Wei Thesis PDF

An Empirical Investigation of Price Impact:
An Agent-based Modelling Approach
By
Wei Cui
B.Sc. Hon (Wuhan University)
M.Eng. Hon (Dublin City University)
Thesis presented for the degree of Ph.D.

to the
Michael Smurfit School of Business
University College Dublin
Research Supervisor:
Prof. Anthony Brabazon
November 2012
Abstract
Understanding price impact is a fundamental task in finance. Many ex-

ecution algorithms, used to execute a large order by dividing and spreading
it over time, are based on price effects and in particular on the way how
volume affects prices. Moreover, the analysis of price impact is helpful for
understanding how financial markets function as price impact is one of the
mechanisms determining price formation.
The thesis is motivated by the recent emergence of algorithmic trading
which requires a good understanding of price impact. This thesis addresses
three questions concerning price impact in order to gain a better understand-
ing on the intraday behaviour of price impact, and the factors affecting price
impact.
The first study examines the intraday behaviours of price impact and
market liquidity. The data is drawn from the NYSE-Euronext TAQ database
and the LSE ROB database. Six stocks from the US markets and six stocks
from the UK markets are analysed. The intraday patterns on price volatility,
bid-ask spread, trading volume and market depth are documented and gener-
ally confirm findings in prior studies on intraday phenomena. In particular,
a reverse S-shaped intraday pattern on price impact is found for both US and
UK stocks for the first time.
The second study investigates whether agent intelligence plays an im-
portant role in determining the magnitude of price impact. This chapter
constructs an artificial stock market composed of zero-intelligence agents,
and calibrates it using the LSE ROB data. The result shows that the price
impact in the artificial market is generally larger than that in the real mar-
ket. This is consistent with the hypothesis that agent intelligence plays an
important role in determining the magnitude of price impact. It supports the
selective liquidity argument in Farmer et al. (2004) & Hopman (2007).
The third study addresses whether order choice affects the price impact
i
of trading a large order. A typical approach in trading a large order is to
devise a strategy which divides it into numerous pieces and spreads it over
time (usually one trading day). In this study, several execution strategies
with various order types, and a number of simple strategies with one order
type as benchmarks are constructed and evaluated by their effects on prices.
Novelly, these strategies are evolved and evaluated in simulated artificial
markets. The results show that the combined strategies outperform the sim-
ple strategies significantly, suggesting that order choice plays an important
role in determining the price impact of trading large orders.
The results in this thesis suggest that time-of-the-day, agent intelligence
and order choice are important factors affecting price impact, and need to
be considered in the theoretical microstructure models and in the design of
trading strategies.
ii
Acknowledgements
I wish to express my greatest appreciation to Prof. Anthony Brabazon, my
academic supervisor, for his strong support and insightful suggestions. He pro-
vided me with valuable guidance and was always accessible and generous with
his comments whenever I had questions. Without his consistent encouragement
and help, this work would be impossible.
A special note of my sincere gratitude goes to Dr. Michael O’Neill, for his
contribution, encouragement and support. A warm thank to the wonderful FMC2
and NCRA group members for their constant support and all their fruitful discus-
sions.
I want to thank also the two reviewers of my dissertation: Dr. Mark Hutchin-
son and Dr. Conall O’Sullivan, for their involvement in reviewing the manuscript.
I dedicate this thesis to my family and friends who all came along for the ride
and without whom none of this would have been possible. Special thanks go to
Dad and Mom.
I would also like to thank Science Foundation Ireland. This research is based
upon works supported by the Science Foundation Ireland under Grant Number
08/SRC/FM1389.
Last but not least, very special thanks to my dear wife for her unflagging love
and support.
iii
Contents
Abstract i
Acknowledgements iii
Table of Contents iv
List of Figures ix
List of Tables xii
Publications xiv
1 Introduction 1
1.1 Limit Order Market and Trade Execution . . . . . . . . . . . . . 2
1.1.1 Limit Order Book . . . . . . . . . . . . . . . . . . . . . . 4
1.1.2 Trade Execution . . . . . . . . . . . . . . . . . . . . . . 6
1.2 Research Aims . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.2.1 Objectives of Research . . . . . . . . . . . . . . . . . . . 10
1.3 Framework of Research . . . . . . . . . . . . . . . . . . . . . . . 10
1.4 Contributions of Thesis . . . . . . . . . . . . . . . . . . . . . . . 12
1.5 Scope Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.6 Structure of Thesis . . . . . . . . . . . . . . . . . . . . . . . . . 15
iv
I Literature Review 17
2 Market Microstructure 19
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.2 Overview of Market Microstructure . . . . . . . . . . . . . . . . 20
2.2.1 Market Liquidity . . . . . . . . . . . . . . . . . . . . . . 21
2.2.2 Theoretical Market Microstructure Models . . . . . . . . 22
2.3 Price Impact . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.3.1 Price Impact and Trade Size . . . . . . . . . . . . . . . . 24
2.3.2 Large Price Fluctuations and the Order Book . . . . . . . 27
2.4 Intraday Patterns . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.4.1 Intraday Price Volatility . . . . . . . . . . . . . . . . . . 31
2.4.2 Intraday Bid-ask Spread . . . . . . . . . . . . . . . . . . 33
2.4.3 Intraday Trading Volume . . . . . . . . . . . . . . . . . . 35
2.4.4 Intraday Market Depth . . . . . . . . . . . . . . . . . . . 37
2.5 Trade Execution Strategies . . . . . . . . . . . . . . . . . . . . . 39
2.5.1 Optimal Trading Strategy . . . . . . . . . . . . . . . . . . 39
2.5.2 Order Submission Strategy . . . . . . . . . . . . . . . . . 43
2.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3 Agent-based Modelling in Finance 51

3.1 Agent-based Financial Markets . . . . . . . . . . . . . . . . . . . 52
3.2 The Core Elements of Agent-based Market Simulation . . . . . . 54
3.2.1 Agents . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.2.2 Trading Mechanism . . . . . . . . . . . . . . . . . . . . 56
3.3 Genetic Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.3.1 Grammar-based Genetic Algorithm . . . . . . . . . . . . 59
3.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
v
II Experiments 64
4 Datasets and Data Processing 66

4.1 Datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.1.1 NYSE Euronext TAQ Database . . . . . . . . . . . . . . 68
4.1.2 LSE ROB Database . . . . . . . . . . . . . . . . . . . . . 76
4.1.3 Sample Data . . . . . . . . . . . . . . . . . . . . . . . . 77
4.2 Data Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
4.2.1 Preprocessing of TAQ data . . . . . . . . . . . . . . . . . 83
4.2.2 Preprocessing of ROB data . . . . . . . . . . . . . . . . . 98
4.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
5 An Analysis of Intraday Behaviours of Price Impact and Market Liq-

uidity 103
5.1 Measures of Market Liquidity . . . . . . . . . . . . . . . . . . . 105
5.1.1 Price-based Measures . . . . . . . . . . . . . . . . . . . . 105
5.1.2 Volume-based Measures . . . . . . . . . . . . . . . . . . 106
5.1.3 Price and Volume Mixed Measures . . . . . . . . . . . . 107
5.2 Empirical Results and Discussions . . . . . . . . . . . . . . . . . 108
5.2.1 Summary Statistics . . . . . . . . . . . . . . . . . . . . . 110
5.2.2 Intraday Price Volatility . . . . . . . . . . . . . . . . . . 116
5.2.3 Intraday Bid-ask Spread . . . . . . . . . . . . . . . . . . 120
5.2.4 Intraday Market Depth . . . . . . . . . . . . . . . . . . . 124
5.2.5 Intraday Trade Size and Volume . . . . . . . . . . . . . . 134
5.2.6 Intraday Price Impact . . . . . . . . . . . . . . . . . . . . 145
5.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
6 An Investigation of Price Impact and Agent Intelligence 151

6.1 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
vi
6.2 A Zero-intelligence based Model of Limit Order Market . . . . . 154
6.2.1 Specification of Model . . . . . . . . . . . . . . . . . . . 155
6.2.2 Algorithm of Model . . . . . . . . . . . . . . . . . . . . 156
6.3 Experimental Setup . . . . . . . . . . . . . . . . . . . . . . . . . 158
6.3.1 Parameters of Model . . . . . . . . . . . . . . . . . . . . 158
6.3.2 Parameter Estimation . . . . . . . . . . . . . . . . . . . . 158
6.3.3 Simulation Setup . . . . . . . . . . . . . . . . . . . . . . 159
6.4 Results and Discussions . . . . . . . . . . . . . . . . . . . . . . . 161
6.4.1 Results of Probabilities of Events and Limit Order Types . 162
6.4.2 Summary Statistics of Order Sizes and Relative Limit Prices162
6.4.3 Results of Price Impact vs. Trade Size . . . . . . . . . . . 165
6.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
7 Dynamic Trade Execution with Limit and Market Orders 170

7.1 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
7.2 Designs of Experiments . . . . . . . . . . . . . . . . . . . . . . . 173
7.2.1 Design of Experiment One . . . . . . . . . . . . . . . . . 177
7.2.2 Design of Experiment Two . . . . . . . . . . . . . . . . . 178
7.2.3 Design of Experiment Three . . . . . . . . . . . . . . . . 181
7.3 Artificial Market Simulators . . . . . . . . . . . . . . . . . . . . 183
7.3.1 Simulator I . . . . . . . . . . . . . . . . . . . . . . . . . 183
7.3.2 Simulator II . . . . . . . . . . . . . . . . . . . . . . . . . 186
7.4 Results and Discussions . . . . . . . . . . . . . . . . . . . . . . . 189
7.4.1 Results of Experiment One . . . . . . . . . . . . . . . . . 189
7.4.2 Results of Experiment Two . . . . . . . . . . . . . . . . . 191
7.4.3 Results of Experiment Three . . . . . . . . . . . . . . . . 195
7.4.4 Discussions . . . . . . . . . . . . . . . . . . . . . . . . . 196
7.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
vii
III Conclusions 199
8 Conclusions and Future Work 200

8.1 Thesis Summary . . . . . . . . . . . . . . . . . . . . . . . . . . 200
8.1.1 Contributions . . . . . . . . . . . . . . . . . . . . . . . . 202
8.2 Limitations of this Thesis . . . . . . . . . . . . . . . . . . . . . . 205
8.3 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
A Explanation of the Chosen Time Period in Chapter 5 208
B More Results of Chapter 5 211
Bibliography 222
viii
List of Figures
2.1 Different Shapes of Intraday Patterns . . . . . . . . . . . . . . . . 30
3.1 Evolutionary Cycle . . . . . . . . . . . . . . . . . . . . . . . . . 58

3.2 An Example GE Individual’s Genome Represented as Integers for
Ease of Reading. . . . . . . . . . . . . . . . . . . . . . . . . . . 62
4.1 Data Selection Screen for TAQ3 Extraction Tool . . . . . . . . . . 70

4.2 Data Filter Screen for TAQ3 Extraction Tool . . . . . . . . . . . . 71
4.3 Unfiltered Time Series for CHDX at 02/02/09 from 9:30-11:00
am. The blue line depicts quotes occurring at NASDAQ; the green
line depicts quotes occurring at all exchanges; the red crosses rep-
resent the transaction prices. . . . . . . . . . . . . . . . . . . . . 84
4.4 Frequencies of Price Changes for CHDX. (a) frequency of bid
price change (b) frequency of ask price change (c) frequency of
transaction price change. . . . . . . . . . . . . . . . . . . . . . . 89
4.5 Frequencies of Transaction Price Changes for Six Stocks . . . . . 90
4.6 Tick-by-tick Price Series on 03 Feb 2009 (9:30-11:00) for CHDX.
The blue line (below) represents the bid price time series; the
green line (above) represents the ask price time series; the red
circles represent the transaction prices. . . . . . . . . . . . . . . . 91
4.7 Raw Data for CHDX on Feb 10, 2009 . . . . . . . . . . . . . . . 96
ix
4.8 Clean Data for CHDX on Feb 10, 2009 . . . . . . . . . . . . . . 97
4.9 ROB Data Preprocessing Chart . . . . . . . . . . . . . . . . . . . 99
5.1 Price of XOM during Feb. 2009 (from Yahoo Finance) . . . . . . 114
5.2 Intraday Price Volatility (US stocks) . . . . . . . . . . . . . . . . 118
5.3 Intraday Price Volatility (UK stocks) . . . . . . . . . . . . . . . . 119
5.4 Intraday Proportional Bid-Ask Spread (US stocks) . . . . . . . . . 121
5.5 Intraday Proportional Spread (UK stocks) . . . . . . . . . . . . . 122
5.6 Intraday Time-weighted Proportional Spread (UK stocks) . . . . . 123
5.7 Intraday Bid Depth in Share (US stocks) . . . . . . . . . . . . . . 126
5.8 Intraday Ask Depth in Share (US stocks) . . . . . . . . . . . . . . 127
5.9 Intraday Depth in Share (UK stocks) . . . . . . . . . . . . . . . . 128
5.10 Intraday Bid Depth in Share (UK stocks) . . . . . . . . . . . . . . 129
5.11 Intraday Ask Depth in Share (UK stocks) . . . . . . . . . . . . . 130
5.12 Intraday Near Depth in Share (UK stocks) . . . . . . . . . . . . . 131
5.13 Intraday Near Bid Depth in Share (UK stocks) . . . . . . . . . . . 132
5.14 Intraday Near Ask Depth in Share (UK stocks) . . . . . . . . . . 133
5.15 Intraday Trading Volume in Share (US stocks) . . . . . . . . . . . 136
5.16 Intraday Trading Volume in Value (US stocks) . . . . . . . . . . . 137
5.17 Intraday Trading Volume in Share (UK stocks) . . . . . . . . . . 138
5.18 Intraday Trading Volume in Value (UK stocks) . . . . . . . . . . 139
5.19 Intraday Trade Frequency (US stocks) . . . . . . . . . . . . . . . 140
5.20 Intraday Trade Frequency (UK stocks) . . . . . . . . . . . . . . . 141
5.21 Intraday Trade Size (US stocks) . . . . . . . . . . . . . . . . . . 142
5.22 Intraday Trade Size in Share (UK stocks) . . . . . . . . . . . . . 143
5.23 Intraday Trade Size in Value (UK stocks) . . . . . . . . . . . . . 144
5.24 Intraday Price Impact Ratio (US stocks) . . . . . . . . . . . . . . 147
5.25 Intraday Price Impact Ratio (UK stocks) . . . . . . . . . . . . . . 148
x
6.1 Price Impact vs. Trade Size in the LSE market . . . . . . . . . . . 166
6.2 Price Impact vs. Trade Size in the simulated market . . . . . . . . 167
7.1 Representation of Order’s Relative Price and Lifetime . . . . . . . 178

7.2 Grammars in Experiment Two . . . . . . . . . . . . . . . . . . . 180
7.3 An Example of Trade Execution Strategy in Experiment Two . . . 181
7.4 Grammars in Experiment Three . . . . . . . . . . . . . . . . . . 182
7.5 An Example of Trade Execution Strategy in Experiment Three . . 182
7.6 Place Limit Price . . . . . . . . . . . . . . . . . . . . . . . . . . 184
A.1 Intraday Price Volatility (One-day data) . . . . . . . . . . . . . . 209

A.2 Intraday Price Volatility (One-month data) . . . . . . . . . . . . . 210
B.1 Intraday Time-weighted Depth in Share (UK stocks) . . . . . . . 212

B.2 Intraday Depth in Value (UK stocks) . . . . . . . . . . . . . . . . 213
B.3 Intraday Time-weighted Depth in Value (UK stocks) . . . . . . . 214
B.4 Intraday Bid Depth in Value (UK stocks) . . . . . . . . . . . . . . 215
B.5 Intraday Ask Depth in Value (UK stocks) . . . . . . . . . . . . . 216
B.6 Intraday Time-weighted Near Depth in Share (UK stocks) . . . . 217
B.7 Intraday Near Depth in Value (UK stocks) . . . . . . . . . . . . . 218
B.8 Intraday Time-weighted Near Depth in Value (UK stocks) . . . . 219
B.9 Intraday Near Bid Depth in Value (UK stocks) . . . . . . . . . . . 220
B.10 Intraday Near Ask Depth in Value (UK stocks) . . . . . . . . . . 221
xi
List of Tables
1.1 Order Book 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

1.2 Order Book 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Order Book 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.4 Order Book 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.1 Price Impact Function of Trade Size . . . . . . . . . . . . . . . . 25

2.2 Intraday Pattern: Price Volatility . . . . . . . . . . . . . . . . . . 32
2.3 Intraday Pattern: Bid-Ask Spread . . . . . . . . . . . . . . . . . . 34
2.4 Intraday Pattern: Trading Volume . . . . . . . . . . . . . . . . . 36
2.5 Intraday Pattern: Market Depth . . . . . . . . . . . . . . . . . . . 38
4.1 A Sample of Trade Data for Ford Drawn from 2/2/09 . . . . . . . 72

4.2 A Sample of Quote Data for Ford Drawn from 2/2/09 . . . . . . . 73
4.3 Exchange Codes as at 31 January 2011 . . . . . . . . . . . . . . . 74
4.4 US Daily Market Share Equities for 2 February 2011 (Financial-
Times, 2011a). . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
4.5 A Sample of Order-Detail Data . . . . . . . . . . . . . . . . . . . 78
4.6 A Sample of Order-History Data (Part I) . . . . . . . . . . . . . . 80
4.7 A Sample of Order-History Data (Part II) . . . . . . . . . . . . . 81
4.8 A Sample of Trade-Report Data . . . . . . . . . . . . . . . . . . 82
4.9 Order Action Type (ROB) . . . . . . . . . . . . . . . . . . . . . 83
xii
4.10 Parameters for Removing Outliers . . . . . . . . . . . . . . . . . 92
4.11 Cleaning Quote Data. . . . . . . . . . . . . . . . . . . . . . . . . 93
4.12 Cleaning Trade Data. . . . . . . . . . . . . . . . . . . . . . . . . 94
5.1 Companies Listed in the US markets. Market capitalisation and

share prices as at 01/04/2011 . . . . . . . . . . . . . . . . . . . . 109
5.2 Trading Activity of US Stocks (02/02/2009). . . . . . . . . . . . . 109
5.3 Information of Companies Listed in the LSE (28/09/2012) . . . . 109
5.4 Descriptive Statistics for US stocks. . . . . . . . . . . . . . . . . 112
5.5 Descriptive Statistics for UK stocks. . . . . . . . . . . . . . . . . 115
6.1 Descriptive Statistics of ROB Data at Sep. 6, 2010 for Barclays

Capital . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
6.2 Parameters of Artificial Limit Order Market Simulation . . . . . . 161
6.3 Events and Limit Order Types in Artificial Market . . . . . . . . . 162
6.4 Summary Statistics of Order Sizes and Relative Limit Prices . . . 164
7.1 VWAP Calculation of A Sample Buy Strategy . . . . . . . . . . . 176

7.2 Definitions of Market Variables . . . . . . . . . . . . . . . . . . . 179
7.3 Initial Parameters for Order Book based ASM I . . . . . . . . . . 185
7.4 Initial Parameters for Artificial Limit Order Market II . . . . . . . 188
7.5 Parameters for Genetic Algorithm in Experiment One . . . . . . 189
7.6 Results of Experiment One. . . . . . . . . . . . . . . . . . . . . . 190
7.7 Parameters for Grammatical Evolution in Experiment Two . . . . 191
7.8 Results of Experiment Two (Buy Orders) . . . . . . . . . . . . . 193
7.9 Results of Experiment Two (Sell Orders) . . . . . . . . . . . . . . 194
7.10 Parameters for Grammatical Evolution in Experiment Three . . . 195
7.11 Results of Experiment Three . . . . . . . . . . . . . . . . . . . . 196
xiii
Publications
Parts of this thesis have featured in various publications. These include:
Journal Article:
• Cui, W., Brabazon, A. and O’Neill, M. (2011). Adaptive Trade Execution:

A Grammatical Evolution Approach. International Journal of Financial
Markets and Derivatives. Vol. 2, No. 1/2, pp. 4-31. (Cui et al., 2011)
Book Chapter:
• Cui, W., Brabazon, A. and O’Neill, M. (2010). Evolutionary Computation

and Trade Execution. Natural Computing in Computational Finance (vol-
ume III), pp. 51-69, Springer-Verlag, Berlin. (Cui et al., 2010b)
Conference Papers:
• Cui, W. and Brabazon, A. (2012). An Analysis of Liquidity in London

Stock Exchange. Proceedings of the 25th Irish Accounting and Finance As-
sociation Conference, pp. 1-45, Galway, Ireland. (Cui & Brabazon, 2012b)
• Cui, W. and Brabazon, A. (2012). An Agent-based Modelling Approach to

Study Price Impact. Proceedings of the 2012 IEEE Conference on Compu-
tational Intelligence for Financial Engineering & Economics, pp. 1-8, New
York City, USA. (Cui & Brabazon, 2012a)
• Cui, W. and Brabazon, A. (2011). Ultra High-frequency Financial Data:

Analysing NYSE Euronext TAQ Data. Proceedings of the 24th Irish Ac-
counting and Finance Association Conference, pp. 1-38, Cork, Ireland.
(Cui & Brabazon, 2011)
xiv
• Cui, W., Brabazon, A. and O’Neill, M. (2010). Evolving Efficient Limit Or-
der Strategy Using Grammatical Evolution. Proceedings of the 2010 IEEE
Congress on Evolutionary Computation, pp. 1-6, Barcelona, Spain. (Cui
et al., 2010d)
• Cui, W., Brabazon, A. and O’Neill, M. (2010). Evolving Dynamic Trade

Execution Strategies Using Grammatical Evolution. Proceedings of the
2010 International Conference on Applications of Evolutionary Computa-
tion, Lecture Notes in Computer Science 6025, pp. 192-201. (nominated as
best paper) (Cui et al., 2010c)
• Cui, W., Brabazon, A. and O’Neill, M. (2009). Efficient Trade Execution

