive process, AR(1):
Xt − Xt−1 = θ(µ − Xt−1 )∆t + σ ∆Wt
= µ θ ∆t − Xt−1 θ ∆t + σ ∆Wt
= µ θ ∆t − Xt−1 θ ∆t + t ,
where ∆t → 0 and {t } are iidN (0, σ2 ), σ2 = σ 2 ∆t.
To estimate parameters of the mean-reversion one can use OLS regression
Xt − Xt−1 = A + BXt−1 + t . (2.19)
This estimation of the process {Xt } gives
A B
µ̂ = − , θ̂ = − .
B ∆t
Standard deviation of the process {Xt } is
σ
σ=√ ,
∆t
24
where σ is a standard deviation from the regression residual process.
The above method of parameters estimation was proposed by Dixit and Pindyck (1994).
There are a number of other methods for estimating parameters of the mean-reverting
process in academic literature (Elliott et al., 1999; Franco, 2003).
The strength of the mean-reversion θ, and the process standard deviation σ are the
two most important parameters if the process {Xt } is associated with tradable assets.
A greater parameter θ means that the process crosses zero more often, and so the asset
will be traded more often. Appendix A derives a formula for expected crossing time.
It follows that for greater θ an expected crossing time is smaller and, as a result, mean
crossing would happen more often.
The parameter σ defines how far from the mean the process {Xt } deviates, which deter-
mines a potential profit in each trade, if one opens positions when the process deviates
from its equilibrium and closes when the process returns to equilibrium. So large θ and
σ provide better opportunities to trade more often and with larger potential profit in
each trade.
2.3 Market Data and Software Used
Trading strategies presented in Chapters 3 and 4 of this thesis are tested using daily
prices of the stocks from Australia and the US. This includes:
• Australian market — prices from the Australian Stock Exchange (ASX);
• US market — prices of the shares included in Standard & Poor’s 500 (large capi-
talisation companies), Standard & Poor’s 400 (medium capitalisation companies)
and Standard & Poor’s 600 (small capitalisation companies) traded mostly on the
25
New York Stock Exchange and NASDAQ Stock market.
The study in Chapter 5 uses one hour prices for 80 companies (40 pairs) listed on the
NYSE and stock exchanges of nine Asia-Pacific countries.
All data sets are provided by the Securities Industry Research Centre of Asia-Pacific
(http://www.SIRCA.org.au/)) — the world leading provider of financial data for aca-
demic research and the investment industry.
The period of testing covers the last 10–15 years of stock prices history (it varies for
different markets) and includes the market crash of 2000 and the periods of high volatil-
ity due to the recent GFC. The variety of markets and market conditions allows the
evaluation of robustness of the trading strategies and level of market neutrality.
Statistical analysis, numerical simulations and all tests on the real data have been done
in the MATLAB R2008b software. For cointegration tests I used the freely distributed
Econometrics Toolbox by LeSage and Pace (2009) (www.spatial-econometrics.com/).
26
Chapter 3
Study 1. Performance Evaluation of

the Pairs Trading Strategies
This chapter reviews three methods of pairs trading and evaluates their performance
on the daily market data from the US and Australia. Also, it provides a theoretical
justification for strategy parameters previously taken as arbitrary.1
3.1 Introduction
Pairs trading is a popular trading strategy used by institutional and individual in-
vestors. The idea of a pairs trading strategy is fairly simple: find two stocks that
have historically moved together, when they deviate from each other open positions
towards the historical mean and close them when stocks converge together (Gatev
et al., 2006). The general description and the history of pairs trading strategy can be
found in many articles and books (Gatev et al., 2006; Vidyamurthy, 2004; Chan, 2008;
Whistler, 2004).
The premise of the pairs trading strategy is to hold long and short positions simultane-
ously. This way the trading is market neutral, and any profit or loss generated should
1
Bogomolov, T. ‘Pairs trading on the Land Down Under’, presented at 2nd Finance and Corporate
Governance Conference, Melbourne, April 2011.
27
be attributed to the relative price movements of the two assets, but not the market.
Hence, the total position stays hedged against any market movements.
On the one hand, market neutrality makes pairs trading extremely attractive for insti-
tutional investors such as superannuation funds, insurance companies and risk averse
hedge funds — that is, investors which are more interested in small but steady profit
at a low risk, rather than a high return at a higher risk.
On the other hand, pairs trading is a naturally leveraged strategy as money from the
short sale of one asset could be used to buy long another one. Therefore, less risk
averse investors can earn substantial profit by increasing the level of leverage. That
makes the strategy equally interesting for retail investors and hedge funds focused on
higher returns.
Despite the great interest in pairs trading from practitioners and academics, very little
research has published rigorous tests of pairs trading strategies on the real market data.
To the best of my knowledge, the only published works include tests on the US market
(Gatev et al., 2006; Do and Faff, 2011, 2010), the Brazilian market (Perlin, 2009), and
a sample of FTSE100 (Bowen et al., 2010). All those tests use only one method of
pairs trading — the distance method. Other methods proposed in academic literature
have never been tested on market data.
The purpose of this research is to examine the three most cited pairs trading strategies
and their performance using two data sets: the US companies included in the index
S&P 500 and the Australian Stock Exchange (ASX) market data. The first choice is
straightforward — the US is the biggest market and its top 500 companies are the
most liquid stocks. The second choice aims to test the generalisability and the level
of robustness of the tested methods of pairs trading. It might be reasonable to expect
that a good trading strategy provides similar results for different markets The last
28
one is subject to market regulations, liquidity and transaction costs, which are not so
different for the US and Australia.
This research follows the testing framework used by Gatev et al. (2006). When an
arbitrary decision needs to be made regarding the length of history or trading periods,
number of pairs traded simultaneously, level of deviation, method of excess return
calculation and so on, the same parameters and approaches as in Gatev et al. (2006)
are used to avoid any risk of data mining bias.
3.2 Background
3.2.1 Methods used
Following the general definition of pairs trading, an investor, seeking to create a working
strategy, needs to answer the following three questions:
1. What does it mean to ‘move together’ ? In other words, which assets should form
a long-short portfolio?
2. How ‘far’ should those assets deviate before applying the strategy or when to
open positions?
3. What does it mean that the assets ‘converge together’ and what to do if it never
happens (i.e. what is an exit strategy)?
Research by Do et al. (2006) reviews different approaches to pairs trading described in

academic literature, and separates them into three groups:
1. the distance method used by Gatev et al. (2006);
2. the cointegration method described by Vidyamurthy (2004);
3. the stochastic spread method proposed by Elliott et al. (2005).
29
While all three above mentioned methods of pairs trading have received attention in
academic literature, the distance method is the only one that has been tested on the
different data sets, probably due to its simplicity. The other two methods have never
been tested on real market data.
Each of the above approaches offers its own answers for questions, summarised earlier.
However, the details of the methods are not always clear. I try to follow the proposed
strategies as close as possible to the original description and make reasonable assump-
tions where necessary. All the methods mentioned above have the same structure:
pairs formation based on the analysis of the historical data; and rules about when to
open and unwind positions based on the spread process behavior.
It is worth noting that only the stochastic spread method (Elliott et al., 2005) proposes
a detailed ‘exit’ strategy — a complete set of rules for winning and losing cases. This
is an important development in the theory of pairs trading. The other methods only
suggest keeping the losing positions of the diverging stocks until the end of the trading
period. There are some attempts to improve the distance method by using stop-
loss (Nath, 2003) or empirically estimated 55 days holding period (Herlemont, 2004).
However, none of these improvements have any theoretical justification or published
empirical research to support them.
3.2.2 Distance method
The most cited research about pairs trading is Gatev et al. (2006). It is a purely empir-
ical study. The authors proposed a simple trading strategy, conducted a wide range of
testing over 40 years history, and offered some possible explanations on why one could
profit from pairs trading.
The authors take the following reasonable and straightforward trading rules:
1. Combine pairs by the minimal distance — a sum of squared deviations — between
30
log prices of two stocks xk and yk
n
X
mxy = (log xk − log yk )2 .
k=1
2. Enter into the trade when the spread becomes wider than two standard deviations
r
mxy
| log xn+t − log yn+t | > 2 .
n−1
Each pair was traded dollar-neutral (e.g. the value of the long position matched
the value of the short one).
3. Unwind positions when the spread converges to zero — log prices cross each other
the first time
| log xn+t − log yn+t | = 0.
Obviously, these rules are not the best option. A minimal distance between log prices
means that standard deviation of the spread is also minimal. As a result, the potential
profit from the trading by opening positions when the spread widens to 2σ and closing
when it returns to 0 may be very small. Indeed, Gatev et al. (2006) reported that
the profit on some of pairs trades was less than any reasonable estimation for possible
transaction cost.
Nevertheless, even these simple rules provided a robust profit: an average monthly ex-
cess return was 1.31–1.44% for the different groups of pairs. Through the entire period
of testing the pairs trading strategy performed well regardless of the market condi-
tions. However, the graph of an accumulated profit provided by Gatev et al. (2006)
became flatter after 1980. The plausible explanation for this is a reduced number of
opportunities for pairs trading as a result of an increasing activity of arbitragers and
hedge funds combined with a general decreasing of trading costs.
The results of Gatev et al. (2006) provide some important conclusions:
31
• Companies from different sectors, as well as companies with different capitalisa-
tion, were matched together and successfully traded over the period of testing.
Any restrictions on sector and company size may not be necessary.
• An average duration of trade was 3.75 months, which makes the approach used
by Gatev et al. (2006) a short-to-medium term investment. Traditional under-
standing of pairs trading as a contrarian strategy may not be the best option on
that investment horizon.
Perlin (2009) fully replicates the Gatev et al. (2006) approach on data from the Brazil-
ian stock market and reported similar results. Both studies confirm a good performance
of the strategy and its independence from the market conditions.
3.2.3 Cointegration approach
Vidyamurthy (2004) provides a good introduction into pairs trading and offered a trad-
ing strategy based on cointegration and a common trends model. However, he did not
provide any testing of his trading strategy on real market data.
The Vidyamurthy (2004) method is not restricted to be dollar neutral and trades two
stocks with ratio 1 : γ, where γ is a coefficient of cointegration. Positions should
be opened when the spread diverges from long-run equilibrium on selected ∆, and
unwounded when the spread converges to equilibrium or even to −∆.
log(PtA ) − γ log(PtB ) = µ − ∆ (3.1)

A B
log(Pt+i ) − γ log(Pt+i ) = µ + ∆, (3.2)
where µ is a long-run equilibrium and γ is the cointegration coefficient.
32
Combining the above equations together then one can estimate the (percentage) profit:
A B
) − log(PtA ) − γ log(PtB )

Profit = log(Pt+i ) − γ log(Pt+i
A
) − log(PtA ) − γ log(Pt+i
B
) − log(PtB )

= log(Pt+i
= RA − γRB = µ + ∆ − (µ − ∆)
= 2∆. (3.3)
The potential (percentage) profit from pairs trading is the weighted difference of re-
turns on the stocks A and B, or just 2∆ — twice the value of the selected deviation
from the long-run equilibrium.
Using cointegration provides a number of benefits with the most important two being:
• As the spread between log of stocks prices becomes stationary (at least approx-
imately so), it is possible to estimate parameters and the potential (percentage)
profit. Thus, one can test the suitability of a pair of stocks for pairs formation.
• The coefficient of cointegration should allow a trader to eliminate common trends

and make the strategy market-neutral.
Vidyamurthy (2004) recommends that one chooses pairs of stocks which potentially
could be cointegrated based on fundamentals. This approach is used by many practi-
tioners.
As an example, Herlemont (2004) advises traders to use companies with approximately

the same market β in order to eliminate market risk, from the same sector to eliminate
sector related risk, and take companies with similar capitalisation to reduce risk related
to small/big caps.
3.2.4 Stochastic spread process method
Elliott et al. (2005) propose a stochastic spread method. The authors do not try to
33
address the problem of pairs formation, but rather focus on the spread itself. An
observed spread process is modelled as a noisy representation of a mean-reverting state
process. The state process is
xk+1 = A + Bxk + Cεk+1 (3.4)
and the observation process is

yk = xk + Dωk , (3.5)
where noise processes {εk } and {ωk } are iid Gaussian N (0, 1); and 0 < B < 1. Param-
eters A, B, C, D may be estimated from the observation process {yk } by using Kalman
filter techniques.
The authors step away from trading individual stocks even combined in pairs. They
offer an algorithm for the spread process trading with detailed quantitative approach
on all aspects of it and provide all necessary formulas.
The trading strategy is as follows:
• Run calibration process over the last N observations and estimate the spread
process parameters. If estimated B is greater than 0 but less than 1, then the
spread process is mean-reverting and suitable for pairs trading.
• Based on standard deviation and speed of mean-reversion B, choose some level

of deviation from the mean which would be a trigger to enter the pairs trade.
• Unwind the position when the spread process converges to mean or at time T
later, that could be estimated from the speed of mean-reversion B.
Elliott et al. (2005) provide a new approach in giving explicitly the most likely life-time
limit for holding the pairs trading position open. Gatev et al. (2006) and Vidyamurthy
(2004) keep the open positions until the closing signal is generated or until the end
of the selected investment period. Nath (2003) implements a stop-loss trigger to limit
34
loss/drawdown on the open position. Herlemont (2004) advises an empirically esti-
mated 55 day holding period.
The holding time estimate reflects an important development in pairs trading. Pairs
trading acquired an extra dimension effecting the strategy — time. It is not necessary
for the stocks to have a long-run relationship or a stationary spread to form a pair. It
is good enough for the spread process to be ‘stable’ over the last N observations. If pa-
rameters of the spread process change, then the trading on that spread process should
be stopped, open positions should be closed, and re-calibration should take place.
The stochastic spread process method has been critiqued by Do et al. (2006) as having
very little practical application. The authors claim that it would be very unlikely to
find two stocks that have a mean-reverting spread process (with two stocks having
similar returns). In the cointegration paradigm, it means finding two stocks with the
cointegration vector of the accumulated returns (1, −1). Do et al. (2006) advise that
using companies with dual listed structure or companies listed in multiple exchanges
is the only way for this strategy to work in real world scenarios.
The above critique is partly true. However, Elliott et al. (2005) do not restrict the
spread formation in any way. Indeed, it may not be a spread but a single asset which
has a mean-reverting property, or a spread from a variety of assets that would form
two portfolios — long positions and short positions. The strategy would work the same
way. It is up to an investor to construct the suitable spread process and trade it with
the proposed technique. The potential scope of its application is very wide. Elliott
et al. (2005) provide a numerical test for the proposed technique on generated data
but did not perform any testing on real market data.
The method proposed by Elliott et al. (2005) has a number of followers. Do et al.
(2006) use the same framework and discuss the general approach to modelling relative
35
mispricing of two assets and its connection to Asset Pricing Theory. Triantafyllopoulos
and Montana (2011) introduce time-dependency in the parameters of the Elliott et al.
(2005) model and provide an on-line estimation algorithm which could be suitable for
intraday trading on high frequency data. Unfortunately, no empirical tests on the
market data for any of those methods have been published.
3.3 Computation of returns
3.3.1 Excess returns computation
Return on investment (ROI), or just return, is the ratio of money gained (or lost) rel-
ative to money invested. The calculation of returns in pairs trading is non-trivial. A
short sale of one asset is a way to borrow money to buy long another asset. A dollar
neutral pairs trading portfolio is a zero cost investment (at this stage commissions and
possible margin requirements are ignored). Any profit or loss made from zero invest-
ment would mean infinite positive or negative return, which is not very useful for the
purpose of comparing the performance of different trading strategies.
To avoid this problem I follow Gatev et al. (2006) by trading $1 positions in each stock
(which makes $2 total trading volume for the pair) and calculate a portfolio returns
rP for each trading day t as value-weighted daily cash flows from each pair, which are
considered as excess return:
P
i∈P wi,t ci,t
rP,t = P , (3.6)
i∈P wi,t
where:
ci,t is a daily cash flow from the two positions formed the pair i;
wi,t is a weight of each pair. For each newly opened position on the pair, initial weight
equals 1 and then evolves depending on the returns from previous days as a
constantly re-invested profit
wi,t = wi,t−1 (1 + ci,t−1 ) = (1 + ci,1 ) · · · (1 + ci,t−1 ).
36
The daily cash flow from the pair or a daily return of the pair is
2
X
ci,t = Ij,t vj,t rj,t , (3.7)
j=1
where:
Ij,t is a dummy variable which is equal to:
0 if the position on the stock j is not open,
1 if a long position on stock j is open, and
-1 if a short position on stock j is open;
rj,t is a daily return on stock j;
vj,t is a weight of stock j is used to calculate daily cash flows

vj,t = vj,t−1 (1 + rj,t−1 ) = (1 + rj,1 ) · · · (1 + rj,t−1 ).
The strategies’ daily returns are then compounded to obtain monthly returns.
This method of return calculation is widely used in the pairs trading literature to eval-
uate performance of the long-short portfolios. However, it should be mentioned that
pairs trading is a naturally leveraged product and the above method uses 2:1 leverage.
Hence, one should be careful when comparing the results of the pairs trading strategy
with non-leveraged strategies, for example, the naive buy-and-hold strategy.
For the cointegration method, which is not dollar neutral, I scale the initial weights
for both stocks in the pair to make the total market position of the pair equal to $2.
This allows us to compare the results of this strategy with the dollar neutral ones —
the distance and stochastic spread methods.
Trading is started on the first working day of each month, and is traded for six months.
So for each trading month, except for the first and the last 5 months which are ex-
cluded from final report, I calculate six different estimations of monthly returns and
37
then average them to obtain the final estimation.
Following Gatev et al. (2006), I consider two measures of excess return:
1. The return on committed capital; that is $5 and $20 investments in the pairs
portfolio. This is a conservative estimation as I include in calculation $1 invest-
ment per each pair regardless if the pair has opened any positions during the
trading period. That is, I commit the full amount which might be necessary for
future trading. So, the total return is calculated from this basis.
2. The fully invested return; that is, a cash flow divided by the actual investment
used to generate this cash flow. This measure simulates the situation when the
investor borrows and invests $1 each time he sees an opportunity for trading and
returns money after unwinding positions.
In calculating the fully invested return my approach is different to one that is described
in Gatev et al. (2006). They make a $1 investment committed to trading after the first
trade on the pair and do not include that extra investment capital only if there are no
trades at all on the pair during the entire trading period. I include in the calculation a
$1 investment only if there is a trade currently open; that is, if an investment generates
some cash flow (either positive or negative). If there is no opened positions, then
invested capital is zero. This difference is very important for proper evaluation of the
short-term trading strategies which generates a number of short-living non-overlapping
positions.
3.3.2 Transaction costs
Stock trading involves some direct and indirect transaction costs including brokerage
fees, fees for shorting, bid-ask spread, opportunity, liquidity or market impact costs.
Despite the possible small size, transaction costs could have a serious impact on the
performance of pairs trading strategies. This is especially true for short-term trading
strategies which involve many trades. Bowen et al. (2010) report more than a 50%
38
reduction in the excess returns of the high frequency pairs trading strategy after ap-
plying a 15 basis point transaction fee. Do and Faff (2011) fully replicate the research
by Gatev et al. (2006) and report that the strategy becomes virtually unprofitable after
careful accounting for all transaction costs.
For retail traders the largest part of these costs is the brokerage fees which are paid
by traders each time they buy or sell shares. On average, these fees vary from 5 to 15
basis points of the total amount traded.
In all tests, transaction costs are equal to 0.10% (10 basis points), which is an average
brokerage fee at May 2011 for retail investors (Interactive Brokers 0.08%, CommSec
0.12%, E-Trade 0.11%, Macquare Edge 0.10%). Because the brokerage fee applies to
the full traded volume, I adjust cash flows for all traded pairs using the following rules:
• On the day of the opening positions, I reduce the cash flow from each stock in
the pair by the size of transaction costs; that is, I reduce the total cash flow from
the pair by doubling the size of transaction costs
ci, t → ci, t − 2 b.
• On the day of closing, I reduce cash flow from each pair as follows
2
X
ci, t = (Ij, t vj, t rj, t − b vj, t (1 + rj, t )) ,
j=1
where:
ci,t is a cash flow from the pair i or excess return on the pair i;
b = 0.0010 is a transaction cost;
Ij,t is a dummy variable which is equal to 0 if the position on the stock j is not
open, 1 – if a long position on stock j is open, -1 – if a short position on
stock j is open;
39
rj, t is a daily return on the stock j;
vj, t = vj, t−1 (1 + rj, t−1 ) weight of the stock j.
3.3.3 Data sets
For testing, I use two data sets with daily closing prices of the Australian market and
US companies included in the market index S&P 500. The data are obtained from
the Securities Industry Research Centre of Asia-Pacific (SIRCA) and cover 184 months
starting from January 1, 1996 and finishing on May 21, 2011.
The actual time interval used for trading is shorter — from July 1, 1997 to November
31, 2010 — due to 12 months historical data used for strategy calibration and 5 months
before and after trading interval discarded at averaging.
The testing interval includes the GFC. Pairs trading as a contrarian strategy should
benefit from uncertainty and high volatility of stock markets during the GFC. Aus-
tralian Securities & Investments Commission (ASIC) banned naked short selling from
November 2008 until May 2009. However, I still test pairs trading over that period as
usual. Institutional investors, who hold large diversified portfolios of Australian stocks,
can use pairs trading as a part of the tactical asset allocation strategy. They do not need
short selling to fulfill rules of pairs trading, as they can sell some shares from the exist-
ing portfolio and buy them back when the strategy signals to close position on the pair.
For pairs trading, only stocks with sufficient liquidity are considered. For both data
sets I filter out all stocks which have more than 10 non-trading days during the history
period selected for calibration of the trading strategy. For the Australian data I pre-
select only the top 30% companies by the average daily dollar-valued trading volume
during the last 12 month period. For each testing period the US data set contains only
companies included in the S&P 500 index during the period of calibration. All of the
US firms are actively traded large capitalisation companies, so I use all stocks except
40
those with a high number of non-trading days.
As a result of the screening, the number of stocks available for pairs trading varies
from 41 to 244 for the Australian data and from 392 to 541 for the US data. The total
number of potential pairs varies from 820 to 29,646 and from 76,636 to 146,070 for the
Australian and US data sets respectively.
Number of companies preselected Total Trading

Mean St.Dev. Median Max Min trading days months
Australian market 114.1 51.8 103 244 41 3863 160
US market 455.3 29.0 450 541 392 3853 160
Table 3.1: Statistics on data sets used in the Australian and US markets
This approach makes the research biased towards large capitalisation companies. On
the positive side, I have virtually eliminated a potential liquidity problem. Simultane-
ous trade execution for different assets is crucial for pairs trading. Hence, this approach
can be considered close to a real-life situation, when the investor would prefer limit-
ing his trading universe to large capitalisation companies only to have enough liquidity.
Both data sets include de-listed companies. If the company had been de-listed during
the trading period, I then use the last trading day closing price to unwind position
on the spread with this company. This excludes the survival bias. For the US data
the company is considered for pairs trading only if it is in the S&P 500 index on the
last day of the calibration period. The selected company is traded till the end of the
trading period even if it happens to be excluded from the index during that period.
The opening and closing prices are the results of auctions, which usually attract a large
trading volume. Using the closing prices I can be sure that one could make a trade at
the given price, thus avoiding bid-ask bounce bias.
If a stock has a non-trading day during the trading period (price and/or volume equal
41
to zero), the closing price of the previous day is used to create the spread. However,
positions on the pairs having that stock cannot be opened or closed on that day, even
if the spread process signals to do so.
3.3.4 Testing periods and benchmarks
For the testing, I take 12 months of the historical data to create a list of pairs and
estimate parameters of the spread processes for every pair. After that I select the best
pairs and trade them during the following 6 months using only the parameters calcu-
lated from 12 month historical data. For each of the three strategies, I run tests for
5 and 20 pairs traded simultaneously with and without accounting for transaction costs.
The choice of a benchmark for pairs trading strategy is not obvious. A proper bench-
mark for pairs trading strategy is a highly debatable topic and it does not have an
agreed definition. On the one side, the strategy is claimed to be market neutral and
promises small but consistent profit, so one can argue that the proper benchmark is a
‘risk-free’ asset, like government backed bonds.
On the other side, all traded pairs in the performance tests constitute stocks — portfo-
lio of long and short positions, so potentially there is the same level of risk involved as in
stock trading. ‘Market neutral strategy’ does not mean risk free strategy. Pairs trading
might be neutral to the general market movements but widely exposed to movements
of individual stocks and their relative co-movements. If stocks do not converge as ex-
pected but continue to diverge, then the investor faces potentially unlimited risk due to
having a short position. Hence, it is appropriate to use a market index as a benchmark.
In this research I follow Gatev et al. (2006) using market index as a benchmark mostly
for the risks level assessment purpose. I take S&P/ASX 200 and S&P 500 indexes
as benchmark for the Australian and US markets to compare all strategies against,
respectively.
42
As mentioned above, pairs trading is a naturally leveraged investment. All presented
results of the strategies tests use 2:1 leverage, while the benchmarks are non-leveraged
‘buy-and-hold’ strategy. It’s an approach widely used in the literature on pairs trading,
so I choose to follow the common framework to keep my results comparable to other
studies in the field. De-leveraged returns on each strategy are twice smaller of the
reported results and have twice smaller standard deviations.
3.3.5 Data snooping bias
Important concern of any empirical research on the trading strategies performance is

the risk of the data snooping bias. That is the risk of unintentional bias towards the
most profitable strategies happens when the same data set is used multiple times for
parameters estimation and following testing.
Originally, the problem of data snooping had been raised by Lo and MacKinlay (1990)
and Brock et al. (1992) and later confirmed by many studies. White (2000); Hansen
(2005); Hsu et al. (2010) propose different types of ‘reality check’ or data snooping tests
and correction for data snooping. However, those tests are not applicable for this study
as they require considering the ‘universe’ of all possible trading strategies, while the
focus of this research is the performance evaluation of three individual strategies only.
To address the problem of potential data snooping bias I use each data set only once for
each strategy and employ a true out-of-sample test. In each test I take 12 months his-
tory to calibrate models and estimate all necessary parameters. I select the top 5 or 20
pairs based on the parameters estimation — that is the smallest standard deviation of
the spread process during the 12 month history. Then I trade these top pairs during the
following 6 month trading period using re-defined rules and parameters estimated dur-
ing calibration. Reported strategy performance is a result of the out-of-sample trading.
43
The only arbitrary chosen parameter is a trigger level to initiate trading, which is
defined as two standard deviations. This choice is based on the literature for non-
parametric (distance and cointegration) methods and a statistical model for parametric
method (stochastic spread process). The research does not employ any alternative set
of rules and/or parameters.
3.4 Trading rules
3.4.1 Distance method
I use the following strategy based on the distance method of pairs trading proposed by
Gatev et al. (2006):
1. Pairs formation: I take log prices for all stocks selected for pairs trading over
the 12 month history period and combine stocks in all possible pairs. The total
number of possible pairs is quite large:
 