Using a Genetic Algorithm in an Order Book Based Artificial Stock Market.
Proceedings of the 11th Annual Conference Companion on Genetic and
Evolutionary Computation Conference, pp. 2023-2028. (Cui et al., 2009a)

in Trade Execution. Proceedings of the 22nd Irish Accounting and Finance
Association Conference, pp. 1-10, Dublin, Ireland. (Cui et al., 2009b)
Conference Abstract:
• Cui, W., Brabazon, A. and O’Neill, M. (2010). An Analysis of Price Im-

pact. Proceedings of the 3rd International Conference of European Re-
search Consortium on Informatics and Mathematics, London, UK. (Cui
et al., 2010a)
xv
Chapter 1
Introduction
This thesis is concerned with the effect of trading on the price of a financial se-
curity, that is price impact. Understanding price impact is a fundamental task in
finance, as price impact is one of the basic mechanisms determining price forma-
tion. In one of the seminal papers on price impact, Black (1971) stated that large
order execution would always exert an impact on price, irrespective of the method
of execution or technological advances in market structure. Due to the availabil-
ity of transaction-level financial data, an increasing number of studies have been
devoted to the analysis of price impact in the last two decades (Bouchaud et al.,
2009).
Recent years have seen the emergence of an important technological prod-
uct of electronic trading, namely Algorithmic Trading (AT), which has received
considerable public attention and is commonly defined as the use of computer
programs to automate the trading process which encompasses making trading de-
cisions, submitting orders and managing orders after their submissions (Hender-
shott et al., 2011). AT now makes up a large portion of the market activity in
multiple financial markets ranging from equities to foreign exchange (Chaboud
et al., 2009), to derivative (Mishra et al., 2012) markets. It accounted for more
1
than 70% of equity trades in the U.S. by 2009 (Hendershott et al., 2011). Signif-
icant volumes traded using AT are also found in Asian markets (Decovny, 2008;
Grant, 2012) and European markets (Hendershott & Riordan, 2012).
The widespread use of AT has stressed the importance of studying price im-
pact as understanding price impact is crucial for the design of algorithmic trading
strategies. Moreover, the significance of the study of price impact is also illus-
trated by the “Flash Crash” of May 6, 2010. An important contribution in trig-
gering this event was the extremely rapid execution of a large order of futures
contracts. From this point of view, understanding price impact is helpful for fi-
nancial regulators to monitor markets.
The research objective of this thesis is to empirically investigate price impact,
aiming to gain a better understanding of the intraday behaviours of price impact
and the factors affecting price impact. This thesis contributes to the literature on
market microstructure.
The rest of this chapter is organised as follows. The next section offers some
background information relevant to this thesis, including a description of market
mechanism and a brief introduction to trade execution. The research aims and
objectives of this thesis are then outlined. Following from this is a short overview
of the research framework, along with the contribution of the thesis and its scope
limitations. An overview of the structure of the remainder of this thesis then
completes the chapter.
1.1 Limit Order Market and Trade Execution

Traditional exchanges are organised as dealer markets, where designated market
makers set the quotes and directly trade with other market participants. With
the developments of information technology over the last few decades, electronic
2
trading systems have gradually replaced the role of designated market makers.
Many securities exchanges have set up new limit order markets, or converted tra-
ditional dealer markets to hybrid ones, which allow trading via a blend of an elec-
tronic trading platform and a traditional floor broker system.
• Examples of pure limit order markets include Tokyo Stock Exchange, Hong
Kong Stock Exchange, Paris Bourse, Toronto Stock Exchange, Australian
Stock Exchange, Shanghai Stock Exchange, and Singapore Exchange.
• Examples of newly set-up limit order markets in traditional exchanges in-

clude the Stock Exchange Electronic Trading Service (SETS) system in the
London Stock Exchange (LSE), the Globex trading system in the Chicago
Mercantile Exchange (CME), and the electronic option trading platform in
the International Securities Exchange.
• Examples of hybrid markets include the New York Stock Exchange (NYSE)
and NASDAQ.
The electronic revolution has also brought many off-exchange trading venues,
including electronic communication networks (ECNs) and dark pools. Examples
of ECNs include ArcaEdge, BATS Chi-X, and Turquoise. In contrast to ECNs and
limit order markets where order and quote information are visible, participants in
dark pools cannot see the orders placed by the other participants. Examples of
dark pools include Instinet, Liquidnet, and POSIT.
This thesis focusses attention on limit order markets which are the prevalent
trading mechanism in the world’s major stock exchanges. A brief introduction to
the limit order market mechanism is provided in the next section followed by a
description of the trade execution problem.
3
1.1.1 Limit Order Book
Today more than half of the stock exchanges around the world are limit order
markets which operate a limit order book to match buyers and sellers (Jain, 2003).
One advantage of an electronic limit-order-book is the greater transparency of-
fered by these systems when compared with dealer market settings. Price quotes
and transactions are visible to all participants which generally improves the effi-
ciency of price discovery, thus promoting market confidence (Gould et al., 2011;
Parlour & Seppi, 2008). It also promotes competition as dealers/market makers
are encouraged to post the best prices to attract order flow (CFA, 2009).
In a limit order market, traders can either submit a limit order or a market
order. A market order is an order to buy or to sell a specified number of shares. It
guarantees immediate execution but provides no control over its execution price.
In contrast, a limit order is an order to buy or to sell a specified number of shares at
a specified price. It provides control over its execution price but does not guarantee
its execution.
Table 1.1: Order Book 1 Table 1.2: Order Book 2
Bid Ask Bid Ask
Shares Prices Prices Shares Shares Prices Prices Shares
300 50.19 50.22 200 300 50.19 50.22 200
200 50.18 50.23 300 500 50.18 50.23 300
400 50.17 50.24 100 400 50.17 50.24 100
500 50.16 50.25 300 500 50.16 50.25 300
300 50.15 50.26 200 300 50.15 50.26 200
100 50.14 50.27 400 100 50.14 50.27 400
Table 1.1 shows a sample order book, where all the buy and sell orders are
visible/transparent to traders in the market. It consists of two queues which store
buy and sell limit orders, respectively. Buy limit orders are called bids, and sell
limit orders are called offers or asks. The highest bid price on the order book is
called best bid, and the lowest ask price on the order book is called best ask. The
4
Table 1.3: Order Book 3 Table 1.4: Order Book 4
Bid Ask Bid Ask
Shares Prices Prices Shares Shares Prices Prices Shares
300 50.19 50.22 100 300 50.19 50.23 100
500 50.18 50.23 300 500 50.18 50.24 100
400 50.17 50.24 100 400 50.17 50.25 300
500 50.16 50.25 300 500 50.16 50.26 200
300 50.15 50.26 200 300 50.15 50.27 400
100 50.14 50.27 400 100 50.14 50.28 300
difference between best bid and best ask is called bid-ask spread. Prices on the
order book are not continuous, but rather change in discrete quanta called ticks.
Orders in a limit order market arrive stochastically in time. The price limit of
a newly arrived order is compared to those of orders already held in the system to
ascertain whether there is a match. If so, the trade occurs at the price set by the
first order. The set of unexecuted limit orders held by the system constitutes the
dynamic order book, where limit orders can be cancelled or modified at any time
prior to their execution. Limit orders on the order book are typically (depending
on market rules) executed strictly according to (1) price priority and (2) time pri-
ority. Bid (ask) orders with higher (lower) prices get executed first with time of
placement being used to break ties. A buy (sell) market order is executed at the
best ask (bid) price. The limit order book is highly dynamic, because new limit
orders will be added into the order book, and current limit orders will get executed
or cancelled from the order book throughout the trading day. Table 1.2 shows the
order book after a trader submits a buy limit order with 300 shares placed at price
50.18. Table 1.3 shows the order book after a trader submits a buy market or-
der with 100 shares. Table 1.4 shows the order book after a trader submits a buy
market order with 300 shares.
Apart from market and limit orders, some stock exchanges also offer hid-
den/iceberg orders to allow traders to conceal the total size of a large limit order.
5
Such orders consist of two components, a small component whose size is visi-
ble in the order book and a larger hidden component with a size known only to
the order submitter. The hidden component is exposed to the market gradually
through execution of the visible part of the order (Aitken et al., 2001; Biais et al.,
1995). Many electronic trading platforms have introduced this kind of order, in-
cluding Euronext, the Toronto Stock Exchange, the London Stock Exchange, and
XETRA. Hidden limit orders are often used by large liquidity traders to hide their
intent to trade (Bongiovanni et al., 2006). However, iceberg orders typically ex-
hibit a less favorable time priority compared with pure limit orders (Bessembinder
et al., 2009; Esser & Monch, 2007). After the visible portion of an iceberg order
is completely matched, other visible limit orders at the same limit price that were
entered before the new portion is displayed take priority.
1.1.2 Trade Execution
“The investment process has been described as a three-legged

stool supported equally by securities research, portfolio management,
and securities trading. Of the three, trading is often the least under-
stood and the least appreciated function.”
A. Madhavan, J.L. Treynor and W.H. Wagner (2007)
Trade execution is the implementation of trading decisions. The costs that

arise during the execution process are called trading costs. Understanding trading
costs is important for the design of efficient and effective trade execution strat-
egy, as well as for the assessment of performance of traders, brokers and trading
algorithms.
Trading costs involve commissions paid to brokers, fees paid to exchanges,
spread, market impact, and opportunity cost. The first three items are visible,
6
which can be known before any trade is made. The last two items are not directly
measurable a priori. These five costs are explained as follows.
1. Commission is the payment made to brokers, which varies from broker to

broker. Commissions have been greatly reduced over recent years. Goldstein
et al. (2009) analysing institutional commissions using the data provided by
Greenwich Associates find that the average commission per share in 2004 is
under 5 cents, about one third of that in 1977.
2. Fees are charged by exchanges during the execution process. These are often
bundled into the total commission paid to brokers.
3. Bid-ask spread is the difference between the best available bid and ask price
in markets.
4. Market impact, also known as price impact, is the effect of a trade on the price
of the asset. The most obvious illustration of price impact is that a buying
event always drives the price up and a selling action pushes the price down.
For example, the mid-quote (the average of the best bid and the best ask)
of a stock drops from $200 to $199.6 (supposed) after a market sell order is
executed, where the price decline is the effect caused by trading the sell order.
5. Opportunity cost arises when an order fails to be fully executed. For example,
a trading task of selling 120,000 shares at $200 per share is supposed to be
finished by the end of the trading day, but only 100,000 shares are sold by the
end of the day. At that time, the price drops to $198. The opportunity cost
of failing to execute this entire order is $40,000 which is calculated as 20,000
multiplied by the $2 price decline per share.
Many empirical studies (Chan & Lakonishok, 1995, 1997; Collins, 1991; Kraus
& Stoll, 1972) indicate that the visible components of trading costs are much
7
smaller than the invisible components and price impact always dominates total
trading cost. As markets have become electronic, commissions and fees have
been driven down. In the optimal trading literature (Alfonsi et al., 2010; Almgren
& Chriss, 1999, 2000; Bertsimas & Lo, 1998; Hora, 2006; Obizhaeva & Wang,
2012), price impact cost is equated with trading cost.
A practical problem arising in limit order markets is how to buy or sell large
quantities of an asset with minimum execution costs. Executing a large order at
once will cause significant price impact cost. The most obvious way to reduce this
cost is to break the large order into a number of smaller pieces and spread them
over time. A trade execution strategy is devised to minimise the price impact cost.
1.2 Research Aims

In this thesis, three broad questions are addressed in order to gain a better under-
standing of price impact:
1. Does price impact on the LSE and the NYSE exhibit an intraday pattern? The
literature on market microstructure has discovered many intraday patterns of
a number of interesting variables in financial markets. For example, bid-ask
spread on the NYSE exhibits a U-shaped intraday pattern (Lee et al., 1993):
the spread is relatively higher at market open and market close than during
the middle of the trading day. However, only a few studies pay attention to
the intraday behaviour of price impact. Chan (2000) finds that price impact
displays a U-shaped pattern over the trading day on the Hong Kong Stock
Exchange. This thesis aims to examines the intraday behaviour of price impact
in US and UK markets, using the data drawn from the NYSE-Euronext TAQ
database and the LSE ROB database.
2. Does agent intelligence affect the magnitude of price impact? Previous studies
8
(e.g. Farmer et al. (2004); Hasbrouck (1991); Lillo et al. (2003)) concerning
price impact find that price impact is influenced by trade size and the state
of the limit order book, and is a concave function of trade size. Some of
these studies (Farmer et al., 2004; Hopman, 2007; Weber & Rosenow, 2006)
argue that this concavity is due to agent’s selective liquidity taking, namely
agents conditioning their trade size on market liquidity. The second aim of
this thesis is to investigate whether agent intelligence plays an important role
in determining price impact.
3. Does order choice affect price impact when trading large orders? Controlling
price impact when trading a large order is an important issue in financial mar-
kets. Many studies have been devoted to this problem. However, only market
orders are examined in these studies. The third aim of this thesis is to examine
the influence of order choice on the price impact of trading a large order.
These questions have practical significance. Without an understanding of the in-

fluences of these three factors on price impact, our understanding of price impact
is incomplete. Answering these questions will help us gain a better understanding
of price formation, thus will help us to better understand how securities markets
function, and help practitioners to devise better trading strategies leading to re-
duced execution costs and therefore improved investment returns.
In the first experiment, an intraday pattern on price impact is expected to be
observed in US and UK markets. In the second experiment, the price impact gen-
erated from the artificial market is expected to be higher than the price impact
produced from the real market, which will support the hypothesis that agent in-
telligence is an important determinant of the magnitude of price impact. In the
third experiment, it is expected that the trade execution strategies adopting mul-
tiple order types outperform those using single order type, which indicates that
9
order choice between market order and limit order plays an important role in de-
termining price impact when trading a large order.
1.2.1 Objectives of Research
The following objectives need to be achieved in order to fulfill the research aims
of this thesis.
1. Review relevant literature on market microstructure and price impact.
2. Develop an algorithm for preprocessing TAQ data.
3. Develop an algorithm for rebuilding the limit order book using the LSE ROB
data.
4. Develop an agent-based model of a limit order market.
5. Design trade execution strategies.
6. Analyse price impact.
7. Draw conclusions from the results.
1.3 Framework of Research

To address the first research aim, an empirical analysis of intraday market liq-
uidity and price impact is undertaken using ultra high-frequency datasets from
the London Stock Exchange (LSE) and the NYSE-Euronext. The datasets record
trade and quote information on each trading day. This enables us to identify the
price before and after each single trade, and thus measure the price impact of each
trade. Before its use for analysis, the data has to be preprocessed.
10
To address research aims two and three, an agent-based modelling (ABM)
approach is adopted. This is a computerised simulation consisting of a number
of agents. The emergent properties of an agent-based model are the result of
“bottom-up” processes, where the decisions of individual and interacting agents at
a microscopic level determine the macroscopic behavior of the system (LeBaron,
2006; Samanidou et al., 2007; Tesfatsion, 2006). Unlike classical micro-economic
models, the concept of equilibrium has little meaning in ABM-simulated systems,
as the interactions of the agents result in continual movement between system
states. The collective emergent property of the complex system makes the whole
become more than the sum of its parts.
Agent-based modelling approach has been applied in the fields of economics
and finance for many years. One of the earliest agent-based models in concepts
was developed by the Nobel laureate Thomas Schelling (1971). It is now widely
used in economic and finance as an alternative research methodology and offers
a number of advantages. The first is that most agent-based stock markets exhibit
some stylised facts, like volatility clustering and fat tails of price returns, which
can not be explained by traditional theoretical models (Chen et al., 2012). Another
advantage is that it is very easy to analyse one specific factor by isolating it in the
simulation. On the contrary, it is often difficult to isolate the effects resulting from
a specific factor from others with econometric approaches.
In order to investigate whether agent intelligence and order choice decision
affect price impact, agent-based modelling is used to simulate artificial markets
which isolate these two factors. Chapter 6 simulates an artificial market consist-
ing of intelligence-isolated agents, namely zero-intelligence agents, in order to
investigate the price impact generated by zero-intelligence agents. In Chapter 7,
artificial markets considering agent’s order choice decisions are simulated, as well
as artificial markets isolating agent’s order choice decisions, in order to evaluate
11
the effect of order choice on the price impact when trading a large order. In this
thesis, the artificial market settings possess the salient features of electronic limit
order markets: continuous trading, a visible book of orders, price-time priority
rules, instantaneous trade reporting rules, order cancellation capabilities, and both
limit order and market order functionality.
1.4 Contributions of Thesis

The work presented in this thesis has given rise to a number of contributions as
described below.
• Up-to-date review of relevant literature
This thesis presents an up-to-date survey of relevant literature on market

microstructure, price impact, and agent-based modellings of financial mar-
kets. The review in Chapter 2 covers up-to-date studies on the relation-
ship between price impact and trade size, studies on large price impact are
covered, studies of intraday phenomena on price volatility, bid-ask spread,
trading volume and market depth, studies on optimal trading strategies for
a single asset, and studies on order submission strategies. The literature
review of agent-based financial markets in Chapter 3 covers various appli-
cations of agent-based modelling to financial markets.
• Analysis of intraday behaviours of price impact in UK and US markets
This thesis analyses intraday behaviours of price impact, using the recent
data drawn from the NYSE-Euronext TAQ database and the LSE ROB
database. While many studies analyse intraday behaviours of price volatil-
ity, bid-ask spread, trading volume and market depth, little attention has
been paid to price impact in UK and US markets. This thesis aims to shed
12
light on this, and for the first time documents intraday patterns of price im-
pact in UK and US markets.
• Up-to-date evidence on intraday behaviours of market liquidity in UK

and US markets
This thesis analyses intraday behaviours of market liquidity, using the data
drawn from the NYSE-Euronext TAQ database and the LSE ROB database.
Unlike previous studies on intraday phenomena which only examined the
market depth at the best level of the limit order book, this thesis for the first
time analyses the intraday behaviour of market depth for up to ten levels
of the limit order book in the UK market. Moreover, since the previous
studies on intraday phenomena are dated, this thesis aims to provide up-
to-date evidence on intraday behaviours of price volatility, bid-ask spread,
market depth, trading volume, trading frequency and trade size in UK and
US markets.
• Investigation of the role of agent intelligence in determining the magni-

tude of price impact
Several studies argue that the concavity of price impact is due to selective
liquidity taking. However, no studies have been conducted to investigate
this. This thesis for the first time investigates whether agent intelligence
plays an important role in determining the magnitude of price impact, and
provides evidence supporting this hypothesis.
• Examination of the effect of order choice on the price impact when

trading a large order
Many studies examine how to control price impact when trading a large or-
der. However, these studies only consider market orders. This thesis for the
13
first time analyses whether order choice affects the price impact of trading
a large order, and demonstrates that it does.
• Use of ABM for investigating price impact
Various applications of agent-based modelling to economics and finance can

be found in prior literature, but there has been little attempt to use ABM for
investigating price impact. This thesis makes the first attempt to investigate
price impact using ABM in its two experimental studies. In Chapter 6,
agent-based modelling is used to simulate an artificial limit order market
consisting of zero-intelligence agents, in order to examine the price impact
produced by zero-intelligence agents. In Chapter 7, agent-based artificial
markets are also simulated in order to analyse the price impact of trading a
large order.
1.5 Scope Limitations

This thesis makes a useful contribution to our understanding of price impact.
However, it does not conduct an exhaustive analysis of all possible avenues. There
are a number of areas worthy of further research that fall outside the scope of this
work. The following parts of this section list some of these areas.
When examining the intraday behaviours of price impact, the focus of the
present work is on the price impact caused by market orders. The analysis can
be extended to price impact of other types of orders. For example, Eisler et al.
(2012); Hautsch & Huang (2012) examine the price impact caused by limit orders,
and Gabaix et al. (2006); Keim & Madhavan (1996) investigate the price impact
of upstairs orders.
In this thesis, an agent-based modelling approach is used to model price im-
pact, that is the immediate effect of trading on prices. This could be extended to
14
consider the resilience of price impact. For example, Gatheral et al. (2011) models
that the price impact decays exponentially over time.
While studying the issue of controlling price impact, attention is limited to the
trading of a large order occurring in one trading venue. This could be extended to
the case where a large order is being traded across multiple trading venues. For
example, Foucault & Menkveld (2008) empirically examine the smart order rout-
ing systems, and Rawal (2010) discusses the importance of intelligent decision-
making in order routings.
1.6 Structure of Thesis

The remainder of this thesis is structured as follows:
• Chapter 2 - Review of Relevant Literature
A review of the literature on intraday patterns, price impact, trade execution

strategies is provided. Some research gaps in the current state of the art are
identified at the end of the chapter.
• Chapter 3 - Agent-based Modelling in Finance
A brief introduction to agent-based modelling and its applications to finance

are given in this chapter, along with an introduction to the Genetic Algo-
rithm, a technique often adopted for simulating agent learning.
• Chapter 4 - Datasets and Data Processing
A description of the datasets used to analyse price impact, as well as the

data preprocessing steps required is provided.
• Chapter 5 - Intraday Behaviour of Price Impact
15
Ultra high-frequency data from the LSE and the NYSE-Euronext is used to
analyse the intraday behaviour of price impact, as well as intraday features
of market activity and market liquidity.
• Chapter 6 - An Investigation of Price Impact and Agent Intelligence
An agent-based modelling approach is adopted to simulate the market for a

particular stock in the LSE, where every aspect of the order flow in the LSE
market is imitated in the artificial market except the trader’s intelligence.
Then the price impact reproduced from the artificial market is compared
with that in real market, in order to understand whether agent intelligence
plays an important role in determining the magnitude of price impact.
• Chapter 7 - Dynamic Trade Execution with Limit and Market Orders
A number of trade execution strategies which adopt different types of orders

are devised as well as some benchmark strategies which use single order
type. These strategies are evaluated by their trading impact on the market,
in order to examine the effect of order choice on price impact when trading
a large order.
• Chapter 8 - Conclusions
A summary of the thesis is provided in this chapter, along with an overview

of its contributions and limitations. Opportunities for future work are also
outlined.
16
Part I
Literature Review
17
In Part I the relevant literature is reviewed and the main methodology of re-
search is described.
In Chapter 2, the theoretical background of this thesis, that is market mi-
crostructure, is introduced, and a brief overview is provided. Previous work on
price impacts, intraday patterns and trade execution is reviewed, and gaps of re-
search are revealed.
In Chapter 3, agent-based modelling approach is introduced, and framework
of modelling is described. Prior work on artificial markets modelling is reviewed.
18
Chapter 2
Market Microstructure
This chapter provides a literature review on relevant work, including an overview

of market microstructure, a review of literature on intraday phenomena, a review
of studies concerning price impact, and a review of literature on trade execution
strategies. At the end of the literature review, a few research gaps in the current
state of the art are identified which lead to the three experimental studies of this
thesis.
2.1 Introduction
The theoretical foundation for the studies in this thesis is found primarily in the
field of market microstructure. Market microstructure is formally defined by
O’Hara (1995) as “the study of the process and outcomes of exchanging assets
under explicit trading rules”.
Studies on market microstructure are helpful in understanding the returns to
financial assets and how financial markets become efficient. Moreover, it has
immediate application in understanding the design of trading strategies and the
regulation of financial markets.
19
The domain of market microstructure is wide and broad. An overview of this
area is provided in section 2.2. This thesis is related to several streams of market
microstructure research. The following topics will form the focus of this chapter.
• Price impact.
• Intraday pattern.
• Trade execution strategies.
2.2 Overview of Market Microstructure