N  N!
PN =   = ,
2 2!(N − 2)!
where N is the total number of stocks eligible for pairs trading.
I do not shift individual stocks log price processes at the start from $1 as in
Gatev et al. (2006), because it could result in a bias if the two stocks are in the
phase of divergence at the first day of calibration period. To avoid that bias and
to simplify further the calculations, I adjusted the spread process between two
stocks by its mean.
yi,j (t) = log Pi (t) − log Pj (t) − ȳi,j ,
where Pi (t), Pj (t) are prices of stocks i and j on day t, ȳi,j is the mean of the
44
spread between the two stocks i and j.
Then, all pairs are sorted in ascending order by the size of the standard deviation
of the spread process yi,j (t), which is proportional to the squared distances be-
tween stocks used by Gatev et al. (2006). The pairs with the smallest standard
deviations are used for pairs trading.
Stocks which are picked for a pair are removed from the pool, so the same com-
pany shares cannot be a part of several pairs. The method of stock picking
without replacement resembles the statistical method known as one-level hierar-
chical clustering and should increase diversification of an investment portfolio.
2. Rules to open positions: I choose arbitrarily a trigger level as two standard

deviations of the spread process. If the difference between the log prices (that
is, the spread process) of the selected stocks hits the trigger level, then I open
position on the pair.
• If the spread hits the level 2σspread then trading signal ‘sell spread’ is gener-
ated, one sells the first stock (i) in the pair and buys the second one (j).
• If the spread process hits the level −2σspread then trading signal ‘buy spread’
is generated, one buys the first stock (i) in the pair and sells the second
one (j).
The parameter σspread is determined over the 12 month calibration period and
does not change until the end of the trading period.
3. Rules to close positions: I close open positions when the spread process hits
zero the first time after opening positions or at the end of the 6 month trading
period — whichever occurs first.
45
3.4.2 Cointegration method
The pairs trading strategy proposed by Vidyamurthy (2004) is based on the theory
of cointegration developed by Granger (1981), Granger and Weiss (1983) and Engle
and Granger (1987), and the common trends model by Stock and Watson (1988).
Vidyamurthy (2004) suggested several approaches to the pairs selection and trading. I
adopt the purely quantitative approach in my testing as the most objective and straight
forward method.
1. Pairs formation: similar to the distance method, I take log prices for all stocks
selected for pairs trading over the 12 month history period and combine them in
pairs. I employ ordinary least squares (OLS) regression of the first stock in the
pair on the second one and build a spread process {yi,j }.
yi,j (t) = log Pi (t) − γi,j log Pj (t) − αi,j ,
where Pi (t), Pj (t) are prices of stocks i and j on day t, αi,j and γi,j are an inter-
cept and a slope (cointegration coefficient) of OLS regression of the log prices of
stock i on the log prices of stock j.
The spread process is tested for stationarity by the Dickey–Fuller (DF) test. If
the DF test statistic is greater than the critical value for 5% significance, then
the pair is rejected as non-cointegrated. After that, all accepted pairs are sorted
in ascending order by the value of the spread process standard deviation and the
pairs with the smallest values are used for pairs trading.
Stocks which are selected for a pair are removed from the pool, so any of the
selected stocks cannot be a part of several pairs.
2. Rules to open positions: I build the spread process with parameters (αi,j , γi,j , σi,j )
defined during the calibration period. If the spread hits a trigger level, then I
46
open position on the pair — trade the spread process towards zero. To be con-
sistent with the testing of the other methods, I arbitrarily choose a trigger level
as two standard deviations of the spread process.
The stocks in each pair are traded in proportion 1 : γi,j with the total value of
the open position being $2. This means that the method is not dollar neutral.
3. Rules to close positions: I close positions when the spread process hits zero
the first time after the opening or at the end of the 6 month trading period —
whatever occurs first.
3.4.3 Stochastic spread method
The general idea of the strategy proposed by Elliott et al. (2005) is based on the as-
sumption that if one detects a mean-reverting property of the spread between two
stocks, he/she can expect that the spread process stays mean-reverting for some time
in the future. This means one can exploit those properties to make a profit.
Elliott et al. (2005) consider the spread process as a two equation model and use
Kalman filter to estimate all parameters of the process.
A hidden state equation

xk+1 = A + B xk + C k+1 (3.8)
and observation equation

yk = xk + D ωk , (3.9)
where t and ωk are iid and ∼ N(0, 1).
I employ the following strategy of pairs trading based on the mean-reverting property
of the spread process:
47
1. Pairs formation: I use log prices of the stocks considered for pairs trading to
build spread processes {yi,j }
yi,j (t) = log Pi (t) − log Pj (t). (3.10)
Let {yk } in (3.9) to be defined by {yi,j } in (3.10) and estimate parameters

(A, B, C, D) of the spread processes by the Kalman filter. Then I define the
processes’ means and standard deviations by
A C
µi,j = ; σi,j = √ .
1−B 1 − B2
I then sort all spread processes in ascending order by the value of the process
standard deviation σi,j . This is similar to the sum of squared deviations between
stocks i and j used by Gatev et al. (2006). However, it is a standard deviation
of the invisible ‘true’ spread process but not its observed noisy interpretation.
Pairs that form top 5–20 spread processes are considered for pairs trading.
2. Rules to open positions: I take a trigger level λ as two standard deviations

σi,j of the spread process to be in line with other methods tested.
c+,− = µ ± λ σ,
where λ = 2.
When the spread process yi,j (t) hits level c+,− one opens position on the spread
process towards its mean.
If the level c+ is hit, then one ‘shorts the spread’ — sells stock i and buys stock
j. If the level c− is hit, then one buys ‘long the spread’ — buys stock i and sells
stock j. All trades are ‘dollar neutral’, that is, equal dollar size positions are
opened long and short.
3. Rules to close positions: I unwind positions if the spread process hits its mean
48
µ or t̂ times later, whichever occurs first. The parameter t̂ is the most likely time
for the spread process to hit its mean. The estimation of t̂ can be done using
an approximation of the Ornstein–Uhlenbeck process and depends on λ chosen
before opening the positions.
r !
1 1 λ2 λ4 λ2 9
t̂ = log − + + − + .
2(1 − B) 2 2 4 2 4
It is possible to show that the Ornstein–Uhlenbeck process traded from the level
c+,− corresponding to λ = 2 and closed t̂ times later generates non-negative profit
in 85% of cases (see Appendix A). Hence, if the closed trade profit before transac-
tion costs is less than zero, I then decide that the model is broken and stop trading
this pair. Otherwise, I keep the pair and wait for the next opportunity to trade.
3.5 Empirical results
Tables 3.3–3.14 summarise monthly excess returns trading statistics and Figures 3.1–
3.12 demonstrate historical performance of the different strategies. In the tables, Mean
is an average monthly excess return before transaction costs or after transaction costs
(10 basis points per stock per trade, roughly it is about 0.4% per round trip on the
pair). I employ the bootstrap method (Efron and Tibshirani, 1994) with 10,000 simu-
lations to estimate standard errors for all reported parameters.
In most tests, all trading strategies demonstrate statistically significant returns with
relatively low standard deviations and outperform corresponding market indexes used
as benchmarks (during the period of testing S&P/ASX 200 earns in average 0.31%
per month with standard deviation 3.95% and S&P5̇00 earns 0.11% per month with
standard deviation 4.92%).
Sharpe ratio is close to 1 in the most test for the distance and stochastic spread pro-
cess methods. Modigliani risk adjusted performance (RAP) is an another measure of
49
Mean St.Dev. Median Max Min Kurtosis Skewness
S&P/ASX 200 0.0031 0.0395 0.0108 0.0740 -0.1408 4.0935 -0.9532
S&P5̇00 0.0011 0.0492 0.0086 0.0932 -0.1928 4.0498 -0.7637
Table 3.2: Historical performance of the benchmark indexes
the risk adjusted returns (Modigliani and Modigliani, 1997). It is derived from the
Sharpe ratio and defined as a product of the Sharpe ratio and the standard deviation
of the returns on the benchmark. It provides percent unit returns with consideration
for volatility of the strategy and benchmark returns. That allows direct comparing be-
tween strategy and market returns. For all tested strategies Modigliani RAP is higher
than the actual monthly excess returns. This means that strategies’ returns are earned
at a lower level of risk then the normal market risks. So, for risk averse investors,
an investment in the portfolio of pairs might be more preferable than commonly used
diversified market portfolio.
Despite the similar approach to pairs selection biased towards large capitalisation com-
panies, the tested trading strategies demonstrate very different behavior on the Aus-
tralian and US markets. On the Australian market the distance and cointegration
methods perform reasonably well in the tests on the top 5 pairs (monthly excess re-
turns 0.63% and 0.51%) and relatively bad on the top 20 (monthly excess returns 0.31%
and 0.14%). On the US market, the situation is opposite — top 20 pairs perform better
then the top 5. The reason behind this could be the small number of liquid stocks on
the Australian market available for pairs formation. There is a chance that increasing
the size of the portfolio, I add pairs that have poorer fit for pairs trading. Companies
that form the S&P 500 index are all very actively traded and one could not have that
problem. By increasing the size of portfolio, chances to select more good pairs were
improved. The stochastic spread method demonstrates a different result. It shows
better returns for the larger portfolio on both markets. The method happens to be
‘immune’ to the problem of the smaller size of the Australian market.
The excess returns on the committed capital of the stochastic spread method looks
50
modest compared to other methods. This strategy has an average holding time five
times shorter than the distance and cointegration methods with a similar number of
trades executed during the 6 month trading period (Table 3.15). Therefore, most of
the time the money committed to the trading is not being used. It is an extremely rare
situation when all 5 or 20 pairs are open simultaneously, in contrast to the distance
and cointegration methods where it is a very common scenario. In this case, the excess
return on the actual invested capital is a more appropriate measure of the strategy
performance.
A correlation with the market indexes and the market betas for all methods are very
low and statistically not significant. This allows us to conclude that all considered
methods of pairs trading are market neutral. During the GFC and the dot-com crash
in the US, all strategies show higher levels of returns compared to ‘normal’ times. As
expected, pairs trading strategies benefit from the high volatility and uncertainty of
the financial crises. Hence, there is still some relation between the performance and
market conditions. The performance depends on the market volatility but not on its
direction.
Table 3.15 provides statistics of the individual trades for each method of pairs trading
for both markets. These statistics are based on the trading top 5 and 20 pairs and do
not include transaction costs. On average, each $1 trade on the spread (that is, a short
trade on one asset and long trade on another) earns up to 1.85 cents for distance, 1.32
cents for cointegration and 1.09 cents for stochastic spread methods (top 5 pairs for
the Australian market).
The big difference is in the time required to earn that profit. The stochastic spread
method has an average holding time of only 9.5 days, while the other two methods hold
positions 5–6 times longer. A longer holding time makes the distance and cointegration
methods much riskier investments. The higher level of risk is reflected in the higher
51
standard deviations of returns and much wider ranges between the average maximal
profit and maximal loss in one trade over the 6 month trading period. The average
maximum profit in one trade is 12%, 16% and 8%, and the average maximum loss in
one trade is 15%, 27% and 6% for the distance, cointegration and stochastic spread
methods respectively.
3.6 Conclusions
The purpose of this research is to define the methodology of a practical application of

the three methods of pairs trading described in the academic literature and to evaluate
their performance on real market data from the Australian and US stock markets.
While all three approaches demonstrate true market neutrality and reasonably good
performance before transaction costs in some tests, the real-life profitability of these
strategies in their existing forms is questionable. Some tests show very low excess
return and transaction costs reduce them even further. Level of returns is not con-
sistent over time, there are long periods with zero or even negative returns. The lack
of liquidity limits the number of stocks that can be considered for pairs trading, thus
decreasing potential profit.
However, the general idea of pairs trading is sound. A better criteria of pairs selection
and alternative rules of trading could be developed. The distance and cointegration
methods could be improved in regards to closing rules and control of constantly diverg-
ing pairs. The stochastic spread method is based on the assumption of normality of
the innovation process which is not the case for stocks log prices (Mandelbrot, 1963).
Less restricted models could significantly improve strategy’s performance.
52
Committed capital Fully invested capital
Before trans.cost After trans.cost Before trans.cost After trans.cost
Value S.E. Value S.E. Value S.E. Value S.E.
Mean 0.0063 0.0014 0.0055 0.0014 0.0106 0.0024 0.0091 0.0023
Median 0.0039 0.0017 0.0031 0.0017 0.0087 0.0029 0.0070 0.0028
Standard deviation 0.0182 0.0016 0.0181 0.0016 0.0298 0.0022 0.0295 0.0022
Skewness 0.9496 0.4130 0.9098 0.4107 0.8484 0.2787 0.8382 0.2800
Kurtosis 5.9494 1.6353 5.8175 1.5804 4.6000 0.9027 4.5938 0.9031
Minimum -0.0374 0.0034 -0.0386 0.0036 -0.0466 0.0013 -0.0474 0.0011
Maximum 0.0928 0.0179 0.0902 0.0172 0.1257 0.0154 0.1223 0.0155
Average profitable month 0.0164 0.0015 0.0164 0.0015 0.0280 0.0025 0.0276 0.0025
Average losing month -0.0097 0.0011 -0.0097 0.0011 -0.0154 0.0017 -0.0158 0.0016
Proportion of losing months 0.3875 0.0383 0.4188 0.0389 0.4000 0.0383 0.4250 0.0387
Trades per month 0.2771 0.0074 0.2771 0.0074 0.2771 0.0073 0.2771 0.0074
Correlation with Index -0.1613 0.1083 -0.1591 0.1068 -0.0787 0.1013 -0.0774 0.1014
Sharpe Ratio 0.3454 0.0720 0.3023 0.0727 0.3576 0.0719 0.3096 0.0729
Modigliani RAP 0.0136 0.0032 0.0119 0.0032 0.0141 0.0032 0.0122 0.0032
Jensen’s Alpha 0.0065 0.0015 0.0057 0.0015 0.0108 0.0024 0.0093 0.0024
Market Beta -0.0745 0.0517 -0.0728 0.0504 -0.0593 0.0770 -0.0579 0.0765
Table 3.3: Monthly excess returns statistics and risk measures for trading top 5 pairs
by the distance method of pairs trading on the Australian market before and after
transaction cost.
S.E. is a corresponding standard error for parameters values in the left column.
ASX, top5 pairs

1.8
Market Index
1.6 Strategy Before Tr.Cost − Committed Capital
Strategy After Tr.Cost − Committed Capital
1.4 Strategy Before Tr.Cost − Invested Capital
Strategy After Tr.Cost − Invested Capital
1.2
0.8
0.6
0.4
0.2
−0.2
1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010
Figure 3.1: Historical performance of the distance method of pairs trading on the
Australian market for top 5 pairs — accumulated returns without reinvestment. Axes
are time and returns.
53
Mean 0.0031 0.0011 0.0023 0.0011 0.0052 0.0018 0.0035 0.0017
Median 0.0030 0.0014 0.0021 0.0013 0.0043 0.0017 0.0027 0.0017
Skewness 0.1420 0.3588 0.1151 0.3581 -0.0892 0.2473 -0.1112 0.2435
Kurtosis 4.6053 0.7843 4.5562 0.7582 3.6150 0.4616 3.6066 0.4625
Minimum -0.0391 0.0036 -0.0400 0.0037 -0.0704 0.0110 -0.0722 0.0111
Maximum 0.0568 0.0094 0.0551 0.0093 0.0703 0.0059 0.0677 0.0057
Trades per month 0.2590 0.0054 0.2590 0.0054 0.2590 0.0054 0.2590 0.0053
Correlation with Index 0.0212 0.1105 0.0234 0.1109 0.0493 0.0973 0.0499 0.0979
Sharpe Ratio 0.2272 0.0801 0.1688 0.0801 0.2324 0.0817 0.1592 0.0801
Modigliani RAP 0.0090 0.0032 0.0067 0.0032 0.0092 0.0033 0.0063 0.0032
Jensen’s Alpha 0.0031 0.0011 0.0023 0.0011 0.0051 0.0018 0.0035 0.0018
Market Beta 0.0073 0.0383 0.0081 0.0381 0.0279 0.0547 0.0282 0.0549
by the distance method of pairs trading on the Australian market before and after
transaction cost.
ASX, top20 pairs

1
Market Index
Strategy Before Tr.Cost − Committed Capital
0.8 Strategy After Tr.Cost − Committed Capital
Strategy Before Tr.Cost − Invested Capital
0.6
0.4
0.2
−0.2
−0.4
1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010
Figure 3.2: Historical performance of the distance method of pairs trading on the
54
Mean 0.0030 0.0011 0.0022 0.0011 0.0051 0.0019 0.0037 0.0019
Median 0.0024 0.0012 0.0018 0.0012 0.0044 0.0024 0.0036 0.0023
Skewness 0.2685 0.1684 0.2549 0.1660 0.0638 0.1440 0.0556 0.1506
Kurtosis 3.1482 0.3080 3.1223 0.3038 2.7653 0.2294 2.7762 0.2361
Minimum -0.0327 0.0032 -0.0330 0.0029 -0.0466 0.0013 -0.0487 0.0015
Maximum 0.0456 0.0036 0.0443 0.0035 0.0670 0.0044 0.0654 0.0045
Trades per month 0.2542 0.0059 0.2542 0.0059 0.2542 0.0060 0.2542 0.0058
Sharpe Ratio 0.2073 0.0772 0.1541 0.0780 0.2158 0.0798 0.1567 0.0800
Modigliani RAP 0.0102 0.0038 0.0076 0.0039 0.0106 0.0040 0.0077 0.0040
Jensen’s Alpha 0.0029 0.0011 0.0022 0.0011 0.0051 0.0019 0.0037 0.0019
Market Beta 0.0245 0.0253 0.0249 0.0251 0.0269 0.0399 0.0273 0.0399
by the distance method of pairs trading on the S&P 500 companies before and after
transaction cost.
SPX500, top5 pairs

1
Market Index
0.6
0.4
0.2
−0.2
−0.4
1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010
Figure 3.3: Historical performance of the distance method of pairs trading on the S&P
500 companies for top 5 pairs — accumulated returns without reinvestment. Axes are
time and returns.
55
Mean 0.0042 0.0008 0.0034 0.0008 0.0066 0.0013 0.0049 0.0013
Median 0.0036 0.0006 0.0030 0.0006 0.0053 0.0011 0.0037 0.0011
Skewness 0.5451 0.2233 0.5265 0.2238 0.4633 0.2784 0.4537 0.2815
Kurtosis 3.9476 0.4700 3.9139 0.4673 4.1046 0.5570 4.0841 0.5661
Minimum -0.0266 0.0050 -0.0273 0.0050 -0.0484 0.0113 -0.0499 0.0111
Maximum 0.0400 0.0033 0.0390 0.0033 0.0621 0.0055 0.0604 0.0057
Trades per month 0.2704 0.0043 0.2704 0.0044 0.2704 0.0043 0.2704 0.0043
Sharpe Ratio 0.3906 0.0762 0.3136 0.0750 0.3935 0.0772 0.2932 0.0778
Modigliani RAP 0.0192 0.0039 0.0154 0.0038 0.0194 0.0040 0.0144 0.0039
Jensen’s Alpha 0.0042 0.0008 0.0033 0.0008 0.0065 0.0013 0.0048 0.0013
Market Beta 0.0243 0.0214 0.0243 0.0210 0.0364 0.0331 0.0361 0.0332
by the distance method of pairs trading on the S&P 500 companies before and after
transaction cost.
SPX500, top20 pairs

1.2
Market Index
1 Strategy After Tr.Cost − Committed Capital
0.8 Strategy After Tr.Cost − Invested Capital
0.6
0.4
0.2
−0.2
−0.4
1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010
Figure 3.4: Historical performance of the distance method of pairs trading on the S&P
500 companies for top 20 pairs — accumulated returns without reinvestment. Axes are
time and returns.
56
Mean 0.0051 0.0019 0.0042 0.0019 0.0075 0.0028 0.0059 0.0028
Median 0.0039 0.0017 0.0032 0.0017 0.0055 0.0037 0.0043 0.0036
Skewness 0.2849 0.3546 0.2757 0.3527 0.2444 0.2433 0.2364 0.2421
Kurtosis 4.7141 0.7978 4.7157 0.7988 3.6041 0.4937 3.5872 0.5001
Minimum -0.0789 0.0148 -0.0795 0.0147 -0.0953 0.0131 -0.0960 0.0129
Maximum 0.0979 0.0165 0.0967 0.0166 0.1296 0.0159 0.1279 0.0165
Trades per month 0.3106 0.0079 0.3106 0.0079 0.3106 0.0078 0.3106 0.0079
Sharpe Ratio 0.2135 0.0789 0.1762 0.0779 0.2114 0.0794 0.1682 0.0792
Modigliani RAP 0.0084 0.0032 0.0070 0.0031 0.0083 0.0033 0.0066 0.0032
Jensen’s Alpha 0.0054 0.0020 0.0044 0.0019 0.0079 0.0029 0.0063 0.0029
Market Beta -0.0668 0.0642 -0.0656 0.0637 -0.1242 0.0902 -0.1223 0.0895
by the cointegration method of pairs trading on the Australian market before and after
transaction cost.
ASX, top5 pairs

1.4
Market Index
1 Strategy After Tr.Cost − Invested Capital
0.8
0.6
0.4
0.2
−0.2
1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010
Figure 3.5: Historical performance of the cointegration method of pairs trading on the
57
Mean 0.0014 0.0012 0.0005 0.0012 0.0034 0.0019 0.0016 0.0019
Median 0.0013 0.0013 0.0004 0.0013 0.0020 0.0018 0.0006 0.0018
Skewness 0.0061 0.2998 -0.0054 0.2955 0.0779 0.2064 0.0749 0.2070
Kurtosis 3.8617 0.6401 3.8450 0.6290 3.2459 0.3904 3.2441 0.3860
Minimum -0.0422 0.0028 -0.0430 0.0028 -0.0577 0.0034 -0.0597 0.0035
Maximum 0.0580 0.0112 0.0567 0.0112 0.0804 0.0109 0.0783 0.0110
Trades per month 0.2778 0.0055 0.2778 0.0055 0.2778 0.0055 0.2778 0.0054
Sharpe Ratio 0.0883 0.0807 0.0331 0.0797 0.1413 0.0800 0.0669 0.0800
Modigliani RAP 0.0035 0.0032 0.0013 0.0031 0.0056 0.0032 0.0026 0.0032
Jensen’s Alpha 0.0013 0.0012 0.0005 0.0012 0.0033 0.0019 0.0015 0.0019
Market Beta 0.0067 0.0320 0.0073 0.0319 0.0229 0.0508 0.0230 0.0500
by the cointegration method of pairs trading on the Australian market before and after
transaction cost.
ASX, top20 pairs

1.2
Market Index
1 Strategy After Tr.Cost − Committed Capital
0.8
0.6
0.4
0.2
−0.2
1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010
58
Mean 0.0092 0.0024 0.0084 0.0024 0.0141 0.0039 0.0126 0.0039
Median 0.0062 0.0023 0.0055 0.0023 0.0105 0.0050 0.0089 0.0050
Skewness 1.4923 0.4599 1.4724 0.4598 1.2014 0.3394 1.1826 0.3383
Kurtosis 7.9924 1.8129 7.9305 1.8306 6.0013 1.2967 5.9130 1.2754
Minimum -0.0783 0.0183 -0.0790 0.0184 -0.0870 0.0059 -0.0878 0.0058
Maximum 0.1598 0.0213 0.1582 0.0215 0.2377 0.0285 0.2330 0.0283
Trades per month 0.2517 0.0088 0.2517 0.0089 0.2517 0.0088 0.2517 0.0088
Sharpe Ratio 0.3030 0.0671 0.2782 0.0679 0.2828 0.0689 0.2546 0.0705
Modigliani RAP 0.0149 0.0035 0.0137 0.0035 0.0139 0.0035 0.0125 0.0036
Jensen’s Alpha 0.0091 0.0023 0.0083 0.0023 0.0139 0.0038 0.0124 0.0038
Market Beta 0.1122 0.0603 0.1124 0.0594 0.2057 0.0947 0.2052 0.0920
by the cointegration method of pairs trading on the S&P 500 companies before and
after transaction cost.
SPX500, top5 pairs

2.5
Market Index
2
1.5
0.5
−0.5
1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010
S&P 500 companies for top 5 pairs — accumulated returns without reinvestment. Axes
59
Mean 0.0054 0.0015 0.0045 0.0015 0.0096 0.0024 0.0078 0.0024
Median 0.0014 0.0015 0.0004 0.0014 0.0047 0.0026 0.0027 0.0026
Skewness 1.0128 0.2250 1.0062 0.2224 1.1105 0.1944 1.1045 0.1925
Kurtosis 4.7469 0.6026 4.7387 0.5903 4.6714 0.6338 4.6412 0.6234
Minimum -0.0448 0.0081 -0.0456 0.0080 -0.0484 0.0045 -0.0498 0.0044
Maximum 0.0717 0.0042 0.0705 0.0042 0.1196 0.0066 0.1163 0.0062
Trades per month 0.2707 0.0056 0.2707 0.0057 0.2707 0.0057 0.2707 0.0057
Sharpe Ratio 0.2843 0.0714 0.2409 0.0718 0.3146 0.0688 0.2565 0.0703
Modigliani RAP 0.0140 0.0037 0.0119 0.0037 0.0155 0.0036 0.0126 0.0036
Jensen’s Alpha 0.0053 0.0015 0.0045 0.0015 0.0095 0.0023 0.0077 0.0023
Market Beta 0.0743 0.0395 0.0744 0.0389 0.1405 0.0579 0.1396 0.0585
by the cointegration method of pairs trading on the S&P 500 companies before and
after transaction cost.
SPX500, top20 pairs

1.6
Market Index
1
0.8
0.6
0.4
0.2
−0.2
−0.4
1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010
S&P 500 companies for top 20 pairs — accumulated returns without reinvestment.
Axes are time and returns.
60
Mean 0.0049 0.0008 0.0043 0.0008 0.0160 0.0030 0.0136 0.0030
Median 0.0025 0.0008 0.0021 0.0008 0.0102 0.0031 0.0085 0.0030
Skewness 2.1043 0.5076 2.0907 0.5184 2.3792 0.5678 2.3740 0.5737
Kurtosis 10.6010 2.8574 10.7390 2.9344 12.9308 3.0220 13.0129 3.0503
Minimum -0.0205 0.0041 -0.0213 0.0041 -0.0669 0.0126 -0.0695 0.0124
Maximum 0.0625 0.0100 0.0613 0.0100 0.2262 0.0284 0.2223 0.0283
Trades per month 0.3256 0.0093 0.3256 0.0094 0.3256 0.0094 0.3256 0.0093
Sharpe Ratio 0.4750 0.0598 0.4160 0.0599 0.4235 0.0595 0.3636 0.0596
Modigliani RAP 0.0187 0.0027 0.0164 0.0027 0.0167 0.0026 0.0143 0.0026
Jensen’s Alpha 0.0051 0.0008 0.0044 0.0008 0.0165 0.0031 0.0140 0.0031
Market Beta -0.0459 0.0372 -0.0447 0.0364 -0.1414 0.1429 -0.1377 0.1397
by the stochastic spread process method of pairs trading on the Australian market
before and after transaction cost.
ASX, top5 pairs