The study of market microstructure involves every aspect of the trading process.
There has been a vast number of studies on market microstructure since its emer-
gence. Several papers and books have provided a valuable survey of the literature
in this area. O’Hara (1995) gives a comprehensive review of theoretical work
on market microstructure. Madhavan (2000) provides an overview of market mi-
crostructure literature from the viewpoint of informational economics. Also, he
distinctively summarises the issues examined in market microstructure and clas-
sifies them into the four following categories.
1. Price formation. It is also known as price discovery, is the process by which

prices come to impound new information. This theme concerns every aspect
of the trading process.
2. Market structure and design. This theme focuses on how different market
structures affect the speed and quality of price discovery, liquidity, and the
cost of trading, and concerns how to design a better market. The design of
a market determines its market microstructure, and thus affects the quality of
the market.
20
3. Market transparency. It is defined as the ability of market participants to
observe information about the trading process (O’Hara, 1995). This theme
focuses on how various levels of market transparency influence the process of
price formation.
4. Applications to other areas in finance. This theme concerns the interface of

market microstructure with other areas of finance including corporate finance,
asset pricing and international finance.
2.2.1 Market Liquidity
Market liquidity is an important quality characteristic of a financial market and is

also a significant research topic in the area of market microstructure. Understand-
ing liquidity is beneficial for both market participants and the exchange. O’Hara
(1995) interprets liquidity as a measure of the cost of waiting and a function of the
scale of trading. High liquidity means the ability to trade securities quickly and at
desirable prices. Market liquidity is a crucial consideration when market partic-
ipants make their investment decisions and trading decisions. Moreover, a more
liquid exchange can attract more participants which has a network externality ef-
fect. The more participants the exchange has, the more valuable the exchange
is.
Market liquidity is a multi-dimensional variable. The following four dimen-
sions of liquidity are distinguished by the literature (Black, 1971; Kyle, 1985).
1. Market tightness is the ability to execute a buy order and a sell order of an asset
at the same price at the same time. The proportional quoted spread (Copeland
& Galai, 1983) is often used as an approximation of liquidity costs, measuring
hypothetical trading cost.
21
2. Market depth is the ability to execute a sale / purchase of a certain amount of
an asset without influence on the price.
3. Market immediacy is the ability to execute an order immediately at the pre-

vailing price.
4. Market resiliency measures the speed at which liquidity recovers from shocks.
Market liquidity is affected by market structure. Market makers are the only
suppliers of liquidity in dealer markets. However, the role of market makers
has largely been superceded, as most market have adopted electronic limit or-
der books. In a limit order market, liquidity is supplied by market participants
themselves, and is illustratd by the state of the limit order book. The limit order
book evolves over time as matched limit orders are deleted from the order book
and as new incoming limit orders are added into the order book.
2.2.2 Theoretical Market Microstructure Models
One of the crucial aims of market microstructure is to understand how price is

formed. Theoretical market microstructure models of price formation fall into the
three types of models. This section briefly introduce the three types of models.
1. Inventory based model. These models try to explain security prices from in-
ventory imbalances of market makers. The market maker maintains the bid-
ask spread in order to compensate the risk of holding inventory.
2. Information based model. The second class of models is based on the idea of
asymmetric information. Traders can either be informed or uninformed (Bage-
hot, 1971). If they posses private information on the value of the asset, they
are called informed traders. If they trade in order to rebalance their positions,
22
they are called liquidity or uninformed traders. A consequence of asymmetric
information is that trading itself conveys information.
3. Limit order market model. The third class of models tries to explain the price
formation process on the limit-order-book market. In a limit order market, the
difference between informed traders and uninformed traders is not obvious.
Moreover, unlike the dealer market, the limit order market has no specific
market maker who maintains market liquidity. The roles of supplying and
demanding liquidity are played by the market participants by using either limit
orders or market orders. This category of models has a particular focus on
analysing how order submission strategies and other aspects of trading affect
the price of an asset in limit order markets.
Financial markets are constantly evolving and becoming more complex. From
2005 to 2009, the average trade size on the NYSE has decreased from 724 shares
to 268 shares and the average daily trades have increased from 2.9 million to
22.1 million (Durbin, 2010). Thus, one challenge faced by researchers in market
microstructure is the dramatically growing amount of data.
2.3 Price Impact

The effect of trading on the price of the asset is called price/market impact, which
is considered a dynamic property of liquidity. A sell trade is always associated
with a following fall in market price and a buy trade is always followed by a
rise in market price. Black (1971) states that large order execution would always
exert an impact on price, irrespective of the method of execution or technological
advances in market structure.
The earliest research empirically examining the effects of trading large orders
on a market dates back to Kraus & Stoll (1972), who found that the execution of
23
large orders on the NYSE had a temporary price impact and a permanent price
impact. This fact is later confirmed by Biais et al. (1995), Coppejans et al. (2003),
Evans & Lyons (2002), Holthausen et al. (1987), and Holthausen et al. (1990).
They argue that the temporary impact is related to the trade size and market liq-
uidity, and the permanent impact is associated with information. Recently, a few
studies have different views on price impact. Bouchaud et al. (2004) present the
view that price impact is fixed and temporary. Farmer et al. (2006) argue that price
impact is variable and permanent.
The explanation for price impact in earlier studies asserts that trade affects
price because it contains information, and this effect increases with trade size.
French & Roll (1986) show that price movements are due to private information
conveyed through trading. Hasbrouck (1991) proposes a method to measure the
information content in a trade.
2.3.1 Price Impact and Trade Size
The relationship between price impact and trade volume has been extensively
studied in the financial market microstructure literature. The conclusion of previ-
ous studies is that the price impact of a market order is a concave function of order
size, and this has been validated using transaction-level data in Lillo et al. (2003);
Farmer & Lillo (2004); Farmer et al. (2005b); Lim & Coggins (2005a); Hopman
(2007).
Price impact is typically measured as the difference of mid-quote price before
the order arrives and the mid-quote price after it has been executed. The functional
form is
p = γ ∗ vµ
24
Table 2.1: Price Impact Function of Trade Size
Authors Functions Parameters Market Years

nonlinear, positive,
Hasbrouck (1991) and increasing, NYSE 1989
but concave
Hausman et al. (1992) strongly concave NYSE 1988
Keim & Madhavan (1996) concave upstairs market 1985-1992
Torre (1997) kxβ β = 0.5 NYSE 1994
25
Dufour & Engle (2000) nonlinear NYSE (TORQ) 1-Nov-1990 to 31-Jan-1991
Lillo et al. (2003) sign(x)|x|β /C λ β = 0.2 − 0.5; λ = 0.4 NYSE 1995-1998
Potters & Bouchaud (2003) ℜ(τ )ln(x) Paris Bourse; NASDAQ 2002
Gabaix et al. (2003) kxβ β = 0.5 NYSE 1994-1995
Farmer & Lillo (2004) kxβ β = 0.26 LSE May 2000 to Dec. 2002
β λ
Lim & Coggins (2005a) kx /C ASX 2001-2004
Farmer et al. (2005b) kxβ β = 0.25 LSE (SETS) 1-Aug-1998 to 30-Apr-2000
Hopman (2007) kxβ + ϵ β = 0.37 Paris Bourse 4-Jan-1995 to 19-Sep-1999
Bouchaud et al. (2009) kxβ β = 0.3 LSE 2002
where p is the price impact of a market order with order size v, γ and µ are
constants of the function. The specific constants of the function vary in different
markets and at different time periods. Although previous literature has agreement
on the concavity functional form of price impact, it is inconclusive in the specific
functional form. A comparison of the empirical studies on immediate price impact
is shown in Table 2.1.
Previous empirical studies show that large trades have less marginal price im-
pact than small trades. Three explanations for the concavity result from previous
literature are summarised in Bouchaud et al. (2009).
The first one is from the information context as postulated by Barclay &
Warner (1993). The price impact reflects the informativeness of trades, which
increase with their private information content. The concavity is due to informed
traders’ “stealthy trading”, keeping their orders small in order to avoid revealing
their superior knowledge. Hasbrouck (1991) propose a VAR approach to measure
the information content in each trade.
The second explanation claims that the concavity is due to the shape of the
limit order book. The depth in the order book as a function of the price will
determine the market impact for a market order as a function of its size (Daniels
et al., 2003). Bouchaud et al. (2002) have observed that the average order book
on the Paris Bourse has a maximum away from the best bid/ask. Smaller market
orders are faced with less liquidity than larger ones in such limit order markets.
This explanation is ruled out by Weber & Rosenow (2006) and Farmer & Zamani
(2007). In Bouchaud et al. (2004), the concave shape of the impact as a function
of the volume is understood as an order book effect, where the average size of the
queue increases with depth.
The third explanation is called selective liquidity taking (Farmer et al., 2004).
Traders with large orders to trade are more patient while smaller order traders are
26
less patient. Large order traders may wait for periods of higher liquidity (Weber
& Rosenow, 2006; Hopman, 2007).
2.3.2 Large Price Fluctuations and the Order Book
Market liquidity plays a significant role in the price formation process. The role of
real information appears to be rather thin when there are large price fluctuations.
Price impact is affected by the market liquidity. The price responses to trades of
the same size is highly variable. Many studies show that the random walk nature
of price changes on short timescales is not due to the unpredictable nature of in-
coming news, but appears as a dynamic consequence of the competition between
market participants (Bouchaud et al., 2004). Recent work suggest that liquid-
ity fluctuations play an essential role in price formation (Farmer & Lillo, 2004;
Farmer et al., 2004; Gillemot et al., 2005; Lillo et al., 2005; Weber & Rosenow,
2006). In price formation, liquidity fluctuations dominate over fluctuations in
transaction size. Joulin et al. (2008) provide direct evidence that large transac-
tion volumes are not responsible for large price jumps, and conjecture that most
price jumps are induced by order flow fluctuations close to the point of vanishing
liquidity.
Recently, several papers cast doubts on concavity, and provide evidence that
the concave shape cannot be a property of the true price impact a trader in an actual
market would observe. Weber & Rosenow (2005) found a strong anticorrelation
between price changes and order flow. Weber & Rosenow (2006) found little
evidence that price changes larger than five standard deviations can be explained
by an extreme order flow alone. They found that a low density of limit orders in the
order book, i.e. limited liquidity, was a necessary prerequisite for the occurrence
of extreme price fluctuations. Taking into account both order flow and liquidity,
large stock price changes can be explained quantitatively. There are two reasons to
27
explain why this concave shape (of volume imbalance) cannot be a property of the
true price impact a trader in an actual market would observe (Weber & Rosenow,
2005). Firstly, such a concave price impact would be an incentive to do large
trades in one step instead of breaking them up into many smaller ones as is done
in practice. Secondly, in the presence of bluffers in the market, the enforcement of
a strictly linear price impact is the only way in which a market maker or liquidity
trader can protect herself against suffering losses.
The above review of literature on price impact has shown that price impact is
determined by trade size as well as the state of the limit order book when the trade
occurs. However, the two following questions concerning price impact are still
unanswered.
1. It is still unknown about the up-to-date intraday behaviour of price impact.

Chan (2000) showed that price impact displayed a U-shaped pattern over the
trading day on the HKSE. The review of literature on intraday phenomena
will show that many financial variables exhibit intraday patterns which differ
across markets and time periods.
2. No study has been conducted to investigate whether agent intelligence plays

an important role in determining the magnitude of price impact. Farmer et al.
(2004); Hopman (2007); Weber & Rosenow (2006) all argue that trader’s
choice of trading time affect price impact.
A better understanding of the two questions concerning price impact is helpful

for developing more comprehensive price impact models than those reviewed in
Section 2.3.1.
28
2.4 Intraday Patterns
Intraday behaviour of different market variables have been analysed extensively
in previous literature which dates back to Wood et al. (1985) and Harris (1986).
One of the interesting findings is that many of these variables demonstrate a U-
shaped pattern over the trading day: the levels of these variables are relatively
high at the beginning and at the end of the trading day, and relatively low in the
middle of the day. The empirical evidence on intraday behaviours of these market
variables helps us build theoretical models explaining the intraday behaviour of
the underlying market.
Different shapes of intraday patterns are defined as below.
• U-shaped pattern: values at market open and close are relatively higher than
the values in the middle of the trading day
• Inverted U-shaped pattern: values at market open and close are relatively
lower than the values in the middle of the trading day
• J-shaped pattern: values at market close are relatively higher than the values
at the rest of the trading day
• Reverse J-shaped pattern: values at market open are relatively higher than
the values at the rest of the trading day
• S-shaped pattern: values at market open are relatively lower than the values
in the middle of the trading day, and values at market close are relatively
higher than the values in the middle of the trading day.
• Reverse S-shaped pattern: values at market open are relatively higher than
the values in the middle of the trading day, and values at market close are
relatively lower than the values in the middle of the trading day.
29
(a) U-shaped (b) Inverted U-shaped
(c) J-shaped (d) Reverse J-shaped
(e) S-shaped (f) Reverse S-shaped
Figure 2.1: Different Shapes of Intraday Patterns
30
There have been a number of studies which have analysed intraday patterns
in financial markets. Admati & Pfleiderer (1988) developed the theoretical foun-
dations of intraday patterns found in financial markets. They argue that both in-
formed and liquidity traders prefer to trade when market is thick, that is when
their trades have little impact on price. Thus at periods when market is thick, both
informed and liquidity traders are willing to trade. Also the strategic interaction
between informed and liquidity traders intensifies the concentration of trading at
these periods of the trading day, leading to the observed intraday pattern of trading
volume.
Another explanation is provided by Brock & Kleidon (1992). They explain the
intraday pattern in a dealer market. They relate the intraday patterns to investors’
demand of transactions during the trading period. Due to the accumulation of
overnight information or due to the imminent non-trading period, investors have
a higher demand for transactions in order to achieve an optimal portfolio mix or
transfer overnight risk at market open and close than for the rest of the trading
day. This motivates the specialist’s interest in exploiting the liquidity inelasticity
at these periods, leading to wide spread and high trading volume at open and close.
More recent studies have attributed the intraday pattern to limit order traders.
Ahn & Cheung (1999) & Chung et al. (1999) show that limit order traders keep the
spread wide in order to offset the high risks of trading against informed traders.
2.4.1 Intraday Price Volatility
Numerous studies examine the intraday variation of bid-ask spread and document
a U-shaped intraday pattern (as summarised in Table 2.2). The price volatility is
relatively high at market open and close, and relatively low during the middle of
the day. The Hong Kong Stock Exchange exhibits a double U-shaped intraday
pattern on bid-ask spread as shown by Cheung et al. (1994). This is because that
31
Table 2.2: Intraday Pattern: Price Volatility
Exchange Intraday Pattern Period of Data Authors

Sep. 1971 - Feb. 1972
NYSE U-shaped Wood et al. (1985)
Jan. - Dec. 1982
NYSE U-shaped before 1986 Harris (1986)
NYSE U-shaped Jan. 1979 - Dec. 1983 Jain & Gun-Ho (1988)
NYSE U-shaped 1980,1981,1983,1984 McInish & Wood (1990b)
NYSE U-shaped Jan. 1964 - Feb. 1989 Lockwood & Linn (1990)
NYSE U-shaped Jan. - Jun. 1989 McInish & Wood (1992)
NYSE U-shaped Jan. - Dec. 1988 Foster & Viswanathan (1993)
NYSE U-shaped Jan. - Dec. 1991 Werner & Kleidon (1996)
NYSE U-shaped Jan. - Dec. 1990 Madhavan et al. (1997)
Oct. - Dec. 1991
NASDAQ U-shaped Chan et al. (1995)
Aug. - Sep. 1992
LSE U-shaped Jan. - Dec. 1991 Werner & Kleidon (1996)
LSE U-shaped Jan. - Mar. 1991 Abhyankar et al. (1997)
LSE U-shaped Mar. - May 2001 Cai et al. (2004)
HKSE double U-shaped before 1994 Cheung et al. (1994)
Jan. - Nov. 1991
TkySE U-shaped Lehmann & Modest (1994)
Feb. 1992 - Apr. 1993
FSE reverse J-shaped May 2004 - Sep. 2005 Hussain (2011)
Note: HKSE is short for Hong Kong Stock Exchange.

TkySE is short for Tokyo Stock Exchange.
FSE is short for Frankfurt Stock Exchange.
32
there was a lunchbreak session for two hours in the middle of the trading day at
that time period, dividing the trading day into a morning trading session and an
afternoon trading session, leading to two U-shaped patterns in the morning and in
the afternoon. Hussain (2011), however, observe a reverse J-shaped pattern while
examining the intraday price volatility in the Frankfurt Stock Exchange.
Admati & Pfleiderer (1988) provide an explanation for the U-shaped intraday
pattern on price volatility. Their model predicts that price volatility is related to
informed trading. Prices are more volatile in high-volume periods than in low-
volume periods because high-volume periods attract more informed traders than
low-volume periods which makes prices are more informative at high-volume pe-
riods.
2.4.2 Intraday Bid-ask Spread
A number of studies examine the intraday behaviour of bid-ask spreads in the

NYSE and report a U-shaped intraday pattern (Chung et al., 1999; Lee et al., 1993;
Madhavan et al., 1997). The spread is wider near the market open and close than
during the middle of the day. The U-shaped intraday pattern on bid-ask spread
is also observed in the Toronto Stock Exchange (Vo, 2007), the LSE (Abhyankar
et al., 1997), the HKSE (Ahn & Cheung, 1999) and the Tokyo Stock Exchange
(Lehmann & Modest, 1994).
Some studies have proposed explanations for the U-shaped intraday pattern on
bid-ask spread. Brock & Kleidon (1992) attribute the wide spread at market open
and close to the actions of the specialists who exploit the liquidity inelasticity at
these time periods. Chan et al. (1995) attribute the narrowing of the spread at
market close to the improved quotes from dealers who need to balance their in-
ventories. More recent research (Ahn & Cheung, 1999) argues that the U-shaped
pattern in spreads is driven by limit order traders who keep spreads wide at the
33
Table 2.3: Intraday Pattern: Bid-Ask Spread

NYSE U-shaped Jan. - Dec. 1988 Lee et al. (1993)
NYSE reverse J-shaped Jan. - Jun. 1989 McInish & Wood (1992)
NYSE U-shaped Jan. - Dec. 1990 Madhavan et al. (1997)
NYSE U-shaped Nov. 1990 - Jan. 1991 Chung et al. (1999)
NYSE reverse S-shaped Jan. - Dec. 1999 Henker & Wang (2006)
Oct. - Dec. 1991
NASDAQ reverse S-shaped Chan et al. (1995)
Aug. - Sep. 1992
TrtSE U-shaped Sep. 1999 - Nov 1999 Vo (2007)
LSE reverse S-shaped Jan. - Dec. 1991 Werner & Kleidon (1996)
LSE U-shaped Jan. - Mar. 1991 Abhyankar et al. (1997)
LSE reverse J-shaped Mar. - May 2001 Cai et al. (2004)
HKSE U-shaped Oct. 1996 - Mar. 1997 Ahn & Cheung (1999)
Jan. - Nov. 1991
Feb. 1992 - Apr. 1993
ISE reverse J-shaped May - Jul. 2008 Koksal (2012)
Note: TrtSE is short for Toronto Stock Exchange.

HKSE is short for Hong Kong Stock Exchange.
ISE is short for Istanbul Stock Exchange.
34
open and close to offset high adverse selection costs. Similar arguments are pro-
posed by Chung et al. (1999) who show that the U-shaped intraday pattern of
spread on the NYSE largely reflects the intraday behaviour of the spread estab-
lished by limit order traders.
In contrast to the U-shaped pattern noted above, reverse J-shaped intraday
patterns on bid-ask spread are demonstrated by Cai et al. (2004), Hussain (2011),
Koksal (2012) and McInish & Wood (1992). The spread opens wide, narrows
down in the first few hours and is relatively stable for the rest of the trading day.
As explained by the model of Madhavan (1992), the high spread in the morning
is due to greater uncertainty. As information asymmetry is gradually resolved
over the trading day, more market participants become informed by observing the
market, leading to a decline in the spread during the day.
Besides, some studies (Chan et al., 1995; Henker & Wang, 2006; Werner &
Kleidon, 1996) document that the spread declines over the trading day. As pointed
out by Chan et al. (1995), the difference of spread pattern between various markets
can be attributed to the differences in trading structure. A summary of the studies
documenting intraday patterns of the bid-ask spread is provided in Table 2.3.
2.4.3 Intraday Trading Volume
U-shaped intraday patterns on trading volume are documented in many studies

(summarised in Table 2.4). Trading volume is high at market open, is relatively
low during the day, and rises again as market close approaches. Jain & Gun-Ho
(1988) explain that high trading volumes at market open arise because investors
trade on information received overnight, and high volumes at market close arise
because investors close open positions that they can not monitor overnight. Atkins
& Basu (1995) examine the public announcements after trading hours and have
the same speculation.
35
Table 2.4: Intraday Pattern: Trading Volume

Sep. 1971 - Feb. 1972
NYSE U-shaped Wood et al. (1985)
Jan. - Dec. 1982
NYSE U-shaped Jan. 1979 - Dec. 1983 Jain & Gun-Ho (1988)
NYSE U-shaped before 1990 Lockwood & Linn (1990)
NYSE U-shaped 1980,1981,1983,1984 McInish & Wood (1990b)
NYSE U-shaped Jan. - Jun. 1989 McInish & Wood (1992)
NYSE U-shaped Jan. - Dec. 1988 Foster & Viswanathan (1993)
NYSE U-shaped Jan. - Dec. 1988 Lee et al. (1993)
Oct. - Dec. 1991
NASDAQ U-shaped Chan et al. (1995)
Aug. - Sep. 1992
TrtSE U-shaped before 1990 McInish & Wood (1990a)
TrtSE S-shaped Sep. 1999 - Nov 1999 Vo (2007)
LSE double-humped Jan. - Mar. 1991 Abhyankar et al. (1997)
LSE U-shaped Jan. - Dec. 1991 Werner & Kleidon (1996)
HKSE U-shaped before 1991 Ho & Cheung (1991)
HKSE U-shaped Oct 1996 - Mar 1997 Ahn & Cheung (1999)
TWSE J-shaped Mar. 1995 Lee et al. (2001)
Jan. - Nov. 1991
Feb. 1992 - Apr. 1993
ASX U-shaped Sep. 1990 - Aug. 1993 Aitken et al. (1995)
SSE U-shaped Nov. 1996 - Jan. 1997 Al-Suhaibani & Kryzanowski (2000)
ISE U-shaped May - Jul. 2008 Koksal (2012)
Note: TrtSE is short for Toronto Stock Exchange.

HKSE is short for Hong Kong Stock Exchange.
TWSE is short for Taiwan Stock Exchange.
ASE is short for Australia Stock Exchange.
SSE is short for Saudi Stock Exchange.
ISE is short for Istanbul Stock Exchange.
36
Other shapes of intraday patterns on trading volume are also observed in prior
literature. Abhyankar et al. (1997) find a two-humped pattern with highs at 9:00
a.m. and 3:00 p.m. on the LSE. This confirms the prediction in Admati & Pflei-
derer (1988), who argue that periods of concentrated trading may occur at some
points in the trading day not necessarily at market open or close. Lee et al. (2001)
show that trading volume exhibits a J-shaped pattern in the Taiwan Stock Ex-
change (TSE): the volume is stable during the day, but increases as market close
approaches and reaches a peak at market close. Vo (2007) observes that trading
volume increases over the trading day in the Toronto Stock Exchange (TrtSe), and
attributes the difference of U-shaped pattern in the NYSE and increasing pattern
in TrtSE to the structural difference between the two exchanges. Hussain (2011)
finds a reverse J-shaped pattern of trading volume in the Frankfurt Stock Exchange
(FSE): the volume is highest at market open, decreases in the first ten minutes and
is stable for the rest of the day.
2.4.4 Intraday Market Depth
Market depth has received relatively less attention in the literature on intraday
market phenomena. Both Lee et al. (1993) and Li et al. (2005) document an
inverted U-shaped intraday pattern of market depth on the NYSE. Market depth is
lowest at the opening and then rises monotonically until the close, at which point
it suddenly drops. These findings are corroborated by Ahn & Cheung (1999) and
Vo (2007) who examine the depth on the Hong Kong Stock Exchange and Toronto
Stock Exchange respectively. A summary of these studies is given in Table 2.5.
These studies all report a negative association between intraday spreads and
depth. The depth displays a reverse pattern of the spread: wide spreads are as-
sociated with small depths and narrow spreads are accompanied by large depths.
Lee et al. (1993) show that both depth and spread are associated with trading vol-
37
Table 2.5: Intraday Pattern: Market Depth

NYSE inverted U-shaped Jan. - Dec. 1988 Lee et al. (1993)
NYSE inverted U-shaped Nov. 1990 - Jan. 1991 Li et al. (2005)
HKSE inverted U-shaped Oct. 1996 - Mar. 1997 Ahn & Cheung (1999)
TrtSE inverted U-shaped Sep. 1999 - Nov 1999 Vo (2007)
Note: HKSE is short for Hong Kong Stock Exchange.

TrtSE is short for Toronto Stock Exchange.
ume. In response to increased volume, spread widens and depth drops. This is
consistent with the prediction of the theoretical model of market microstructure in
Easley & O’Hara (1992).
Ahn & Cheung (1999) attribute the reversed U-shaped pattern on depth to the
adverse selection problem for limit order traders. Due to high information asym-
metry at market open and close, limit order traders suffer high expected adverse
selection cost, and thus place passive limit orders in order to protect themselves
from informed trades, leading to an inverted U-shaped intraday pattern on market
depth, as well as a U-shaped intraday pattern on spread. Li et al. (2005) and Vo
(2007) also find limit order traders play an important role in forming the intraday
pattern on market depth.
The above review of literature on intraday phenomena shows that intraday
patterns have been observed in different markets and in different time periods.
However, the shapes of intraday patterns differ across markets and time periods.
Thus, it is necessary to provide up-to-date evidences on intraday phenomena as
the above literature is dated.
38
2.5 Trade Execution Strategies
Trade execution strategies are predefined sets of instructions for trade execution,
which include how best to break up a large order for execution and the actual
mechanisms for placing and managing the orders. When deciding whether or not
to submit an order to the market place an algorithm must decide on an order’s:
• Submission Time - when should the order be placed?
• Size - what size order should be sent to the market?
• Type - should the order be a market, limit or a hidden order?
• Pricing - at what price should the order be placed?
• Destination - there are many market destinations and types, which one will
provide the best conditions of execution for the order?
• Management - if a limit order has been submitted, how should this order be
managed post submission?
The literature on trade execution strategies can be classified into two cate-
gories: optimal trading strategies and order submission strategies. The first cat-
egory of studies focuses on how to split a large order in order to minimise price
impact cost. The other investigates how to choose order type and order aggres-
siveness. The following sections review these two streams of research separately.
2.5.1 Optimal Trading Strategy
Bertsimas & Lo (1998) are the first to analyse how to control the implicit execution
cost of trading a large order. They argue that large-order execution problem is a
39
dynamic optimisation problem, and form the basic framework to solve for this
problem.
Supposing that an investor wants to trade a large order S of shares within
1
a fixed finite number of periods T , and denoting St as the number of shares
traded in period t and Pt as the corresponding execution price, the optimal trade
execution problem is how to split the order in order to achieve minimised expected
[∑ ] ∑T
execution cost, namely E T
P S
t=1 t t , where t = 1, ..., T and t=1 St = S..
Subsequent studies are largely extensions of their work.

Recent years have seen a burgeoning range of studies on this issue. Some
have extended this model to the case of portfolios (e.g., Moazeni et al. (2010),
Kharroubi & Pham (2010)). Some have considered the case of FX markets (e.g.,
Schmidt (2010)). In this thesis only the studies focusing on single asset are re-
viewed.
The previous work on optimal trading execution differs in the way of mod-
elling price/liquidity dynamics. One stream of research led by Bertsimas & Lo
(1998) and Almgren & Chriss (1999) decomposes the price impact into a tempo-
rary component and a permanent component. The other stream of research led by
Obizhaeva & Wang (2012) and Alfonsi et al. (2010) models the price impact in
the limit order book and explicitly models the resilience of price impact. The first
approach is denoted by Almgren approach, and the other is denoted by Obizhaeva
approach.
Almgren Approach
Within this approach, the design of optimal execution strategies are primarily con-
ducted in the context of markets in which the law of one price holds. Trades have
two different types of price impact on the price process of the market. One is
1
The length of a ‘period’ can be some fraction of a single day, e.g., 30 minute interval (Bert-
simas & Lo, 1998).
40
permanent price impact which alters the market price permanently. The other is
temporary price impact which affect only the execution price of the corresponding
trade.
Bertsimas & Lo (1998) assume a linear and permanent price impact function
depending on the trade size, with the objective of minimising the expected cost
of execution. They derive optimal trade execution strategies using stochastic dy-
namic programming. In the case when the asset price process follows a zero-drift
arithmetic random walk, they show that the optimal execution strategy is to break
the trade equally across time. This optimal strategy is independent of the func-
tion of permanent price impact and the dynamic of the asset price, because the
dynamic of the asset price is only affected by the whole trade size and indepen-
dent of the divided pieces. Although the optimal execution strategy derived in this
work seems counter-intuitive, it provides a basic framework to solve the optimal
trade execution problem.
Based on the work of Bertsimas & Lo (1998), Almgren & Chriss (1999, 2000)
introduce a temporary price impact function of trade size. The temporary price
impact depends on the pace of trades. Fast trading causes larger impact on the
dynamic of the asset price than slow trading. In other words, the temporary im-
pact function is used to penalize speedy trades. Thus, the optimal trade execution
strategy depends on the dynamic of the asset price, the permanent impact function
and the temporary impact function. Their objective function is a mean-variance
criterion. With a linear specification of price impact, Almgren & Chriss (1999)
show that the optimal execution strategy quickly finishes all trades over the first
several periods, and Almgren & Chriss (2000) demonstrate that the optimal strat-
egy becomes more aggressive after the volatility increases, and becomes passive
after the volatility decreases.
Although Almgren’s model (Almgren & Chriss, 1999, 2000) is tractable, the
41
assumption of a linear price impact function is inconsistent with empirical find-
ings. A number of studies have extended the model of Almgren/Chriss by con-
sidering nonlinear price impact functions. With a nonlinear specification of tem-
porary price impact, Almgren (2003) shows that the optimal execution strategy is
independent of the scale of trading period and the whole trade size.
Obizhaeva Approach
While Bertsimas & Lo (1998) and Almgren & Chriss (1999, 2000) make compu-
tations feasible and lead to relatively nice and robust trading strategies, they do
not adequately model the empirically observed transience of price impact.
The price-impact-function based models do not take into account the dynamic
property of supply and demand of market liquidity. This property is explicitly
obvious in an limit order market, where the order book is able to rebuild itself
after being hit by a trade, which is referred to as the resilience of the order book.
Obizhaeva & Wang (2012) are the first to model the dynamic supply and de-
mand in a limit-order-book market. They consider a limit order market with an
order book which has a constant density, namely the financial securities are uni-
formly distributed at different levels of the order book. In the case when the price
impact of a trade decays exponentially over time, they show that the optimal ex-
ecution strategy crucially depends on shape of the limit order book and the speed
of its resilience.
Alfonsi et al. (2010) extend Obizhaeva & Wang (2012)’s model by allowing
for nonuniformly distributed shares within the limit order book. In other words,
the density of limit orders is not constant, but varies as a function of price. They
demonstrate that trading is discrete and for an optimal purchasing strategy all
purchases except the first and last are of the same size. Furthermore, the sizes of
the intermediate purchases are chosen so that the price impact of each purchase is
42
exactly offset by the order book resiliency before the next purchase.
A continuous-time generalization of the volume impact version of the model
has been introduced by Predoiu et al. (2011). They restrict trading to buy-only or
sell-only strategies, so price manipulation is excluded by definition. Predoiu et al.
(2011) permit the order book shape to be completely general. All price impact is
transient. Assuming exponential decay of price impact, Fruth et al. (2011) anal-
yse the specific form and regularity of optimal order execution strategies when
liquidity can be time-dependent or even stochastic. Gatheral (2010) analyse the
existence and the behaviour of optimal order execution strategies for certain mar-
ket microstructure parameters. As shown by Alfonsi & Schied (2010) and further
discussed in Gatheral et al. (2011), these results depend strongly on the way in
which nonlinearity of price impact is modelled.
2.5.2 Order Submission Strategy
Market and limit orders can be categorised into various aggressivenesses in terms
of their quantities and prices (Biais et al., 1995). For example, a buying limit
order’s level of aggressiveness increases with its limit price and its quantity.
Specifying an order’s aggressiveness is a critical decision which a trader has
to make when he wants to transact an order. This decision directly affects the
quality of trade execution, and indirectly influences the final investment return.
An understanding of the dynamic order submission behaviours will contribute to
our understanding of the dynamic price formation process.
Market orders and limit orders both have their advantages and disadvantages.
Although market orders guarantee order executions, they suffer the cost of paying
higher execution prices. While limit orders provide price improvement over mar-
ket orders, their submissions involve two risks (Handa & Schwartz, 1996). The
first risk is adverse selection risk, also known as free trading option risk (Glosten
43
& Milgrom, 1985), arising from adverse information events which trigger an un-
desirable execution. The other risk is nonexecution risk, which arises when the
market price moves away from the limit price and results in the limit order not
executing. The outcome of the second risk is that the limit order traders suffer
opportunity cost as they have to accept inferior prices.
Empirical studies have shown that different types of market participants have
distinguished preferences concerning order aggressiveness. Keim & Madhavan
(1995a) find that value investors are more likely to use limit orders instead of mar-
ket orders. Aitken et al. (2007) find that limit orders placed by proprietary trading
desks and hedge funds are more aggressive than mutual funds, index funds and in-
surance companies. While Glosten (1994) and Seppi (1997) argue that informed
traders are more likely to use market orders, Bloomfield et al. (2005), Chakravarty
& Holden (1995), Kaniel & Liu (2006), Harris (1998) and Wald & Horriga (2005)
suggest that informed traders use limit orders more often than market orders.
The execution performance of aggressive and passive orders have been in-
vestigated in the previous studies. Generally, passive orders outperform aggres-
sive orders in terms of investment return (Handa & Schwartz, 1996) and trading
cost (Harris & Hasbrouck, 1996). Aggressive orders have larger price impact but
smaller opportunity cost than passive orders (Griffiths et al., 2000). Moreover, the
execution performance of institutional traders and individual traders is compared
by Anand et al. (2005), which finds that aggressive orders placed by institutional
traders outperform those placed by individual traders.
Order Aggressiveness and State of The Order Book
The state of the order book is a critical determinant of traders’ order aggressive-
ness. This assertion has been suggested by recent theoretical developments (Fou-
cault, 1999; Foucault et al., 2005; Goettler et al., 2005; Handa et al., 2003; Liu,
44
2009; Menkhoff et al., 2010; Parlour, 1998; Rosu, 2009), and is supported by
many studies in different markets and over different sample periods. The rest of
this section discusses how various order-book related variables affect the aggres-
siveness of incoming orders.
Bid-ask Spread
Bid-ask spread has a positive relationship with incoming order’s aggressive-
ness (Al-Suhaibani & Kryzanowski, 2000; Beber & Caglio, 2005; Bae et al., 2003;
Biais et al., 1995; Cao et al., 2008; Duong et al., 2009; Ellul et al., 2007; Griffiths
et al., 2000; Hall & Hautsch, 2007; Pascual & Verdas, 2009; Ranaldo, 2004; Ver-
hoeven et al., 2004; Xu, 2009). A decrease in the bid-ask spread makes the use
of aggressive orders more attractive for incoming traders, and an increase in the
spread makes the use of passive orders more tempting. However, this relationship
is influenced by the proportion of patient traders in the trading population (Fou-
cault et al., 2005). If patient (impatient) traders dominate the trading population,
more (less) aggressive orders are submitted than passive orders when the spread
is wide.
Volatility
Prior empirical research is inconclusive on the effect of transient volatility
on order aggressiveness. Most studies (Ahn et al., 2001; Beber & Caglio, 2005;
Bae et al., 2003; Chung et al., 1999; Hall & Hautsch, 2007; Ranaldo, 2004; Ver-
hoeven et al., 2004) support an inverse relation between order aggressiveness and
volatility: when market volatility increases, passive orders are more attractive than
aggressive orders for incoming traders. Ahn et al. (2001) examine the relationship
between order placement and the transitory volatility from different sides of the
markets, and find that more buy (ask) limit orders are placed than buy (ask) market
orders when the volatility arises from the bid (ask) side.
This inverse relation can be explained by the high probability of mispricing an
45
asset in a volatile market which thus raises the cost of market order trading (Fou-
cault, 1999). In other words, higher price volatility means a greater opportunity
of execution an order at a better price. Consistently, Lo et al. (2002) show that
in periods of high volatility, the execution probability of limit orders increases,
which makes passive orders more tempting than aggressive orders.
Other studies provide different results. A positive relationship between order
aggressiveness and volatility is reported in Hasbrouck & Saar (2002); Lo & Sapp
(2010); Wald & Horriga (2005). This positive relation is only documented for
institutional traders in large cap stocks in Duong et al. (2009). Hall & Hautsch
(2006) observe an increase of all kinds of order submission during periods of
high volatility. Cao et al. (2008) find that volatility has a minimal effect on order
aggressiveness. Pascual & Verdas (2009) find that the relationship is affected by
market capitalisation: higher historic volatility suggests limit order submission in
mid-cap stocks, but the opposite phenomenon is observed in large-cap stocks.
Some theoretical models (Foucault, 1999; Handa & Schwartz, 1996) of limit
order market have predicted that price volatility determines the choice between
market and limit orders. Foucault (1999) and Handa & Schwartz (1996) predict
that market order is less attractive than limit orders and thus more limit orders are
submitted to the market than market orders when price is volatile. They attribute
this to asymmetric information between investors. When the market is volatile,
the probability of trading against informed investors increases, leading to higher
expected loss.
Market Depth
Market depth has also been found to matter. The limit order book at the best
quotes generates a crowding out effect which affects subsequent order submis-
sions (Parlour, 1998). The incoming trader is more likely to submit a market
order if the book has a deep depth on the same side, and is more likely to submit
46
a less aggressive order if the book is deep on the opposite side (Al-Suhaibani &
Kryzanowski, 2000; Beber & Caglio, 2005; Cao et al., 2009; Duong et al., 2009;
Ellul et al., 2007; Foucault, 1999; Griffiths et al., 2000; Hall & Hautsch, 2007;
Omura et al., 2000; Parlour, 1998; Peterson & Sirri, 2002; Potters & Bouchaud,
2003; Ranaldo, 2004; Verhoeven et al., 2004; Xu, 2009) . Especially when both
sides of the book are thick, ‘fleeting’ limit orders may happen (Rosu, 2009): some
limit orders placed inside the spread are immediately accepted by some traders
from the other side of the book. These studies reveal that time-to-execution is a
significant concern of limit order traders when execution cost less matters.
In contrast, depth beyond the best quotes has an inverse relation with order
aggressiveness (Goettler et al., 2005). Higher depth above the best ask signals that
the best ask may be too low compared with the fundamental value of the asset, so
that incoming buyers are more likely to use aggressive orders and incoming sellers
are more likely to use passive orders, and vice versa for the case of higher depth
below the best bid.
Eisler et al. (2012) show that market orders on one side of the book attract
compensating limit orders, leading to smaller conditional impact of subsequent
market orders on the same side.
Time of the Day
Order submission strategy is also affected by time of the day (Biais et al., 1995;
Ellul et al., 2007). The placement of less aggressive orders tend to be concentrated
in the morning, whereas more aggressive orders tends to occur late in the trading
day. Passive orders are more likely to happen at the end of the day than in an
earlier period (Foucault et al., 2005), because spread improvement via the use of
limit order is small.
Controlling price impact is important in financial markets. An increasing num-
ber of studies have been denoted to this problem. However, these studies, as
47
reviewed in Section 2.5.1, only adopt market orders, although limit orders are
equally important when traders make decisions on order choice as shown in Sec-
tion 2.5.2. It is still unknown whether order choice plays an important role in
determining price impact when trading a large order.
2.6 Summary
The aim of this chapter was to provide a brief overview of market microstructure
literature and a review of relevant work to this thesis.
Section 2.2 provided a brief overview of the literature on market microstruc-
ture. Market microstructure concerns every aspect of the trading process. The-
oretical models on market microstructure can be classified into three categories:
inventory-based models, information-based models, and limit-order-book models.
As more and more markets adopt limit order books, the number of studies on the
trading process in limit order markets is significantly increasing in recent years.
Section 2.3 offered a review of the literature on price impact. Most prior stud-
ies on price impact have been devoted to the relationship between price impact
caused by market orders and their trade sizes. A nonlinear relationship was dis-
covered by these studies. However, the order size does not tell the whole story
about price impact. Recently, a number of studies have found that most of the ex-
treme price changes are not caused by large trading size, but by the gaps between
different levels of the order book.
Section 2.4 reviewed the literature on intraday phenomena in financial mar-
kets. Previous studies concentrate on four variables measuring market liquidity
and trading activity, which are price volatility, bid-ask spread, trading volume and
market depth. A number intraday patterns with different shapes were observed in
prior literature and were summarised in this section.
48
Section 2.5 reviewed the literature on trade execution strategies. This review
focuses on the literature concerning large-order trading strategies and order sub-
mission strategies. The studies on large-order trading strategies differ in how to
model price impact. One stream of studies decomposes price impact into a tempo-
rary component and a permanent component, while the other stream specifically
considers price impact in the limit order book and models both the immediate ef-
fect of trading on the order book and the resilience of price impact. Most studies
on order submission strategies focus on the relation between order aggressiveness
and the state of the order book.
The review in this chapter identifies a number of research gaps. These are
summarised as follows:
1. Analysis of the intraday behaviour of price impact
It is unknown about up-to-date intraday behaviour of price impact. An analysis

of intraday behaviour of price impact is of great interest to both academics and
practitioners. It is helpful for developing a more comprehensive price impact
model than those reviewed in Section 2.3.1, as well as the design of trade
execution strategies as reviewed in Section 2.3.2. Chapter 5 conducts such
analysis using the data drawn from the NYSE-Euronext TAQ database and the
LSE ROB database.
2. Examination of the relation between agent intelligence and price impact
No study has been conducted to investigate whether agent intelligence plays an

important role in determining the magnitude of price impact. This examination
is of great importance to understand the driving factors of the concavity of
price impact. An examination of this is undertaken in Chapter 6.
3. Investigation of the effect of order choice on price impact of trading large

orders
49
It is unknown that whether order choice plays an important role in determining
price impact when trading a large order. This gap is addressed in Chapter 7.
The next chapter will present the methodology adopted in two experimental
studies of this thesis.
50
Chapter 3
Agent-based Modelling in Finance
“... the possibility of building a mathematical theory of a system

or of simulating that system does not depend on having an adequate
microtheory of the natural laws that govern the system components.
Such a microtheory might indeed be simply irrelevant.”
Nobel laureate in Economics: Simon (1996) (p. 19)
This thesis adopts an agent-based modelling approach in two of its experimen-

tal chapters. This chapter reviews studies on agent-based modelling in financial
markets, and describes the core elements of agent-based market simulation as well
as the Genetic Algorithm as a learning algorithm in agent-based modelling.
Agent-based modeling (ABM) is a computer modelling approach to construct
complex systems consisting of a number of decision makers (agents) and insti-
tutions interacting through prescribed rules (Farmer & Foley, 2009). It creates a
virtual universe in which agents act in complex and realistic ways.
The potential scope of applications of ABM is wide. Economics and finance
are among the most promising areas where ABM can be applied. Financial mar-
kets are always considered as complex systems, which are constantly evolving
51
and becoming more complex, and thus difficult to analyse and understand. ABM
has provided a powerful tool to analyse market behaviours and the effects of the
market mechanism in financial markets. The applications of ABM to financial
markets are vast. The review here can not be exhaustive. For general reviews,
see for example Chen (2007, 2012); Chen et al. (2012); Chiarella et al. (2009a);
Consiglio (2007); Cristelli et al. (2011); Hommes & Wagener (2009); LeBaron
(2000, 2001, 2006); Lux (2009); Samanidou et al. (2007); Tesfatsion (2006).
The rest of this chapter is organised as follows. Section 3.1 briefly reviews the
literature on agent-based financial markets, followed by a description of the core
elements of agent-based market simulation in Section 3.2. An introduction to a
learning algorithm in agent-based modelling, the Genetic Algorithm, is given in
Section 3.3. A summary of this chapter is provided in Section 3.4.
3.1 Agent-based Financial Markets