3
Market Index
2
1.5
0.5
−0.5
1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010
Figure 3.9: Historical performance of the stochastic spread process method of pairs
trading on the Australian market for top 5 pairs — accumulated returns without rein-
vestment. Axes are time and returns.
61
Mean 0.0019 0.0005 0.0013 0.0005 0.0122 0.0030 0.0083 0.0030
Median 0.0021 0.0003 0.0016 0.0003 0.0138 0.0025 0.0104 0.0026
Skewness 0.0634 0.5723 -0.0176 0.5606 -0.3275 0.4148 -0.3511 0.4096
Kurtosis 6.6936 1.2041 6.6346 1.1399 5.4098 0.8021 5.4008 0.8195
Minimum -0.0203 0.0030 -0.0211 0.0031 -0.1316 0.0214 -0.1357 0.0219
Maximum 0.0274 0.0048 0.0261 0.0045 0.1290 0.0083 0.1221 0.0074
Trades per month 0.2877 0.0042 0.2877 0.0041 0.2877 0.0041 0.2877 0.0041
Sharpe Ratio 0.3216 0.0850 0.2186 0.0827 0.3247 0.0888 0.2202 0.0846
Modigliani RAP 0.0127 0.0034 0.0086 0.0033 0.0128 0.0037 0.0087 0.0034
Jensen’s Alpha 0.0019 0.0005 0.0013 0.0005 0.0123 0.0030 0.0084 0.0030
Market Beta -0.0085 0.0207 -0.0076 0.0205 -0.0405 0.1049 -0.0392 0.1046
by the stochastic spread process method of pairs trading on the Australian market
ASX, top20 pairs

2
Market Index
0.5
−0.5
1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010
trading on the Australian market for top 20 pairs — accumulated returns without
reinvestment. Axes are time and returns.
62
Mean 0.0004 0.0006 -0.0001 0.0006 0.0012 0.0018 -0.0009 0.0018
Median 0.0001 0.0004 -0.0004 0.0004 0.0009 0.0011 -0.0006 0.0011
Skewness 3.8591 1.9666 3.6363 1.9141 1.0016 0.8098 0.9085 0.7806
Kurtosis 34.6450 14.5307 32.5731 13.5797 9.7468 3.1332 9.3875 2.8965
Minimum -0.0167 0.0016 -0.0174 0.0013 -0.0749 0.0106 -0.0770 0.0106
Maximum 0.0632 0.0213 0.0606 0.0203 0.1317 0.0315 0.1256 0.0297
Trades per month 0.2775 0.0075 0.2775 0.0075 0.2775 0.0075 0.2775 0.0075
Sharpe Ratio 0.0592 0.0751 -0.0184 0.0861 0.0513 0.0780 -0.0388 0.0809
Modigliani RAP 0.0029 0.0037 -0.0009 0.0042 0.0025 0.0038 -0.0019 0.0040
Jensen’s Alpha 0.0004 0.0006 -0.0001 0.0006 0.0011 0.0018 -0.0009 0.0018
Market Beta 0.0028 0.0202 0.0038 0.0201 0.0402 0.0513 0.0415 0.0505
by the stochastic spread process method of pairs trading on the S&P 500 companies
SPX500, top5 pairs

0.6
Market Index
0.3
0.2
0.1
−0.1
−0.2
−0.3
−0.4
1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010
trading on the S&P 500 companies for top 5 pairs — accumulated returns without
63
Mean 0.0012 0.0003 0.0006 0.0003 0.0078 0.0026 0.0036 0.0026
Median 0.0008 0.0002 0.0003 0.0002 0.0065 0.0019 0.0024 0.0019
Skewness 1.3698 0.4453 1.2709 0.4284 0.5799 0.4620 0.5614 0.4606
Kurtosis 7.7777 1.9584 7.3464 1.8141 6.5291 1.2226 6.4874 1.1809
Minimum -0.0091 0.0006 -0.0096 0.0006 -0.0939 0.0088 -0.0976 0.0089
Maximum 0.0234 0.0042 0.0219 0.0039 0.1528 0.0227 0.1469 0.0221
Trades per month 0.2938 0.0043 0.2938 0.0042 0.2938 0.0043 0.2938 0.0043
Sharpe Ratio 0.2653 0.0699 0.1300 0.0742 0.2384 0.0773 0.1105 0.0777
Modigliani RAP 0.0131 0.0036 0.0064 0.0037 0.0117 0.0038 0.0054 0.0038
Jensen’s Alpha 0.0012 0.0004 0.0006 0.0003 0.0077 0.0026 0.0035 0.0025
Market Beta 0.0023 0.0110 0.0029 0.0106 0.1142 0.0619 0.1141 0.0619
by the stochastic spread process method of pairs trading on the S&P 500 companies
SPX500, top20 pairs

1.4
Market Index
1 Strategy After Tr.Cost − Invested Capital
0.8
0.6
0.4
0.2
−0.2
−0.4
1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010
trading on the S&P 500 companies for top 20 pairs — accumulated returns without
64
Top 5 pairs
ASX SP500
distance cointegr. st.spread distance cointegr. st.spread
Average Profit 0.0185 0.0132 0.0109 0.0077 0.0250 -0.0018
St.Dev 0.0947 0.1465 0.0462 0.0832 0.1568 0.0389
Negative trades 31.9% 32.8% 42.0% 36.3% 38.1% 55.1%
Average Positive 0.0737 0.0991 0.0439 0.0611 0.1277 0.0323
Average Negative -0.1019 -0.1577 -0.0346 -0.0866 -0.1506 -0.0290
Average Max Profit 0.1210 0.1673 0.0826 0.0926 0.2215 0.0609
Average Max Loss -0.1559 -0.2753 -0.0616 -0.1394 -0.2294 -0.0575
Trades per 6 months 8.2 9.3 9.7 7.7 7.4 8.3
Av. Holding time 49.6 49.3 9.5 54.7 56.3 9.3
Av. Max time 110.0 112.5 15.5 112.2 110.1 15.4
Av. Min time 7.6 8.0 5.2 11.8 14.5 5.5
Top 20 pairs
ASX SP500
distance cointegr. st.spread distance cointegr. st.spread
Mean 0.0111 0.0050 0.0059 0.0153 0.0177 0.0029
St.Dev. 0.1388 0.1754 0.0657 0.1052 0.1751 0.0470
Negative trades 36.1% 37.3% 44.6% 34.9% 37.3% 47.9%
Average Positive 0.0953 0.1119 0.0508 0.0767 0.1239 0.0370
Average Negative -0.1383 -0.1737 -0.0497 -0.0997 -0.1659 -0.0342
Max Profit 0.2055 0.2798 0.1470 0.1906 0.3236 0.1110
Max Loss -0.3855 -0.5027 -0.1593 -0.2855 -0.4737 -0.1115
Trades per 6 months 30.9 33.2 34.3 32.6 32.1 35.3
Holding time 54.8 52.0 11.2 51.6 52.5 9.0
Max time 124.2 123.2 24.4 123.2 122.8 20.3
Min time 3.8 3.8 3.7 3.8 3.7 3.2
Table 3.15: Individual trade statistics based on trading top 5 and top 20 pairs for the
distance, cointegration and stochastic spread process methods
Mean – an average cash flow in cents from $1 trade on the spread process (that is, an excess re-
turn); St.Dev. – an average standard deviation of the individual trades excess returns during the 6
month trading period; Negative trades – an average proportion of trades with negative returns before
transaction costs; Average Positive / Average Negative – an average cash flow or excess return of the
profitable / losing trades; Max Profit / Max Loss – an average maximal profit / loss from one trade;
Trades per 6 months – an average number of trades made during 6 month trading period by the top 5
or top 20 pairs; Holding time – an average time from opening position on the spread till closing; Max
time / Min time – an average maximal and minimal time of the holding an open position.
65
Chapter 4
Study 2. Pairs Trading Based on

Statistical Variability of the Spread
Process
This chapter reviews a method of time-series statistical analysis proposed by Pastukhov

(2005) and applied to Brownian motion. Then it adapts the method to the Ornstein–
Uhlenbeck process and proves the theorem about new properties of the Ornstein–
Uhlenbeck process. This chapter then proposes a pairs trading strategy based on the
developed theory and tests it on four data sets from the US and Australian mar-
kets1,2 .
4.1 Introduction
Pairs trading is a form of technical analysis strategy known since the 1990s and popular
amongst institutional and individual investors. The strategy is believed to be market
neutral and provide small but constant returns with low standard deviations. A de-
scription of pairs trading and different approaches to it can be found in many articles
1
Bogomolov, T. ‘Pairs trading based on statistical properties of the volatility of the spread process’,
presented at Quantitative Methods in Finance Conference, Sydney, December 2011.
2
Bogomolov, T. ‘Pairs trading based on statistical properties of the volatility of the spread process’,
under review in Quantitative Finance.
66
and books, see among others Gatev et al. (2006); Vidyamurthy (2004); Elliott et al.
(2005); Do et al. (2006); Herlemont (2004).
The general idea of pairs trading is simple: (1) find two assets that historically have
moved together; (2) when they move apart, take a short (sell) position on the higher
priced asset (‘winner’) and long position (buy) on the lower priced asset (‘loser’); (3)
unwind the positions when the assets converge together.
The pairs trading strategy might be viewed as a trading of the synthetic asset (spread
process or long-short portfolio) formed by a short position on one stock and a long po-
sition on another. It is also possible to create a spread process by using more than two
assets or construct a spread between shares portfolio and a market index (Alexander
and Dimitriu, 2005).
To implement a pairs trading strategy, one should answer the following three questions:
1. What stocks should be combined in pairs? — The stage of pairs formation.
2. How far should stocks deviate from each other to initiate a trade? In other words,
how far should the spread process move away from its mean before one opens
positions? — Rules to open positions.
3. To what degree should stocks converge to unwind positions and what is the
strategy if convergence does not occur? — Rules to close positions.
Each method of pairs trading in the literature provides its own set of rules for pairs
formation and trading. Despite some differences, all methods are based on the same
idea of price (or returns) equilibrium: two similar assets must provide similar returns.
Therefore, any deviations from the equilibrium are the result of market over- or under-
reaction on some news and/or market mispricing of one or both stocks in a pair. The
general assumption of pairs trading strategies is that these deviations are temporary
and will be corrected over time.
67
A similar idea of prices or returns equilibrium comes from cointegration theory (Engle
and Granger, 1992): a spread process between returns of the two cointegrated stocks
should be a stationary process. If the spread process deviates from its long-run mean,
it should return back to the mean.
However, in reality this reversion does not always happen. In practice, it can be ob-
served that the parameters of a spread process may change dramatically and shift the
process up or down from its previous long-term mean. This could happen as a result
of some news or events related to only one of the stocks from a pair. Recovery from
that shock may take longer than the investment horizon or may never happen at all.
As a result, a pairs trading strategy based only on the idea of the return to equilibrium
may be unprofitable. Testing existing pairs trading strategies on the market data (Do
and Faff, 2010) confirms this observation — returns after accounting of transaction
costs are minimal and are not consistent over time.
Two approaches exist to target the problem of non-constant mean. One of them uses
moving averages (MA) with some fixed period instead of the long-term mean. This
idea is used by many practitioners, but has a limitation which is common for all models
based on MAs — too large a lag between an event and reaction of the model to that
event.
Another approach is a regime switching model (Bock and Mestel, 2009; Wu and Elliott,
2005) which allows the mean to jump between different levels. Limited research has
been done in this area, and it is not clear if it is possible to recognise switches and new
parameters of the spread process model quickly enough to adapt the trading strategy.
I propose a new nonparametric method of pairs trading based on some statistical prop-
68
1
0.9
0.8
0.7
0.6
0.5
May 96 Jul 97 Sep 98 Oct 99 Dec 00 Feb 02 Mar 03 May 04 Jul 05 Aug 06 Oct 07 Nov 08 Jan 10 Mar 11
Figure 4.1: Log prices spread process between two major Australian banks — Com-
monwealth (CBA) and Westpac (WBC).
The red line is 200 days moving average and blue lines are the same moving average
shifted 10% up and down from the true location.
erties of the spread process. It does not try to find and follow a mean of the process.
Instead, it utilises information about variability of the spread process, and the only
assumption made is that the level of variability remains reasonably constant.
The general idea is simple. Suppose one trades an asset which is suspected to have
some mean-reverting property. The further the asset price moves in one direction, the
higher the probability it reverses. The question is to define how far the price should
move in one direction before trading in the opposite direction becomes potentially prof-
itable. Obviously this depends on a number of parameters, but the most important
one is variability of the asset price or, in the case of pairs trading, the spread process.
Figure 4.1 illustrates an idea of relatively constant variability of the spread process and
uncertainty of the mean location.
The tool used to measure the variability of the spread process is based on renko and
kagi constructions proposed by Pastukhov (2005). Renko and kagi are types of charts
originating from 19th century Japan and are well-known to all adepts of technical anal-
ysis on the financial markets. Both charts are concerned with price movements greater
than some given threshold and do not include information about time or trading vol-
69
umes. This approach is believed to filter out the trading noise — small changes in asset
prices — and focus only on significant price movements. Pastukhov (2005) introduces
renko and kagi to the world of academic research and provides a mathematical basis
for these methods of technical analysis. His research describes two possible trading
strategies based on the statistical properties of the renko and kagi constructions built
on a real asset.
I consider the use of the renko and kagi constructions for different types of processes
and extend their use from real assets to the pairs trading spread processes. I then
provide theoretical proofs of the profitability of the proposed method for the case of
the Ornstein–Uhlenbeck process and test it on real market data from the US and Aus-
tralian stock exchanges.
Section 4.2 provides a brief review of the method proposed by Pastukhov (2005) — the
renko and kagi constructions, their properties, constructions on the Wiener process and
two possible trading strategies. Section 4.3 considers the renko and kagi constructions
on the Ornstein–Uhlenbeck process and on discrete time processes. Section 4.4 provides
details of the practical implementation and real data tests of the proposed pairs trading
strategy. Section 4.5 reports the results of the testing and section 4.6 presents the
conclusions.
4.2 Method of renko and kagi constructions
4.2.1 Renko construction
Let P (t) be a time series of the actual asset prices or asset cumulative returns on the
time interval [0, T ]. At this stage, I assume that P (t) is continuous. Let τi , i = 0, 1, ..., N
be an increasing sequence of random times such that for some arbitrary H > 0
H ≤ max P (t) − min P (t) (4.1)

t∈[0,T ] t∈[0,T ]
70
4
4
3.5 3.5
3 3
2.5 2.5
2 2
1.5 1.5
1 1
0 1000 2000 3000 4000 5000 0 1 2 3 4 5 6 7 8
time, t counts, i
(a) Asset price process P (t) with points at τi (b) Classical renko process X(i)
Figure 4.2: Renko chart
and for τ0 = 0, P (τ0 ) = P (0)
τi = inf{u ∈ [τi−1 , T ] : |P (u) − P (τi−1 )| = H}. (4.2)
The process X(i) : X(i) = P (τi ), i = 0, 1, ..., N is a ‘classical’ renko chart (Figure 4.2)
or renko process.
Another sequence of random times moments {(τna , τnb ), n = 0, 1, ..., M } is now created
based on the sequence {τi }. The sequence {τna } defines time moments when the renko
process X(i) has a local maximum or minimum, that is, the process X(i) = P (τi )
changes its direction, and the sequence {τnb } defines the time moments when the local
maximum or minimum is detected.
More precisely, when take τ0a = τ0 and τ0b = τ1 then
τnb = min{τi > τn−1

b
: (P (τi ) − P (τi−1 ))(P (τi−1 ) − P (τi−2 )) < 0}, (4.3)
τna = {τi−1 : τnb = τi }. (4.4)
71
4
local extremum, an
a
stopping time when we recognise local extremum, n
3.5
2.5
1.5
a b a b
n n n+1 n+1
1
time, t
Figure 4.3: Renko construction
b
If τnb = τi then τna = τi−1 . In some cases τn−1 and τna may be equal to each other,
as they are derived from discrete process X(i) and the point when we detect a local
maximum may happen to be a next local minimum (Figure 4.3).
4.2.2 Kagi construction
The kagi construction is similar to the renko construction with the only difference be-
ing that to create the sequence of time moments {(τna , τnb ), n = 0, 1, ..., M } for the kagi
construction local maximums and minimums of the actual asset price process P (t) is
used rather than the process X(i) derived from it.
The sequence {τna } then defines the time moments when the price process P (t) has a
local maximum or minimum and the sequence {τnb } defines the time moments when
that local maximum or minimum is recognised, that is, the time when the process P (t)
moves away from its last local maximum or minimum by a distance equal to H.
72
More precisely, for some arbitrary H > 0 satisfying (4.1) we define
τ0b = inf{u ∈ [0, T ] : max P (t) − min P (t) = H} (4.5)

t∈[0,u] t∈[0,u]
and
τ0a = inf{u < τ0b : |P (u) − P (τ0b )| = H}. (4.6)
It is important to know whether τ0a defines a local maximum or a minimum. The

variable
S0 = sign(P (τ0a ) − P (τ0b )) (4.7)
can take two values: 1 for a local maximum and −1 for a local minimum.
I then define (τna , τnb ), n > 0 recursively. If at time τ0a I have a local maximum (S0 = 1)
then all odd numbered time moments (τna , τnb ), n = 1, 3, 5, 7, ... relate to local minimums
where Sn = −1, n = 1, 3, 5, 7, ... and should be defined by
τnb = inf{u ∈ [τn−1

a
, T ] : P (u) − min P (t) = H}
a
t∈[τn−1 ,u]
τna = inf{u < τnb : P (u) = min

a
P (t)} (4.8)
t∈[τn−1 ,τnb ]
and all even numbered time moments (τna , τnb ), n = 2, 4, 6, ... relate to local maximums
where Sn = 1, n = 2, 4, 6, ... and should be defined by
τnb = inf{u ∈ [τn−1

a
, T ] : max
a
P (t) − P (u) = H}
t∈[τn−1 ,u]
τna = inf{u < τnb : P (u) = max

a
P (t)}. (4.9)
t∈[τn−1 ,τnb ]
The construction of the full sequence {(τna , τnb ), n = 1, 2, 3, ..., N } is done in the induc-
tive manner alternating steps (4.8) and (4.9).
As the sequence {(τna , τnb )} is derived from the continuous process {P (t)} the probability
b
of τn−1 = τna is zero, even though they can be close to each other (Figure 4.4).
73
4
local extremum, an
a
stopping time when we recognise local extremum, n
3.5
2.5
2
a
n+1
b
1.5 n+1
a b
n n
1
time, t
Figure 4.4: Kagi construction
4.2.3 Some properties of renko and kagi constructions
For the following discussion the term H-construction is used when refering to either
renko or kagi constructions as their properties are similar.
The process P (t), t ∈ [0, T ] will be defined on some probability space (Ω, F, P), tak-
ing values in R. For some arbitrary H, the increasing time sequence {(τna , τnb ), n =
0, 1, ..., N } defined as above (see sections 4.2.1 and 4.2.2).
Obviously, τna are not stopping times as the local maximum or minimum can be defined
only post-factum at times τnb which are stopping times. To simplify the calculation, it
is assumed that T = τNb for some arbitrary N , which means that any stopping time τnb
might be considered as the end of the trading period, that is, trading can be stopped
at this time.
I now list some useful variables introduced by Pastukhov (2005):
H-inversion counts the number of times the process P (t) changes its direction for
74
selected H, T and P (t) and is given by
NT (H, P ) = max{n : τnb = T } = N. (4.10)
H-volatility of order p is a measure of the variability of the process P (t) for selected H
and T and is given by
VTp (H, P )
ξTp (H, P ) = , (4.11)
NT (H, P )
where VTp (H, P ) is a sum of vertical distances between local maximums and minimums
to the power p
N
X
VTp (H, P ) = a
|P (τna ) − P (τn−1 )|p . (4.12)
n=1
H -volatility of order 2 is similar to variance and can be used to describe the process
P (t). However, for the purpose of this research it is sufficient to know H -volatility of
order 1 only:
VT (H, P )
ξT (H, P ) = . (4.13)
NT (H, P )
Pastukhov (2005) shows that for a Wiener process {W (t)} the condition ξT (H, W ) =
2H holds for any value of H, subject to (4.1). More specifically
(p)
lim ξT (H, σW ) = RW (p)H p , (4.14)
T →∞
where 
 P∞ nnp ,

for renko construction;
n=1 2
RW (p) =
 ∞ (1 + x)p e−x dx, for kagi construction.
 R
0
For both constructions RW (1) = 2. So, ξT (H, W ) = 2H.
4.2.4 Trading strategies
Corresponding to the definition of H -construction, the term H-strategy is defined with-

out specifying renko or kagi H -construction, as the differences between renko and kagi
strategies are minor and are not important for this research.
75
There are two possible H -strategies — momentum and contrarian.
1. The trend following or momentum strategy: here, the investor buys (sells)
an asset at a stopping time τnb when he or she recognises that the process passed its
previous local minimum (maximum) and the investor expects a continuation of the
movement. There are two types of trading signals which are equivalent:
P (τnb ) − P (τna ) > 0 or P (τn−1

a
) − P (τna ) > 0 buy signal
P (τnb ) − P (τna ) < 0 or P (τn−1

a
) − P (τna ) < 0 sell signal.
The profit from one trade according to the trend following H -strategy over time from
b
τn−1 to τnb is
Yτnb = (P (τnb ) − P (τn−1
b
)) · sign(P (τna ) − P (τn−1
a
)) (4.15)
and the total profit from time 0 till time T is
YT (H, P ) = (ξT (H, P ) − 2H) · NT (H, P ). (4.16)
2. The contrarian strategy: here the investor sells (buys) an asset at a stopping time
τnb when he or she decides that the process has passed far enough from its previous local
minimum (maximum), and the investor expects a movement reversion. The trading
signals are
P (τnb ) − P (τna ) > 0 or P (τn−1

a
) − P (τna ) > 0 sell signal
P (τnb ) − P (τna ) < 0 or P (τn−1

a
) − P (τna ) < 0 buy signal.
The profit from the one trade according to the contrarian H -strategy over time from
b
τn−1 to τnb is the same as (4.16) but with a negative sign.
Yτnb = (P (τnb ) − P (τn−1

b a
)) · sign(P (τn−1 ) − P (τna )) (4.17)
76
and the total profit till time T is
YT (H, P ) = (2H − ξT (H, P )) · NT (H, P ). (4.18)
As it can be seen, trading signals for both strategies are the same, but point in different
directions. The investor constantly stays in the market for any strategy, only changing
the direction of the trade. At this stage, it is assumed that the investor can trade long
and short with no restrictions, and that the transaction costs are zero.
It clearly follows from (4.16) and (4.18) that the choice of H -strategy depends on the
value of H -volatility, ξT (H, P ). If ξT (H, P ) > 2H, then to achieve a positive profit
the investor should employ a trend following H -strategy; if ξT (H, P ) < 2H then the
investor should use a contrarian H -strategy.
From Pastukhov (2005), it is known that for the Wiener process, the H -volatility
ξT (H, W ) = 2H and, as a result, it is impossible to profit from trading H -strategy on
the Wiener process. It seems that the same result is true for any Lévy processes which
have independent increments symmetrically distributed around zero, regardless of the
shape of their distribution.
We can induce from the above that H -volatility ξT (H, P ) = 2H is a property of a mar-
tingale. Likewise, ξT (H, P ) > 2H could be viewed as a property of a sub-martingale
or a super-martingale or a process regularly switching over time from a sub-martingale
to a super-martingale and back. It is unlikely that these sorts of processes exist in
financial markets. Pastukhov (2005) does not provide any examples of processes for
which ξT (H, P ) > 2H.
From a practical point of view, a more interesting situation occurs if H -volatility is

less than 2, that is ξT (H, P ) < 2H. The obvious example of such a process could be
an Ornstein–Uhlenbeck process (Uhlenbeck and Ornstein, 1930; Finch, 2004) and by
77
extension, any mean-reverting process regardless of the distribution of its innovations.
The condition H -volatility less than 2H is a statistical property of a process P (t) and
can be considered a very mild restriction. It does not require the process P (t) to be
mean-reverting in the formal definition and have a constant mean and variance. To
create a profitable trading strategy one simply needs the process P (t) to have a mean-
reverting property over certain time intervals. A perfect candidate for such process
P (t) is a pairs trading stochastic spread process (Elliott et al., 2005).
4.3 More on the renko and kagi constructions
4.3.1 Properties of H -constructions on the Ornstein–Uhlenbeck
process
I now consider H -constructions made over the Ornstein–Uhlenbeck process:
dXt = −ρ(Xt − µ) dt + σ dBt , (4.19)
where {Bt : t ≥ 0} is a standard Brownian motion and ρ > 0, σ > 0, µ are constants.
In most situations without loss of generality, we can assume µ = 0 and σ = 1 (by using
Xt − µ rather then Xt and by time scaling). Then, (4.19) takes the form
dXt = −ρXt dt + dBt . (4.20)
The Ornstein–Uhlenbeck process has finite variance Var(Pt ) = σ 2 /2ρ. If we take
H = max P (t) − min P (t),
then the local minimum and local maximum in (4.5) are equal to time series global (or
absolute) minimum and maximum correspondingly. As a result, the rule of building
78
kagi construction (4.5) can be satisfied just once, and the rule for renko construction
would never be satisfied. Hence, we get no more than one swing between maximum
and minimum and it equals H, then the H-volatility
ξT (H, P ) = H.
If we increase the value of H, then (4.1) does not hold and we cannot build the H-
construction. Furthermore, if we take H → 0, then it is equivalent to ρ → 0 for fixed
H. As a result, the Ornstein–Uhlenbeck process converges to the Wiener process and
the H-volatility ξT (H, P ) → 2H as H → 0. Hence, for the Ornstein–Uhlenbeck process
ξT (H, P )
∈ [1, 2). (4.21)
H
I have executed multiple numerical simulations, and all of them support the above
intuition. Some concerns regarding the H-constructions over the simulations of the
Ornstein–Uhlenbeck process and their H-volatility are presented in the section 4.3.2.
The following theorem provides a theoretical justification for the upper bound of the
H-volatility of the Ornstein–Uhlenbeck process.
Theorem 4.1. Let P be an Ornstein–Uhlenbeck process. Then for any positive H

satisfying (4.1), the H-volatility is less than 2H
lim ξT (H, P ) < 2H. (4.22)

T →∞
Proof. It is provided in Appendix B.
Hence, trading the Ornstein–Uhlenbeck process by the contrarian H-strategy is prof-

itable for any choice of H (4.18). The same is true for any mean-reverting process
regardless of the distribution of its innovations.
79
4.3.2 H -construction on the discrete process
Most financial data are discrete. Therefore, it is important to consider the properties
of H -construction on the discrete process.
Random Walk
Let {X(t)}, t = 0, 1, 2, ... be an independent and identically distributed increments

process, X(t) ∼ N (0, σ 2 ). Then the sum of increments is a random walk
t
X
Y (t) = X(i).
i=0
If we build the H -construction on the discrete process Y (t) then conditions (4.2), (4.8),
(4.9) do not hold, as
P(|Y (t) − Y (t + n)| = H) = 0 ∀ n, t ≥ 0.
For the discrete process Y (t) we have the following condition at the stopping time τnb
|Y (τna ) − Y (τnb )| = H̃n ≥ H, (4.23)
where H̃n is an independent random variable.
This means that we get an overshot and as a result the ratio of H -volatility to the
parameter H gets inflated.
ξT (H, Y ) = 2E[H̃n ] ≥ 2H. (4.24)
Hence, when applied to real world scenarios, it is quite possible to observe H -volatility
greater than 2H. However, this does not imply that the underlying process is not a
martingale and that it would be possible to trade it with a trend-following H -strategy.
80
The size of overshot depends on the value of H and on the standard deviation of the
increments process X(t). If H/σ → ∞ then E[H̃n ] → H.
AR(1) process
Now I consider AR(1) process which is a discrete representation of the Ornstein–

Uhlenbeck process.
Y (t) = αY (t − 1) + X(t),
where α ∈ [0, 1) and X(t) ∼ N (0, σ 2 ).
The AR(1) process has the same problem as the random walk. As the AR(1) process
is not continuous, then
P(|Y (t) − Y (t + n)| = H) = 0 ∀ n, t ≥ 0.
As a result we get an overshot at each stopping time
|Y (τna ) − Y (τnb )| = H̃n ≥ H. (4.25)
The difference is: the value of H for AR(1) cannot go to infinity as the AR(1) process
is bounded by some global minimum and maximum (Novikov and Kordzakhia, 2008).
Hence, while, similar to the random walk case, the value of E[H̃n ] gets closer to H as
the value of H increases, the value of E[H̃n ] does not converge to H. It follows that
it is possible to observe a discrete mean-reverting process with ξT (H, Y ) ≥ 2H and it
does not contradict Theorem 4.1.
The true value of the ratio of the H -volatility to the parameter H is
ξ(H, P )
R(H, P ) = ∈ [1, 2), (4.26)
E[H̃n (H, P )]
1
PN
where E[H̃n (H, P )] = N n=0 |P (τna ) − P (τnb )|.
81
The renko H -strategy is the most effected by the overshot, as it has to generate a
renko process — a sequence of stopping times with the fixed price step H. Obviously,
it is quite problematic to do it with the discrete time series, especially if the standard
deviation of the increments is comparable to the value of H. That is why I ran the
test for kagi H -strategy only, which is more appropriate for the daily data that I had.
However, overshot is not always a bad thing. It might improve the profitability of the
contrarian trading strategy. The overshot always happens in the direction of the price
movement and in a contrarian strategy, one always trades in the direction opposite to
the last movement of the spread process. It means that in the case of the overshot, the
investor sells at the higher price and buys at lower price than one would if trades as
per the continuous process.
Choice of parameters H and T for AR(1) process
It is clear that for any given process P (t) the investor can control only two parameters:
H and T . There are several important considerations related to their choice.
1. Value of H should be reasonably large to minimise the overshot problem and im-
prove ration ξT (H, P )/H, which one would like to be close to 1.
2. At the same time, H cannot be too large. I have demonstrated the profit from the
trading of the Ornstein–Uhlenbeck process {P (t)} by the contrarian H -strategy (4.18).
If we consider the discrete process and some transaction cost λ then the profit is
YT (H, P ) = (2E[H̃n (H, P )] − ξT (H, P ) − λ) · NT (H, P )

λ
= 2 − R(H, P ) − · NT (H, P ) · E[H̃n (H, P )]. (4.27)
E[H̃n (H, P )]
On the one hand, from equation (4.27), it follows that if the value of H increases, the
82
expectation E[H̃n (H, P )] increases, the ratio R(H, P ) decreases (4.26) and the ratio of
transaction costs to the expectation of H decreases. As a result, the profit increases.
On the other hand, as H increases the H -inversion NT (H, P ) — the number of trades
— decreases and dramatically reduces profit. Hence, there exists some optimal H
which maximises the profit for any given P (t), T and transaction costs λ.
3. The H -strategy exploits a statistical property of the process, so we need a sample

size to be large enough to give us some confidence that the observed process behaviour
is not a result of random fluctuation. Number of observations equals H -inversion of
the spread process. So, the length of the history period T used to calibrate the trad-
ing strategy needs to be quite large and/or the value of H small enough to maximise
H -inversion NT (P, H).
For testing, I arbitrarily choose H equal to one standard deviation of the spread process
and the length of the calibration period of one year. This provides around 30 trades
over the history period for the best pairs.
A more rigorous approach to the selection of H is possible. The first idea is not to use
the constant but adaptive H. As the H -volatility ξT (P, H) is a measure of variability
of the process P (t), it looks promising to use a GARCH model to predict the size of
the swing between the next local maximum and minimum and adjust H accordingly.
Another possibility is to analyse the distribution of {Xn }
Xn = |P (τna ) − P (τn−1
a
)|. (4.28)
It is clear that {Xn } ≥ H and the distribution is skewed to the right. It follows from
(4.18) that Xn > 2H means negative profit in the nth trade. So, during pairs formation
we can select pairs with ‘lighter’ right tale and minimal number of observations with
Xn > 2H.
83
4.4 Pairs trading by the contrarian H -strategy
Each pairs trading strategy defines three steps: pairs formation, rules to open position
on the spread, and rules to close position. For the H -strategy the last two points are
combined into one. The signal to close position acts as the signal to open a new one
in the opposite direction.
This section provides a description of the trading strategy based on the kagi construc-
tions as well as details of the data set and testing method. Important consideration
regarding potential data snooping bias and choice of the benchmarks discussed in sec-
tions 3.3.5 and 3.3.4 are equally applicable to this study.
4.4.1 Data sets
I used the ASX, NYSE and NASDAQ daily closing prices to test H -strategies of pairs
trading. The data are obtained from the Securities Industry Research Centre of Asia-
Pacific (SIRCA). I use market indexes as benchmark for the reasons discussed in section
3.3.4. There are four data sets:
1. The ASX data set covers 3,863 trading days starting from January 2, 1996 and
finishing on April 6, 2011 and includes more than 3,000 shares traded at different
times. I use an index S&P/ASX 200 as a benchmark for this data set.
2. The top 500 American companies by market capitalisation included in the S&P
500 index, which is also used as a benchmark. The data set covers 3,853 trading
days from January 2, 1996 to April 29, 2011.
Number of companies preselected Total Trading

Mean St.Dev. Median Max Min trading days months
Australian market 114.1 51.8 103 244 41 3863 160
S&P 500 market 455.3 29.0 450 541 392 3853 160
S&P 400 Mid Cap 383.7 16.1 383 479 352 2592 100
S&P 600 Small Cap 563.1 36.7 562 766 475 2842 112
Table 4.1: Statistics on data sets used in the Australian and US markets
84
3. The medium capitalisation companies included in the S&P 400 Mid Cap index,
which is also used as a benchmark. The data set covers 2,592 trading days from
January 2, 2001 to April 29, 2011.
4. The small capitalisation companies included in the S&P 600 Small Cap index,
which is also used as a benchmark. The data set covers 2,842 trading days from
January 3, 2000 to April 29, 2011.
Following the testing methodology in Gatev et al. (2006), a 12 month history is used
to calibrate the system and construct pairs and the next 6 months to trade selected
pairs. I start trading at the first working day of each month and trade until the last
working day of the trading period. In each month, except for the first and the last
five months, I get six different estimations of the monthly returns which are averaged
to get the final estimation. Hence, the actual time interval used for testing is shorter
than the length of the data sets due to the 12 month historical data used for strategy
calibration and the first and last five months of the actual testing period disregarded
due to averaging.
The testing interval includes the Global Financial Crisis (GFC). Short selling was
banned at some period of time during the GFC. However, I run the test of pairs
trading over that period as usual. Institutional investors, who hold large diversified
portfolios, could still use pairs trading as a part of the tactical asset allocation strategy.
They do not need short selling to fulfill the rules of pairs trading; they can sell some
shares from the existing portfolio and buy them back when the strategy signals to close
position on the pair.
4.4.2 Stocks pre-selection
From the ASX data set, I only pre-select the top 30% of companies by their dollar
valued trading volume during the 12 month history period used for system calibration.
The S&P 500 data set also includes only large cap stocks. That ensures that I run the
strategy test over the most liquid companies on the day of the start of trading period,
85
which, most probably, could be traded at the prices used for testing. Thus, I expect
to avoid a potential liquidity problem.
On the downside, this approach makes the testing biased towards large cap companies.
It could be expected that medium and small cap companies have a higher probabil-
ity of being mispriced than the large cap companies which attract greater attention
of institutional and individual traders. As a result, potential profit could be higher.
Testing of the same pairs trading strategy on the ASX data with inclusion of a broader
range of stocks shows much higher returns. However, the results can be unreliable due
to liquidity problems, thus, are not reported here.
The last two data sets — companies from S&P 400 Mid Cap and S&P 600 Small Cap
— are included to test if medium and small capitalisation companies provide more
opportunities for pairs trading. It looks to be a reasonable compromise between the
opportunity to test the strategy on stocks with smaller capitalisation and possible liq-
uidity issues. As the American market is the most liquid stock market in the world, I
expect that the effect of the liquidity problem is minor if any. However, one should be
aware of the possible limitation of the presented results.
I put a restriction on not more than 10 non-trading days during the calibration period.
It is a milder condition than ‘zero non-trading days’ in Gatev et al. (2006). I believe it
is reasonable as even the strongest and most liquid stocks can halt trading once or twice
a year. I aim to employ a pure quantitative approach to the testing with a minimal
number of constrain, so I do not put any extra restrictions on the pairs selection, for
example, sectors.
The opening and closing prices are the results of auctions, which usually attract a large
trading volume. In many instances, the volume of the opening and closing auctions
exceeds 50% of the total daily trading volume. By using the closing and/or opening
86
prices I can be sure that I could make a trade at a given price and at the same time
avoid bid-ask bounce bias.
If a stock has a non-trading day during the trading period (price and/or volume equals
to zero), I use the closing price of the previous day to create the spread. However, the
trading positions on the pairs having that stock cannot be opened or closed on that
day, even if the spread process signals to do so.
4.4.3 Pairs formation
I take log prices of all stocks pre-selected for pairs trading based on the 12 month
history, combine them in all possible pairs and build spread process for each pair.
yi,j (t) = log Pi (t) − log Pj (t),
where Pi (t), Pj (t) are prices of stocks i and j on day t.
For each spread process, I calculate its standard deviation Ci,j . I arbitrarily set pa-
rameter Hi,j for my H -strategy equal parameter Ci,j
Hi,j = Ci,j .
I make H -construction for each spread process and calculated H -volatility ξi,j (Hi,j )
and H -inversion Ni,j (Hi,j ). Then all pairs are sorted in descending order by the size of
the H -inversion Ni,j (Hi,j ). I anticipate that pairs with the highest H -inversion over the
period of history used for calibration will tend to have statistically similar behaviour
in the future and provide higher profit.
H -inversion acts as a good proxy for a number of parameters. The spread process with
smaller standard deviation (which is equivalent to a smaller squared distance in Gatev
et al. (2006)) has smaller H, and as a result tends to have higher H -inversion. For two
87
spread processes with the same H and ξ(H), the higher value of H -inversion means
higher profit by (4.27). At the same time, higher H -inversion means a larger sample
size. This provides us more confidence in the statistical power of the calibration results.
The top 5 and 20 pairs with the highest H -inversion Ni,j (Hi,j ) are used for pairs trading.
In contrast to previous research on pairs trading, I remove the company selected for
the pair from the pool of the pre-selected stocks. Therefore, each company could be
selected just once and be a part of the one pair only. I expect this approach improves
the diversification of the portfolio. Also it allows avoidance of a situation where the
same stock, being a part of different pairs, can be traded short and long simultaneously.
4.4.4 Trading rules
I start trading all pairs selected during the stage of pairs formation from the first day
of the trading period and constantly stay in the market until the very last day of the
trading period when I close all positions.
To define the direction of trades for each pair on the first day, I make the H -construction
over the history using the parameter Hi,j defined during calibration and take the di-
rection of the trade at the end of the calibration period and the value of the last local
extremum. Hence, the virtual trading on the history extends beyond the end of the
calibration period and becomes real trading on the 6 month trading period.
It is known for each spread if the spread has a local maximum or minimum at the end
of the calibration period. That is, just before the first day of real trading. Let’s assume
it is a maximum. It means the following:
1. on the first day of the trading period, one should have a long position on the
spread process yi,j (t), so one buys long stock i and sells short stock j regardless
of the current price levels or the spread process current value;
2. then one follows the spread process waiting for the sell signal — the first moment
88
t after time τ0b (the last stopping time of the calibration period) such that
yi,j (t) − min yi,j (n) ≥ Hi,j ,

τ0b ≤n≤t
that is the time when one recognises the next local minimum — the spread process
moves away from the previous local minimum on the distance greater than Hi,j .
When it happens, one sets τ1b = t and reverses position on the spread from ‘long on the
spread’ to ‘short on the spread’. To do so, one closes existing positions and sells short
stock i and buys stock j. One then keeps following the spread process yi,j (t) waiting
for the signal to buy the spread again — the first moment t after time τ1b such that
max yi,j (n) − yi,j (t) ≥ Hi,j .

τ1b ≤n≤t
This procedure should be repeated again and again. The investor is constantly staying
in the market and alternating ‘buying the spread’ and ‘selling the spread’ until the last
day of the predefined trading period when one just closes all open positions.
If on the first trading day the spread process has a minimum as its last historical local
extremum, then one does the opposite for the first trade — sells the spread, that is,
sells short the stock i and buys the stock j. After that, one follows the same way as
above changing ‘selling the spread’ and ‘buying the spread’.
4.4.5 Excess return calculation and transaction costs
To calculate strategy excess return I follow the procedure common for pairs trading
literature (Gatev et al., 2006; Do and Faff, 2010) which is described in detail in Sec-
tion 3.3. The proposed strategy is dollar neutral, so I trade $1 in each leg of the pair.
I calculate value-weighted daily market-to-market cash flows from each pair which are
considered as excess return:
P
i∈P wi,t ci,t
rP,t = P , (4.29)
i∈P wi,t
89
where:
ci,t is the daily cash flow from the two positions from the pair i;
wi,t is the weight of each pair. For each newly opened position on the pair initial
weight equals 1 and then evolves by the formula wi,t = wi,t−1 (1 + ci,t−1 ) = (1 +
ci,1 ) · · · (1 + ci,t−1 ).
The daily cash flow from the pair or a daily return on the pair is
2
X
ci (t) = Ij (t) vj (t) rj (t), (4.30)
j=1
where:
Ij (t) is a dummy variable which is equal to 1 if a long position on stock j is open and
rj (t) is a daily return on stock j;
vj (t) is a weight of stock j and is used to calculate daily cash flows

vj (t) = vj (t − 1) · (1 + rj (t − 1)) = (1 + rj (1)) · · · (1 + rj (t − 1)).
Then the strategies’ daily returns are compounded to obtain monthly returns.
It is reported in academic literature that transaction costs make a serious impact on

the profitability of the pairs trading strategies. Bowen et al. (2010) record more than a
50% reduction in the excess returns of the high frequency pairs trading strategy after
applying a 15 basis point transaction fee. Do and Faff (2011) replicate the research
by Gatev et al. (2006) and demonstrate that the strategy became unprofitable after
detailed accounting for all transaction costs.
For the tests, I choose transaction costs equal to 0.10% (10 basis points) per trans-
action. It is the average brokerage fee as of April 2011 on the Australian market for
retail investors (Interactive Brokers 0.08%, CommSec 0.12%, E-Trade 0.11%, Macquare
90
Edge 0.10%). Commission on the trading US stocks is calculated based on the number
of stocks traded rather than dollar volume and starts from US$0.005 per share. In
most cases, it is cheaper than 0.10% per trade as most companies in the S&P 500
and even in the S&P 600 Small Cap are priced higher than $5 per share. The se-
lected level of transaction costs of 0.10% per trade means about 0.20% per round trip
per stock and about 0.40% per round trip for the pair or the spread as a synthetic asset.
To account for transaction costs, I employ the following rules. At stopping time τnb ,
when one changes direction of the trade on the spread process, I reduce the cash flow
from the current day by the weighted size of transaction costs
2
X
ci (τnb ) = (Ij, t vj, t rj, t − λ vj, t (1 + rj, t )) , τnb = t;
j=1
and reduce the next day’s cash flow from the new position on the spread by doubling
the size of the transaction costs (as one trades $1 short and $1 long — total trading
volume $2)
ci (τnb + 1) = ci (τnb + 1) − 2λ,
where:
ci,t is a cash flow or excess return on the pair i calculated as (4.30);
λ = 0.001 is a brokerage fee;
Ij,t is a dummy variable which is equal to 1 if a long position on stock j is open and
rj, t is a daily return on the stock j;
vj, t = vj, t−1 (1 − rj, t−1 ) weight of the stock j.
My estimation of transaction costs is quite conservative. Nowadays, the price of trad-

ing for the most institutional investors has reduced dramatically. I present the strategy
performance before transaction costs as well. One can make their own estimation of
91
the impact of transaction cost by multiplying the chosen level of transaction cost per
trade per stock by four and by the average number of trades per month as reported in
the results section.
It is important to remember that pairs trading is a naturally leveraged product and the
above method of excess return computation uses a 2:1 leverage. One should be careful
when comparing the results of the pairs trading strategy with possible transaction
costs or performance of the non-leveraged strategies, for example, the naı̈ve buy-and-
hold strategy. For example, the brokerage fee applies to the full traded volume, which is
$2 for pairs trading, while reported excess returns are based on a $1 dollar investment.
4.5 Results
The results of the testing presented in Tables 4.3 and 4.4 and on Figures 4.5–4.12 show
monthly excess returns for the strategy and its historical performance together with
proper benchmark indexes. Monthly excess returns are statistically significant at 99%
confidence level for all scenarios except one — S&P 400 MidCaps top 20 pairs after
transaction costs. Strategy excess returns before transaction costs vary from 1.42% to
3.65% per month at standard deviations from 2.0% to 5.67%. For the same period of
time the benchmark performances are: S&P/ASX 200 — 0.31% per month at 3.95%
standard deviations; S&P 500 — 0.11% per month at 4.92%; S&P 400 MidCaps —
0.44% per month at 5.49%; S&P 600 SmallCaps — 0.34% per month at 5.97%. The
proposed strategy outperforms the market indexes for each data set for the top 5 and
top 20 pairs portfolios.
Mean St.Dev. Median Max Min Kurtosis Skewness

S&P/ASX 200 0.0031 0.0395 0.0108 0.0740 -0.1408 4.0935 -0.9532
S&P5̇00 0.0011 0.0492 0.0086 0.0932 -0.1928 4.0498 -0.7637
S&P 400 MidCaps 0.0044 0.0549 0.0122 0.1395 -0.2404 6.0031 -1.0468
S&P 600 SmallCaps 0.0034 0.0597 0.0138 0.1625 -0.2250 4.5517 -0.7971
Table 4.2: Historical performance of the benchmark indexes
92
Increasing the number of pairs traded simultaneously from 5 to 20 pairs reduces the
profit but at the same time reduces the standard deviations of returns. That is ex-
pected from an increased diversification of the portfolio. As a result, the Modigliani
Risk-Adjusted Performance and Sharpe ratio for the portfolio of the top 20 pairs are
higher than for the top 5 pairs for all data sets.
Also as expected, testing on the small capitalisation stocks demonstrates higher returns
than the large cap stocks. One possible explanation for this is the different levels of
market efficiency. The stocks from the S&P 500 are the most liquid and most efficient
stock market in the world attracting great attention of domestic and international
institutional and individual investors. There is a very low chance of mispricing any
companies under normal market conditions. It can be clearly seen in Figures 4.7 and
4.8 that the strategy is highly profitable from 2000 to 2003 (dot-com crash, September
11) and from mid 2008 to 2010 (GFC) — the returns are from 3% to 6% per month.
But the strategy barely covers transaction costs between those two periods. Hence, the
strategy generates more profit in times of uncertainty and high volatility, which is not
a surprise for a contrarian strategy.
The small capitalisation companies attract less attention and less research. It is more
difficult to estimate their future risks and returns. Therefore, they have a higher chance
of being mispriced. This provides more opportunities for pairs trading. Strategy test-
ing on the S&P 600 SmallCap data set shows the highest and the most stable returns.
Even the GFC does not change anything in the pattern of returns compared to the
‘normal’ periods.
The strategy performance on all data sets has a very low correlation with chosen bench-
marks, and market betas close to zero. The plot of historical performance for each data
set is very close to a straight line regardless the market conditions for the same period
of time. The only difference is an increase in the slope of that line during the periods
93
of higher financial uncertainty. All of this allows me to conclude that the proposed
strategy is truly market neutral.
Transaction cost is a big issue for all methods of pairs trading. The strategy makes
about 2.5 trades on the spread per month. It means that the average transaction cost
per month at the chosen level of 10 basis points per trade per stock is about 1%
0.1% · 2.5 (trades) · 2 (stocks in pair) · 2 (in and out) ≈ 1%
As a result, the strategy loses from 40% to 50% of its monthly excess returns. However,
even after accounting for the transaction costs, the strategy demonstrates quite an
impressive performance of about 1% per month.
4.6 Conclusions
This chapter proposes a new method of pairs trading based on the volatility of the
spread process. The novelty of this approach is in its flexibility and the less restrictive
nature of the method, as compared to other methods in academic literature. I do not
expect the spread process to be mean-reverting in its formal definition with constant
mean, variance and coefficient of mean-reversion. To generate the profit from pairs
trading, it is sufficient that the process, while not mean-reverting globally, locally has
mean-reverting properties statistically more often than not. As it is a non-parametric
method, it is free from possible problems with model misspecification.
Testing on real market data shows statistically significant profits from using the same
strategy across different markets. Differences in the level of returns can be explained
in the framework of the efficient-market hypothesis.
The H -constructions are a very effective way to measure the variability of the process
and can be used successfully as a basis of the pairs trading strategy. However, it is not
the only method to tackle variability and other approaches can be developed.
94
Market ASX S&P 500 S&P 400 MidCap S&P 600 SmallCap
Number of pairs traded top 5 pairs top 20 pairs top 5 pairs top 20 pairs top 5 pairs top 20 pairs top 5 pairs top 20 pairs
Distribution of monthly excess returns before transaction costs
Mean 0.0269 0.0185 0.0228 0.0190 0.0239 0.0142 0.0365 0.0238
Standard error 0.0027 0.0016 0.0032 0.0023 0.0057 0.0032 0.0047 0.0024
t-Statistics 10.0348 11.6954 7.0183 8.3539 4.2161 4.3956 7.8292 9.8084
P-value 0.0000 0.0000 0.0000 0.0000 0.0001 0.0000 0.0000 0.0000
Median 0.0264 0.0184 0.0160 0.0132 0.0134 0.0116 0.0328 0.0242
Standard deviations 0.0339 0.0200 0.0410 0.0288 0.0567 0.0324 0.0493 0.0257
Skewness 0.9833 0.5514 1.6244 1.2752 2.6557 2.6096 1.1913 0.9283
Kurtosis 8.0619 5.8379 7.6551 6.2938 12.5225 13.0219 6.8876 5.0155
Minimum -0.0761 -0.0377 -0.0751 -0.0565 -0.0507 -0.0599 -0.0667 -0.0259
Maximum 0.1931 0.1016 0.2124 0.1373 0.3225 0.1802 0.2686 0.1116
Average losing month -0.0196 -0.0108 -0.0142 -0.0113 -0.0200 -0.0116 -0.0201 -0.0107
Negative observations 16.3% 15.0% 31.9% 22.5% 31.0% 30.0% 23.2% 17.0%
95
Distribution of the monthly excess returns after transaction costs
Mean 0.0161 0.0098 0.0124 0.0094 0.0143 0.0057 0.0253 0.0140
Standard error 0.0026 0.0015 0.0030 0.0021 0.0053 0.0030 0.0044 0.0023
t-Statistics 6.3098 6.5384 4.0824 4.4415 2.7048 1.9016 5.7690 6.0861
P-value 0.0000 0.0000 0.0001 0.0000 0.0080 0.0601 0.0000 0.0000
Median 0.0168 0.0097 0.0063 0.0037 0.0049 0.0028 0.0224 0.0135
Standard deviations 0.0323 0.0190 0.0384 0.0267 0.0528 0.0298 0.0464 0.0243
Skewness 0.8772 0.3251 1.5523 1.0882 2.5974 2.4183 1.1020 0.7981
Kurtosis 8.0292 5.5916 7.4354 6.0756 12.3564 12.4902 6.3577 4.7958
Minimum -0.0829 -0.0457 -0.0816 -0.0725 -0.0574 -0.0710 -0.0726 -0.0420
Maximum 0.1763 0.0838 0.1891 0.1156 0.2919 0.1572 0.2365 0.0976
Average losing month -0.0187 -0.0120 -0.0171 -0.0123 -0.0237 -0.0141 -0.0254 -0.0133
Negative observation 28.1% 26.3% 45.0% 40.0% 37.0% 43.0% 26.8% 27.7%
Table 4.3: Monthly excess returns of the kagi pairs trading strategy with and without transaction costs (0.10% per one trade per stock)
Market ASX S&P 500
Number of pairs traded top 5 pairs top 20 pairs top 5 pairs top 20 pairs
Trades per month per pair 2.6 2.1 2.5 2.3
Maximum holding time, days 116 150 124 127
Average holding time, days 8.76 10.79 9.16 9.71
Correlation with the benchmark 0.13 0.06 0.03 0.02
Sharpe ratio 0.7933 0.9246 0.5548 0.6604
Annualised Sharpe ratio 2.7481 3.2029 1.9219 2.2877
Modigliani RAP 0.0313 0.0365 0.0273 0.0325
Jensen’s alpha 0.0266 0.0184 0.0227 0.0190
Market beta 0.1099 0.0289 0.0246 0.0128
Market S&P 400 MidCap S&P 600 SmallCap
Number of pairs traded top 5 pairs top 20 pairs top 5 pairs top 20 pairs
Trades per month per pair 2.2 2.0 2.6 2.3
Maximum holding time, days 115 162 122 122
Average holding time, days 10.17 11.05 8.72 9.48
Correlation with the benchmark -0.13 -0.15 -0.12 -0.03
Sharpe ratio 0.4216 0.4396 0.7398 0.9268
Annualised Sharpe ratio 1.4605 1.5228 2.5627 3.2105
Modigliani RAP 0.0231 0.0241 0.0442 0.0553
Jensen’s alpha 0.0245 0.0146 0.0368 0.0239
Market beta -0.1351 -0.0886 -0.1030 -0.0116
Table 4.4: Risk measures based on the monthly excess returns before transaction costs
Sharpe ratio is a ratio of the strategy excess return to its standard deviation.
Modigliani RAP is a risk-adjusted performance of the strategy defined as a product of
the Sharpe ratio and the standard deviation of the returns on the benchmark.
Jensen’s alpha is a strategy abnormal return.
Market beta is a measure of relation between strategy and market volatility. It is
defined as a ratio of the covariance between strategy and benchmark returns to the
variance of the benchmark returns.
96
5
Strategy Before Tr.Cost
Strategy After Tr.Cost
4 Market Index
−1
1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010
Figure 4.5: Strategy historical performance on the Australian market data set for top
5 pairs portfolio before and after transaction cost — accumulated returns without
3
2.5
Market Index
1.5
0.5
−0.5
1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010
Figure 4.6: Strategy historical performance on the Australian market data set for top
97
4
3.5 Strategy After Tr.Cost
Market Index
3
2.5
1.5
0.5
−0.5
1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010
Figure 4.7: Strategy historical performance on the S&P 500 data set for top 5 pairs
portfolio before and after transaction cost — accumulated returns without reinvest-
ment. Axes are time and returns.
3.5
3 Strategy After Tr.Cost
Market Index
2.5
1.5
0.5
−0.5
1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010
Figure 4.8: Strategy historical performance on the S&P 500 data set for top 20 pairs
portfolio before and after transaction cost — accumulated returns without reinvest-
ment. Axes are time and returns.
98
2.5
2 Market Index
1.5
0.5
−0.5
2003 2004 2005 2006 2007 2008 2009 2010
Figure 4.9: Strategy historical performance on the S&P 400 MidCap data set for top
1.6
1.4 Strategy After Tr.Cost
Market Index
1.2
0.8
0.6
0.4
0.2
−0.2
2003 2004 2005 2006 2007 2008 2009 2010
Figure 4.10: Strategy historical performance on the S&P 400 MidCap data set for top
99
5
4 Market Index
−1
2002 2003 2004 2005 2006 2007 2008 2009 2010
Figure 4.11: Strategy historical performance on the S&P 600 SmallCap data set for
top 5 pairs portfolio before and after transaction cost — accumulated returns without
3
2.5
Market Index
1.5
0.5
−0.5
2002 2003 2004 2005 2006 2007 2008 2009 2010
Figure 4.12: Strategy historical performance on the S&P 600 SmallCap data set for
top 20 pairs portfolio before and after transaction cost — accumulated returns without
100
Chapter 5
Study 3. Arbitrage Strategies