Agent-based modelling has been applied to economics and finance since 1971
(Schelling, 1971). As defined by Tesfatsion (2006), agent-based computational
economics is “the computational study of economic processes modelled as dy-
namic systems of interacting agents”. It offers an alternative approach to mod-
elling and studying financial markets. It differs from the traditional equilibrium
models which can not always provide analytical solutions (Parlour & Seppi, 2008).
Agent-based models of financial markets can help us to understand how markets
operate and can guide policy formulation.
Agent-based markets have been used to study price bubbles and market crashes
in a number of studies, like Feldman & Friedman (2010), Kim & Markowitz
(1989), Miller (2008) and Palmer et al. (1994). The Nobel laureate Harry Markowitz
(Kim & Markowitz, 1989) is among the earliest researchers using an agent-based
52
market simulation. Their study proposes an agent-based simulation to explain the
1987 crash. The agent-based artificial market with agents that learn their trad-
ing behaviours using genetic algorithm in Palmer et al. (1994) exhibit financial
bubbles. Feldman & Friedman (2010) investigate how human traders and robot
traders behave and interact in a market prone to bubbles and crashes. They find
that human traders have little impact on market crashes as they are more expe-
rienced. Miller (2008) finds that the bubbles and crashes seen in experiments
conducted with robot traders are much less drastic than those with human sub-
jects. This difference is caused by the restriction that robot agents are prohibited
from engaging in actions that could result in a loss and are thus not allowed to
speculate.
Some studies have used agent-based financial markets to study the statistical
regularities, often know as stylised facts, observed in real markets. Examples of
stylised facts are fat tails of price return and volatility clustering (Cont, 2001).
An foreign exchange market with chartist and fundamentalist agents is simulated
in Lux (1998) and exhibits fat tails of price return. LiCalzi & Pellizzari (2003)
produces a leptokurtic distribution of short-term log-returns in an artificial order-
driven market consisting of fundamentalist and Zero Intelligence (ZI) value-based
agents under budget constraints. Their results support the conjecture in Bouchaud
et al. (2002) and Daniels et al. (2003) that the emergence of some of the statistical
properties of order-driven markets is mostly due to their microstructure. Chiarella
& Iori (2002) produce volatility clustering in an artificial markets composed of
fundamentalist, chartist and noise traders. Chiarella et al. (2009b) show that the
chartist strategy is mainly responsible of the fat tails and clustering in their ar-
tificial markets. Their results offer evidence that large price impacts are mostly
caused by the presence of large gaps in the order book.
In addition, ABM has been applied to other problems in finance, like tests of
53
economic theories (Chen & Yeh, 2002), market design (Darley & Outkin, 2007),
option pricing (Suzuki et al., 2009), evaluation of automated trading strategies
(Izumi et al., 2009), and analysis of liquidity costs (Huang et al., 2012). Chen &
Yeh (2002) use an agent-based stock market with a number of evolving agents to
test both the efficient market hypothesis and the rational expectations hypothesis,
and show that their agent-based model can replicate some economic behaviours
empirically. Darley & Outkin (2007) examine the effect of the tick-size reduc-
tion in the NASDAQ market using an agent-based artificial market. Suzuki et al.
(2009) analyse implied volatility smile and the skewness premium using an agent-
based modelling approach. Izumi et al. (2009) use an artificial market to evaluate
the risks and returns of automated trading strategies in various market environ-
ments and test the market impact of the trading strategies. Huang et al. (2012)
report a higher liquidity costs of market orders in their agent-based order-driven
stock market than in the Taiwan Stock Market.
3.2 The Core Elements of Agent-based Market Sim-

ulation
The core elements of constructing an agent-based artificial market are
• agents,
• market mechanism.
3.2.1 Agents
Agents are the essential elements of an agent-based market (LeBaron, 2001). The
agents can range from “active data-gathering decision-makers with sophisticated
54
learning capabilities to passive world features with no cognitive functioning”, ac-
cording to Tesfatsion (2006). Chen (2012) undertakes a comprehensive review of
various types of agents designs.
Intelligent Agents
At one end of the spectrum are intelligent agents who can learn, adapt and evolve.
A variety of techniques from artificial intelligence have been used to model agent
learning. The celebrated example is the Sante Fe Institute Artificial Stock Mar-
ket (SFI ASM) (Arthur et al., 1997). In their model, agents are heterogeneous
in the way they form their expectations. Each agent, possessing a number of
linear forecasting models at any time, learns and evolves by discovering which
forecasting model works best as well as developing new ones from time to time,
via an inductive algorithm, namely the Genetic Algorithm. They show that both
herd effects and systematic speculative profits are possible in an endogenous-
expectations market.
Zero-intelligence Agents
The zero-intelligence (ZI) agents are the other extreme of the spectrum of agent
types. Zero-intelligence modelling was originally pioneered by the Nobel laure-
ate Becker (1962), who showed that some aspects of supply and demand curves
could be understood without any reliance on strategic thinking. Later, it was pop-
ularised by Gode & Sunder (1993). As characterised in the work of Gode &
Sunder (1993), these agents are random behaving agents who randomly generate
buying and selling orders subject to budget constraints. The orders submitted by
impatient traders are executed against patient orders previously submitted to the
market. The patient orders are placed in the order book according to price and
time priorities. Despite their simplicity, ZI agents are able to get a remarkable
55
allocation efficiency in the artificial double auction market. A good review of ZI
agents based artificial markets is given by Iori et al. (2003); Ladley & Schenk
(2009); Zovko & Farmer (2002).
Despite their simplicity, the zero-intelligence artificial markets are able to gen-
erate many non-trivial behaviours seen in real markets. Bak et al. (1997) show that
their zero-intelligence model is able to reproduce phenomena seen in actual mar-
kets, including larger than Gaussian fluctuations at short time scales and some of
the power-law forms of price variations. The zero-intelligence model in Maslov
(2000) can produce approximate several regularities observed in actual markets,
including the autocorrelation of the absolute value of price changes and short-
range correlations in the signs of price movements, as well as the fat tails of price
changes.
3.2.2 Trading Mechanism
Trading Mechanism is a critical design decision which the builder has to con-
sider when creating an agent-based financial market. It relates the agents’ trading
demands represented by their orders and the prices of the market.
Most artificial financial markets implement simplified market mechanisms,
which omit some institutional details of trading, but serve their research needs
sufficiently. One example is the SFI market (Arthur et al., 1997), which imple-
ments a simple market-clearing mechanism. At the end of each trading period, the
market specialist collects the accumulated buy and sell orders of all the agents, and
clears the market at a new market price. If the demand excesses the supply, the
specialist increases the market price, otherwise he decreases the price.
A number of studies have implemented a fairly realistic market, which incor-
porates explicit market microstructure, like a continuous double auction mech-
anism with limit order books. As pointed out in LeBaron (2001), this type of
56
trading mechanism is “most appealing for modelling high-frequency data” and is
only easy to implement when there is no human intervention (like specialists or
market makers) in the price formation process. All ZI artificial markets replicat-
ing a limit order market are of this type. Some examples are Chan et al. (2001);
Raberto & Cincotti (2005); Yang (2002).
The next section presents the Genetic Algorithm, which is often used to simu-
late agent learning in agent-based modellings.
3.3 Genetic Algorithm

The genetic algorithm (GA) is an optimisation algorithm, founded upon the evo-
lutionary principle, “survival of the fittest”. It provides a systematic search pro-
cess directed by performance feedback. In financial markets, GA methodologies
have been successfully used for solving a broad selection of problems, ranging
from predicting movements in current values to optimizing equity portfolio com-
position. Examples of applications of GA in economics and finance are Allen
& Karjalainen (1999), Arifovic (1995), Arifovic (1996), Brabazon et al. (2008),
Chavez-Demoulin & McGill (2012), Chen & Yeh (1997), Dempster & Jones
(2001), Dworman et al. (1996), Kozhan & Salmon (2011), Neely et al. (1997),
Neely & Weller (1999), Neely & Weller (2001), Neely & Weller (2003), and Se-
fiane & Benbouziane (2012).
The evolutionary processes in GA represent an archetype, whose application
transcends their biological root. In biological evolution, species are positively or
negatively selected depending on their relative success in surviving and reproduc-
ing in the environment. Differential survival, and variety generation during repro-
duction, provide the engine for evolution (Darwin, 1859; Spencer, 1864) (Figure
3.1).
57
Selection
Parents
Recombination
Population &
Mutation
Offspring
Replacement
Figure 3.1: Evolutionary Cycle
These concepts have metaphorically inspired the field of evolutionary com-

putation (EC). Algorithm 1 outlines the evolutionary meta-algorithm. There are
many ways of operationalising each of the steps in this meta-algorithm, conse-
quently, there are many different, but related, evolutionary algorithms (EA). Just
as in biological evolution, the selection step is a pivotal driver of the algorithm’s
workings. The selection step is biased in order to preferentially select better (or
‘more fit’) members of the current population. The generation of new individuals
creates children which bear some similarity to their parents but are not identical
to them. Hence, each individual represents a trial solution in the environment,
with better individuals having increased chance of influencing the composition of
individuals in future generations. This process can be considered as a ‘search’ pro-
cess, where the objective is to continually improve the quality of individuals in the
population. In financial markets, EC methodologies including Genetic Algorithm
58
(GA), Genetic Programming (GP) and Grammatical Evolution (GE), have been
used for solving a broad selection of problems, ranging from prediction, asset se-
lection, portfolio optimization, derivatives pricing and credit risk assessment. An
overview of some EC applications in finance can be seen in Brabazon & O’Neill
(2006) and Brabazon et al. (2008).
Initialise the population of candidate solutions;

repeat
Select individuals (parents) for breeding from the current population;
Generate new individuals (children) from these parents;
Replace some / all of the current population with the newly-generated individuals;
until terminating condition;
Algorithm 1: Evolutionary Algorithm
3.3.1 Grammar-based Genetic Algorithm
The grammar-based genetic algorithm, which is Grammatical Evolution (Dempsey

et al., 2009; O’Neill & Ryan, 2003), is an evolutionary methodology, and can be
used to evolve structures, programs or ‘rule sets’. These rule sets can be as gen-
eral as a functional expression which produces a good mapping between a series
of known input-output data vectors.
GE is an Evolutionary Automatic Programming (EAP) technique which al-
lows the generation of computer programs (or ‘rule sets’) in an arbitrary language.
GE can conduct an efficient exploration of a search space, and notably permits
the incorporation of existing domain knowledge in order to generate ‘solutions’
with a desired structure. In finance (for example), this allows the users to seed
the evolutionary process with their current trading strategies in order to see what
improvements the evolutionary process can uncover. Recently GE has been suc-
cessfully applied to a number of financial problems. These include financial time
series modelling, intraday financial asset trading, corporate credit rating, and the
59
uncovering of technical trading rules (Brabazon & O’Neill, 2006).
A particular strength of the methodology is that the form of the model need
not be specified a priori by the modeler. This is of particular utility in cases,
such as in this study, where there is a theoretical or intuitive idea of the nature
of the relevant explanatory variables, but a weak understanding of the functional
relationship between the explanatory and the dependent variable(s). GE does not
require that the model form is linear, nor does the method require that the measure
of model error used in model construction is a continuous or differentiable func-
tion. Another useful feature of a GE approach is that it produces human-readable
rules that have the potential to enhance understanding of the problem domain.
Interestingly, GP/GE methods enjoy wide application beyond finance (O’Neill &
Brabazon, 2009), and routinely produce human-competitive performance, with
some solutions being patentable in their own right (Koza, 2010).
Genotype-phenotype Mapping
A genotype-phenotype mapping is employed such that each individual’s variable

length binary string, contains in its codons (groups of 8 bits) the information to
select production rules from a Backus Naur Form (BNF) grammar. GE is there-
fore a grammar-based approach to GP (McKay et al., 2010). The user can tailor
the grammar to produce solutions that incorporate domain knowledge by biasing
the grammar to produce very specific forms of sentences. BNF is a notation that
represents a language in the form of production rules. It is comprised of a set of
non-terminals that can be mapped to elements of the set of terminals (the primitive
symbols that can be used to construct the output program or sentence(s)), accord-
ing to the production rules. A simple example BNF grammar is given below,
where <expr> is the start symbol from which all programs are generated. The
grammar states that <expr> can be replaced with either <expr><op><expr>
60
or <var>. An <op> can become either +, -, or *, and a <var> can become
either x, or y.
<expr> ::= <expr><op><expr> (0)

| <var> (1)
<op> ::= + (0)
| - (1)
| * (2)
<var> ::= x (0)
| y (1)
The grammar is used in a developmental process to construct a program by apply-

ing production rules, selected by the genome, beginning from the start symbol of
the grammar. In order to select a production rule in GE, the next codon value on
the genome is read, interpreted, and placed in the following formula:
Rule = Codon V alue M od N um. Rules
where Mod represents the modulus operator. Given the example individual’s
genome (where each 8-bit codon has been represented as an integer for ease of
reading) in Fig.3.2, the first codon integer value is 20, and given that there are 2
rules to select from for <expr> as in the above example, 20 M od 2 = 0 is
obtained. <expr> will therefore be replaced with <expr><op><expr>.
Beginning from the left hand side of the genome codon integer values are
generated and used to select appropriate rules for the left-most non-terminal in the
developing program from the BNF grammar, until one of the following situations
arise:
61
• A complete program is generated. This occurs when all the non-terminals
in the expression being mapped are transformed into elements from the ter-
minal set of the BNF grammar.
• The end of the genome is reached, in which case the wrapping operator
is invoked. This results in the return of the genome reading frame to the
left hand side of the genome once again. The reading of codons will then
continue unless an upper threshold representing the maximum number of
wrapping events has occurred during this individual’s mapping process.
• In the event that a threshold on the number of wrapping events has oc-
curred and the individual is still incompletely mapped, the mapping process
is halted, and the individual assigned the lowest possible fitness value.
Returning to the example individual, the left-most <expr> in <expr><op><expr>

is mapped by reading the next codon integer value 124 and used in 124 M od 2 =
0 to become another <expr><op><expr>. The developing program now looks
like <expr><op><expr><op><expr>. Continuing to read subsequent codons
and always mapping the left-most non-terminal the individual finally generates
the expression y*x-x-x+x, leaving a number of unused codons at the end of the
individual, which are deemed to be introns (unused codons) and simply ignored.
Figure 3.2: An Example GE Individual’s Genome Represented as Integers for

Ease of Reading.
In the context of this study, GE is used to evolve trade execution strategies.
62
A population of strategies is maintained and iteratively improved via a simulated
evolutionary process. The structure of these rules is governed by a choice of
grammar and the utility of evolved strategies is assessed by testing them in an
artificial stock market environment. Both the choice of grammar and the design
of the stock market environment are discussed in later chapters.
3.4 Summary
This chapter presented the ABM methodology and briefly reviewed its applica-
tions to the area of finance. A description of the three essential components for
constructing agent-based artificial markets is also given. Moreover, a popular
learning algorithms, the Genetic Algorithm, was introduced. Although ABM has
been applied to economics and finance for a long time, no previous studies have
investigated price impact using ABM.
In Part II of this thesis, the three empirical studies are presented. Chapter 4
initially provides a description of the data used and of the necessary data prepro-
cessing steps.
63
Part II
Experiments
64
In Part II, several empirical investigations of price impact are presented. The
literature review in Part I has identified a number of research gaps concerning our
understanding of intraday behaviour of price impact, whether agent intelligence
affects the magnitude of price impacts, and the effects of order choice on the price
impact of trading large-orders.
The analysed data is drawn from the NYSE-Euronext TAQ database and the
LSE ROB database. In Chapter 4, the data is introduced and a description of the
data preprocessing steps is given.
Three studies are presented in Chapters 5-7 to address the three research ques-
tions presented in Chapter 1. Chapter 5 studies the intraday behaviour of price im-
pacts and market liquidity using TAQ data and the ROB data. Chapter 6 presents a
study investigating whether agent intelligence determines the magnitude of price
impact. In Chapter 7, the effect of order choice on the price impacts of trading
large-orders is examined.
65
Chapter 4
Datasets and Data Processing
Parts of this chapter have appeared in:


More than ever before researchers today face the challenges of working with
high-frequency datasets from financial markets. Examples of this data include
NYSE’s TAQ data, and LSE’s ROB data. These datasets are very large and very
difficult to process. The data records all visible order activities, i.e. limit or-
der submissions, cancellations and executions, and can be used to reconstruct the
historical limit order book up to any required precision. However, this creates a
challenge as it is necessary to reconstruct the limit order book using the same rules
66
that were used by the matching algorithm applied by the exchange. Although sim-
ple in principle, such algorithms need to take into account market-specific issues
and considering the large volume of data, work is extremely difficult.
The previous two chapters conducted a literature review in the field of market
microstructure relating to the work in this thesis. This chapter gives an introduc-
tion to the two datasets used in this thesis. One is the NYSE Euronext trade &
quote dataset (TAQ), and the other is the LSE rebuild order book dataset (ROB).
Both of these are intraday high-frequency data, recording information on both
trades and quotes throughout each trading day. The ultra high-frequency data pro-
vided by NYSE-Euronext is unfiltered and contains errors and noise. Analysing
the raw data can produce inaccurate results which lead to imprecise conclusions.
Filtering the raw NYSE-Euronext data is important before data analysis. Like-
wise, the ultra high-frequency data provided by LSE is also raw data, for example,
the original LSE data is out of time sequence. Preprossing the raw LSE data is
necessary before data analysis.
The remainder of this chapter is organised as follows. Section 4.1 introduces
the datasets. Section 4.2 discusses the issues in data processing and gives the steps
used to process the data. A summary of this chapter is provided in Section 4.3.
4.1 Datasets
This section presents descriptions of the two databases, namely the NYSE Eu-
ronext TAQ database and the LSE ROB database.
67
4.1.1 NYSE Euronext TAQ Database
Source of the Data
The NYSE Euronext Trade and Quote (TAQ) database contains intra day infor-
mation on trades and quotes for equity securities listed on most of the major US
exchanges. The database commenced in 1993, and the information contained is
in turn drawn from the ‘Consolidated Tape’ which takes an information flow on
real-time trades and quotes from all the participating markets and ‘consolidates’
this to provide traders with a complete picture of trades and quotes across all mar-
kets. The current participants in the system include all major US equity trading
venues (further background information on these venues is provided in Sect. 4.1.1
below).
Technically, there are two discrete elements of the consolidated tape, the Con-
solidated Tape Plan, which governs trades and the Consolidated Quotation Plan,
which governs quotes. Since the late 1970s, all SEC-registered exchanges and
market centers that trade NYSE or AMEX-listed securities send their trades and
quotes to a central consolidator where the Consolidated Tape System (CTS) and
Consolidated Quote System (CQS) data streams are produced and distributed
worldwide.
For investors and traders, the CQS provides pre-trade transparency as it dis-
plays the best bid and offer price and volume (and all revisions of quotes), posted
by specialists (NYSE, AMEX) and by market makers (NASDAQ), for all stocks
for a particular security on each exchange on which it trades, allowing investors
and traders to decide how best to route their orders. In contrast, the CTS provides
post-trade transparency allowing investors and traders to compare the execution
price of their trades against that of other trades in the market place which took
place at the same time.
68
The way that trade and quote information reaches the CTS and CQS have
changed over the years, moving from largely manual systems in traditional stock
exchanges (such as use of floor reporters by the NYSE until 2001) with the atten-
dant risk of delay and errors in recording,1 to largely automated recording of this
data in modern markets. As noted above, the TAQ database is extracted from the
consolidated tape and consequently includes all transactions data, time-stamped
to the nearest second, reported during the Consolidated Tape hours of operation
(currently 4am EST until 6.30pm EST). The data does not include information on
the identity of the trading parties or whether the trade is buyer or seller initiated.
TAQ data (along with other tick data products) is available for purchase from
NYSE Euronext in a number of delivery formats. The academic research version
of the TAQ data is supplied via monthly data DVDs. In the following sections, the
scale of trade and quote activity for a number of equities will be illustrated, but
in advance of that, it is worth noting the large number of equity securities that are
covered by the consolidated tape (as at 27 January 2011, there were 3,243 firms
listed on the NYSE, 2,875 listed on Nasdaq and 551 listed on AMEX).
Accessing the Data
The academic version of the data is available for purchase approximately four
weeks after the end of each month. The data is accessible via a supplied front-end
extraction program. Figure 4.1 illustrates the first screen wherein the user is able
to select the desired date ranges and trade/quote information. A number of other
screens (including Figure 4.2) allow the user to select the relevant ticker codes
and exchanges of interest, along with a choice of data output file format. Users
1
For example, using data drawn from July 1994 to June 1995, Blume and Goldstein Blume &
Goldstein (1997) reported a median delay of 16 seconds between the execution and the consequent
reporting of NYSE trades, with Peterson and Sirri Peterson & Sirri (2003) reporting a median delay
of 2 seconds using 1997 NYSE data from the NYSE System Order database File.
69
Figure 4.1: Data Selection Screen for TAQ3 Extraction Tool
70
Figure 4.2: Data Filter Screen for TAQ3 Extraction Tool
71
can also access the raw data files directly via custom written scripts.
Sample Data
A sample of the trade data available is provided in Table 4.1. Starting from the
left-most field, there are the stock ticker code (here ‘F’ for Ford), the exchange
on which the trade occurred (here ‘N’ for NYSE), the date and time of the trade,
the trade price and trade size, and a series of data flags which indicate the trade
condition, the correction indicator and a flag which indicates whether the trade is
a ‘G’ trade or a rule 127 transaction.2
Table 4.1: A Sample of Trade Data for Ford Drawn from 2/2/09
Ticker Exchange Date Time Price Size Cond Corr G127