Between Listed Asian-Pacific Stocks
and Their NYSE ADRs When
There Is No Overlap of Trading
This chapter considers arbitrage-style trading between markets without overlap in trad-
ing hours. It proposes a method of analysis of market data based on two spread pro-
cesses — observable but not tradable, and tradable but not observable. The method
can be used for markets segmentation analysis. This chapter then proposes two meth-
ods of trading and tests them on market data.1
5.1 Introduction
Technological progress and the liberalisation of capital market policies greatly facili-
tate the capital market internationalisation and mitigate market segmentation. The
first wave of internationalisation of markets began during the 1970s, with investors and
firms investing funds in foreign equity markets to diversify their portfolios and earn
1
Bogomolov, T., Lui, L., Kalev, P. ‘Can Time Difference Deter Arbitrage Opportunities? An
Examination of Cross-listed Asia-Pacific ADRs and Underlying Shares’ under review in Journal of
Asset Management
101
higher returns than were possible with only a domestic portfolio (Foerster and Karolyi,
1993, 1999). Then, another globalisation phenomenon saw firms actually cross-listing
their stocks on foreign capital markets in the early 1980s. Although the past years
have witnessed a significant slowdown in the pace of new international cross-listings
with numbers worldwide down from 4,700 to 3,065 (World Federation of Exchanges
2010), London, NASDAQ, New York and Singapore markets still attract the most for-
eign company listings both from the emerging and developed economies. Among those
cross-listings, most of them have been both listed and traded in the home and foreign
markets.
Theoretically, if international financial markets are perfectly integrated and the law
of one price (LOP) holds, stocks cross-border traded in different markets should be
perfect substitutes for the investors. As a result, the prices adjusted for exchange rate
should be identical. When the markets are segmented from each other and LOP is
violated, the arbitrage opportunities can then exist, which in turn ensures that the
LOP is upheld. The price disparities are readily observed for the cross-listed stocks
that are traded in both their local and US markets, but for different reasons, not many
of them result in the arbitrage opportunities.
The literature on the multi-market arbitrage is abundant. However, until recently,

very little research has been done to investigate the existence of the arbitrage op-
portunities for the Asia-Pacific stocks, which have been cross-listed on the New York
Stock Exchange (NYSE) in the form of American Depositary Receipt (ADR). When
studing the Asia-Pacific companies, besides the common impediments limiting arbi-
trage (e.g. transaction costs, holding costs, taxes), one needs to consider the effect of
non-overlapping trading hours of the US and domestic markets.
Strictly speaking, trading of pairs ADRs and Asia-Pacific shares cannot be considered
as a true arbitrage, but rather an arbitrage-style trading. Due to the time gap between
102
markets, the trading becomes not risk-free.
The purpose of this study is to examine whether the time gap between the Asia-Pacific
and US markets would deter arbitrage. I propose a statistical approach to tackle the
question of non-overlapping trading hours and differentiate the observed price discrep-
ancies and their effect on the possibility of arbitrage. I also demonstrate that the nature
of price disparity evolves over time and the same deviations from the parity may create
an arbitrage opportunity or may not, depending on the stage on the company and/or
market life cycle.
This chapter proceeds as follows: Section 5.2 briefly reviews the theory of arbitrage for
cross-listed securities and proposes the method to analyse arbitrage opportunities for
markets with non-overlapping trading hours; Section 5.3 formulates research questions;
Section 5.4 describes the data and design of the empirical tests; Section 5.5 reports and
discusses the results; Section 5.6 concludes.
5.2 Background
5.2.1 Review of arbitrage for cross-listings
In a perfect capital market scenario where there is no transaction cost, no tax and
perfect information, financial markets are undoubtedly integrated and hence, are per-
fect substitutes for firm listings. Furthermore, if two markets are integrated and one
security is listed in both markets, there should not be any disparity in terms of asset
prices and volatility. However, Werner and Kleidon (1996) showed that price disparity
and volatility in asset prices exist for firms with multiple listings.
Cross-listings may not be perfect substitutes for each other for a number of important
reasons. Firstly, they trade in different markets with different trading hours (Rosen-
thal, 1983). Secondly, when a security is traded in different markets, information lags
103
between different trading venues produce short-term disparities in the prices at which
the security trades at different locations at any given time (Chowhdry and Nanda,
1991). This provides more opportunities for the informed investors to exploit their
private information. It is also predicted that a cross-listing does not change return
variance. Thirdly, international market imperfections could result in some price differ-
ence, vis-à-vis, the underlying securities.
Under the no-arbitrage condition, the price of the stocks traded on both the home mar-
ket and US market as ADR must equal each other after the consideration of exchange
rate, which is presented in equation (5.1) below:
PU S
PHome = F XHome/U S . (5.1)
C
where PHome is the stock price at home market, PU S is the ADR price at the US market,
C is the ADR ratio, which represents the number of shares per one ADR, F XHome/U S
is the exchange rate between the home and US currencies, quoted as a number of home
currency units per one US dollar.
The above equation could be presented as the following log price relation
log PHome = log PU S − log C + log F XHome/U S . (5.2)
The return spread or spread between log prices of the home and US markets at any
time-moment (t) is
s(t) = log PHome (t) − log PU S (t) − log F XHome/U S (t) + log C. (5.3)
If the deviation of s(t) from zero is greater than total transaction cost, an arbitrage
opportunity would be considered as a risk-free profit.
104
Most studies support the LOP for internationally traded stocks after adjustment for ex-
change rates and transactions costs (Maldonado and Saunders, 1983; Miller and Morey,
1996). While for 98% of observations, price disparities (5.3) stay within economically
small values, large deviations from the parity happen, suggesting that arbitrage trading
profits could be made on these large disequilibria when transaction costs were consid-
ered (Wahab et al., 1992; Suarez, 2005; De Jong et al., 2009; Gagnon and Karolyi,
2010). These studies also assert that markets are disintegrated and not fully efficient,
as measured by the presence of arbitrage opportunities.
Similar research was also conducted in other markets with overlapping trading hours
rather than US market. Ding (1999) examines the cross-listed stocks that were traded
on the Stock Exchange of Singapore and the Kuala Lumpur Stock Exchange. The
results showed that the two markets were well linked in terms of their returns and
volatility, the LOP reasonably confirmed, but arbitrage opportunities appeared to ex-
ist when stock-broking houses trade for their own accounts.
On the critical side, observed large deviations could be a result of imperfections of

the market data used for analysis. If we assume that any potentially profitable arbi-
trage opportunity should be quickly exploited by arbitragers, then only high frequency
market data, carefully matched with exchange rates, should be used for arbitrage anal-
ysis. Lok and Kalev (2006) examine tick data of the cross-listed Australia and New
Zealand shares and found no significant price disparity, hence, no obvious arbitrage
opportunities exist.
5.2.2 Review of arbitrage for non-overlapping
traded cross-listings
The examined above studies do not have the time issues with the simultaneously
traded cross-listed securities. Conversely, the Asia-Pacific and US markets have non-
overlapping trading time. True arbitrage between those markets is impossible. The
105
investor has to bear a significant risk of holding just one leg of the arbitrage pair during
the time gap between the US and domestic market.
The recent study of Gagnon and Karolyi (2010) provides an extensive analysis of multi-
market trading and arbitrage of 506 cross-listed companies from 35 countries including
Asia-Pacific companies. For all markets, Gagnon and Karolyi (2010) use US prices and
exchange rates synchronised with home markets. Asia-Pacific companies were treated
in the same manner as companies from other countries. However, the authors admit
that those pairs are perfectly non-synchronised as there is no time overlap between
markets trading hours.
Lack of framework to handle the time gap and associated risks could be the reason
why an arbitrage trading of the Asia-Pacific shares and corresponding ADRs did not
attract a similar level of attention as a true arbitrage until recently. However, the
growing importance of Asian markets facilitates interest in the topic.
Hsu and Wang (2008) study the effect of trading volume and macro events on the price
spreads on the sample of 37 cross-listed firms of six Far Eastern countries (China, Hong
Kong, Japan, Singapore, Korea, Taiwan) and reported change in the markets segmen-
tation as a result of the liberalisation of capital control in Korea and Taiwan, and some
arbitrage opportunities between Hong Kong stocks and ADRs. Hsu and Wang (2008)
also try to use previous values of the spread and the US and domestic stock indexes
as predictors in the regression to estimate the future spread. Similar research by Dey
and Wang (2012) focuses on Chinese H-shares listed in Hong Kong and New York and
the effect of liquidity, trading volume and turnover on the return spread. Both studies
found that changes in trading volume shifted price spreads and the influence of other
parameters was negligible.
Again, these studies use research methods traditional to the arbitrage literature. My
106
goal is to extend their research and propose an alternative approach to the arbitrage
and markets integration/segmentation analysis for the markets with non-overlapping
trading hours.
It is impossible to execute all trades at the same time as we observe a price disparity
due to time gap between markets trading hours. Then we have to consider more than
one spread s(t) (5.3). The first spread s1 (t) is to look for trading signals or price
discrepancies greater than some predefined level. Therefore, equation (5.3) is modified
as shown below
C
s1 (t) = log PHome (t) − log PUOS (t) − log F XUOS (t) + log C (5.4)
C
where PHome is the closing price at home market on day t; PUOS is an opening price in
US market on the same day t; F XUOS is an exchange rate at the day t open time of
the US market, quoted as a number of local (home) currency units per US$1; C is an
ADR ratio (number of share per one ADR).
If there is a deviation from the long-term mean on the distance greater than total
transaction costs, it can be considered as an arbitrage opportunity. However, that
opportunity cannot be exploited as the home market is already closed. So, the spread
s1 (t) indicates a potential profit but not the real one.
Suppose there is an investor who makes a decision to trade. He or she executes one
leg of the arbitrage trade (ADRs) and makes a commitment to execute the other leg
(domestic stocks) on the opening of the local market at any price available. The
resultant spread s2 (t) is then an actual return process as shown below
O
s2 (t + 1) = log PHome (t + 1) − log PUOS (t) − log F XHome
O
(t) + log C (5.5)
O
where PHome (t + 1) is the next day opening price of the home market; PUOS is an open
107
O
price of the US market; F XHome is an exchange rate at the day t + 1 open time of the
home market, quoted as a number of local currency units per US$1; C is an ADR ratio
(number of share per one ADR).
Due to the time issue, there are four possible scenarios which can be considered as
the most logical ways to organise trading in two markets without overlapping trading
hours:
1. Analyse the market and make a trading decision on US market open based on
known prices of Asia-Pacific close and US open. Trade according to trading
signals at US open and Asia-Pacific open next day.
2. Analyse the market and make a trading decision on US market close based on
known prices of Asia-Pacific close and US close. Trade according to trading
signals at US close and Asia-Pacific open next day.
3. Analyse the market and make a trading decision on Asia-Pacific market open
based on known prices of US close of the previous day and Asia-Pacific open.
Trade according to trading signals at Asia-Pacific open and US open same day.
4. Analyse the market and make a trading decision on Asia-Pacific markets close
based on known prices of US close of the previous day and Asia-Pacific close.
Trade according to trading signals at Asia-Pacific close and US open same day.
Under the international capital market setting, one would expect the home market of
multi-traded securities to be the dominant market because the markets around the
world are more likely to be segmented, and information about the underlying com-
pany is more likely to stem from the home market (Licht, 1998). A significant number
of studies have provided evidence for the price-discovery prediction. Eun and Shim
(1989); Hamao et al. (1990); Neumark et al. (1991); deB. Harris et al. (1995); Has-
brouck (1995); Lieberman et al. (1999); Grammig et al. (2005); Pascual et al. (2006);
Lok and Kalev (2006) all support the idea that price discovery occurs in the home
country for cross-listed stocks. Preliminary research over different scenarios showed
108
the same results — the dominance of the domestic market. Therefore, in this research,
only scenario 1 is considered, which is reflected in equations (5.4) and (5.5).
It is important to stress that the investor can observe the spread s1 (t) but cannot trade
it, as the time has already pased and the home market is closed. It is an observable but
not tradable process. Also, the investor can trade the spread s2 (t) but he or she can
observe it only after making all trades. So, it is a tradable but not observable process.
The spreads s1 (t) and s2 (t) look similar but obviously they are far from being iden-
tical. Due to the time differences between markets and, as a result, between trades
execution, this type of trading is not risk-free. The deviation from the mean detected
on the spread s1 (t) can be fixed overnight and result in a loss in the trade that the
investor has committed himself/herself to execute on the next day.
Observed price disparity between a closing price of the stock on the Asia-Pacific market
and an opening price of the ADR might be contributed to one of two possible reasons:
1. Observed price in US market is a new fair price. That is, extra information
appeared during the time gap between markets working hours and has digested
by investors in the right way. Therefore, the price disparity is a reflection of
the price discovery process and true change in the evaluation of the company
happened overnight.
2. Mispricing. This is a result of wrong interpretation of the information available

before and/or arising during the time gap by the domestic and/or US investors.
If the price disparity is caused by the first reason, it is not possible to make an arbi-
trage profit because markets are efficient or close to being efficient. The current price
appears to be a fair price. New information arrives in a random order and future values
of the spread cannot be predicted. Large deviations from the parity can be observed
on the observable spread s1 (t) but one cannot benefit from trading those deviations as
109
they have no connection to the tradable spread s2 (t).
If the price disparity is a result of mispricing, then it might not exhibit an objective
economical ground and does not mean the true price changes under the changes of the
fundamental factors. One can expect that over time the mispricing will be fixed and
prices will return to parity. Hence, an arbitrage-style trading can be profitable.
5.3 Research questions
Before engaging in trading between markets with non-overlapping trading hours, one
should answer the following major questions:
1. Is the tradable spread s2 (t) stationary? If it is, then one can expect that any
deviation from the mean is temporary, and the LOP holds.
2. Does the observable spread s1 (t) predict the tradable spread s2 (t + 1)? And if it
does, to what extent?
For the first question I consider lag one auto-regressive model of the tradable spread
s2 (t) as shown below:
s2 (t) = αs2 (t − 1) + ε(t) (5.6)
If coefficient α is statistically significant and close to 1, then the spread should be

considered as a unit-root process. One can use any unit-root tests, like Augmented
Dickey–Fuller (ADF) or Phillips–Perron, to test for the stationarity. However, for prac-
tical purposes, an ordinary least square (OLS) regression is sufficient and one should
avoid trading the spread with coefficient α close to 1. An ideal situation is to have α
close to 0, which guarantees instantaneous return to price equilibrium.
To answer the second question, I regress the tradable spread s2 (t + 1) on the observable
spread s1 (t) as shown in equation (5.7) below:
s2 (t + 1) = βs1 (t) + ω(t). (5.7)
110
If coefficient β is significantly different from 0, then the spread s1 (t) can make some
prediction regarding the spread s2 (t + 1). If the value of β is close to 1, then the
observed deviation from the mean directly translates into the similar size deviation of
the tradable spread. Ideally, I would like to see this coefficient equal to 1 or larger.
These two coefficients discriminate companies (that is all return spreads between ADRs
and Asia-Pacific stocks) into three distinctive groups:
• Group 1. If α is close to 1, then the spread is a unit root process or close to

it. The LOP does not hold and trading that company involves unlimited risk.
Companies from this group should not be considered for arbitrage-style trading.
• Group 2. If α is less than 1 (ideally close to 0) and β is greater than 0 (ideally

close to 1) then the LOP holds, and the observed spread does predict the tradable
spread. The investor can be confident that risk in the each trade is limited due to
the LOP. The expected profit in each trade is positive as the deviation from the
mean and mean-reversion of the observed spread guarantees the similar deviation
and following mean-reversion of the tradable spread.
• Group 3. If α is less than 1 (close to 0) and β is close to 0 then the LOP holds,
but the observed spread does not predict the tradable spread. Observed deviation
from the mean will be fixed during the time gap due to the high level of market
efficiency.
These three groups can be viewed as three stages of the company and market life cycle
and correspondingly three levels of market efficiency or international markets integra-
tion. This classification is in line with claims about dependance of the market efficiency
from the degree of market maturity (Hsu and Kuan, 2005).
Group 1 can be considered as a market with a very low level of market efficiency.
The spread between ADR and stock prices close to a random walk, the LOP does not
hold. The reason can be attributed to a number of factors which includes government
111
regulation, general lack of market transparency and restricted access for international
investors. Company specific factors should be included as well (e.g. a level of familiar-
ity of the US investors with the company).
Group 2 can be considered a medium efficient market. The LOP holds; the spread
between ADR and stock prices is close to the parity most of the time. However, dis-
parities regularly occur and stay long enough to allow one to make all necessary trades
and profit from the correction of the mispricing in the near future.
Group 3 is considered a high efficient market. The LOP holds; any observed disparities
are results of the price discovery which already happened in the past and do not predict
disparities on the tradable spread. Obviously, it is impossible to profit from trading
those disparities.
The coefficient β has another important meaning. It defines the value of the trigger
level to start trading – distance δ. The following ratio is derived from equation (5.7):
δs2 = βδs1 + ω(t) (5.8)
where δs1 and δs2 are deviations from the mean on the spreads s1 (t) and s2 (t) respec-
tively.
Hence, if one wants to trade the deviations equal or greater than the transaction cost
c then the trigger level should be: δs1 = c/β. For small values of β the distance δs1
might become too large and the chance to observe that deviation would be very small.
112
5.4 Research Design
5.4.1 Sample and data description
I investigate the possible predictability of the log prices (return) spread, international
markets integration and arbitrage opportunities in the sample of the companies cross-
listed on the NYSE, representing nine countries (or districts) from the Asia-Pacific
region, which includes companies from Australia, China (Hong Kong), India, Indone-
sia, Japan, New Zealand, Philippines, South Korea, and Taiwan. All these countries
and districts do not have overlapping trading hours with the NYSE. All of the trading
data are obtained from the SIRCA.
I take all the Asia-Pacific based companies traded on the NYSE at the end of 2004
with an average dollar-valued trading volume above US$1,000,000 per day to satisfy
the liquidity requirement, which left 40 cross-listed stocks. The data set used for this
research covers the time period from January 1, 2005 to August 20, 2011. It includes
the period of the GFC which allows me to study the process of price discovery in the
associated high volatility conditions.
As the trading hours of US and Asia-Pacific markets do not overlap, there is no need
to use high frequency data. However, I collect one-hour intraday prices for all of the
40 cross-listed stocks to carefully match them to currency exchange rates to the time
of the stock prices. Holidays in the Asia-Pacific and US markets are excluded from
analysis and there is no trading to be done on those days. To minimise a possible bias
due to dividends, no trading can be done at the ex-dividend day, day before and after.
5.4.2 Trading strategy
There are two possible strategies that can be utilised to exploit an arbitrage over the
cross-listed securities in the multi-market environment with non-overlapping trading
hours. The first one involves the conversion of ADR into shares or shares into ADR:
113
• If the spread s1 (t) is above zero — buy ADRs for US dollars and convert them
into shares; on the next day, sell shares in the home market and convert local
currency into US dollars on the FX market.
• If the spread s1 (t) is below zero — sell short ADRs; on the next day buy local
currency for US dollars, buy shares on the home market and convert them in to
ADRs to cover an obligation on the short sell.
The above procedures can be available for large institutional investors only. However,
small retail investors can profit from these arbitrage opportunities as well. For example,
they can adapt a pairs trading approach (Gatev et al., 2006) and make the following
three trading operations if the spread s1 (t) deviates from zero on the distance greater
than transaction costs:
• sell (buy) stock on the home market for (with) local currency,
• sell (buy) equal volume of local currency for US dollars on the FX market and
• buy (sell) equal US dollar volume of ADRs on the US market.
When the spread s1 (t) returns to zero, the investor closes all open positions.
Transaction costs for the second strategy include brokerage fees, stamp duties, transfer
levies, currency conversion fees and taxes, if any. I arbitrarily choose the transaction
cost per one trade equals 0.1% for Japan, Hong Kong, Australia and New Zealand and
0.5% for other countries. I take transaction cost for trading ADRs 1 US cent per unit
(per share). That makes the total transaction costs for two groups of countries around
0.25–0.3% and 1.1–1.3% respectively. It is a conservative estimation of the transaction
cost. It is based on the current fees for retail investors (August, 2011). Large institu-
tional investors might have much lower transaction costs.
The first strategy takes less market trading and, as a result, it attracts a lower trans-
action fee. However, it requires the conversion of ADRs into shares and back, which
114
involves conversion/cancelation fees up to $0.05 per share depending on the price, hold-
ing fee, custodian fee, and some others (Gagnon and Karolyi, 2010). Overall it would
be safe to estimate the total transaction costs of both strategies as equal and take it
as defined above.
For the empirical test of the profitability, I employ the following approach: take the
spread s1 to analyse the situation and make trading decisions. If the spread is above or
below its 6 month mean on a distance δ, then it is a signal to initiate trading/conversion
for the first strategy and trade the spread towards its mean for the second one. The
distance δ is defined as a ratio of the transaction costs to the coefficient β estimated
over the 6 month history.
Short the spread means: buy the ADR on the US market, sell short the stock on local
market at the next day market open and simultaneously sell an equal amount of foreign
currency. The long position on the spread means the opposite: sell short the ADR on
the US market, then buy the stock on the local market at the next day market open
and buy an equal amount of foreign currency. When the spread s1 crosses the mean
for the first time, it is a signal to close all open positions, which can be executed on
the current day in the US leg and on the next day in the home market.
Trading decisions are governed by the values of the observable spread s1 at the time
moments t (one decides to open an arbitrage position) and t+n (one decides to close an
arbitrage position). Trading actions result in the profit or loss defined by the difference
between values of the tradable spread s2 at the time moments t + 1 (one opens position
on the spread) and t + n + 1 (one closes position on the spread).
5.4.3 Excess returns calculation
In theory, an arbitrage is a zero-cost investment. That is money from the short position
pays for the long position. In practice, some investment is still required to finance a
115
margin account and for brokerage, conversion and other fees. Following the general
approach used in the literature on contrarian trading strategies (Gatev et al., 2006;
Do and Faff, 2010), I trade US$1 in each leg of the arbitrage. Daily cash flow from
that trading is considered an excess return. Then, daily returns for each company are
compounded to obtain monthly returns.
5.5 Empirical Results
Table 5.1 depicts the descriptive statistics of the observable spread s1 for all 40 Asia-
Pacific cross-listed stocks as well as the values of the coefficients α and β for the spread
s2 — lag one auto regression and predictive power of the spread s1 respectively.
The coefficient of auto regression α segregates the countries and companies. Almost all
companies from India, along with some companies from South Korea and Taiwan, form
Group 1 (Figure 5.1). The coefficient of auto regression α is large and the spread be-
haves quite wildly. In some cases, the pairs trading strategy (Gatev et al., 2006), based
on the deviation from the long-term mean, can generated significant profit; however,
the level of risk is high and unacceptable for arbitrage-style trading. These markets
are highly disintegrated. The price of ADR has a very limited connection to the price
of the corresponding stock.
Another distinctive group is Japan. All Japanese companies demonstrate mean-stationary

spread; in most cases the coefficient α is statistically not different from 0 (Figure 5.3).
At the same time, the coefficient β is close to 0 as well, so the observable spread does
not predict the tradable spread. These companies are the most representative example
of Group 3 — old, developed, highly integrated markets. The Australian company
BHP Billiton Limited is in this group as well. Observable spread has some predictive
power; however, it is not large enough to guarantee profit.
Chinese companies traded in Hong Kong as well as New Zealand, Indonesia, Philippines
116
Company name Symbol Local Country Mean Median St.Dev 1% 99% α p-Value β p-Value Group
BHP Billiton Limited BHP BHP.AX Australia 0.000 0.000 0.015 -0.015 0.016 -0.044 0.084 0.162 0.000 3
Westpac Banking Corporation WBK WBC.AX Australia -0.001 -0.001 0.008 -0.009 0.007 -0.049 0.052 0.658 0.000 2
Aluminum Corporation of China Ltd ACH 2600.HK China 0.000 -0.001 0.013 -0.013 0.013 -0.026 0.320 0.688 0.000 2
CNOOC Limited CEO 0883.HK China 0.000 0.000 0.009 -0.010 0.010 0.062 0.017 1.052 0.000 2
China Telecom Corporation Limited CHA 0728.HK China 0.000 0.000 0.008 -0.009 0.009 0.072 0.005 0.624 0.000 2
China Mobile Limited CHL 0941.HK China -0.001 -0.001 0.009 -0.010 0.008 -0.040 0.127 0.713 0.000 2
China Unicom (Hong Kong) Limited CHU 0762.HK China 0.001 0.001 0.010 -0.010 0.012 0.014 0.601 0.799 0.000 2
Huaneng Power International, Inc. HNP 0902.HK China -0.001 -0.002 0.012 -0.012 0.010 0.036 0.165 0.793 0.000 2
China Life Insurance Company Limited LFC 2628.HK China 0.004 0.003 0.012 -0.008 0.019 0.051 0.046 0.841 0.000 2
PetroChina Company Limited PTR 0857.HK China 0.000 0.000 0.009 -0.009 0.010 -0.022 0.392 0.866 0.000 2
Sinopec Shanghai Petrochemical Company Limited SHI 0338.HK China 0.000 0.000 0.011 -0.013 0.012 0.004 0.891 0.723 0.000 2
China Petroleum & Chemical Corporation SNP 0386.HK China 0.000 0.000 0.009 -0.009 0.010 0.039 0.137 0.867 0.000 2
Yanzhou Coal Mining Co. Ltd. YZC 1171.HK China -0.001 -0.001 0.012 -0.015 0.012 0.009 0.717 0.835 0.000 2
HDFC Bank Limited HDB HDBK.BO India -0.069 -0.064 0.055 -0.147 0.000 0.816 0.000 0.996 0.000 1
ICICI Bank Ltd. IBN ICBK.BO India -0.027 -0.007 0.052 -0.109 0.014 0.813 0.000 0.987 0.000 1
Dr. Reddy’s Laboratories Limited RDY REDY.BO India -0.001 -0.001 0.020 -0.024 0.020 0.428 0.000 0.956 0.000 2
Tata Motors Limited TTM TAMO.BO India -0.048 -0.007 0.106 -0.245 0.013 0.946 0.000 0.990 0.000 1
Wipro Limited WIT WIPR.BO India -0.266 -0.246 0.107 -0.404 -0.134 0.945 0.000 1.004 0.000 1
PT Indosat Tbk IIT ISAT.JK Indonesia 0.003 0.002 0.016 -0.014 0.021 0.212 0.000 0.830 0.000 2
P.T. Telekomunikasi Indonesia TLK TLKM.JK Indonesia 0.002 0.002 0.014 -0.014 0.018 0.021 0.415 0.791 0.000 2
Canon, Inc. CAJ 7751.T Japan 0.001 0.000 0.009 -0.008 0.011 0.092 0.000 0.116 0.007 3
Hitachi, Ltd. HIT 6501.T Japan 0.000 0.000 0.008 -0.009 0.010 0.086 0.001 0.232 0.000 3
Honda Motor Co., Ltd. HMC 7267.T Japan 0.000 -0.001 0.010 -0.009 0.010 0.046 0.082 0.122 0.005 3
Kyocera Corporation KYO 6971.T Japan 0.000 0.000 0.009 -0.009 0.009 0.029 0.269 0.378 0.000 3
117
Mitsubishi UFJ Financial Group, Inc. MTU 8306.T Japan 0.002 0.001 0.014 -0.010 0.015 0.081 0.003 0.564 0.000 3
Nomura Holdings, Inc. NMR 8604.T Japan 0.002 0.001 0.012 -0.009 0.013 0.068 0.010 0.443 0.000 3
Nippon Telegraph and Telephone Corporation NTT 9432.T Japan 0.001 0.000 0.010 -0.008 0.010 0.132 0.000 0.388 0.000 3
Sony Corporation SNE 6758.T Japan 0.001 0.000 0.010 -0.009 0.011 0.041 0.121 0.022 0.540 3
Toyota Motor Corporation TM 7203.T Japan 0.000 -0.001 0.009 -0.009 0.009 0.077 0.003 0.157 0.000 3
Telecom New Zealand NZT TEL.NZ New Zealand 0.000 -0.001 0.013 -0.013 0.014 0.087 0.001 0.688 0.000 2
Philippine Long Distance Telephone Co. PHI TEL.PS Philippines 0.001 0.000 0.012 -0.013 0.015 0.124 0.000 0.768 0.000 2
Korea Electric Power Corporation KEP 015760.KS South Korea -0.005 -0.002 0.017 -0.029 0.013 0.516 0.000 0.946 0.000 1
LG Display Co., Ltd. LPL 034220.KS South Korea 0.002 0.001 0.011 -0.011 0.014 0.065 0.011 0.763 0.000 2
POSCO PKX 005490.KS South Korea 0.002 0.002 0.011 -0.010 0.014 0.021 0.409 0.714 0.000 2
SK Telecom Co., Ltd. SKM 017670.KS South Korea -0.033 -0.016 0.046 -0.101 0.010 0.922 0.000 1.002 0.000 1
Advanced Semiconductor Engineering, Inc. ASX 2311.TW Taiwan -0.001 -0.001 0.019 -0.021 0.020 0.254 0.000 0.802 0.000 2
AU Optronics Corp. AUO 2409.TW Taiwan -0.001 -0.001 0.013 -0.015 0.014 0.129 0.000 0.629 0.000 2
Chunghwa Telecom Co., Ltd. CHT 2412.TW Taiwan -0.013 -0.002 0.028 -0.057 0.013 0.860 0.000 0.931 0.000 1
Taiwan Semiconductor Manufacturing Company Ltd. TSM 2330.TW Taiwan -0.061 -0.048 0.052 -0.145 -0.012 0.915 0.000 0.968 0.000 1
United Microelectronics Corporation UMC 2303.TW Taiwan -0.196 -0.126 0.144 -0.434 -0.072 0.979 0.000 0.994 0.000 1
Table 5.1: Descriptive statistics of the visible return spreads s1 (t), coefficient of lag one auto-regression of the spread s2 (t) and coefficient
of predictive power of s1 (t)
Symbol and Local – trading tickers on NYSE and domestic exchange; Mean, Median, St.Dev – descriptive statistics of the spread s1 ; α – coefficient of lag
one auto regression of the spread s2 ; β – coefficient of regression of s2 on s1 ; p-Value indicates level of statistical significance of the corresponding coefficients
NYSE Country Monthly profit Sharpe Trades per Individual trades profit
symbol or origin Mean p-Value Median St.Dev Skew. Kurt. Max Min ratio Total month Mean p-Value Median St.Dev Skew. Kurt.
BHP Australia 0.011 0.004 0.006 0.035 -0.135 3.831 0.111 -0.091 1.105 492 6.7 0.001 0.053 0.002 0.020 0.418 9.848
WBK Australia 0.026 0.001 0.000 0.070 2.065 8.591 0.346 -0.101 1.293 588 8.1 0.003 0.000 0.002 0.020 0.809 7.637
ACH China 0.077 0.000 0.064 0.085 0.961 3.609 0.336 -0.080 3.127 897 12.3 0.006 0.000 0.006 0.024 1.076 16.098
CEO China 0.070 0.000 0.040 0.122 3.563 20.097 0.805 -0.072 1.987 841 11.5 0.006 0.000 0.004 0.029 13.720 304.735
CHA China 0.028 0.001 0.025 0.073 -0.069 3.547 0.228 -0.163 1.331 780 10.7 0.003 0.000 0.003 0.022 -0.953 10.073
CHL China 0.046 0.000 0.035 0.052 1.110 5.389 0.214 -0.076 3.015 798 10.9 0.004 0.000 0.004 0.015 -2.380 34.168
CHU China 0.049 0.000 0.033 0.082 0.746 3.438 0.279 -0.139 2.065 745 10.2 0.005 0.000 0.005 0.022 -0.828 12.427
HNP China 0.058 0.000 0.046 0.089 1.784 9.651 0.488 -0.119 2.250 853 11.7 0.005 0.000 0.004 0.022 1.705 15.795
LFC China 0.061 0.000 0.046 0.104 1.769 9.717 0.582 -0.105 2.018 870 11.9 0.005 0.000 0.004 0.024 7.054 130.300
PTR China 0.060 0.000 0.048 0.068 0.621 3.406 0.288 -0.058 3.066 932 12.8 0.005 0.000 0.004 0.016 0.711 12.827
SHI China 0.071 0.000 0.067 0.092 0.447 4.120 0.354 -0.163 2.693 976 13.6 0.005 0.000 0.005 0.023 -0.335 29.887
SNP China 0.054 0.000 0.041 0.077 0.218 3.069 0.215 -0.143 2.426 951 13.2 0.004 0.000 0.004 0.019 -0.324 8.850
YZC China 0.080 0.000 0.054 0.102 1.135 4.522 0.421 -0.090 2.708 878 12.0 0.007 0.000 0.005 0.023 0.753 9.780
HDB India 0.139 0.119 0.060 0.999 0.673 3.633 3.061 -1.655 0.482 1040 14.2 0.009 0.000 -0.001 0.080 0.231 1.977
IBN India 0.076 0.089 0.029 0.476 1.568 9.303 2.016 -1.242 0.552 697 9.5 0.008 0.000 0.002 0.053 0.540 3.418
RDY India 0.087 0.000 0.042 0.143 1.991 6.993 0.677 -0.070 2.106 589 8.1 0.011 0.000 0.008 0.023 1.439 14.768
TTM India 0.382 0.010 0.037 1.363 1.604 7.543 5.061 -3.804 0.970 821 11.2 0.034 0.000 0.005 0.120 0.742 4.264
WIT India 0.285 0.197 0.320 2.841 -0.035 2.316 6.777 -5.869 0.348 1045 14.3 0.017 0.012 -0.068 0.239 0.020 1.434
IIT Indonesia 0.042 0.000 0.026 0.079 1.686 8.244 0.326 -0.173 1.851 462 6.4 0.007 0.000 0.006 0.023 1.745 19.273
TLK Indonesia 0.034 0.000 0.020 0.051 1.930 7.491 0.241 -0.056 2.315 394 5.4 0.006 0.000 0.006 0.019 0.509 6.394
CAJ Japan 0.002 0.325 0.000 0.029 0.239 7.898 0.113 -0.095 0.187 165 2.3 0.001 0.355 0.000 0.021 0.009 5.434
HIT Japan 0.002 0.382 0.000 0.058 -2.727 15.847 0.134 -0.286 0.124 364 5.1 0.000 0.367 0.001 0.021 -2.087 20.770
HMC Japan -0.007 0.890 0.000 0.052 2.378 20.217 0.305 -0.124 -0.502 183 2.5 -0.003 0.956 -0.004 0.026 1.666 15.897
118
KYO Japan 0.012 0.033 0.003 0.057 4.536 32.370 0.406 -0.077 0.758 419 5.7 0.002 0.005 0.001 0.018 1.446 14.031
MTU Japan 0.015 0.021 0.000 0.056 1.551 6.573 0.209 -0.107 0.909 332 5.3 0.003 0.011 0.002 0.022 0.829 5.654
NMR Japan 0.007 0.133 0.000 0.050 1.635 9.614 0.216 -0.111 0.458 282 3.9 0.002 0.127 0.001 0.025 1.742 15.041
NTT Japan 0.009 0.008 0.005 0.032 0.125 5.549 0.105 -0.099 1.001 356 4.9 0.002 0.008 0.002 0.015 0.086 13.333
SNE Japan -0.004 0.906 0.000 0.027 -3.984 27.510 0.052 -0.184 -0.538 64 0.9 -0.005 0.937 -0.001 0.024 -1.095 5.813
TM Japan 0.003 0.216 0.000 0.034 -0.032 12.553 0.135 -0.154 0.322 228 3.2 0.001 0.219 0.001 0.017 -0.016 9.904
NZT New Zealand 0.047 0.000 0.034 0.078 2.379 11.596 0.434 -0.131 2.054 669 9.2 0.005 0.000 0.004 0.023 5.800 84.635
PHI Philippines 0.018 0.000 0.010 0.038 2.586 12.774 0.217 -0.049 1.662 304 4.2 0.004 0.000 0.004 0.017 1.868 17.792
KEP South Korea 0.023 0.055 0.000 0.123 1.727 9.301 0.549 -0.336 0.655 339 4.6 0.005 0.000 0.002 0.023 1.130 7.205
LPL South Korea 0.013 0.007 0.000 0.044 3.057 15.564 0.248 -0.056 1.025 180 2.5 0.005 0.008 0.004 0.026 0.732 8.261
PKX South Korea 0.015 0.055 0.000 0.079 7.072 56.351 0.638 -0.060 0.665 194 2.7 0.005 0.004 0.004 0.028 1.266 10.769
SKM South Korea 0.342 0.000 0.009 0.753 1.744 5.681 3.133 -0.823 1.572 844 11.6 0.029 0.000 0.019 0.054 0.405 2.658
ASX Taiwan 0.035 0.000 0.021 0.064 1.169 4.482 0.228 -0.095 1.881 362 5.0 0.007 0.000 0.006 0.022 0.502 6.062
AUO Taiwan 0.006 0.065 0.000 0.031 0.745 7.988 0.132 -0.095 0.622 170 2.3 0.002 0.101 0.003 0.021 -0.008 5.473
CHT Taiwan 0.047 0.094 0.000 0.300 2.216 10.881 1.447 -0.542 0.539 413 5.7 0.008 0.000 0.000 0.036 0.393 2.107
TSM Taiwan 0.119 0.173 -0.096 1.072 1.062 4.914 3.451 -2.473 0.384 905 12.4 0.009 0.000 -0.009 0.080 0.334 2.747
UMC Taiwan -0.526 0.924 -0.439 3.088 0.203 3.674 8.100 -7.248 -0.590 940 13.1 -0.041 1.000 -0.081 0.219 0.455 2.335
Table 5.2: Monthly and individual trades excess returns of the strategy with conversion between ADR and local shares
Mean, Median, St.Dev, Skew. (skewness), Kurt. (kurtosis) – descriptive statistics of the monthly returns and returns in each individual trades; p-Value is
a result of single-sided t-Test and indicates level of statistical significance of the null hypothesis that the mean of returns equal zero, is p-Value less than 0.05,
then I reject the null hypothesis; Max – the highest monthly profit; Min – the highest monthly loss.
NYSE Country Monthly profit Sharpe Trades per Hold. Individual trades profit
symbol or origin Mean p-Value Median St.Dev Skew. Kurt. Max Min ratio Total month time Mean p-Value Median St.Dev Skew. Kurt.
BHP Australia 0.001 0.366 0.000 0.033 -0.077 2.905 0.071 -0.073 0.139 244 3.3 2.1 0.000 0.459 -0.001 0.022 0.296 9.882
WBK Australia 0.024 0.000 0.008 0.056 1.600 7.705 0.263 -0.087 1.482 252 3.5 2.4 0.002 0.004 -0.001 0.023 0.153 13.329
ACH China 0.063 0.000 0.047 0.076 0.972 3.846 0.298 -0.060 2.881 344 4.7 2.3 0.004 0.000 -0.001 0.037 0.310 122.938
CEO China 0.043 0.000 0.031 0.074 1.742 10.970 0.409 -0.181 1.999 341 4.7 2.1 0.003 0.021 -0.001 0.048 2.227 356.042
CHA China 0.020 0.018 0.027 0.080 -0.579 5.334 0.253 -0.280 0.868 348 4.8 2.1 0.001 0.032 -0.001 0.025 -2.657 36.557
CHL China 0.031 0.000 0.023 0.052 1.265 8.294 0.270 -0.110 2.039 319 4.4 2.5 0.002 0.000 -0.001 0.018 -0.584 13.459
CHU China 0.048 0.000 0.035 0.060 0.976 4.190 0.264 -0.049 2.779 292 4.0 2.5 0.004 0.000 -0.002 0.027 0.327 12.486
HNP China 0.042 0.000 0.029 0.067 2.575 14.779 0.415 -0.052 2.160 283 3.9 2.9 0.003 0.000 -0.001 0.024 1.015 15.745
LFC China 0.055 0.000 0.047 0.086 1.418 8.075 0.452 -0.134 2.222 333 4.6 2.5 0.003 0.000 -0.001 0.025 3.628 54.889
PTR China 0.049 0.000 0.037 0.060 0.426 3.905 0.214 -0.125 2.800 375 5.1 2.1 0.003 0.000 -0.001 0.019 1.157 12.250
SHI China 0.044 0.000 0.043 0.071 0.252 3.662 0.238 -0.131 2.143 368 5.1 2.3 0.003 0.000 -0.001 0.027 -0.291 14.838
SNP China 0.049 0.000 0.040 0.063 0.674 3.460 0.244 -0.084 2.691 370 5.1 2.3 0.003 0.000 -0.001 0.023 0.852 11.393
YZC China 0.062 0.000 0.046 0.081 0.326 4.157 0.288 -0.202 2.637 358 4.9 2.2 0.004 0.000 -0.001 0.033 -0.098 47.556
HDB India 0.034 0.000 0.025 0.064 0.199 3.507 0.177 -0.164 1.860 126 1.7 9.0 0.002 0.015 -0.001 0.033 0.237 7.745
IBN India 0.037 0.000 0.020 0.088 1.698 6.876 0.392 -0.082 1.453 159 2.2 5.1 0.003 0.003 -0.003 0.032 0.980 9.918
RDY India 0.037 0.000 0.019 0.065 1.923 7.906 0.319 -0.056 1.958 160 2.2 5.3 0.003 0.001 -0.001 0.026 0.366 12.629
TTM India 0.039 0.001 0.031 0.098 -0.328 7.666 0.354 -0.369 1.377 142 1.9 6.6 0.003 0.008 0.000 0.037 0.202 10.880
WIT India 0.041 0.000 0.030 0.069 0.746 3.550 0.265 -0.105 2.063 138 1.9 8.1 0.002 0.006 0.000 0.035 -0.202 16.075
IIT Indonesia 0.028 0.000 0.016 0.046 1.373 8.012 0.239 -0.083 2.110 173 2.4 4.2 0.002 0.005 -0.003 0.025 1.088 19.980
TLK Indonesia 0.035 0.000 0.024 0.044 1.505 7.205 0.233 -0.029 2.757 184 2.5 3.2 0.003 0.000 -0.003 0.024 0.598 7.679
CAJ Japan 0.000 0.480 0.000 0.030 -1.878 11.096 0.059 -0.147 0.021 106 1.5 2.1 0.000 0.490 -0.001 0.021 -0.299 8.836
HIT Japan 0.002 0.360 0.000 0.050 -1.991 16.341 0.155 -0.280 0.147 185 2.6 2.2 0.000 0.354 -0.001 0.021 -1.161 24.422
HMC Japan -0.005 0.764 0.000 0.063 4.014 31.790 0.427 -0.155 -0.293 98 1.3 2.3 -0.001 0.784 -0.001 0.029 0.661 13.456
KYO Japan 0.016 0.001 0.009 0.043 3.116 18.528 0.272 -0.061 1.326 213 2.9 2.2 0.002 0.007 -0.001 0.019 2.213 25.869
119
MTU Japan 0.013 0.014 0.001 0.047 0.158 4.638 0.135 -0.136 0.977 148 2.3 2.9 0.001 0.091 -0.002 0.026 0.752 8.623
NMR Japan 0.011 0.005 0.004 0.034 0.710 4.558 0.134 -0.063 1.080 155 2.2 2.3 0.002 0.096 -0.002 0.026 0.602 18.835
NTT Japan 0.011 0.006 0.005 0.035 0.482 5.376 0.138 -0.081 1.048 180 2.5 2.5 0.001 0.027 -0.001 0.016 0.999 21.716
SNE Japan -0.004 0.904 0.000 0.026 -2.723 21.951 0.082 -0.160 -0.533 43 0.6 2.3 -0.002 0.895 -0.001 0.019 -0.897 7.184
TM Japan 0.001 0.385 0.000 0.031 0.048 9.324 0.110 -0.119 0.120 134 1.9 2.2 0.000 0.424 -0.001 0.017 1.019 17.482
NZT New Zealand 0.039 0.000 0.034 0.056 1.004 7.473 0.257 -0.151 2.379 278 3.8 2.6 0.003 0.001 -0.002 0.028 -0.179 132.145
PHI Philippines 0.023 0.000 0.017 0.036 1.661 8.121 0.190 -0.041 2.210 155 2.2 3.1 0.003 0.000 -0.002 0.018 0.886 7.306
KEP South Korea 0.010 0.001 0.000 0.025 2.039 10.438 0.135 -0.038 1.375 88 1.2 6.1 0.001 0.078 -0.001 0.021 0.851 12.612
LPL South Korea 0.016 0.000 0.012 0.034 0.771 4.729 0.136 -0.066 1.623 122 1.7 2.2 0.003 0.010 -0.005 0.026 0.625 10.134
PKX South Korea 0.016 0.019 0.000 0.063 4.471 31.510 0.442 -0.112 0.872 123 1.7 2.2 0.003 0.028 -0.005 0.028 1.089 10.539
SKM South Korea 0.001 0.361 0.000 0.030 -0.156 3.332 0.074 -0.076 0.145 56 0.8 17.7 0.000 0.436 -0.001 0.020 0.039 16.032
ASX Taiwan 0.033 0.000 0.020 0.053 1.302 5.584 0.227 -0.070 2.150 159 2.2 3.9 0.003 0.000 -0.003 0.025 0.734 6.948
AUO Taiwan 0.011 0.007 0.000 0.036 2.312 13.981 0.207 -0.074 1.012 89 1.2 3.9 0.002 0.080 -0.003 0.024 0.434 7.765
CHT Taiwan 0.007 0.002 0.003 0.019 0.769 5.562 0.084 -0.035 1.202 58 0.8 10.5 0.001 0.085 -0.001 0.014 -0.043 5.922
TSM Taiwan 0.001 0.372 0.004 0.037 -0.850 5.107 0.085 -0.120 0.133 79 1.1 13.8 0.000 0.424 -0.001 0.022 0.006 5.457
UMC Taiwan 0.008 0.173 0.003 0.069 0.034 10.028 0.265 -0.292 0.387 70 1.0 16.5 0.000 0.290 -0.001 0.030 -0.450 7.625
Table 5.3: Monthly and individual trades excess returns of the pairs trading style strategy
Mean, Median, St.Dev, Skew. (skewness), Kurt. (kurtosis) – descriptive statistics of the monthly returns and returns in each individual trades; p-Value is
a result of single-sided t-Test and indicates level of statistical significance of the null hypothesis that the mean of returns equal zero, is p-Value less than 0.05,
then I reject the null hypothesis; Max – the highest monthly profit; Min – the highest monthly loss; Hold. time – average holding time, time the investor stays
in the market and keep open positions.
and some Taiwanese and South Korean companies represent Group 2 (Figure 5.2). Low
coefficient of lag one auto regression ensures stability of the spread, and high value of
β guarantees predictability of the tradable spread. It is the best combination for the
investor — a moderate level of markets segmentation provides opportunities to profit
at low level of risk.
Groupings presented in Table 5.1 are made by a simple rule: if the one lag auto re-
gression coefficient α is greater than 0.5, it is categorised in Group 1; if the coefficient
α is less than 0.5 and the coefficient β is greater than 0.5, it is in Group 2; all other
companies are classified into Group 3.
Companies/countries do not stay in one group permanently. Over time all markets
and companies evolve and move from Group 1 into Group 2 and then to Group 3. A
few reasons could be attributed to this transition. At the country level, there is glob-
alisation and market liberalisation, as that is demonstrated by Hsu and Wang (2008).
On the individual company level, one has to consider the degree of familiarity of the
investors (predominantly from the US) with the company. That transition can be seen
with the example of ICICI Bank Ltd as shown in Figure 5.4. ICICI Bank Ltd moved
from Group 1 of low efficient segmented markets, to Group 3 of high efficient integrated
markets during the period of only five years.
I report the results of the trading strategies after transaction costs in Table 5.2 (the
strategy with ADR conversion) and Table 5.3 (the adaptation of pairs trading strategy).
As could be expected from the analysis of Table 5.1, Group 3 makes no profit at all.
They are developed, highly integrated and highly efficient markets. Average return per
each trade is negative and close to transaction costs for all companies in that group.
Group 1 provides mixed and unpredictable results: some companies are profitable, oth-
ers are not. Most of them have a higher variance of returns and much longer average
120
holding time in the pairs trading style strategy than other groups which indicates them
as more risky.
Only Group 2 generates consistent profits from the both strategies of arbitrage-style
trading. Countries from Group 2 are moderately segmented markets. Levels of market
liberalisation and transparency are high enough to ensure free flow of international
capitals in and out of the country. At the same time, market inefficiencies provide
regular opportunities for profit.
Sharpe ratios for Groups 1 and 3 are equally low but for very different reasons: Group 1
has a high variance of returns, Group 3 has a low level of returns. Group 2 has Sharpe
ratios in excess of 2 for most companies. That demonstrates an economical significance
and relatively low risk of the arbitrage trading between the markets without overlap
in trading hours as long as those markets are from Group 2.
Results shown in Tables 5.2 and 5.3 demonstrate a high correlation relationship be-
tween groups and average profit. Similarly, there is a correlation relationship between
groups and average holding time. Group 1 has the longest holding time, which can be
directly connected to the risk level of the trading.
121
HDB = HDFC Bank Limited = India HDB = HDFC Bank Limited = India HDB = HDFC Bank Limited = India
0.15 20 3
0.1
15 2.5
0.05
10 2
0
−0.05 5 1.5
−0.1
0 1
−0.15
−5 0.5
−0.2
−0.25 −10 0
Jan05 Jan06 Jan07 Jan08 Jan09 Jan10 Jan11 Jul05 Jul06 Jul07 Jul08 Jul09 Jul10 Jul11 Jul05 Jul06 Jul07 Jul08 Jul09 Jul10 Jul11
(a) Spread process (b) Historical profit with ADR conversion (c) Historical profit pairs trading style
Figure 5.1: Example of a Group 1 company
122
SHI = Sinopec Shanghai Petrochemical Company Limited = China SHI = Sinopec Shanghai Petrochemical Company Limited = China SHI = Sinopec Shanghai Petrochemical Company Limited = China
0.06 6 3.5
0.04
3
5
0.02
2.5
0 4
2
−0.02
3
−0.04
1.5
−0.06 2
1
−0.08
1
0.5
−0.1
−0.12 0 0
Jan05 Jan06 Jan07 Jan08 Jan09 Jan10 Jan11 Aug05 Aug06 Aug07 Aug08 Aug09 Aug10 Aug05 Aug06 Aug07 Aug08 Aug09 Aug10
(d) Spread process (e) Historical profit with ADR conversion (f) Historical profit pairs trading style

SNE = Sony Corporation = Japan SNE = Sony Corporation = Japan SNE = Sony Corporation = Japan
0.1 0.05 0.05
0.08 0 0
0.06 −0.05 −0.05
0.04 −0.1 −0.1
0.02 −0.15 −0.15
0 −0.2 −0.2
−0.02 −0.25 −0.25
−0.04 −0.3 −0.3
−0.06 −0.35 −0.35
−0.08 −0.4 −0.4