F N 2/2/09 11:20:48 1.8700 100 0 0
F N 2/2/09 11:20:51 1.8800 100 0 0
F D 2/2/09 11:20:54 1.8784 1000 0 0
F D 2/2/09 11:20:55 1.8800 3000 0 0
F Z 2/2/09 11:20:55 1.8800 100 0 0
F D 2/2/09 11:21:04 1.8700 145 0 0
The quote information indicates the best (quoted) bid / offer available on an
exchange at a point in time. This changes with each alteration of the best bid/offer
or any changes to the number of shares on offer at either of these prices. As would
be expected, a very large quantity of quote information is generated each day
for the more actively traded stocks. Taking the sample information illustrated in
Table 4.2, the leftmost column indicates the ticker code, followed by the exchange
generating the quote (‘C’ = Cincinnati,3 ‘T’=Nasdaq, ‘N’=NYSE etc.), date and
time of quote, best bid/offer, and the volume (in 100 share lots) at each of these
prices.
2
“G” trade: A member firm trading for its own account must publicly identify that the order
is principal. Rule 127: An NYSE trade reported as having been executed as a block position.
3
Known as the National Stock Exchange since 2003 and headquartered in New Jersey.
72
Table 4.2: A Sample of Quote Data for Ford Drawn from 2/2/09
Ticker Exchange Date Time Bid Size Offer Size Mode MMID
F C 2/2/09 11:20:56 1.8700 175 1.8800 75 12
F T 2/2/09 11:20:56 1.8700 458 1.8800 597 12
F C 2/2/09 11:20:56 1.8700 170 1.8800 75 12
F N 2/2/09 11:20:56 1.8700 425 1.8800 162 12
F T 2/2/09 11:20:56 1.8700 458 1.8800 517 12
F N 2/2/09 11:20:56 1.8700 426 1.8800 162 12
F I 2/2/09 11:20:58 1.8700 1005 1.8800 555 12
F N 2/2/09 11:20:58 1.8700 426 1.8800 182 12
TAQ User Guide
The TAQ 3 User’s Guide NYSE (2006) provides a description of the operation of
the extraction software and the associated meanings for the possible values in each
field. However, it should be noted that the guide has not been updated since 30 Oc-
tober 2006 and hence a data user should examine the expanded list of field values
for both the quotes and trades files which can be found on http://www.nyxdata.com/Data-
Products/Daily-TAQ. For example, the range of trade venues participating in the
consolidated tape and the range of related exchange identifier codes in the TAQ
database has increased since 2006. The current listing is shown in Table 4.3.
As can be seen from the listing (shown in Table 4.3), the US equity market is
spread across a large number of trading venues. Recent years have seen a consol-
idation and a renaming of some of these trading venues as discussed below.
The main variants of the NYSE are the traditional NYSE floor (now a hybrid
of an auction and an automated market) and NYSE Arca (formerly known as the
Archipelago Exchange and the Pacific Exchange). NYSE Arca is a fully electronic
stock exchange. The American Stock exchange (historically known as the New
York Curb Exchange) was acquired by the NYSE in 2008 and is now known as
NYSE Amex Equities.
Nasdaq (formerly, the ‘National Association of Securities Dealers Automated
73
Table 4.3: Exchange Codes as at 31 January 2011
Code Exchange
A American Stock Exchange
B Boston Stock Exchange (now Nasdaq OMX BX)
C National (Cincinnati) Stock Exchange
D National Association of Securities Dealers (ADF)
E Market Independent (SIP - Generated)
I ISE (International Securities exchange)
M Chicago Stock Exchange
N New York Stock Exchange
P ARCA
T/Q NASDAQ Stock Exchange
S Consolidated Tape System
W CBOE
Z BATS
J DirectEdge A
K DirectEdge X
X NASDAQ OMX PSX
Y BATS Y-Exchange Inc.
Quotations’) is an electronic screen-based market. Like the NYSE, the Nasdaq

consists of a cluster of trade venues. For example, the Boston stock exchange was
formerly a regional stock exchange but was acquired by Nasdaq in 2007. It now
trades as Nasdaq OMX BX. Likewise, the Philadelphia stock exchange was also
acquired by Nasdaq in 2007, and trades as Nasdaq OMX PHLX. One interesting
recent development is the creation of a new Nasdaq market, the Nasdaq OMX
‘PSX’ platform (launched in October 2010). In most electronic markets, priority
for limit orders in the book is decided by price and then time of placement. In
other words, priority in the limit order book is given to orders with the best price
and then on a time basis (in a price tie, the oldest orders are executed first). In
contrast, Nasdaq OMX ‘PSX’ provides priority to limit orders on a price and
volume (displayed) basis (i.e. large limit orders are executed first). The object of
this is to attract large block trades to the exchange, away from dark pool trading
74
venues. Since its launch, the average trade size on PSX platform is 30% higher
than on the main Nasdaq market, and the new venue already accounts for about
1% of US equity trading (FinancialTimes, 2011b)
The National Association of Securities Dealers (now the Financial Industry
Regulatory Authority - FINRA) (ADF) is not technically an exchange as it does
not offer trade execution facilities. Instead it is an ‘alternative display facility’
which collects and disseminates quotes and trade information, for example from
some electronic communication networks (ECNs).4
The national stock exchange is an electronic stock exchange based in Chicago.
Prior to 2003 it was known as the Cincinnati stock exchange.
Table 4.4: US Daily Market Share

Equities for 2 February 2011 (Finan-
cialTimes, 2011a).
Market % Daily Turnover

OTC 27.6
Nasdaq 18.1
NYSE Arca 15.4
NYSE 13.8
BATS 12.3
Direct Edge 9.5
Boston 1.3
Other 2.0
Note in this table, OTC includes trades
not undertaken on regulated exchanges and
therefore includes dark pool venues and
electronic communication networks.
BATS is an electronic stock market, founded in 2005, trading approximately

10% of US equities. In 2010, BATS launched a second US equities exchange
(BATS Y-Exchange). Direct Edge is also an electronic stock market and was
4
ECNs are regulated by the SEC but are not recognised as stock exchanges.
75
granted SEC approval to become a stock exchange in 2010. Formerly it had been
an ECN.The International Securities Exchange (ISE) took a shareholding in Direct
Edge in 2008. The ISE formerly traded equities and options. The ISE stock
exchange (although not the ISE options exchange) was decommissioned in 2010.
A guide to the relative sizes of each exchange is provided in Table 4.4.
4.1.2 LSE ROB Database
Compared to the NYSE-Euronext TAQ database, which contains the best quotes
and the corresponding depths, the LSE Rebuild Order Book data has richer in-
formation. The LSE is the fourth largest exchange in the world and the largest
in Europe in terms of market capitalisation. It operates an electronic order book
platform, Stock Exchange Electronic Trading System (SETS). This trading system
has three sessions on each trading day:
• opening auction session, from 07:50 to 08:00, where buy and sell orders are
matched at 08:00,
• continuous trading session, from 08:00 to 16:30, where buy and sell orders
are continuously matched at price/time priority, and
• closing auction session, from 16:30 to 16:35, where buy and sell orders are
matched at 16:35.
The ROB data records order activities, which pooled together comprise the
quote prices and depths. It can be used to reconstruct the limit order book up
to any quote level. The data consists of every order and trade that occurred at
the LSE. It contains information on order submissions, order cancelations, order
modifications and executions. Every trade is linked to the corresponding orders
through a trade code. Therefore, if a large order is executed against many small
76
orders resulting in many small trades, one can trace back to each small order using
corresponding trade code. With this information one is able to reconstruct the
limit order book for each millisecond of the trading day. The LSE provides three
separate files to record trade and order information for each trading day. These
files are:
• ‘order details’ file, which contains details of every new persistent order5
entering the electronic order book,
• ‘order history’ file, which contains information on changes to each persis-

tent order, including order partial/full matching, order deletion, order expi-
ration and order modification, and
• ‘trade report’ file, which contains details of every trade execution.
4.1.3 Sample Data
A sample of the Order Detail file is shown in Table 4.5. As shown in the table,
each order has an order code which is unique to the order. The stock code in the
second column of the table corresponding to a stock listed in the LSE. In the third
column, “B” stands for a buy order and “S” represents a sell order. The next four
columns show the price, visible size, submission date and time for each submitted
order. The last column is the message sequence number which is unique for each
event (e.g. limit order submission, order cancellation, order matching, and order
modification) occurring in the market. The message sequence number is usefull to
distinguish the time priorities of two events which occur at the same time (accurate
to one millisecond). It also appears in the Order History file and the Trade Report
file.
5
Persistent orders are the orders which are stored on the electronic order book after they are
submitted to the market.
77
Table 4.5: A Sample of Order-Detail Data
OrderCode StockCode BuySellInd Price AggregateSize Date Time MessageSequenceNumber

B101ZLMIUO 3989 B 620.00 1000 30072010 09:01:47.661 568016
78
B1021HZOGE 8210 S 11.50 5000 20082010 10:02:13.942 1366931
B1021PXYRW 3990 S 0.05 50000 24082010 07:55:43.761 5519
B1022A4QAL 1490 S 110.00 2640 31082010 07:50:53.980 6914
B1022AL8NY 3989 S 721.00 654 31082010 08:43:28.465 445616
A sample of the Order History file is shown in Tables 4.6 & 4.7. As shown
in Table 4.6, there is an order action type attached to each entry. It records the
action made to the order submitted previously and still listed in the order book.
There are six order action types, which are explained in Table 4.9. After actions
“D” or “E”, the order will be removed from the order book. Partial match means
that part of the order is executed, and the unexecuted part is still left in the order
book. Full match means that the order is fully executed and thus its aggregate size
is zero as shown in the second column of Table 4.7. If it is either partial matched
or fully matched, the counterparty’s code is recorded as shown in the third column
of Table 4.6.6 Action “T” rarely appears in the database.
A sample of the Trade Report file is shown in Table 4.8. As shown in Table
4.6, each order matching entry has a unique trade code.7 This trade code corre-
sponds to an entry in the Trade Report file. As shown in Table 4.8, each entry
records details of each transaction, including the trading price, size, date and time
(accurate to one second).
As shown in Tables 4.5, 4.6, 4.7 and 4.8, the information contained in these
three files are connected by order codes and trade codes. The order code maps
the new persistent orders listed on ‘order details’ file and their order trajectories
(full execution, partial execution, deletion, expiration and modification) recorded
on ‘order history’ file. Similarly, the trade code is used to track the trading de-
tails contained on the ‘trade report’ file for each order matching record on ‘order
history’ file.
6
The counterparty of a buy order is a sell order, and vice versa.
7
A limit order executed against two or more market orders will have two different trade codes.
79
Table 4.6: A Sample of Order-History Data (Part I)
OrderCode OrderActionType MatchingOrderCode TradeSize TradeCode

ZO00UO59SD P ZO00UO5CNV 441 ZF001UD7AU
80
ZO00UL7QO1 M ZO00UL8QKK 1407 ZF001UCFA1
ZO00ULIX1H M ZO00ULIXCB 1000 ZF001UCHP0
ZO00UNKBZE P ZO00UNKCD4 1331 ZF001UCZVZ
B700NPHAYC P B700NPO2CH 3800 A7001XEFZ1
Table 4.7: A Sample of Order-History Data (Part II)
StockCode AggregateSize BuySellInd MessageSequenceNumber Date Time

4154 759 B 2694979 04012010 08:00:20.087
81
4154 0 S 240963 04012010 13:50:37.040
4154 0 S 466808 04012010 16:19:48.828
4154 1089 B 2130431 04012010 15:01:17.559
9460 3820 S 463254 04012010 14:51:13.631
Table 4.8: A Sample of Trade-Report Data
MessageSequenceNumber StockCode TradeCode TradePrice TradeSize TradeDate TradeTime

5272 9056 A2005L4TG0 10.32 492 01092010 10:36:38
82
644861 2570 A90022YXOG 146.58 7291 01092010 09:02:02
913100 9454 A700223A2U 43.94 5500 01092010 09:08:15
1281027 1437 ZA0022NJXM 2550.00 2 01092010 09:39:08
99 5657 A90022YQPO 1.67 488 31082010 17:45:33
Table 4.9: Order Action Type (ROB)
Order Action Type Description

D Deletion
E Expiry
P Partial Match
M Full Match
T Transaction Limit ( 150 trades)
Z Modification (Size decrease)
4.2 Data Processing

The course of data preprocessing is presented in this section.
4.2.1 Preprocessing of TAQ data
This section initially introduces the issues in data preprocessing and then presents
the steps used to preprocess the TAQ data.
Issues in Data Preprocessing
Although the data contained in the TAQ database is a valuable resource for re-
search purposes, it is supplied in a raw (unprocessed) format and therefore re-
quires pre-processing in order to prepare it for analysis. Figure 4.3 shows the un-
filtered time series for CHDX. It clearly illustrates that there are often extremely
abnormal prices in the non-NASDAQ quote time series (green line). Thus it is
very necessary to filter the raw time series before using it for analysis. The clean-
ing of high-frequency financial data have been given special attention in many pre-
vious studies, such as Lee & Ready (1991) Bessembinder (2003) Vergote (2005)
and Henker & Wang (2006). Issues which can arise in raw high-frequency data
include:
Quote data only
83
18
16
14
12
10
price
0
9:30 10:00 10:30 11:00
time
Figure 4.3: Unfiltered Time Series for CHDX at 02/02/09 from 9:30-11:00 am.
The blue line depicts quotes occurring at NASDAQ; the green line depicts quotes
occurring at all exchanges; the red crosses represent the transaction prices.
84
Q1. Quotes which are generated by non normal market activity
Trade data only
T1. Trades which are subsequently modified or cancelled
T2. Trades of which the report time is delayed
All data
A1. Ticks occurring outside normal trading period
A2. Ticks with non-positive prices, volumes or bid-ask spreads
A3. Ticks of which the prices appear inconsistent with their neighbouring prices
(i.e. ‘outliers’)
Typically it is relatively easy to identify and correct for items Q1, T1-2 and
A1-2 as ‘flags’ in the database and the relevant ‘timestamp’ can be used to help
filter these items. Items falling into category A3 are trickier to handle as a series
of modeller decisions are required in order to classify a data point as an ‘error’
which therefore should be removed in data preprocessing. In this process, the data
analyst is attempting to balance out the risks of under scrubbing vs over scrubbing
the data. Under scrubbing the data risks the embedding of noise in the data being
analysed whereas over scrubbing the data filters out potentially vital information.
The process must consider issues such as the change in price between one trade
record and the next simultaneously with the level of trading activity. A trade which
is, for example, $0.10 off the previous recorded trade price would be unusual for
a highly-liquid stock which trades many times a second, but may be less suspect
for a thinly-traded security which only trades every 10 minutes.
Depending on the research questions of interest, and the associated required
analyses (for example, trade classification, calculating effective spreads, and es-
timating the information context of trades (Henker & Wang, 2006)), it may also
85
be necessary to align the trades and quotes data chronologically into a single time
series. This task is not as simple as it appears at first glance as there is no field
in the separate trade / quote files which make up the TAQ database which links
trades with their corresponding quote update. The problem is made more complex
due to the potentially differing levels of latency in data from differing sources in
reaching the consolidated tape.
The degree of latency in the recording of both quotes and trades has varied
over the years and hence, the most appropriate procedure to adopt in aligning
these time series has also varied with obvious implications for research study de-
sign, depending on the dates from which the data is drawn. Historically, there has
been greater latency in the recording of quote updates than trade updates on the
consolidated tape, and consequently, lining up the time series based on time stamp
may produce an unintended outcome where a trade is recorded before the associ-
ated change in the best quote for that exchange (obviously, any trade will impact
on the current best quote for an exchange as it will usually alter some element
of best bid/ask price or volume at those prices). For example, under the manual
system operated in the NYSE in the 1980s, if a specialist assistant was faster in
recording a quote revision than the floor reporter was in recording a trade, the
corresponding quote update could be recorded before the trade that triggered it.
As a result of this problem, Lee & Ready (1991) suggested, based on NYSE data
drawn from 1988, that the prevailing quote five seconds before a trade should be
considered to be the actual quote at the time of the trade. This rule was used in
many academic studies up to and including the early 2000s.
Of course, with the increasing use of technology there have been many struc-
tural changes in the way that trade and quote information has been captured over
the past 20 years. With the introduction of ‘auto-quoting’ for all stocks in 2003
on the NYSE, the best bid/offer was automatically updated whenever a limit order
86
was posted to the Display Book at a better price than the previous best bid or of-
fer, and trades taking place at the best bid/offer automatically led to an updating of
the best bid/offer (Vergote, 2005). A study after the introduction of the autoquote
system, using data drawn from October-December 2003, indicated that a trade -
quote change lag still existed, although it had been reduced to about 2 seconds
(Vergote, 2005), with another study (Henker & Wang, 2006) (which used earlier
data) suggesting that the appropriate lag was one second.8
The increasing use of ‘unlit’ trading venues such as dark pools has resulted in
an growing portion of trades taking place ‘off exchange’. While these trades are
recorded on the CTS, there will not be a corresponding ‘quote change’ recorded on
the CQS. Another issue that should be noted is that the lag from a trade execution
to the recording of that trade on the CTS is permitted to range up to 90 seconds
and hence, it cannot be assumed that actual trade time and the reporting time of
a trade are identical (Hasbrouck et al in a 1993 study (Hasbrouck et al., 1993)
reported a median delay of 14 seconds for NYSE trades between the SuperDot
execution report time and the CTS print time).
Actual Data Preprocessing
Only the quotes and trades ticks which happen at their main listed exchange are in-
cluded. So the ticks happened outside NYSE for DAR, F, MDS and XOM, and the
ticks happened outside NASDAQ for CHDX and GOOG are excluded. For NYSE
traded stocks, it is easy to obtain their records by choosing ’N-NYSE’ when using
the TAQ3 extraction tool as shown in Figure 4.2. On obtaining NASDAQ traded
ticks, this is a little more complicated. As shown in Table 4.3, there are two ex-
8
While increased use of electronic trading would be expected to enhance the quality and time-
liness of trading data capture, this is not always so. A recent case in point was the introduction of a
new matching engine on the London Stock Exchange (14 February 2011) where due to computer
glitches, price data vendors were displaying incorrect prices, blank prices and incorrect trading
volumes (King, 2011).
87
change codes corresponding to NASDAQ, which are ’T’ and ’Q’. But as shown
in Figure 4.2, only ’T-NASDAQ’ can be chosen when using the TAQ3 extraction
tool, where ’Q-NASDAQ’ is not listed. In order to obtain all ticks which happen
at NASDAQ, analyst needs to select all exchanges when using the TAQ3 extrac-
tion tool. By doing this, all ticks from NASDAQ can be included in the extracted
file labeled with ’Q’ or ’T’ in the ’Exchange’ column.
In preprocessing the raw data this study follows the lead of Brownlees & Gallo
(2006). For both quote and trade data, initially all ticks recorded outside the nor-
mal trading day are removed. Following Brownlees & Gallo (2006), an expanded
trading period spanning between 9:30 AM and 4:05 PM is adopted, in order to in-
sure that closing prices possibly recorded with a delay are accounted for. For trade
data, all incorrect and delayed trades are initially discarded, which are indicated
by the CORR field differing from 0 and the COND field equaling ‘Z’ respectively
in the trade file. Typically, this removes only a small amount of data from the
initial dataset (usually less than < 0.5%). On quote data, all records generated by
non normal activities are deleted, as indicated by the MODE field values 1,2,3,6
or 18 in the quote file. Then all quote and trade ticks with non-positive prices or
volumes and all quotes with non-positive bid-ask spreads are deleted.
Removing Outliers
To remove outliers, observations are omitted when the absolute difference be-
tween the current price and the average neighbour price is outside three standard
deviations plus a parameter that controls for the minimum price variation. To il-
N
lustrate, let pi be a value of the price time series Pi=1 . The proposed method in
Brownlees & Gallo (2006) to remove outliers is:
|pi − pi (k, δ)| < 3si (k, δ) + λ (1)
88
Here pi (k) and si (k) denote respectively the δ-trimmed sample mean and sample
standard deviation of a neighborhood of k observations around i and λ is a granu-
larity parameter9 . The k neighbor observations are chosen based on the following
rules: the neighborhood of an observation in the middle of the day are the first
preceding k/2 observations and the following k/2 ones; the neighborhood of one
of the first k/2 observations for the trading day are the first k observations of the
same trading day; the neighborhood of one of the last k/2 observations for the
trading day are the last k observations of the same trading day. For any given ob-
servation price pi , if the condition above is true, this observation is kept, otherwise
discarded.
A very heuristic procedure is introduced in Brownlees & Gallo (2006) for
choosing these three parameters. The parameter k is chosen based on the level of
trading intensity, the more active the trading, the larger the k. The parameter δ
is chosen on the basis of the frequency of outliers, the higher the frequency, the
higher the δ. The choice of the granularity parameter λ is guided by the frequen-
cies of the price changes which is always multiples of the minimum price variance
for that stock.
0.9 0.7 0.5
0.8 0.45
0.6
0.7 0.4
0.5
0.35
0.6
0.3
0.4
Percentage
Percentage
0.5
Percentage
0.25
0.4
0.3
0.2
0.3
0.2 0.15
0.2
0.1
0.1
0.1 0.05
0 0 0
<=−6 −5 −4 −3 −2 −1 0 1 2 3 4 5 >=6 <=−6 −5 −4 −3 −2 −1 0 1 2 3 4 5 >=6 <=−6 −5 −4 −3 −2 −1 0 1 2 3 4 5 >=6
Bid Price Change in Ticks Ask Price Change in Ticks Trade Price Change in Ticks
(a) (b) (c)
Figure 4.4: Frequencies of Price Changes for CHDX. (a) frequency of bid price
change (b) frequency of ask price change (c) frequency of transaction price
change.
9
The parameter λ is used to avoid zero variances from the time series of k equal prices around
the observation price pi .
89
0.8 0.7 0.7
0.7
0.6 0.6
0.6
0.5 0.5
0.5
0.4 0.4
Percentage
Percentage
Percentage
0.4
0.3 0.3
0.3
0.2 0.2
0.2
0.1 0.1
0.1
0 0 0
<=−6 −5 −4 −3 −2 −1 0 1 2 3 4 5 >=6 <=−6 −5 −4 −3 −2 −1 0 1 2 3 4 5 >=6 <=−6 −5 −4 −3 −2 −1 0 1 2 3 4 5 >=6
Trade Price Change in Ticks Trade Price Change in Ticks Trade Price Change in Ticks
(a) F (b) DAR (c) XOM

0.5 0.35 0.35
0.45
0.3 0.3
0.4
0.25 0.25
0.35
0.3
0.2 0.2
Percentage
Percentage
Percentage
0.25
0.15 0.15
0.2
0.15
0.1 0.1
0.1
0.05 0.05
0.05
0 0 0
<=−6 −5 −4 −3 −2 −1 0 1 2 3 4 5 >=6 <=−6 −5 −4 −3 −2 −1 0 1 2 3 4 5 >=6 <=−6 −5 −4 −3 −2 −1 0 1 2 3 4 5 >=6
Trade Price Change in Ticks Trade Price Change in Ticks Trade Price Change in Ticks
(d) CHDX (e) MDS (f) GOOG
Figure 4.5: Frequencies of Transaction Price Changes for Six Stocks
Figure 4.4 illustrates the frequencies of bid, ask and transaction price changes
for CHDX, which guides the choice of parameter λ. Figure 4.5 depicts the fre-
quencies of transaction price changes for the six stock. As shown in Figure 4.5(f),
there is a significant fraction of price changes equal or larger than six minimum
ticks.10 In contrast, there is almost no price change larger than two ticks for F as
shown in Figure 4.5(a). The distributions of price changes for the other four stocks
are between these two extremes. It is possible that more outliers are identified in
the price time series with higher dispersion level of price change. The parameter
δ is fixed at 10%, the same as Brownlees & Gallo (2006). Several sets of parame-
ters (k, λ) are tested. The quality of the cleaning is judged by a visual inspection
of the clean tick-by-tick price series graph. Figure 4.6 depicts the time series on
03 Feb 2009 from 9:30 to 11:00 for CHDX. The plot in Figure 4.6(a) shows the
time series of raw data. The time series in Figures 4.6(b), 4.6(c) & 4.6(d) demon-
strate the clean data with three different sets of parameters (k, λ). 4.6(d) provide
10
One minimum tick equals 0.01.
90
6.9 6.9
6.8 6.8
6.7 6.7
6.6 6.6
price
price
6.5 6.5
6.4 6.4
6.3 6.3
6.2 6.2
6.1 6.1
9:30 10:00 10:30 11:00 9:30 10:00 10:30 11:00
time time
(a) raw data (b) clean data with kq = 40 λq = 0.06 for

quote data, and kt = 10 λt = 0.08 for trade
data
6.9 6.9
6.8 6.8
6.7 6.7
6.6 6.6
price
price
6.5 6.5
6.4 6.4
6.3 6.3
6.2 6.2
6.1 6.1
9:30 10:00 10:30 11:00 9:30 10:00 10:30 11:00
time time
(c) clean data with kq = 80 λq = 0.02 for (d) clean data with kq = 40 λq = 0.02 for
quote data, and kt = 10 λt = 0.04 for trade quote data, and kt = 10 λt = 0.04 for trade
data data
Figure 4.6: Tick-by-tick Price Series on 03 Feb 2009 (9:30-11:00) for CHDX.
The blue line (below) represents the bid price time series; the green line (above)
represents the ask price time series; the red circles represent the transaction prices.
91
the most satisfactory result. The sets of most satisfactory parameters (k, λ) are
reported in Table 4.10.
Table 4.10: Parameters for Removing Outliers
Quotes Data Trades Data