Jan06 Jan07 Jan08 Jan09 Jan10 Jan11 Jul05 Jul06 Jul07 Jul08 Jul09 Jul10 Jul11 Jul05 Jul06 Jul07 Jul08 Jul09 Jul10 Jul11
(a) Spread process (b) Historical profit with ADR conversion (c) Historical profit pairs trading style
123
IBN = ICICI Bank Ltd. = India IBN = ICICI Bank Ltd. = India IBN = ICICI Bank Ltd. = India
0.1 6 3
0.05 5
2.5
4
0
2
3
−0.05
1.5
2
−0.1
1
1
−0.15
0
0.5
−0.2
−1
0
−0.25 −2
−0.3 −3 −0.5
Jan05 Jan06 Jan07 Jan08 Jan09 Jan10 Jan11 Jul05 Jul06 Jul07 Jul08 Jul09 Jul10 Jul11 Jul05 Jul06 Jul07 Jul08 Jul09 Jul10 Jul11
(d) Spread process (e) Historical profit with ADR conversion (f) Historical profit pairs trading style
Figure 5.4: Quick evolution of ICICI Bank Ltd from Group 1 to Group 3
5.6 Conclusion
The objective of this study was to ascertain if price parity is maintained between the
cross-listed Asia-Pacific stocks and their ADRs traded on NYSE. This study has shown
that while these assets may be traded freely across international borders, the outcome
of the arbitrage trading varies dramatically for different countries and companies. Price
deviations from the parity on the distance greater than transaction cost are observed on
all markets. However, they do not guarantee profitability of the arbitrage-style trading.
On the other hand, this study demonstrates that after careful consideration of the
degree of markets segmentation, markets transparency and familiarity with the com-
panies, it is possible to generate an economically significant profit at a low level of risk
from the arbitrage between the markets with non-overlapping trading hours.
124
Chapter 6
Conclusion
This chapter summarises the results and outlines practical implications of the findings
from all three studies presented in this thesis for academics and industry practitioners.
Also, it discusses general limitations and suggests future research directions.
6.1 Contributions to theory and knowledge
The overall aim of this research is to examine different approaches to market neutral
trading, provide improvements and theoretical justification for existing practices and
propose new methods of trading based on the statistical analysis of historical prices.
This thesis adds to the body of knowledge for finance and mathematics. In financial
terms it contributes to the studies of market efficiency, markets segmentation/integration
and market neutral trading strategies. In mathematical terms, it extends the theory
of autoregressive processes and study of the Ornstein–Uhlenbeck process.
Traditional financial theories, Capital Asset Pricing Model and Arbitrage Pricing The-
ory, assume markets to be efficient and, as a result, no abnormal return can be archived
systematically over a long period of time. While all three independent studies pre-
sented in the thesis demonstrate statistically significant abnormal returns from the
trading based on the historical prices only, they do not contradict or challenge existing
125
financial theories.
Contrarily, it is clearly shown that presented trading strategies generate economically

significant profit only on low efficient markets and do not earn anything on the truly
efficient markets. The reasons for market inefficiencies could be different: uncertainty
and high volatility of the GFC, less developed markets, small capitalisation companies
or a combination of those factors. Investors employing examined strategies (pairs trad-
ing or arbitrage style trading) look for market inefficiencies. The profit from trading
can be considered as reward for improving market efficiency and enforcing financial
theories. Hence, performance of the proposed methods, in particular pairs trading
based on statistical variability of the spread process, can be considered as a test on
market efficiency.
In the analysis of the Ornstein–Uhlenbeck process, I prove that its h-volatility is always
less than 2H for any H > 0. The consequence of this is as follows. The Ornstein–
Uhlenbeck process can be represented as a monotonic piece-wise approximation with
some fixed step H (in particular, renko and kagi constructions). Contrarian trading
strategy over that construction (that is, changing direction after each process move-
ment greater or equal to H) is profitable in the long-term for any choice of step H. This
result might look surprising and even counterintuitive. Trading towards the mean of
the Ornstein–Uhlenbeck process has positive expectation, while the proposed method
does not consider current position and trades disregarding the mean. However, this
sort of trading has positive expectation.
After the generalisation of the idea for auto-regressive processes and random walk, I
apply the method to a more general case of discrete processes. They have some mean-
reversion properties over the history but I do not assume they are auto-regressive
processes. This is a probabilistic–statistical approach to pairs trading and it forms a
method of pairs trading based on statistical variability of the spread process.
126
6.2 Research findings
The first study (Chapter 3) reviews and tests three most cited methods of pairs trading
in academic literature. Two of those methods have never been tested on real market
data before. I provide improvement and theoretical justification for the choice of pa-
rameters for the stochastic spread process method (Elliott et al., 2005).
The finding of the study is that all three methods stay market neutral, that is, very
low correlation with market indexes and market β close to 0. Level of risk, defined as
a standard deviation of returns, is low. However, the level of return is low and not
consistent over time.
The second study (Chapter 4) proposes a new non-parametric method of pairs trading
based on statistical variability of the spread process. I provide an extensive theoretical
background for potential profitability of the method under a very mild assumption of
maintaining the spread with some mean-reverting properties during the out-of-sample
period. The actual mean and variance of the spread process do not constrain the
method in any manner and can vary significantly.
The empirical test on the US and Australian market data demonstrate that the new
method outperforms the methods studied in Chapter 3 in terms of profitability, gener-
ating statistically significant monthly excess returns 1–3% after transaction costs, and
stability. Different levels of returns observed in the test are direct reflections of the
level of market efficiency.
The third study (Chapter 5) investigates an arbitrage style trading scenario for the
markets with no overlap in trading hours. I consider the Asia-Pacific stocks listed
in the home markets and ADR traded on the NYSE. Due to the time gap between
the markets, it is impossible to use arbitrage or pairs trading strategies. The investor
should take a potentially unlimited risk of holding just one leg of the trading pair.
127
The main finding of the study is that it is possible to trade cross-listed assets from
non-overlapping markets with economically significant returns with relatively low levels
of risk. This research proposes a new approach to market data analysis based on two
spreads, where one spread process can be a regressor for another one. The results
of such regression analysis work as discriminants allowing to segregate markets and
companies, and then to identify moderately segmented markets with a weak form of
market efficiency, which have a higher probability to generate profit from shares/ADR
trading.
6.3 Practical implications
There are a number of practical ramifications important for the finance industry.
Firstly, all studies in this thesis demonstrate that pairs trading is a viable way of in-
vesting, able to generate economically significant profit at a reasonable level of risk,
which can be controlled by the investor. Including pairs trading in a tactical asset
allocation strategy allows investors to diversify risks and improve the general perfor-
mance of their investments. This could be seen clearly in the example of the recent
GFC, when traditional methods of investment lost significant money but pairs trading
performed exceptionally well.
Secondly, this research shows that these strategies look for market inefficiencies and
profit from fixing them. Widespread of proposed strategies would be beneficial for fi-
nancial markets, as it improves markets efficiency. An investor employing pairs trading
strategy buys falling (or just under-performing stocks) and sells over-performing ones.
So, the investor increases market liquidity and reduces volatility.
Thirdly, the proposed methods might work as a tool for market efficiency analysis.
Based on the result of such analysis, investors can decide when it is the most advan-
128
tageous to enter the market or stay away from it depending on the evaluation of the
current market conditions and prevailing level of market efficiency.
Also, this study provides theoretical consideration for the potential use of the stop-loss
barriers. The stop-loss barrier is an order to close position if the trading process moves
in unfavorable direction on the distance exceeding some predefined level. The position
will be closed with some financial loss. However, it is assumed that stop-loss would
‘protect’ capital from even larger losses and, as a result, reduce risk.
The stop-loss barriers are acknowledged by many investment practitioners as an essen-

tial part of any trading strategy. However, in most cases, the practitioners base the
decision about implementation of the stop-loss barriers on their individual preferences
and personal level of risk aversion rather than mathematical models or market analysis,
which are the focus of this thesis.
For example, the Capital Assets Pricing Model considers a log-price process to be a
random walk. Using the Doob’s optional stopping theorem it is possible to show that
there are no stopping rules that can improve the expected return of the trading strategy
on the random walk process. Hence, the use of the stop-loss barriers in the context of
random walk (martingale process) does no make any difference and is not necessary.
If the trading process is a submartingale or a supermartingale, then one can expect

a continuation of the price up or down movement respectively. Hitting the stop-loss
barrier means that one trades against the market and might be a signal to reverse the
trading position. Under this scenario use of the stop-loss barriers would be essential
and can improve the strategy.
This thesis studies the spread processes between pairs of assets. These processes are
assumed to be mean-reverting or, at least, to have some mean-reverting properties.
129
The third study (Chapter 5) and, in particular, Theorem B.1 demonstrate that for this
kind of process, after a movement in one direction, we can expect an approximately
equal movement in the opposite direction. Hence, the loss would be recovered if the
investor keeps holding the position. So, the use of the stop-loss barriers can not be
justified theoretically and would be unreasonable.
6.4 Limitations and future research
The individual limitations of each study have been discussed in the previous chapters.
The common limitation for all empirical tests is the use of daily market prices. All
possible precautions are taken and reasonably conservative estimations are used for
transaction costs to ensure reliability of the testing process. Each test reported in this
thesis is a true out-of-sample test. However, the results of real-time tests might differ
from the historical tests.
A pure quantitative approach to pairs formation, parameters estimation and trading

allows formulation of unambiguous rules and testing those rules over long history data
sets. Including some fundamental data into consideration might improve strategy per-
formance, but this analysis lays outside the scope due to the quantitative nature of
this research.
The proposed method of pairs trading based on statistical variability of the spread
process opens a new direction in the analysis of mean-reverting and near mean-reverting
processes. This study can be extended by more rigorous rules for step size selection,
which does not need to be constant. It can be variable — stochastic or deterministic.
130
Appendix A
Crossing Time for the

Ornstein–Uhlenbeck Process
A.1 Most likely time to hit the mean for the Ornstein–
Uhlenbeck process
Define the Ornstein–Uhlenbeck process as Jeanblanc and Rutkowski (2000)
p
drt = (φ − λrt )dt + βdWt , r0 = r.
Then for any ρ > r the probability density function for the first passage time (tρ =
inf{t ≥ 0 : rt ≥ ρ}) for Ornstein–Uhlenbeck process
3/2 " 2 2 !#
ρ − r0 λt λ φ φ
f (t) = p eλt/2 exp − ρ− − r0 − + (ρ − r0 )2 coth λt .
2βπt3 sinh λt 2β λ λ
Consider a special case of hitting mean of the Ornstein–Uhlenbeck process, which

equals to zero, and change some variables to make equation in line with Finch (2004)
and Elliott et al. (2005)
φ = 0; ρ = 0; β = σ 2 ; r0 = −c; λ = ρ.
131
Hence the new Ornstein–Uhlenbeck process is
dX(t) = −ρX(t) dt + σ dW (t)
and its probability density function is
3/2
c ρt h ρ i
f (t) = √ eρt/2 exp − 2 − (−c)2 + (c)2 coth ρt
2σ 2 πt3 sinh ρt 2σ
3/2
ρ c2

c ρ ρt/2
= √ e exp − 2 (coth ρt − 1)
σ 2π sinh ρt 2σ
3/2
ρ c2 eρt + e−ρt

c 2ρ ρt/2
= √ e exp − 2 −1
σ 2π eρt − e−ρt 2σ eρt − e−ρt
ρ3/2 e−ρt ρ c2 e−2ρt

2c
= √ exp − 2
σ π (1 − e−2ρt )3/2 σ 1 − e−2ρt
.
√
If we take ρ = 1 and σ = 2 then
r
c e−t c2 e−2t

2
f (t) = exp − ,
π (1 − e−2t )3/2 2(1 − e−2t )
which is the same as in Finch (2004) and Elliott et al. (2005).
Next, we find t which maximizes f (t). First, take log f (t)
2 c ρ3/2 e−ρt ρ c2 e−2ρt

log f (t) = log √ + log + − 2
σ π (1 − e−2ρt )3/2 σ 1 − e−2ρt
2 c ρ3/2 ρ c2 e−2ρt

3 −2ρt
= log √ − ρt − log (1 − e )+ − 2
σ π 2 σ 1 − e−2ρt
2 c ρ3/2 3 ρ c2
= log √ − ρt + log (e−2ρt − 1) + 2 (e2ρt − 1)−1
σ π 2 σ
132
We take the first derivative and make it equal to zero
d log f (t) 3 2 ρ e−2ρt 2ρ2 c2 e2ρt

= −ρ + −
dt 2 (e−2ρt − 1) σ 2 (1 − e2ρt )2
3ρ 2ρ2 c2 e2ρt
= −ρ + − =0
1 − e2ρt σ 2 (1 − e2ρt )2
2ρ2 c2 2ρt
0 = −ρ(1 − e2ρt )2 + 3 ρ(1 − e2ρt ) − e
σ2
2ρ2 c2 2ρt

= ρ e4ρt + ρ − e − 2 ρ.
σ2
Then we solve quadratic equation for e2ρt

 s 
2 2 2 2
2
1  2ρ c 2ρ c
e2ρt = −ρ + 2
± ρ− + 8ρ2 
2ρ σ σ2
r
1 ρ c2 ρ2 c4 ρ c2 9
=− + 2 ± − 2 + .
2 σ σ4 σ 4
Hence
r !
1 1 ρ c2 ρ2 c4 ρ c2 9
T = log − + 2 + − 2 + .
2ρ 2 σ σ4 σ 4
√
If we consider a special case for ρ = 1 and σ = 2, then we get the same formula for t̂
as in Elliott et al. (2005).
Alternatively, we take a barrier

σ
c = λ√
2ρ
and calculate T for general case
s !
1 1 ρ λ2 σ 2 ρ2 λ4 σ 4 ρ λ2 σ 2 9
T = log − + 2 + − 2 +
2ρ 2 σ 2ρ σ 4 4ρ2 σ 2ρ 4
r !
1 1 λ2 λ4 λ2 9
= log − + + − + . (A.1)
2ρ 2 2 4 2 4
133
A.2 Estimation for optimal trigger level to start
trading for the Ornstein–Uhlenbeck process
Suppose that we open position on the spread at time t = 0 and plan to close it at time
t̂, which is the most likely time to cross mean of the Ornstein–Uhlenbeck process. We
want to be confident at some level P > 50% that we get a non-negative profit. What
is the optimal level to open position — Y (0)?
Without loss of generality we can assume that the mean of the spread process is zero.
Then there are two possible scenarios to profit from the expected mean-reversion: open
a short position if Y (0) > 0 or open a long position if Y (0) < 0. Solution for these two
cases are the similar due to the symmetry of the Ornstein–Uhlenbeck process, so we
consider only the first one with the short position.
Y (0) − Y (t̂) > 0

s
σ2
Y (0) − Y (0)e−ρt̂ − kP (1 − e−2ρt̂ ) > 0
2ρ
σ p
Y (0)(1 − e−ρt̂ ) − kP √ 1 − e−2ρt̂ > 0
2ρ
s
σ σ 1 + e−ρt̂
λ √ − kP √ >0
2ρ 2ρ 1 − e−ρt̂
s
1 + e−ρt̂
λ − kP > 0,
1 − e−ρt̂
where starting level λ represents a distance from the mean measured in the number of
√
standard deviations σ/ 2ρ of the stationary process with the same parameters as the
process y(t) and
r !
1 1 λ2 λ4 λ2 9
t̂ = log − + + − + .
2ρ 2 2 4 2 4
134
Hence
v
u q −1/2
u1 + −1 + λ2 λ4 λ2 9
u
u 2 2
+ 4
− 2
+ 4
λ − kP u
u q −1/2 > 0, and
λ2 λ4 λ2
1 − − 12 + 9
t
2
+ 4
− 2
+ 4
v
u 2
λ − kP u
u1 + q 1/2 > 0. (A.2)
λ2 λ4 λ2
− 21 + 9
t
2
+ 4
− 2
+ 4
−1
2
Profit
−2 P = 70%
P = 75%
P = 80%
P = 85%
−4
P = 90%
P = 95%
P = 97.5%
−6
1 1.5 2 2.5 3 3.5 4 4.5 5
λ −− number of std.dev. from the mean
Figure A.1: Starting point to trade
Figure A.1 demonstrates expected profit calculated by (A.2) for different values of kP ,
which are defined by the desirable probability to guarantee non-negative profit. Each
graph represents some probability P and shows the profit as a function of the level λ
above or below mean to open position, which is measured in standard deviations of the
Ornstein–Uhlenbeck process. All positions are closed at the most likely time to cross
mean t̂ calculated by (A.1).
135
As we can see from Figure A.1, to be 97.5% confident that the trade will bring non-
negative profit, when we close position at time t̂, the starting level Y (0) should be more
than three standard deviations away from the mean. Surprisingly, it does not depend
on the speed of mean reversion as we might expect.
However, as the Ornstein–Uhlenbeck process has a normal distribution, there is less

than 0.5% chances that the process hit a barrier at three standard deviations from the
mean. Waiting time can be too long. Moreover, for the real-life applications, the fact
of hitting that level can be considered as evidence that the process has changed and it
is probably not mean-reverting any more.
The starting level at two standard deviations away from the mean guaranties non-
negative profit in 85% of trades, which looks to be a reasonably good probability of
success.
A.3 Numerical simulation for the first passage time
of AR(1)
The simulation runs for B = 0.9, normally distributed increments with σ = 1 and
barrier/starting point c equals to two standard deviations of the mean-reverting process
yt , which is estimated as
σ2
Var(yt ) = .
1 − B2
Time that maximizes f (t) is 6-7. The above method (A.1) estimates the most likely
time for the process with the same parameters as 6.4. Therefore, it is reasonable to use
formula (A.1) for the most likely crossing time estimation. However, from the graphs
it’s obvious that less than a quarter of processes crossed mean over that time.
136
0.06
0.05
0.04
0.03
0.02
0.01
0
0 20 40 60 80 100 120 140 160 180
Figure A.2: Empirical PDF for the first passage time

1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0 20 40 60 80 100 120 140 160 180
Figure A.3: Empirical CDF for the first passage time
137
Appendix B
Proofs of the Theorem and Axillary

Lemmas
Before proving the theorem, we need some axillary lemmas.
Lemma B.1. Let {xt } be the Ornstein–Uhlenbeck process with mean µ, standard de-
viation σ and λ > 0 defined by
dxt = λ(µ − xt ) dt + σ dBt , (B.1)
where Bt is a standard Brownian motion.

Then it can be represented as a time change of another Brownian motion W :
σ
xt = x0 e−λt + µ(1 − e−λt ) + √ e−λt W (e2λt − 1).
2λ
Proof. The solution of (B.1) is
Z t Z t
−λt −λ(t−s)
xt = x0 e + e λµ ds + σ e−λ(t−s) dBs .
0 0
138
Then
Z t
λt λt
eλs dBs .

e xt = x0 + µ e − 1 + σ
0
Now we apply a time-change τ ,
Z τ (t)
λτ (t) λτ (t)
eλs dBs .

e xτ (t) = x0 + µ e −1 +σ (B.2)
0
The last integral is
Z τ (t) Z t Z t p
λs λτ (s)
e dBs = e dBτ (s) = eλτ (s) τ 0 (s) dW
fs (B.3)
0 0 0
where W
f is an another Brownian motion,
Z t
1
W
ft = p dBs .
0 τ 0 (s)
We choose τ so that
p
eλτ (s) τ 0 (s) = 1
1
τ (t) = log (2λt + 1).
2λ
Take an inverse
1 2λt
τ −1 (t) = (e − 1).
2λ
Now plug this into (B.2)
f (τ −1 (t))
eλt xt = x0 + µ eλt − 1 + σ W

−1 1 2λt
W
f (τ (t)) = W f (e − 1)
2λ
1
= √ W (e2λt − 1),
2λ
139
√ 1

where W (t) = 2λ W
f
2λ
is also a standard Brownian motion.
Hence, the Ornstein–Uhlenbeck process can be represented as a time-change of the
Brownian motion
σ
xt = x0 e−λt + µ(1 − e−λt ) + √ e−λt W (e2λt − 1).
2λ
Lemma B.2. Let {Yt } be the Ornstein–Uhlenbeck process with mean zero, variance
one and λ > 0 on the time interval [0, T ]
dYt = −λYt dt + dBt . (B.4)
We make a H-construction on the Ornstein–Uhlenbeck process for some H as in Section

4.2. Then the H-inversion goes to infinity as time goes to infinity, that is
NT (H, Y ) → ∞ (a.s.) as T → ∞.
Proof. Let ε > 0 and Y0 = −ε. We take the Ornstein–Uhlenbeck process as a time-
changed Brownian motion (Lemma B.1) and find the probability that the Ornstein–
Uhlenbeck process is above ε, that is P(Yt > ε).
The solution of (B.4) is
1
Yt = −εe−λt + √ e−λt W (e2λt − 1).
2λ
140
Hence,

−λt 1 −λt 2λt
P(Yt > ε) = P −εe + √ e W (e − 1) > ε
2λ

2λt
√ λt

= P W (e − 1) > 2λ 1 + e ε
√ !
W (e2λt − 1) 2λ(1 + eλt )ε
=P √ > √
e2λt − 1 e2λt − 1
√ !
2λ(1 + eλt )ε
=1−Φ √
e2λt − 1
√
= 1 − Φ( 2λε) as t → ∞
6= 0,
where
x u2
e− 2
Z
Φ(x) = √ du.
−∞ 2π
In a similar, way we can show that the probability of the Ornstein–Uhlenbeck process
with an initial value ε to be below −ε does not equal zero either. Therefore, the
Ornstein–Uhlenbeck process never converges completely to its mean but fluctuates
between −ε and ε. Then, if we take H ≤ 2ε
NT (H, P ) → ∞ (almost surely)
as T → ∞.
Lemma B.3. Limiting state probability of the recombining binomial tree approximation
of the Ornstein–Uhlenbeck process {yn } being on the level m is
1 −λ m(m−1) −2λ m
Q(m) = Q(0) e e +1 ,
2
141
where
∞
!−1
X
e−λ i(i−1) e−2λ i + 1

Q(0) = 1+ .
i=1
Proof. Let {xt } be the Ornstein–Uhlenbeck process with mean zero and ρ > 0
dxt = −ρxt dt + σdBt
and let {yn } be a recombining binomial tree approximation of the Ornstein–Uhlenbeck

process {xt } with the probability of moving up from the state yn
ρ(−yn ) √

↑ 1 1
P (yn ) = + tanh ∆t (B.5)
2 2 σ
and the size of step up or down
√
H = σ ∆t. (B.6)
If we take ∆t = 1, ρ = λ and σ = 1, then the size of up or down movement equals 1

and the process {yn } takes integer values yn = m, m ∈ [−n, . . . , −2, −1, 0, 1, 2, . . . , n].
The probability of moving up from the level m is
1 1
P↑ (m) = + tanh (−λ m) . (B.7)
2 2
We are interested in the limiting probability P(yn = m) as n → ∞, that is the process

{yn } is on any given level m. We use a brief notation Q(m) = P(yn = m) as n → ∞
and P↑ (m) and P↓ (m) for the probability of up and down movements from the level
m.
The process {yn } is symmetrical around zero. Then
Q(0) = P↓ (1) Q(1) + P↑ (−1) Q(−1) = 2 P↓ (1) Q(1). (B.8)
142
Working in the similar way and taking P↓ (0) = P↑ (0) = 1/2
Q(1) = P↓ (2) Q(2) + P↑ (0) Q(0)
= P↓ (2) Q(2) + P↑ (0) 2 P↓ (1) Q(1)
= P↓ (2) Q(2) + P↓ (1) Q(1)
Q(1)(1 − P↓ (1)) = P↓ (2) Q(2)

P↑ (1) P↓ (2)
Q(2) = Q(1) and Q(1) = Q(2) .
P↓ (2) P↑ (1)
Now we employ the same approach for the next level m = 2
Q(2) = P↓ (3) Q(3) + P↑ (1) Q(1)

P↓ (2)
= P↓ (3) Q(3) + P↑ (1) Q(2)
P↑ (1)
= P↓ (3) Q(3) + P↓ (2) Q(2)
Q(2)(1 − P↓ (2)) = P↓ (3) Q(3)

P↑ (2)
Q(3) = Q(2) .
P↓ (3)
We can repeat this exercise for the following levels and get a recursive relation for the
limiting probability of being on level m
P↑ (m − 1)
Q(m) = Q(m − 1) . (B.9)
P↓ (m)
To prove the claim (B.9) for the general case, we assume that it is true for m ≤ k and
check if it holds for m = k + 1. It is clear that the probability of being on level k is
Q(k) = Q(k − 1)P↑ (k − 1) + Q(k + 1)P↓ (k + 1).
143
By the formula (B.9) we get
P↑ (k − 1)
Q(k) = Q(k − 1)
P↓ (k)
⇒ Q(k − 1)P↑ (k − 1) = Q(k)P↓ (k).
Hence
Q(k) = Q(k)P↓ (k) + Q(k + 1)P↓ (k + 1)
Q(k)[1 − P↓ (k)] = Q(k + 1)P↓ (k + 1)
Q(k)P↑ (k) = Q(k + 1)P↓ (k + 1)

P↑ (k)
Q(k + 1) = Q(k) .
P↓ (k + 1)
Therefore, claim (B.9) is true for m = k + 1. We can then conclude that (B.9) holds
for all m by the Principle of Mathematical Induction.
The probability Q(m) = Q(−m) as the process {yn } is symmetrical. Then for m > 0
m−1
Y P↑ (j)
Q(m) = Q(0) .
j=0
P↓ (j + 1)
We know the probability of up or down movement, then
1
P↑ (i) 2
+ 12 tanh(−λi)
= 1
P↓ (i + 1) 2
− 12 tanh(−λ(i + 1))
e−2λi − 1
1+ −2λ(i+1)
= e−2λi + 1 = e−2λi e +1
.
−2λ(i+1) e −2λi +1
e −1
1 − −2λ(i+1)
e +1
It follows that
m−1
Y P↑ (j) −2λ0 e
−2λ1
+ 1 −2λ1 e−2λ2 + 1 −2λ2 e−2λ3 + 1 −2λ(m−1) e−2λm + 1
= e e e · · · e
j=0
P↓ (j + 1) e−2λ0 + 1 e−2λ1 + 1 e−2λ2 + 1 e−2λ(m−1) + 1
1
= exp (−λ m(m − 1)) (exp (−2λ m) + 1) .
2
144
Also, due to the symmetry of the Ornstein–Uhlenbeck process, the limiting probability
of being on level m = 0 can be calculated as
Q(0) = 1 − 2(Q(1) + Q(2) + Q(3) + . . . )

∞ m−1
!
X Y P↑ (j)
= 1 − 2 Q(0)
m=1 j=0
P↓ (j + 1)
∞
X
= 1 − Q(0) exp (−λ m(m − 1)) (exp (−2λ m) + 1)
m=1
1
= ∞ .
X
1+ exp (−λ m(m − 1)) (exp (−2λ m) + 1)
m=1
Hence, the limiting probability that the Ornstein–Uhlenbeck process {yn } is on level
m is
1 −λ m(m−1) −2λ m
Q(m) = Q(0) e e +1 ,
2
where
∞
!−1
X
e−λ i(i−1) e−2λ i + 1

Q(0) = 1+ .
i=1
Remark: If we consider a more general case with an arbitrary value of the step
up or down H, that is a classical renko chart, to get a recombining binomial tree
approximation, the value ∆t in (B.5) should be scaled
2
H
∆t = .
σ
Alternatively, one can keep ∆t = 1 and uses (B.7) with the unit increments and scaled
coefficient of mean-reversion
H
λ=ρ .
σ
145
Lemma B.4. The Ornstein–Uhlenbeck process satisfies the strong mixing condition
(α-mixing).
Proof. For any two sequences {ξ} = U 0 and {η} = U 00 with finite second moments, we
have the following index (Kolmogorov and Rozanov, 1960):
|E[(ξ − E[ξ])(η − E[η])]|

ρ(U 0 , U 00 ) = sup p .
ξ,η E[(ξ − E[ξ])2 ]E[(η − E[η])2 ]
If U 0 and U 00 are respectively the collections of all random variables which are measur-
able with respect to the σ-algebras M0 and M00 , then
ρ(M0 , M00 ) = ρ(U 0 , U 00 )
is the maximal correlation coefficient between the σ-algebras M0 and M00 .
Let x(t) be the Ornstein–Uhlenbeck process
dx(t) = −λ x(t) dt + σ dBt ,
then its stationary solution is
Z t
x(t) = σ e−λ(t−u) dBu .
−∞
For the process x(t), we have the following two measures of dependence α(τ ) =
α(Mt−∞ , M∞ t ∞ t
t+τ ) and ρ(τ ) = ρ(M−∞ , Mt+τ ), where Ms is the σ-algebra of events
which is determined by x(u), s ≤ u ≤ t (Kolmogorov and Rozanov, 1960).
For the Ornstein–Uhlenbeck process, the maximal correlation coefficient between σ-

algebras Mt−∞ and M∞
t+τ equals the module of the correlation coefficient between two
146
closest points from the above σ-algebras x(t) and x(t + τ ) and it depends only on τ
ρ(τ ) = ρ(Mt−∞ , M∞
t+τ ) = |ρ(x(t), x(t + τ ))| = e
−λτ
. (B.10)
The correlation coefficient between any linear combination of any other random vari-
ables from σ-algebras Mt−∞ and M∞
t+τ is less than (B.10). We can see it in the following
example.
Example: Let x(t) be the Ornstein–Uhlenbeck process and s ≤ t < z, z = t + n. Find

the correlation between x(t) + x(s) and x(z).
The covariance between x(t) + x(s) and x(z) is
Cov (x(t) + x(s), x(z)) =

Z t Z s Z z
−λ(t−u) −λ(s−u) −λ(z−u)
=E σ e dBu + e dBu σ e dBu
−∞ −∞ −∞
Z t Z t+n Z s Z t+n
2 −λ(t−u) −λ(t+n−u) −λ(s−u) −λ(t+n−u)
=σ E e dBu e dBu + e dBu e dBu
−∞ −∞ −∞ −∞
Z t Z s
2 −λ(t+t+n) 2λu −λ(s+t+n) 2λu
=σ e e du + e e du
−∞ −∞
2 σ 2 −λn
σ
e−λn + e−λ(t−s+n) = 1 + e−λ(t−s) .

= e
2λ 2λ
The variance of x(z) is σ 2 /2λ and variance of x(t) + x(s) is
" Z 2 #
t s
σ2
Z
e−λ(t−u) dBu + σ e−λ(s−u) dBu 1 + e−λ(t−s) .

Var(x(t)+x(s)) = E σ =
−∞ −∞ λ
Then the correlation between x(t) + x(s) and x(z) is
Cov (x(t) + x(s), x(z))

Corr (x(t) + x(s), x(z)) = p
Var(x(t) + x(s)) Var(x(z))
r
1 + e−λ(t−s)
= e−λn ≤ e−λn as t ≥ s.
2
The maximal correlation coefficient ρ(τ ) between σ-algebras Mt−∞ and M∞

t+τ goes to
147
zero as τ → ∞. This equivalents to α(τ ) → 0 as τ → ∞ (Kolmogorov and Rozanov,
1960; Bradley, 2005), where α(τ ) is a measure of dependence (Rosenblatt, 1956)
\
α(τ ) = α(M0 , M00 ) = sup |P(A0 A00 ) − P(A0 )P(A00 )|.
A0 ∈M0 ,A00 ∈M00
Hence, the Ornstein–Uhlenbeck process {xt } possesses the property of strong mixing.
Theorem B.1. H-volatility of the Ornstein–Uhlenbeck process.

Let P (t) be an Ornstein–Uhlenbeck process with mean zero and ρ > 0
dP (t) = −ρP (t) dt + σdBt .
Then for any positive H satisfying (4.1), the H-volatility is less than 2H
lim ξT (H, P ) < 2H. (B.11)

T →∞
Proof. Let {(τna , τnb ), n = 0, 1, ..., N } be a time sequence defined on the Ornstein–
Uhlenbeck process P (t) as in Section 4.2.
The H -inversion is a number of times H -process changes its direction and equals the
number of stopping times τnb when that change of direction manifests itself. Then by
Lemma B.2
N = NT (H, P ) → ∞ (almost surely) (B.12)
as T → ∞.
148
We define the distance between the two sequential local extremums
cn = |P (τna ) − P (τn−1
a
)|
= (P (τna ) − P (τn−1
a
)) · sign(P (τna ) − P (τn−1
a
))
= (P (τna ) − P (τnb ) + P (τnb ) − P (τn−1

a b
) + P (τn−1 b
) − P (τn−1 ))
a
· sign(P (τna ) − P (τn−1 ))
= P (τna ) − P (τnb ) − P (τn−1 a b

) + P (τnb ) − P (τn−1
b

) − P (τn−1 )
· sign P (τna ) − P (τn−1

a

) .
The distance between P (τna ) and P (τnb ) is equal to H by the rules of renko and kagi
constructions, but we need to know the sign for that distance. There are two possible
cases:
a
1. P (τna ) is a local maximum and P (τn−1 ) is a local minimum, then
P (τna ) − P (τnb ) − P (τn−1

a b
) · sign P (τna ) − P (τn−1
a

) − P (τn−1 )
= [H − (−H)] · 1 = 2H;
a
2. P (τna ) is a local minimum and P (τn−1 ) is a local maximum, then
P (τna ) − P (τnb ) − P (τn−1

a b
) · sign P (τna ) − P (τn−1
a

) − P (τn−1 )
= [−H − H] · (−1) = 2H.
It follows that
a
cn = |P (τna ) − P (τn−1 )|
= 2H + (P (τnb ) − P (τn−1
b
)) · sign(P (τna ) − P (τn−1
a
)) (B.13)
The value of sign(P (τna ) − P (τn−1

a
)) is completely defined by the process {P (t), t ∈
b
[τn−1 , τnb ]} and known at the stopping time τnb of the Ornstein–Uhlenbeck process, but
cn = |P (τna ) − P (τn−1
a
)| are not independent.
149
However, they are ‘nearly’ (or almost) independent. The sequence {cn } is stationary,
as the distribution of the random vector (cn , cn+1 , ..., cn+k ) does not depend on n, and
α-mixing with αn = 0 for large n by Lemma B.4. Hence, by the Central Limit Theorem
for Dependant Variables (Billingsley (1995), Theorem 27.4).
N
1 X
lim ξT (H, P ) = lim cn
T →∞ N →∞ N
n=1
N
1 X
= lim |P (τna ) − P (τn−1
a
)|
N →∞ N
n=1
→ E[|P (τ1a ) − P (τ0a )|] (a.s.) as T → ∞. (B.14)
Now we have to separate the proofs for renko and kagi constructions.
First we prove (B.11) for renko construction. We consider a sequence of random vari-
ables {dk , k = 1, 2, ...} such that

 1,

pk
dk = (B.15)
 −1, 1 − pk

Define the process

n
X
γn = dk , n = 1, 2, ... (B.16)
k=1
It is clear that the process {γn } is a recombining binomial tree approximation of the
Ornstein–Uhlenbeck process (van der Hoek, 2009) which has the following general
formula for the probability of moving up
ρ(µ − P (n)) √

1 1
pn = + tanh ∆t .
2 2 σ
For the process {γn } we take the probability in (B.15) as
1 1
pn = + tanh (−λγn ) .
2 2
150
Under the probability pn , the process {γn } defined by (B.15) and (B.16) is a recombin-
ing binomial tree approximation of the Ornstein–Uhlenbeck process (4.19) with λ = ρ Hσ
and µ = 0.
It follows from the definition of the stopping times τi for the renko process in (4.2) that
P (τi ) Law
= γn (B.17)
H
P (τi ) − P (τi−1 ) Law
= dn . (B.18)
H
We define a random variable
ν = min{n ≥ 1 : γn = n − 2} (B.19)
or equivalently
ν = min{n ≥ 1 : max (γt ) − γn = 1}, (B.20)
t∈[0,n]
which is a time of the first downfall of {γn }. We assume that γn−1 is a local maximum
discovered at time ν = n. The case with local minimum works in similar way due to
the symmetry of the Ornstein–Uhlenbeck process.
From (B.13) we have
Law
|P (τna ) − P (τn−1
a
)| = (2H + γν H)
= (2H + (ν − 2)H)
= νH (B.21)
E[|P (τ1a ) − P (τ0a )|] = HE[ν]. (B.22)
As the variable ν is a time of the first downfall after the number of raises then its
1
probability follows a geometric distribution with probability of ‘success’ pn = 2
−
151
1
2
tanh (−λγn ). The expected value of ν is
∞ n−1 1
X 1 1 1 1
E[ν] = n + tanh(−λγn ) − tanh(−λγn ) , (B.23)
n=1
2 2 2 2
where the current value of the process γn = γ0 + n − 1.
An initial value of the process γ0 can take any integer value from the minimal to
maximal value of the process {γn }.
Then
∞ ∞
X X (n + 1)
E[ν] = P(k) (1 + tanh(−λ(k + n)))n (1 − tanh(−λ(k + n)))
k=−∞ n=0
2n+1
∞ ∞
X (n + 1) X
= P(k) (1 + tanh(−λ(k + n)))n (1 − tanh(−λ(k + n))) , (B.24)
n=0
2n+1 k=−∞
where k takes integer values from (−∞, ∞) and P(k) is the probability that the initial
value equals k
P(k) = P(γ0 = k).
The density function of γ0 is provided by Lemma B.3.
1 −λ k(k−1) −2λ k
P(k) = P(0) e e +1 ,
2
where
∞
!−1
X
e−λ i(i−1) e−2λ i + 1

P(0) = 1+ .
i=1
152
Consider the second summation in (B.24)
∞
X
P(k) (1 + tanh(−λ(k + n)))n (1 − tanh(−λ(k + n))) =
k=−∞
∞
e−λ k(k−1) e−2λ k + 1

X
= P∞ −λ i(i−1) −2λ i ·
k=−∞
2 (1 + i=1 e (e + 1))
· (1 + tanh(−λ(k + n)))n (1 − tanh(−λ(k + n))) < 1. (B.25)
It looks impossible to get a closed form solution for this equation. However, numerical
simulations show that (B.25) is less than 1 for any λ > 0 and n ≥ 0. Hence, (B.24)
takes form
∞
X (n + 1)
E[ν] < = 2. (B.26)
n=0
2n+1
Therefore, from (B.22) and (B.26), we conclude that for the renko construction on the
Ornstein–Uhlenbeck process the H -volatility is less than 2H
ξT (H, P ) < 2H.
We now prove (B.11) for the kagi construction. Let
θ = min{u ≥ 0 : max P (t) − P (u) = H} (B.27)

t∈[0,u]
By Lemma B.1 P (θ) can be represented as a time-changed Wiener process defined by

the mean-reverting property of the Ornstein–Uhlenbeck process. Then it follows
1
|P (τ1b ) − P (τ0b )| = |P (τ0b )e−ρθ + √ e−ρθ W (e2ρθ − 1)| (B.28)
2ρ
and by (B.13)
E[|P (τ1a ) − P (τ0a )|] = E [2H + OUθ ] , (B.29)
153
where
OUθ = OU (τ0b , τ1b )

b −ρθ 1 −ρθ
= P (τ0 )e + √ e W (e − 1) sign(P (τ1a ) − P (τ0a )).
2ρθ
(B.30)
2ρ
As the process P (t) is the Ornstein–Uhlenbeck process then the initial point P (τ0b ) is
√
normally distributed with mean zero and standard deviation σ/ 2ρ. Hence
√
OU (τ0b , τ1b )
Z
−ρθ 1 −ρθ 2ρθ
2
− ρx2 ρ
= xe + √ e W (e − 1) e σ √ dx
sign(P (τ1a ) − P (τ0a )) R 2ρ πσ 2
1
= √ e−ρθ W (e2ρθ − 1)
2ρ
1
= √ W (1 − e−2ρθ )
2ρ
D
W (1 − e−2ρθ ) < W (2ρθ) for all ρ > 0. (B.31)
Hence, the Ornstein–Uhlenbeck process is smaller in distribution than the Wiener

process
OUθ < Wθ .
We can get the same result from the maximal inequalities for the Ornstein–Uhlenbeck
process (Graversen and Peskir, 2000). The Ornstein–Uhlenbeck process starting from
p √
its mean in average behaves as log(1 + t), while the Wiener process behaves as t.
Therefore, the Ornstein–Uhlenbeck process is smaller in distribution than the Wiener
process for any t > 0.
154
Then it follows from (B.29)
E[|P (τ1a ) − P (τ0a )|] = E [2H + OUθ ]
< E[2H + Wθ ]

Wθ
=HE 1+ 1+
H

Wθ
=H 1+E 1+
H
Z ∞
−x
=H 1+ x e dx = 2H (B.32)
0
So, for the kagi construction over the Ornstein–Uhlenbeck process the H -volatility is
less than 2H.
ξT (H, P ) < 2H.
155
Bibliography
Alexander, C. (2001). Market Models: A Guide to Financial Data Analysis. Wiley,

New York, NY.
Alexander, C. and Dimitriu, A. (2005). Indexing and statistical arbitrage: tracking

error or cointegration? Journal of Portfolio Management, 32(2):50–63.
Banerjee, A. (1999). Panel data unit roots and cointegration: An overview. Oxford
Bulletin of Economics and Statistics, 61(S1):607–629.
Billingsley, P. (1995). Probability and Measure. Wiley-Interscience, 3 edition.
Bock, M. and Mestel, R. (2009). A regime-switching relative value arbitrage rule. In

Fleischmann, B., Borgwardt, K.-H., Klein, R., and Tuma, A., editors, Operations
Research Proceedings 2008, pages 9–14. Springer Berlin Heidelberg.
Boguslavsky, M. and Boguslavskaya, E. (2003). Optimal arbitrage trading. SSRN

eLibrary. Available at: http://ssrn.com/abstract=446382.
Bowen, D., Hutchinson, M. C., and O’Sullivan, N. (2010). High frequency equity pairs
trading: Transaction costs, speed of execution and patterns in returns. Journal of
Trading, 5(3):31–38.
Bradley, R. C. (2005). Basic properties of strong mixing conditions. a survey and some
open questions. Probability Surveys, 2:107–144.
Brock, W., Lakonishok, J., and LeBaron, B. (1992). Simple technical trading rules and
the stochastic properties of stock returns. Journal of Finance, 47(5):1731–1764.
Brown, S. J. and Goetzmann, W. N. (2003). Hedge funds with style. Journal of

Portfolio Management, 29(2):101–112.
Chan, E. (2008). Quantitative Trading: How to Build Your Own Algorithmic Trading
Business. Wiley.
Chowhdry, B. and Nanda, V. (1991). Multimarket trading and market liquidity. Review
of Financial Studies, 4(3):483–511.
156
De Jong, A., Rosenthal, L., and Van Dijk, M. A. (2009). The risk and return of
arbitrage in dual-listed companies. Review of Finance, 13(3):495–520.
deB. Harris, F. H., McInish, T. H., Shoesmith, G. L., and Wood, R. A. (1995). Coin-
tegration, error correction, and price discovery on informationally linked security
markets. Journal of Financial and Quantitative Analysis, 30(04):563–579.
Dey, M. K. and Wang, C. (2012). Return spread and liquidity: Evidence from Hong
Kong ADRs. Research in International Business and Finance, 26(2):164–180.
Ding, D. K. (1999). Price parities of stocks listed on both the Kuala Lumpur stock
exchange and Singapores CLOC international, volume 6, pages 21–44. JAI Press
Inc.
Dixit, A. K. and Pindyck, R. S. (1994). Investment under Uncertainty. Princeton

University Press.
Do, B. and Faff, R. (2010). Does simple pairs trading still work? Financial Analysts
Journal, 66(4):83–95.
Do, B. and Faff, R. (2011). Are pairs trading profits robust to trading costs? SSRN
eLibrary. Available at: http://ssrn.com/abstract=1707125.
Do, B., Faff, R., and Hamza, K. (2006). A new approach to modeling and estimation for
pairs trading. In Proceedings of 2006 Financial Management Association European
Conference, Stockholm.
Efron, B. and Tibshirani, R. J. (1994). An Introduction to the Bootstrap. CRC Mono-

graphs on Statistics and Applied Probability. Chapman and Hall/CRC, 1 edition.
Elliott, R., Fischer, P., and Platen, E. (1999). Filtering and parameter estimation for
a mean reverting interest rate model. Technical Report 17, Quantitative Finance
Research Centre, University of Technology, Sydney.
Elliott, R. J., van der Hoek, J., and Malcolm, W. P. (2005). Pairs trading. Quantitative
Finance, 5(3):271–276.
Engle, R. F. and Granger, C. W. J. (1987). Co-integration and error correction: Rep-

resentation, estimation, and testing. Econometrica, 55(2):251–276.
Engle, R. F. and Granger, C. W. J. (1992). Long-Run Economic Relationships Readings

in Cointegration. Advanced Texts in Econometrics. Oxford University Press.
Eun, C. S. and Shim, S. (1989). International transmission of stock market movements.

Journal of Financial and Quantitative Analysis, 24(02):241–256.
157
Fama, E. F. and French, K. R. (1993). Common risk factors in the returns on stocks
and bonds. Journal of Financial Economics, 33(1):3–56.
Finch, S. (2004). Ornstein-Uhlenbeck process. Unpublished Note. Available online at:

http://en.scientificcommons.org/43397274.
Foerster, S. R. and Karolyi, G. A. (1993). International listings of stocks: The case of

Canada and the U.S. Journal of International Business Studies, 24(4):763–784.
Foerster, S. R. and Karolyi, G. A. (1999). The effects of market segmentation and

investor recognition on asset prices: Evidence from foreign stocks listing in the United
States. Journal of Finance, 54(3):981–1013.
Franco, J. C. G. (2003). Maximum likelihood estimation of mean reverting pro-

cesses. Unpublished Note. Available online at: http://www.investmentscience.
com/Content/howtoArticles/MLE_for_OR_mean_reverting.pdf.
Gagnon, L. and Karolyi, G. A. (2010). Multi-market trading and arbitrage. Journal

of Financial Economics, 97(1):53–80.
Galenko, A., Popova, E., and Popova, I. (2007). Trading in the presence of cointegra-
tion. SSRN eLibrary. Available at: http://ssrn.com/abstract=1023791.
Gatev, E., Goetzmann, W., and Rouwenhorst, K. (2006). Pairs trading: Performance
of a relative-value arbitrage rule. Review of Financial Studies, 19(3):797–827.
Grammig, J., Melvin, M., and Schlag, C. (2005). Internationally cross-listed stock
prices during overlapping trading hours: price discovery and exchange rate effects.
Journal of Empirical Finance, 12(1):139–164.
Granger, C. (1981). Some properties of time series data and their use in econometric
model specification. Journal of Econometrics, 16(1):121–130.
Granger, C. W. J. and Weiss, A. A. (1983). Time series analysis of error correction

models. In Studies in economic time series and multivariate statistics, pages 255–278.
Academic Press, New York.
Graversen, S. E. and Peskir, G. (2000). Maximal inequalities for the Ornstein-

Uhlenbeck process. Proceedings of the American Mathematical Society, 128(10):3035–
3041.
Hamao, Y., Masulis, R., and Ng, V. (1990). Correlations in price changes and volatility
across international stock markets. Review of Financial Studies, 3(2):281–307.
158
Hansen, P. R. (2005). A test for superior predictive ability. Journal of Business and
Economic Statistics, 23(4):365–380.
Hasbrouck, J. (1995). One security, many markets: Determining the contributions to

price discovery. Journal of Finance, 50(4):pp. 1175–1199.
Herlemont, D. (2004). Pairs trading, convergence trading, cointegration. Technical

report. Available online at: http://www.yats.com/doc/cointegration-en.pdf.
Hsu, J. and Wang, H.-Y. (2008). Why do price spreads between domestic shares and
their ADRs vary over time? Pacific Economic Review, 13(4):473–491.
Hsu, P.-H., Hsu, Y.-C., and Kuan, C.-M. (2010). Testing the predictive ability of
technical analysis using a new stepwise test without data snooping bias. Journal of
Empirical Finance, 17(3):471–484.
Hsu, P.-H. and Kuan, C.-M. (2005). Reexamining the profitability of technical analysis
with data snooping checks. Journal of Financial Econometrics, 3(4):606–628.
Ineichen, A. and Silberstein, K. (2008). AIMA’s roadmap to hedge funds. Technical

report, Alternative Investment Management Association. Available at: http://www.
aima.org/download.cfm/docid/6133E854-63FF-46FC-95347B445AE4ECFC.
Jeanblanc, M. and Rutkowski, M. (2000). Mathematical Finance: Theory and Practice,

chapter Modelling of Default Risk: An Overview, pages 171–269. Higher Education
Press, Beijing.
Johansen, S. and Juselius, K. (1990). Maximum likelihood estimation and inference

on cointegration–with applications to the demand for money. Oxford Bulletin of
Economics and Statistics, 52(2):169–210.
Kolmogorov, A. N. and Rozanov, Y. A. (1960). On strong mixing conditions for

stationary gaussian processes. Theory of Probability and its Applications, 5(2):204–
208.
LeSage, J. and Pace, K. (2009). Introduction to Spatial Econometrics. Statistics: A

Series of Textbooks and Monographs. CRC Press.
Licht, A. (1998). Regulatory arbitrage for real: International securities regulation in a

world of interacting securities markets. Virginia Journal of International Law, pages
563–636.
Lieberman, O., Ben-Zion, U., and Hauser, S. (1999). A characterization of the price
behavior of international dual stocks: An error correction approach. Journal of
International Money and Finance, 18(2):289–304.
159
Lo, A. W. and MacKinlay, A. C. (1990). Data-snooping biases in tests of financial
asset pricing models. Review of Financial Studies, 3(3):431 – 467.
Lok, E. and Kalev, P. S. (2006). The intraday price behaviour of Australian and
New Zealand cross-listed stocks. International Review of Financial Analysis, 15(4 -
5):377–397. Asian Market Microstructure.
Loomis, C. J. (1966). The Jones nobody keeps up with. Fortune, April:237–247.
Maldonado, R. and Saunders, A. (1983). Foreign exchange restrictions and the law of
one price. Financial Management, 12(1):19–23.
Mandelbrot, B. (1963). The variation of certain speculative prices. Journal of Business,

36(4):394–419.
Miller, D. P. and Morey, M. R. (1996). The intraday pricing behavior of international

dually listed securities. Journal of International Financial Markets, Institutions and
Money, 6(4):79–89.
Modigliani, F. and Modigliani, L. (1997). Risk-adjusted performance. Journal of

Portfolio Management, 23(2):45–54.
Nath, P. (2003). High frequency pairs trading with U.S. treasury securities: Risks and
rewards for hedge funds. SSRN eLibrary. Available at http://ssrn.com/abstract=
565441.
Neumark, D., Tinsley, P. A., and Tosini, S. (1991). After-hours stock prices and post-
crash hangovers. Journal of Finance, 46(1):159–178.
Novikov, A. and Kordzakhia, N. (2008). Martingales and first passage times of AR(1)
sequences. Stochastics: An International Journal of Probability and Stochastic Pro-
cesses, 80(2-3):197–210.
Pascual, R., Pascual-Fuster, B., and Climent, F. (2006). Cross-listing, price discov-
ery and the informativeness of the trading process. Journal of Financial Markets,
9(2):144–161.
Pastukhov, S. V. (2005). On some probabilistic-statistical methods in technical anal-

ysis. Theory of Probability and Its Applications, 49(2):245–260.
Patton, A. J. (2009). Are ‘market neutral’ hedge funds really market neutral? Review
of Financial Studies, 22(7):2495–2530.
Perlin, M. S. (2009). Evaluation of pairs-trading strategy at the Brazilian financial

market. Journal of Derivatives and Hedge Funds, 15(2):122–136.
160
Rosenblatt, M. (1956). A central limit theorem and a strong mixing condition. Pro-
ceedings of the National Academy of Science, 42:43–47.
Rosenthal, L. (1983). An empirical test of the efficiency of the ADR market. Journal
of Banking and Finance, 7(1):17–29.
Sharpe, W. F. (1964). Capital asset prices: A theory of market equilibrium under

conditions of risk. Journal of Finance, 19(3):425–442.
Stock, J. H. and Watson, M. W. (1988). Testing for common trends. Journal of the
American Statistical Association, 83(404):1097–1107.
Suarez, E. D. (2005). Arbitrage opportunities in the depositary receipts market: Myth

or reality? Journal of International Financial Markets, Institutions and Money,
15(5):469–480.
Treynor, J. L. (1962). Toward a theory of market value of risky assets. Unpublished

manuscript.
Triantafyllopoulos, K. and Montana, G. (2011). Dynamic modeling of mean-reverting

spreads for statistical arbitrage. Computational Management Science, 8:23–49.
10.1007/s10287-009-0105-8.
Uhlenbeck, G. E. and Ornstein, L. S. (1930). On the theory of the Brownian motion.

Physical Review, 36(5):823–841.
van der Hoek, J. (2009). Recombining binomial tree approximations for diffusions. In
Ciarlet, P., editor, Special Volume: Mathematical Modeling and Numerical Methods
in Finance, volume 15 of Handbook of Numerical Analysis, pages 361–368. Elsevier.
Vidyamurthy, G. (2004). Pairs trading: Quantitative methods and analysis. J. Wiley,

Canada.
Wahab, M., Lashgari, M., and Cohn, R. (1992). Arbitrage opportunities in the Ameri-
can depository receipts market revisited. Journal of International Financial Markets,
Institutions and Money, 2(3-4):97–130.
Werner, I. and Kleidon, A. (1996). U.K. and U.S. trading of British cross-listed stocks:
an intraday analysis of market integration. Review of Financial Studies, 9(2):619–
664.
Whistler, M. (2004). Trading Pairs: Capturing Profits and Hedging Risk with Statistical
Arbitrage Strategies. Wiley.
White, H. (2000). A reality check for data snooping. Econometrica, 68(5):1097–1126.
161
Wu, P. and Elliott, R. J. (2005). Parameter estimation for a regime-switching mean-
reverting model with jumps. International Journal of Theoretical and Applied Fi-
nance, 8(6):791–806.
162

Trading Strategies Used by Hedge Funds

Загружено:

Сведения о документе

Оригинальное название

Авторское право

Доступные форматы

Поделиться этим документом

Поделиться или встроить документ

Параметры публикации

Этот документ был вам полезен?

Это неприемлемый материал?

Авторское право:

Доступные форматы

Trading Strategies Used by Hedge Funds

Загружено:

Авторское право:

Доступные форматы

Trading Strategies

Used by Hedge Funds

Associate Professor John van der Hoek

Professor Robert Elliott

Professor Petko Kalev

School of Mathematics and Statistics

Division of Information Technology,

University of South Australia

Table of Abbreviations viii

3 Study 1. Performance Evaluation of the Pairs Trading Strategies 27

4 Study 2. Pairs Trading Based on Statistical Variability of the Spread

5 Study 3. Arbitrage Strategies Between Listed Asian-Pacific Stocks

A Crossing Time for the Ornstein–Uhlenbeck Process 131

B Proofs of the Theorem and Axillary Lemmas 138

3.1 Historical performance of the distance method of pairs trading on the

5.1 Example of a Group 1 company . . . . . . . . . . . . . . . . . . . . . . 122

A.1 Starting point to trade . . . . . . . . . . . . . . . . . . . . . . . . . . . 135

3.1 Statistics on data sets used in the Australian and US markets . . . . . 41

4.1 Statistics on data sets used in the Australian and US markets . . . . . 84

5.1 Descriptive statistics of the visible return spreads s1 (t), coefficient of

ADF Augmented Dickey–Fuller test

ADR American Depositary Receipt

APT Arbitrage Pricing Theory

ASIC Australian Securities & Investments Commission

ASX Australian Stock Exchange

AUM Assets Under Management

CAPM Capital Asset Pricing Model

GARCH Generalized Autoregressive Conditional Heteroskedasticity model

GFC Global Financial Crisis

LOP Law of One Price

NASDAQ National Association of Securities Dealers Automated Quotations

NYSE New York Stock Exchange

OLS Ordinary Least Squares

PCA Principal Component Analysis

RAP Risk Adjusted Performance

ROI Return On Investment

SIRCA Securities Industry Research Centre of Asia-Pacific

1.1.1 Hedge funds

1.1.2 Pairs trading

1. The strategy is market-neutral. This means that it is immune to market fluctu-

2. The strategy is self-financing (dollar-neutral ). This means that the investors do

Market neutrality is appealing for superannuation funds, insurance companies and

Do et al. (2006) review different approaches to pairs trading described in academic

1. the distance method used by Gatev et al. (2006);

2. the cointegration method described by Vidyamurthy (2004); and

3. the stochastic spread method proposed by Elliott et al. (2005).

• the natural benefits of pairs trading being market-neutral and dollar-neutral.

A trading style closely related to pairs trading is arbitrage — simultaneously buying

In this thesis I have the following research objectives:

2. provide improvements and theoretical justification for the existing methods of

1.2.1 Study 1. Performance evaluation of the pairs trading

1.2.2 Study 2. Pairs trading based on statistical variability of

the spread process

1.2.3 Study 3. Arbitrage strategies between listed Asia-Pacific

stocks and their NYSE ADRs when there is no overlap

A large amount of academic literature is dedicated to research of arbitrage between

This research also contributes to the theory of auto-regressive processes, mean-reverting

Definition. A time series {xt ; t = 0, 1, 2, . . . } is said to be weakly stationary (also

• Cov[xt , xt−s ] depends only on the lag s

Definition. A time series is integrated of order n, written yt ∼ I(n), if its stochastic

The first difference of the process {yt } is

4yt = yt − yt−1 = εt ∼ I(0).

Definition. A set of I(1) time series is cointegrated if there is a linear combination

where ∆t → 0 and {t } are iidN (0, σ2 ), σ2 = σ 2 ∆t.

Xt − Xt−1 = A + BXt−1 + t . (2.19)