δ k λ δ k λ
MDS 0.1 30 0.02 0.1 10 0.06
CHDX 0.1 40 0.02 0.1 10 0.04
DAR 0.1 50 0.02 0.1 20 0.03
F 0.1 60 0.01 0.1 40 0.02
GOOG 0.1 120 0.01 0.1 60 0.02
XOM 0.1 120 0.01 0.1 60 0.02
There are three steps to clean the raw quote data: the first step is to remove
all quotes which are generated by non normal market activities; the second step
is to remove all quotes with non-positive prices, volumes or bid-ask spreads; the
third step is to remove the outliers. Similarly, there are three steps to clean the raw
trade data: the first step is to remove all trade ticks which are modified, cancelled
or delayed; the second step is to remove all trades with non-positive prices or
volumes; the third step is to remove the outliers.
Tables 4.11 & 4.12 summarize the results of cleaning raw quote and trade data
respectively. The values in bold represent of the percentage of the final filtered
data in the raw data. Most of these percentages are larger than 99%, indicating
that the filtering process generally does not alter the raw data. From the tables,
one can also find that non-positive values did not appear in either raw quote data
or trade data, in contrast that some zero values were found in Barndorff-Nielsen
et al. (2009). This indicates that the consolidated tape system has been improved
and zero values are avoided.11 Among the three steps, the last step removes a
11
In Barndorff-Nielsen et al. (2009), the period of analysed data is from January 3 to June 29,
2007. The time of the data analysed here is February 2009.
92
Table 4.11: Cleaning Quote Data.
MDS CHDX DAR F GOOG XOM

No. Perc. No. Perc. No. Perc. No. Perc. No. Perc. No. Perc.
Raw Data 34891 1.0000 47365 1.0000 160759 1.0000 508852 1.0000 1786150 1.0000 1951197 1.0000
93
Non-normal Quotes 25 0.0007 0 0.0000 25 0.0002 45 0.0001 0 0.0000 42 0.0000
Non-positive Values 0 0.0000 0 0.0000 0 0.0000 0 0.0000 0 0.0000 0 0.0000
Outliers 711 0.0204 452 0.0095 59 0.0004 99 0.0002 42254 0.0237 13832 0.0071
Final Sample 34155 0.9789 46913 0.9905 160675 0.9995 508708 0.9997 1743896 0.9763 1937323 0.9929
Note: No. means the number of observations.

Perc. means the percentage of the number of observations in the raw data.
Table 4.12: Cleaning Trade Data.
MDS CHDX DAR F GOOG XOM

No. Perc. No. Perc. No. Perc. No. Perc. No. Perc. No. Perc.
Raw Data 1339 1.0000 3128 1.0000 14588 1.0000 54874 1.0000 443140 1.0000 511462 1.0000
94
Incorrect Trades 0 0.0000 0 0.0000 0 0.0000 0 0.0000 0 0.0000 0 0.0000
Non-positive Values 0 0.0000 0 0.0000 0 0.0000 0 0.0000 0 0.0000 0 0.0000
Outliers 5 0.0037 8 0.0026 16 0.0011 1 0.0000 4063 0.0092 796 0.0016
Final Sample 1334 0.9963 3120 0.9974 14572 0.9989 54873 1.0000 439077 0.9908 510666 0.9984
Note: No. means the number of observations.

Perc. means the percentage of the number of observations in the raw data.
larger fraction of the observations than the other two steps. Among the six stocks,
F removes a smaller fraction of the observations while GOOG removes a larger
fraction of the observations. It can be found that there is a positive relationship
between the number of removed observations and the dispersion level of price
changes. The more dispersive the price change, the more the observations identi-
fied as outliers. Figures 4.7 & 4.8 illustrate the raw data and the filtered data for
CHDX on 10 Feb. 2009 respectively.
95
1000
900
800
700
600
price
500
400
300
200
100
0
9:30 10:00 10:30 11:00 11:30 12:00 12:30 13:00 13:30 14:00 14:30 15:00 15:30 16:00
time
Figure 4.7: Raw Data for CHDX on Feb 10, 2009
96
7
6.8
6.6
6.4
price
6.2
5.8
5.6
5.4
9:30 10:00 10:30 11:00 11:30 12:00 12:30 13:00 13:30 14:00 14:30 15:00 15:30 16:00
time
Figure 4.8: Clean Data for CHDX on Feb 10, 2009
97
Matching Trade and Quote
To match trade ticks with quote ticks, this study follows the one-second rule in
Henker & Wang (2006). Any quote less than one second prior to the trade is
ignored and the first one at least one second prior to the trade is retained as the
prevailing quote. If the transaction occurs above the prevailing mid-quote, it is
regarded as a buyer-initiated trade; otherwise if the transaction occurs below the
prevailing mid-quote, it is a seller-initiated trade. If the transaction occurs exactly
at the mid-quote, it is signed using the previous transaction price, which is called
’tick test’ in Lee & Ready (1991). If the transaction price is higher (lower) than
the price of the previous trade, it is classified as a buyer(seller)-initiated trade.
If the transaction price is the same as the price of the previous trade, it is signed
using the previous transaction price change. If the previous price change is upward
(downward), the trade is a buy (sell). Sometimes, the trade does not fall into any
cases mentioned above, for example the trade is the first trade of the trading day.
An alternative test called ’reverse tick test’ in Lee & Ready (1991) can be applied
which compares the current trade price with the trade prices immediately after the
current trade. If the following trade price is higher (lower) than the current trade
price, the trade is classified as a sell (buy).
4.2.2 Preprocessing of ROB data
The data contained in the ROB database is supplied in a raw (unprocessed) format
and therefore requires preprocessing in order to prepare it for analysis. Unlike the
raw data in the TAQ database, the ROB data rarely contains non-positive values,
but the ROB data is out of time sequence and more complicated than the TAQ
data. The information needed to analyse the order book is not direct to obtain and
the some preprocessing steps are needed. For example, market order information
98
can not be directly found in the data files provided and needs to be inferred using
the available data information.
Figure 4.9: ROB Data Preprocessing Chart
Figure 4.9 illustrates how to preprocess the ROB data in order to extract the
order book information for the 1st of Aril in 2010 which is denoted by “401”.
There are three steps explained as follows.
1. The first step is to extract the data for each trading day from the three original
files (Order-Detail, Order-History and Trade-Report), because each original
file contains information for numerous (about 11) continuous days. As shown
in Figure 4.9, the information for 1 Oct. 2010 is then written to 401 Order-
Detail, 401 Order-History and 401 Trade-Report.
2. The second step is to infer the information on undisclosed orders, and ex-
tracting the information on limit orders with various levels of aggressiveness
99
(which will be used in the study in Chapter 6). Detail of this step is explained
in the next section.
3. The third step is to reconstruct the order book using the information extracted
in the last step, and record the information of the order book up to any quote
level. As shown in Figure 4.9, the order book information is recorded in
401 OrderBook. Detail of this step is described below.
Inferring Undisclosed Orders
In order to get the full information needed for the experiments in Chapters 5 &
6, non-persistent market orders and three missing events from the given data files
have to be inferred. Market orders can be inferred from the ‘order history’ file as
it records information on the matching side of each transaction. In this study, the
reconstructing process is implemented in JAVA.
The first missing event in ROB data is the execution of iceberg limit order.
LSE allows traders to place iceberg limit orders, part of which are hidden in the
order book and not recorded in the ROB data. When the visible part of the iceberg
limit order is matched by a market order with larger size, the hidden part will be
executed against the rest of the market order. The traded hidden part of the iceberg
limit order can be inferred from the records of the limit order whose transaction
size is larger than its original size.
The second set of missing events are crossing limit orders which cause imme-
diate trade after submission. The ‘order details’ file only records the unexecuted
part of the crossing limit orders. The traded part can be found from the ‘order
history’ file in that each crossing limit order is matched by one or more persistent
limit orders previously submitted to the market.
The third missing event is the limit order which was submitted to the market
a few days ago but is executed today. The information on details of these orders
100
needs to be recovered from older data files.
After this step, seven data files are obtained as shown in Figure 4.9. Four
types of limit orders with various aggressiveness are distinguished for the use in
Chapter 6, which are crossing limit order, in-spread limit order, spread limit order,
and off-spread limit order. Details of these order types are explained in Chapter 6.
Reconstructing The Limit Order Book
In order to get information on the limit order book up to any quote level, the order
book on each trading day needs to be reconstructed using all the available data,
which is already acquired in the second step.
At first, the opening time of the normal trading session needs to be inferred.
This time is a random point between 8:00:00 a.m. and 8:00:30 a.m. At this
point, all the orders submitted during the opening sessions are executed at the
price which can maxmise the trading volume. Therefore, the opening time is the
time point at which the number of trades is largest in the period from 8:00:00
a.m. to 8:00:30 a.m. Then all the orders occurring during the same trading day
are sorted in time sequence. For those events occurring at the same time, they are
distinguished using their sequence numbers. All the rules as in the real market are
applied for reconstructing the limit order book, like the price & time priorities. In
this study, the reconstructing process is implemented in Matlab.
4.3 Summary
The objective of this chapter was to provide an introduction to two popular high-
frequency databases, and using this data, explore some of the issues faced by
researchers seeking to use ultra-high frequency financial databases. With the pre-
processed data, three studies on price impact are conducted in the following chap-
101
ters.
102
Chapter 5
An Analysis of Intraday Behaviours

of Price Impact and Market
Liquidity
Much of the content of this chapter has been in:


Intraday phenomena in financial markets have been extensively studied in the

literature on market microstructure as shown in Section 2.4 of Chapter 2. How-
ever, many of these studies are dated. For example, the most recent studies of
103
intraday phenomena in the US markets and the UK markets are analysed with the
1999 data (Henker & Wang, 2006) and 2001 data (Cai et al., 2004) respectively.
Up-to-date evidence is useful for constructing market microstructure theories and
is crucial for developing intraday trading strategies. It can also help regulators to
design an appropriate measure of liquidity for an efficient and transparent trading
system.
There are two objectives of this chapter. The first objective is to examine
the intraday behaviour of market liquidity. In this chapter, intraday behaviours
of price volatility, bid-ask spread, trading volume, trading frequency, trade size,
market depth (cover the best level of the order book) and near market depth (covers
the ten best levels of the order book) are investigated.
The second objective of this chapter is to investigate the intraday behaviour
of price impact. Several studies have examined the intraday behaviour of price
impact. Dufour & Engle (2000); Engle & Patton (2004); Hasbrouck (1999) find
that price impact measured as changes of mid-quote, bid and ask prices caused by
trades does not have an intraday pattern on the NYSE. In contrast, Chan (2000)
shows that price impact displays a U-shaped pattern over the trading day on the
HKSE. However, no previous paper examines the intraday behaviour of price im-
pact on the LSE or the NASDAQ. Therefore, it is interesting to know whether the
price impact exhibits an intraday pattern in these markets.
In this chapter, the analysed data is drawn from two databases, the NYSE-
Euronext TAQ (trade and quote) database and the LSE ROB (rebuild order book)
database, and covers three markets, the NYSE, the NASDAQ and the SETS of
LSE. This chapter contributes to the literature in the following ways. It documents
up-to-date evidences on the intraday behaviours of a number of interesting vari-
ables including the price volatility, trading size, trading volume, bid-ask spread,
market depths and price impact, in the UK market as well as the US markets,
104
which complements the literature on intraday phenomena.
The chapter proceeds as follows. Section 5.1 introduces various measures of
market liquidity which are examined in this chapter. Section 5.2 presents empiri-
cal results and discussions. A summary of this chapter is given in Section 5.3.
5.1 Measures of Market Liquidity

There is no single all-encompassing measure of market liquidity as liquidity is
a multi-dimensional variable. In this chapter, a number of interesting metrics
are used to investigate the intraday pattern of market liquidity, including bid-ask
spread, market depth, and price impact ratio. Each trading day is partitioned into
successive 15-minutes intervals, and the measure these metrics for each stock dur-
ing each of the 15-minute intervals on each trading day, which are then averaged
over the whole sample period (19 trading days for US stocks and 127 trading days
for UK stocks).1 The methods used to estimate these metrics are described as
follows.
5.1.1 Price-based Measures
Following Abhyankar et al. (1997) and Cai et al. (2004), absolute mid-quote re-
turns2 are used as the return volatility measure. The mid-quote return is calculated
as
Rt = [log(Pt ) − log(Pt−1 )]
1
This thesis chooses a reasonable time length of the data (19 trading days for US stocks and
127 trading days for UK stocks), which is enough to generate intraday phenomena. During these
time periods, there were no disturbances in the markets.
2
The use of mid-quote prices rather than transaction prices should reduce spurious volatility
due to the bid-ask bounce.
105
where P is the mid-quote price at the end of each minute. At first the price returns
for each 1-minute interval is calculated and then the average absolute return is
used as the return volatility at each 15-minute interval.
The bid-ask spread is the difference between the prevailing best ask price
(ASK) and the prevailing best bid (BID). In this chapter the proportional bid-
ask spread (P roSpr)3 is used to measure the spread. It is defined as the ratio of
the prevailing bid-ask spread and the prevailing mid-quote price (0.5 ∗ (ASK +
BID)):
ASK − BID
P roSpr = .
0.5 ∗ (ASK + BID)
The 15-minute proportional quoted spread is calculated as the average quoted
spread during the period.
5.1.2 Volume-based Measures
Previous work typically only measures the depth at the best available prices (best
bid & ask price). However, the quotes and volume beyond the spread also have
effects on the market liquidity. Cao et al. (2008) found that the liquidity beyond
the spread is informative about future short-term returns. Hopman (2007) has
shown that order imbalance inside the order book is an important determinant of
simultaneous price changes. Farmer et al. (2004) found that large price impacts
caused by market order executions are not always associated with large trading
sizes, but are due to gaps at different levels of the order book. In this chapter the
near market depth is also investigated which accounts for the 10 highest levels of
the order book. The market depth (Dep) is measured as the average value of all
shares at the best bid price and all shares at the best ask price:
3
For the purpose of convenient comparison of different stocks.
106
SharesBestBid + SharesBestAsk
DepShare = .
2
The near market depth (N Dep) is measured as the average value of all shares at
the best 10 bid prices and all shares at the best 10 ask prices:
∑10 ∑10
i=1 Sharesbid,i + i=1 Sharesask,i
N DepShare =
2
where i is the level of the order book. The depth in value is also measured. The
depth in value is measured as
BID ∗ SharesBestBid + ASK ∗ SharesBestAsk

DepV alue =
2
The near depth in value is measured as
∑10 ∑10
i=1 BIDi ∗ Sharesbid,i + i=1 ASKi ∗ Sharesask,i
N DepV alue = .
2
The depth is calculated for each time recorded in the data. It is averaged over each
15-minute interval and over the whole sample period.
The 15-minute trading size is the average trading size at this period and it is
then averaged over the whole sample period. The trading size in value which is
trading size timed by its corresponding trading price is also measured.
5.1.3 Price and Volume Mixed Measures
More recently, with the increased availability of ultra high-frequency data from
stock exchanges and other alternative trading platforms, several papers have looked
at measures of liquidity other than bid-ask spread, including Sandas (2001); Ranaldo
(2004); Hall & Hautsch (2007); Large (2007).
107
In this study, price impact ratio is proposed to measure liquidity cost. Price
impact is the mid-quote price change caused by a trading execution. The price im-
pact ratio (P iRatio) is measured as the ratio of price impact and its corresponding
trading size:
P riceaf ter − P ricebef ore

P iRatio =
T radeSize
where P rice is the mid-quote price, P riceaf ter is the market price immediately
after this trading execution and P ricebef ore is the market price just before the
trading execution. Price impact ratio measures the liquidity cost of each traded
share.
5.2 Empirical Results and Discussions

The aim of this experiment is to investigate whether price impact exhibits an intra-
day pattern and whether this intraday pattern is consistent across different stocks
in US and UK markets. Six firms from the TAQ database are selected for analysis
(Table 5.1). The firms selected are of varying market capitalisation and industry
sector. Table 5.2 reports a number of data descriptives concerning these com-
panies drawn from 2nd February 2009. All of the companies are listed on US
markets.
For the ROB database, the six most actively traded stocks, whose market cap-
italisations exceed 20 billion pounds are selected for analysis. The companies are
Barclays, British American Tobacco, BP, Glaxo Smith Kline, HSBC Holdings and
Vodafone Group. They are from five different industries: consumer goods, basic
materials, healthcare, financial and technology sector (shown in Table 5.3). Also
all of them are British multinational companies headquartered in London, United
Kingdom and primarily listed in the London Stock Exchange.
108
Table 5.1: Companies Listed in the US markets. Market capitalisation and share
prices as at 01/04/2011
Name Ticker Primary Mkt. Mkt. Cap. Share Price

Exxon XOM NYSE $419.55 Bn $84.79
Google GOOG Nasdaq $191.22 Bn $589.01
Ford F NYSE $57.02 Bn $15.03
Darling Int’l DAR NYSE $1.85 Bn $15.48
Chindex Int’l CHDX Nasdaq $0.25 Bn $16.05
Midas Inc. MDS NYSE $0.11 Bn $7.76
Table 5.2: Trading Activity of US Stocks (02/02/2009).
Name Ticker Consol. Vol Consol. $ Vol Consol. Trades Av. Trade Size
Exon XOM 36,128,159 $2,761.699m 183,030 197.39
Google GOOG 5,271,804 $1,789.620m 38,653 136.38
Ford F 32,404,587 $60.672m 26,246 1,234.65
Darling Int’l DAR 1,091,619 $4.817m 5,514 197.97
Chindex Int’l CHDX 39,739 $244,389 264 150.53
Midas Inc. MDS 39,200 $376,421 284 138.02
Note: Consol. Vol is short for Consolidated Volume.

Consol. $ Vol is short for Consolidated Dollar Volume.
Consol. Trades is short for Consolidated Trades.
Av. Trade Size is short for Average Trade Size.
Table 5.3: Information of Companies Listed in the LSE (28/09/2012)
Market
Ticker Full Name Sector
Capitalisation
BARC Barclays PLC Financial 26.42b
BATS British American Tobacco PLC Consumer Goods 62.29b
BP BP PLC Basic Materials 83.47b
GSK Glaxo Smith Kline PLC Healthcare 70.19b
HSBC HSBC Holdings PLC Financial 93.16b
VOD Vodafone Group PLC Technology 87.23b
109
The next section provides summaries of descriptive statistics for both datasets.
Then the empirical results on the intraday patterns are presented graphically in
Figures 5.2-5.25, and are discussed in Sections 5.3.2-5.3.5.
5.2.1 Summary Statistics
Summary statistics for the companies are provided. Analysed variables include:
• Mid-quote price, which is the average price over the sample time period;
• Number of trades per day, which is the average number of trades over the
sample time period;
• Trade size, which is the average number of shares per trade for all the trades
in the sample data;
• Buyer-initiated trade size, which is the average number of shares per trade
for all the buyer-initiated trades in the sample data;
• Seller-initiated trade size, which is the average number of shares per trade
for all the seller-initiated trades in the sample data;
• Bid depth, which is the average number of shares available at the best bid
price for the time period in the sample data;
• Ask depth, which is the average number of shares available at the best ask
price for the time period in the sample data;
• Buy/sell ratio, which is the average daily ratio of daily number of buyer-
initiated trades (DailyN umberbuys ) to daily number of seller-initiated trades
(DailyN umbersells )
110
DailyN umberbuys
daily buy/sell ratio = ;
DailyN umbersells
• Probability of reversal, which is the average daily probability of sign rever-

sal in consecutive trades, for instance a buy trade is followed by a sell trade,
and is calculated as
DailyN umBS + DailyN umSB

daily probability of reversal =
DailyN umtrades − 1
where DailyN umtrades is the daily number of trades, DailyN umBS is the
daily number of reversal from a buy to a sell, and DailyN umSB is the daily
number of reversal from a sell to a buy.
Summary Statistics for US Stocks
A summary of descriptive statistics of the six US stocks is reported in Table 5.4.

As shown in this table, GOOG has the highest midquote price among the six
stocks, because Google has never split its stock. The midquote price of stock F is
the smallest which is 1.84.
On the proportional bid-ask spread, the bigger companies (GOOG, XOM)
have narrower spreads than smaller companies (MDS, CHDX, DAR, F), because
bigger stocks are more liquid to trade than smaller stocks.
The average number of quotes per day of the six stocks is 38,960, which im-
plies that there is more than one quote per second.4 Especially, there are more
than four transactions occurring per second for GOOG and XOM, and more than
one for F.
4
One trading day, from 9:30 to 16:05 equals to 23700 seconds or 395 minutes.
111
Table 5.4: Descriptive Statistics for US stocks.
MDS CHDX DAR F GOOG XOM mean

Midquote price 8.64 5.66 4.83 1.84 349.00 74.21 74.03
Proportional bid-ask spread 0.0137 0.0101 0.0051 0.0062 0.0007 0.0005 0.0060
Number of trades per day 81 168 767 2888 23109 26877 8981
Trade size (no. of shares) 177 207 245 2016 127 331 517
112
Buyer-initiated trade size 157 189 248 1843 129 335 483
Seller-initiated trade size 200 223 241 2237 124 329 559
Bid depth (no. of shares) 205 436 1025 57839 227 1584 10219
Ask depth (no. of shares) 202 247 918 46268 223 1254 8185
Buy/Sell ratio 1.1726 0.9295 1.0904 1.2761 0.9873 0.7943 1.0417
Probability of reversal 0.4255 0.4500 0.6135 0.9430 0.1816 0.4297 0.5072
Note: The numbers reported in this table are the average values over 19 trading days (Feb. 2009).
The average number of trades per day for the six stocks is 8,981, which implies
that more than 22 transactions are occurring per minute. About one transaction
happens per second for bigger stocks (GOOG and XOM). In contrast, about one
transaction happens every five minutes for thinly traded stock MDS. The stocks
with higher market capitalisation are traded more frequently than those with lower
market capitalisation.
The average trade size of the six stocks is 517, about half of the trade size ten
years ago.5 There are five stocks with an average trade size less than 400. The
average buyer-initiated trade size of the six stocks is 483, smaller than the average
seller-initiated trade size (559).
The ask(bid) depth is larger than buyer(seller)-initiated trade size for each
stock. The buy/sell ratio is the ratio of the number of buyer-initiated trades to
the number of seller-initiated trades.
The average buy/sell ratio of the six stocks is 1.0417. Three of the stocks,
MDS, DAR and F, have buy/sell ratios larger than one which reflect that there are
more buyer-initiated trades than seller-initiated trades, and the other three (CHDX,
GOOD and XOM) are smaller than one which suggest that there are more seller-
initiated trades than buyer-initiated trades. It is interesting to see how the stock
price moves in the sample period. Figure 5.1 plots the price movement during
the sample period (Feb. 2009) for stock XOM. Since the buy/sell ratio of XOM
is 0.7943, it is not surprising to see that the price falls over the sample period in
Figure 5.1: the price rises from 76.69 and reaches a peak (80.34) at the end of the
first week in the sample period, and then drops for the rest of the sample period
(ending at 67.90).
There are two stocks with a trade reversal probability larger than 0.5, and four
5
Henker Henker & Wang (2006) investigated the average trade size for 401 stocks and more
than 200 trading days in 1999. They found that the average buyer-initiated trade size is 1166 and
the average seller-initiated trade size is 1096.
113
stocks with a trade reversal probability less than 0.5. While F has a very high
reversal probability 0.9430, GOOG has a very low reversal probability 0.1816.
The average probability of trade reversal is very close to 0.5, suggesting that in
general the probability that a trade is followed by another trade with the same sign
is similar to the probability that it is followed by a trade with reverse sign for these
six stocks.
Figure 5.1: Price of XOM during Feb. 2009 (from Yahoo Finance)
Summary Statistics for UK Stocks
A summary of descriptive statistics of the six UK stocks is reported in Table 5.5.

As shown in the table, the proportional bid-ask spreads of these six stocks are
almost the same, two are 0.0007, three are 0.0006 and one is 0.0005.
The rest of the variables reported in Table 5.5 reflect trading activity during
the period from April 2010 to September 2010 for the six UK stocks. The third
row is the values of number of trades per day. Higher values indicate that the stock
is more actively traded. As seen from the table, BARC, BP and HSBC are more
actively traded during the period than BATS, GSK and VOD.
The fourth row is the values of trade size in shares. While the average trade
size of BATS is less than 1,000, the average trade size of VOD is more than 10,000.
114
Table 5.5: Descriptive Statistics for UK stocks.
BARC BATS BP GSK HSBC VOD mean

Midquote price 313.49 2187.29 417.60 1193.99 649.67 146.44 818.08
Proportional bid-ask spread 0.0007 0.0006 0.0006 0.0007 0.0005 0.0006 0.0006
Number of trades per day 9388 3211 12161 3498 8449 5783 7082
Trade size (no. of shares) 3852 657 3339 1573 2434 11176 3839
115
Buyer-initiated trade size 3853 660 3382 1598 2463 11173 3855
Seller-initiated trade size 3856 656 3320 1569 2403 11426 3872
Bid depth (no. of shares) 8220 2274 8593 9378 6622 60499 15931
Ask depth (no. of shares) 8394 2388 8428 10096 6783 65373 16910
Buy/Sell ratio 1.0010 1.0036 1.0048 1.0402 1.0059 1.1405 1.0327
Probability of reversal 0.4825 0.4857 0.4759 0.4742 0.4762 0.4672 0.4769
Note: The numbers reported in this table are the average values over 126 trading days (April-September 2010).
The fifth and the sixth rows are the values for buyer-initiated trade size and sell-
initiated trade size respectively. These values indicate that buyer-initiated trade
size almost equals seller-initiated trade size. This suggests that it is not necessary
to assume that buyer-initiated trade size differs from seller-initiated trade size in
theoretical models.
The seventh and eighth rows are the values of bid depth and ask depth sepa-
rately. Observing these values, the differences between bid depth and ask depth
are very small for the six stocks. This suggests that the limit order book is sym-
metric at least at the best levels in these six stocks.
The ninth row is the values of buy/sell ratio. The buy/sell ratio nearly equals
to one in all the six stocks.
The last row in the table reports the values of probability of reversal. All of the
six stocks have a trade reversal probability very close to 0.5. This suggests that
in general the probability that a trade is followed by another trade with the same
sign is similar to the probability that it is followed by a trade with reverse sign for
these six stocks. This result is consistent with the finding above for the US stocks.
5.2.2 Intraday Price Volatility
Price volatility is a direct measure of risk and an indirect measure of the level of
information (French & Roll, 1986).
Intraday Price Volatility for the US stocks
Figure 5.2 shows all the six stocks have U-shaped intraday pattern on price volatil-
ity, which is consistent with the findings in prior literature (shown in Table 2.2 in
Chapter 2). The results support the predictions of the Brock & Kleidon (1992)
model which predicts higher volatility at market open and close.
116
Intraday Price Volatility for the UK stocks
As shown in Figure 5.3, the volatility is relatively higher at market open and close,
and relatively lower in the middle of the day for all the six UK stocks. This finding
is in line with the U-shaped intraday pattern on price volatility found in previous
literature (shown in Table 2.2 in Chapter 2).
117
−3
x 10
12
0.018
11
0.016
10
0.014
9
Absolute Log Return
8 0.012
Absolute Log Return

7
0.01
6
0.008
0.006
4
0.004
3
2 0.002
9:30 10:30 11:30 12:30 13:30 14:30 15:30 9:30 10:30 11:30 12:30 13:30 14:30 15:30
Time Time
(a) MDS (b) CHDX

−3
x 10
14
0.02
12
10
0.015
Absolute Log Return
Absolute Log Return
6 0.01
2 0.005
9:30 10:30 11:30 12:30 13:30 14:30 15:30 9:30 10:30 11:30 12:30 13:30 14:30 15:30
Time Time
(c) DAR (d) F

−3 −3
x 10 x 10
9 7
8
6
5
6
Absolute Log Return
Absolute Log Return
5 4
4
3
2
2
1 1
9:30 10:30 11:30 12:30 13:30 14:30 15:30 9:30 10:30 11:30 12:30 13:30 14:30 15:30
Time Time
(e) GOOG (f) XOM
Figure 5.2: Intraday Price Volatility (US stocks)
118
Price Volatility Price Volatility
0.16 0.1
0.09
0.14
0.08
0.12
0.07
Price Volatility
Price Volatility
0.1 0.06
0.05
0.08
0.04
0.06
0.03
0.04 0.02
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(a) BARC (b) BATS

0.25 0.09
0.08
0.2
0.07
0.06
0.15
Price Volatility
Price Volatility
0.05
0.1
0.04
0.03
0.05
0.02
0 0.01
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(c) BP (d) GSK

0.075 0.1
0.07 0.09
0.065
0.08
0.06
0.07
Price Volatility
Price Volatility
0.055
0.06
0.05
0.05
0.045
0.04
0.04
0.03
0.035
0.03 0.02
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(e) HSBC (f) VOD
Figure 5.3: Intraday Price Volatility (UK stocks)
119
5.2.3 Intraday Bid-ask Spread
Intraday Bid-ask Spread for the US stocks
As shown in Figure 5.4, the proportional bid-ask spreads in five stocks (MDS,
DAR, F, GOOG and XOM) do not exhibit the U-shaped intraday pattern as ob-
served by many empirical studies (Chung et al., 1999; Lee et al., 1993; Madhavan
et al., 1997) which examine the intraday pattern for the NYSE market. How-
ever, the finding is consistent with Henker & Wang (2006), Chan et al. (1995) and
Werner & Kleidon (1996), who examine intraday patterns on the NYSE, NAS-
DAQ and the LSE markets respectively, and all demonstrate a reverse S-shaped
intraday pattern on bid-ask spread: it is relatively high at market open, is stable in
the middle of the trading day, and drops at market close. As shown in Figure 5.4,
only CHDX has a U-shaped intraday pattern on bid-ask spread consistent with the
findings in Abhyankar et al. (1997), Ahn & Cheung (1999) Chung et al. (1999),
Lee et al. (1993), Lehmann & Modest (1994), Madhavan et al. (1997) and Vo
(2007).
Intraday Bid-ask Spread for the UK stocks
The intraday proportional bid-ask spreads and intraday time-weighted propor-

tional spreads for the UK stocks are shown in Figure 5.5 and in Figure 5.6 re-
spectively. They both demonstrate that the bid-ask spreads of these six UK stocks
have reverse S-shaped intraday pattern: bid-ask spread is highest at market open,
stable at the middle of the trading day, and lowest at market close. The result con-
firms the finding in Werner & Kleidon (1996) who examine British stocks, and is
also consistent with the reversed S-shaped pattern on the NYSE market and the
NASDAQ market demonstrated by Henker & Wang (2006) and Chan et al. (1995)
separately.
120
0.035 0.035
0.03 0.03
Proportional Bid−Ask Spread

0.025 0.025
0.02 0.02
0.015 0.015
0.01 0.01
0.005 0.005
9:30 10:30 11:30 12:30 13:30 14:30 15:30 9:30 10:30 11:30 12:30 13:30 14:30 15:30
Time Time
(a) MDS (b) CHDX

−3 −3
x 10 x 10
10 6.8
9
6.6
8
6.4
6.2
6
5
5.8
4
3 5.6
9:30 10:30 11:30 12:30 13:30 14:30 15:30 9:30 10:30 11:30 12:30 13:30 14:30 15:30
Time Time
(c) DAR (d) F

−3 −4
x 10 x 10
1.6 7
6.5
1.4
6
1.2
5.5
0.8
4.5
0.6
4
0.4 3.5
9:30 10:30 11:30 12:30 13:30 14:30 15:30 9:30 10:30 11:30 12:30 13:30 14:30 15:30
Time Time
(e) GOOG (f) XOM
Figure 5.4: Intraday Proportional Bid-Ask Spread (US stocks)
121
−4
x 10 Proportional Bid−Ask Spread −4
x 10 Proportional Bid−Ask Spread
14 14
13 13
12 12
11 11
10 10
Pro. Spread
Pro. Spread
9 9
8 8
7 7
6 6
5 5
4 4
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(a) BARC (b) BATS

−4
x 10 Proportional Bid−Ask Spread x 10
−3 Proportional Bid−Ask Spread
14 1.5
1.4
12
1.3
1.2
10
1.1
Pro. Spread
Pro. Spread
8 1
0.9
6
0.8
0.7
4
0.6
2 0.5
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(c) BP (d) GSK
−4
x 10 Proportional Bid−Ask Spread −4
x 10 Proportional Bid−Ask Spread
10 13
12
9
11
8
10
7
Pro. Spread
Pro. Spread
8
6
7
5
4
5
3 4
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(e) HSBC (f) VOD
Figure 5.5: Intraday Proportional Spread (UK stocks)
122
x 10
−10 Time−weighted Proportional Bid−Ask Spread −10
x 10 Time−weighted Proportional Bid−Ask Spread
13 12
12 11
11 10
10 9
TW Pro. Spread
TW Pro. Spread
9 8
8 7
7 6
6 5
5 4
4 3
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(a) BARC (b) BATS
x 10
−9 Time−weighted Proportional Bid−Ask Spread −10
x 10 Time−weighted Proportional Bid−Ask Spread
3 12
11
2.5
10
2
TW Pro. Spread
TW Pro. Spread
1.5
1
7
0.5
6
0 5
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(c) BP (d) GSK
x 10
−10 Time−weighted Proportional Bid−Ask Spread −10 Time−weighted Proportional Bid−Ask Spread
8 x 10
11
7 10
9
6
TW Pro. Spread
TW Pro. Spread
4
6
3
5
2 4
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(e) HSBC (f) VOD
Figure 5.6: Intraday Time-weighted Proportional Spread (UK stocks)
123
5.2.4 Intraday Market Depth
Intraday Market Depth for the US stocks
Figures 5.7 & 5.8 demonstrate the intraday behaviours of market depths for the
US stocks. As shown in these figures, bid depth and ask depth exhibit an almost
identical shape of intraday pattern. There are four stocks (MDS, CHDX, GOOG
and XOM) exhibiting a U-shaped intraday pattern on both bid and ask depths.
DAR has a J-shaped intraday pattern on market depth, while F has an S-shaped
intraday pattern on market depth. The results are inconsistent with the inverted
U-shaped pattern on market depth observed by Ahn & Cheung (1999); Lee et al.
(1993); Li et al. (2005); Vo (2007). These studies also report a negative association
between intraday spread and market depth, which is not observed either in this
thesis.
Intraday Market Depth for the UK stocks
Intraday patterns on market depths for the UK stocks are demonstrated in Figures
5.9-5.11. There are three stocks (BATS, GSK and VOD) exhibiting an S-shaped
intraday pattern on market depths. HSBC has a U-shaped intraday pattern on
market depths. For BARC and BP, it is clear to see that market depth is very
high at market close. As shown in the Figures 5.10-5.11, bid depth has an almost
identical intraday pattern with ask depth.
Intraday patterns on near market depths for the UK stocks are demonstrated
in Figures 5.12-5.14. All of the six stocks exhibit an S-shaped intraday pattern
on near market depth: it is relatively lower at market open, and relatively higher
at market close. As seen from the Figures 5.13-5.14, near bid depth has a nearly
identical intraday pattern with near ask depth.
The results on both measures of market depths are inconsistent with the in-
124
verted U-shaped intraday pattern observed by Ahn & Cheung (1999); Lee et al.
(1993); Li et al. (2005); Vo (2007). However, the near depth (Figure 5.12) shows
a negative association with the bid-ask spread (Figures 5.5-5.6), confirming the
findings on the relationship between spread and depth in Ahn & Cheung (1999);
Lee et al. (1993); Li et al. (2005); Vo (2007). While the near depth is relatively
higher during the trading day, that is at market close for the six UK stocks, the
spread is relatively lower; while the near depth is relatively lower during the trad-
ing day, that is at market open, the spread is relatively higher.
125
4
4 x 10
x 10 8
4
7
3.5
3 6
2.5 5
Bid Depth
Bid Depth
2 4
1.5 3
1 2
0.5 1
0 0
9:30 10:30 11:30 12:30 13:30 14:30 15:30 9:30 10:30 11:30 12:30 13:30 14:30 15:30
Time Time
(a) MDS (b) CHDX

6 7
x 10 x 10
3 14
2.5 12
2 10
Bid Depth
Bid Depth
1.5 8
1 6
0.5 4
0 2
9:30 10:30 11:30 12:30 13:30 14:30 15:30 9:30 10:30 11:30 12:30 13:30 14:30 15:30
Time Time
(c) DAR (d) F

6 6
x 10 x 10
1.8 16
1.6 14
1.4 12
1.2 10
Bid Depth
Bid Depth
1 8
0.8 6
0.6 4
0.4 2
9:30 10:30 11:30 12:30 13:30 14:30 15:30 9:30 10:30 11:30 12:30 13:30 14:30 15:30
Time Time
(e) GOOG (f) XOM
Figure 5.7: Intraday Bid Depth in Share (US stocks)
126
4
4 x 10
x 10 7
3.5
6
3
2.5 5
2 4
Ask Depth
Ask Depth
1.5 3
1 2
0.5 1
0 0
9:30 10:30 11:30 12:30 13:30 14:30 15:30 9:30 10:30 11:30 12:30 13:30 14:30 15:30
Time Time
(a) MDS (b) CHDX

6 7
x 10 x 10
2.5 14
12
10
1.5
8
Ask Depth
Ask Depth
6
1
0.5
2
0 0
9:30 10:30 11:30 12:30 13:30 14:30 15:30 9:30 10:30 11:30 12:30 13:30 14:30 15:30
Time Time
(c) DAR (d) F

6 6
x 10 x 10
1.8 14
1.6
12
1.4
10
1.2
Ask Depth
Ask Depth
6
0.8
4
0.6
0.4 2
9:30 10:30 11:30 12:30 13:30 14:30 15:30 9:30 10:30 11:30 12:30 13:30 14:30 15:30
Time Time
(e) GOOG (f) XOM
Figure 5.8: Intraday Ask Depth in Share (US stocks)
127
Depth in Share Depth in Share
10000 3000
2800
9500
2600
9000
Depth (Share)
Depth (Share)
2400
8500
2200
8000
2000
7500
1800
7000 1600
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(a) BARC (b) BATS

Depth in Share Depth in Share
10000 13000
12000
9500
11000
10000
9000
Depth (Share)
Depth (Share)
9000
8000
8500
7000
6000
8000
5000
7500 4000
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(c) BP (d) GSK

x 10
4 Depth in Share
Depth in Share 10
8500
8000
7500
Depth (Share)
7
Depth (Share)
6
7000
6500
4
6000 3
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(e) HSBC (f) VOD
Figure 5.9: Intraday Depth in Share (UK stocks)
128
Bid Depth in Share Bid Depth in Share
10000 2800
9500 2600
9000 2400
Bid Depth (Share)

Bid Depth (Share)
8500 2200
8000 2000
7500 1800
7000 1600
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(a) BARC (b) BATS

Bid Depth in Share Bid Depth in Share
10000 13000
12000
9500
11000
10000
Bid Depth (Share)
Bid Depth (Share)
9000
9000
8000
8500
7000
6000
8000
5000
7500 4000
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(c) BP (d) GSK

x 10
4 Bid Depth in Share
Bid Depth in Share 9
8500
8
8000
7
7500
Bid Depth (Share)
Bid Depth (Share)
6
7000
6500
5
6000 4
5500 3
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(e) HSBC (f) VOD
Figure 5.10: Intraday Bid Depth in Share (UK stocks)
129
Ask Depth in Share Ask Depth in Share
10500 3200
3000
10000
2800
9500
2600
Ask Depth (Share)

Ask Depth (Share)
9000
2400
8500
2200
8000
2000
7500
1800
7000 1600
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(a) BARC (b) BATS

Ask Depth in Share Ask Depth in Share
10000 13000
12000
9500
11000
9000
10000
Ask Depth (Share)
Ask Depth (Share)
8500 9000
8000
8000
7000
7500
6000
7000 5000
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(c) BP (d) GSK

x 10
4 Ask Depth in Share
Ask Depth in Share 11
8500
10
8000
9
8
Ask Depth (Share)
Ask Depth (Share)
7500
7000
6
5
6500
6000 3
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(e) HSBC (f) VOD
Figure 5.11: Intraday Ask Depth in Share (UK stocks)
130
5
x 10 Near Depth in Share 4
x 10 Near Depth in Share
1.6 7
6.5
1.5
6
1.4
5.5
1.3
Near Depth (Share)
Near Depth (Share)

5
1.2 4.5
4
1.1
3.5
1
3
0.9
2.5
0.8 2
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(a) BARC (b) BATS

5
1.8 3
1.7
1.6 2.5
1.5
Near Depth (Share)
Near Depth (Share)
1.4 2
1.3
1.2 1.5
1.1
1 1
0.9
0.8 0.5
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(c) BP (d) GSK
5
1.4 2.2
2
1.3
1.8
1.2 1.6
Near Depth (Share)
Near Depth (Share)
1.4
1.1
1.2
1 1
0.8
0.9
0.6
0.8 0.4
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(e) HSBC (f) VOD
Figure 5.12: Intraday Near Depth in Share (UK stocks)
131
5
x 10 Bid Near Depth in Share 4
x 10 Bid Near Depth in Share
1.5 7
6.5
1.4
1.3
5.5
Bid Near Depth (Share)

5
1.2
4.5
1.1
4
3.5
1
0.9
2.5
0.8 2
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(a) BARC (b) BATS

5
1.8 3
1.7
1.6 2.5
1.5
1.4 2
1.3
1.2 1.5
1.1
1 1
0.9
0.8 0.5
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(c) BP (d) GSK
5
1.4 2
1.3
1.2 1.5
1.1
1 1
0.9
0.8 0.5
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(e) HSBC (f) VOD
Figure 5.13: Intraday Near Bid Depth in Share (UK stocks)
132
5
x 10 Ask Near Depth in Share 4
x 10 Ask Near Depth in Share
1.6 7
6.5
1.5
6
1.4
5.5
Ask Near Depth (Share)

1.3
5
1.2 4.5
4
1.1
3.5
1
3
0.9
2.5
0.8 2
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(a) BARC (b) BATS

5
1.7 3
1.6
2.5
1.5
1.4
2
1.3
1.2
1.5
1.1
1
1
0.9
0.8 0.5
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(c) BP (d) GSK
5
1.4 2.2
2
1.3
1.8
1.2
1.6
1.1 1.4
1.2
1
0.9
0.8
0.8 0.6
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(e) HSBC (f) VOD
Figure 5.14: Intraday Near Ask Depth in Share (UK stocks)
133
5.2.5 Intraday Trade Size and Volume
Volume of trading is a direct measure of trading activity.
Intraday Trading Frequency, Size and Volume for the US stocks
The intraday patterns on trading volume for the six US stocks are shown in Figures
5.15 & 5.16. All of the six stocks have a U-shaped intraday pattern on trading
volume. The result confirms the finding in Foster & Viswanathan (1993); Jain &
Gun-Ho (1988); Lee et al. (1993); Lockwood & Linn (1990); McInish & Wood
(1990b, 1992); Wood et al. (1985) who examine the intraday patterns on the NYSE
and the finding in Chan et al. (1995) who examine the intraday patterns in the
NASDAQ.
It is interesting to find that trading frequency and trade size also both exhibit
a U-shaped intraday pattern as shown in Figures 5.19 & 5.21.
Intraday Trading Frequency, Size and Volume for the UK stocks
Intraday patterns on trading volume for the six UK stocks are shown in Figures
5.17 & 5.18. The six stocks all have a U-shaped intraday pattern on trading vol-
ume. Trading frequency shown in Figure 5.20 also exhibits a U-shaped intraday
pattern. The results are consistent with the finding of Werner & Kleidon (1996)
who examine the intraday pattern on the LSE and the findings for the NYSE mar-
ket in Foster & Viswanathan (1993); Jain & Gun-Ho (1988); Lee et al. (1993);
Lockwood & Linn (1990); McInish & Wood (1990b, 1992); Wood et al. (1985).
It is interesting to observe that there is a sharp increase of trading volume and
trading frequency between 2:30 p.m. and 3:00 p.m. for all the six stocks in Figures
5.17-5.20. This sharp increase in trading volume in the afternoon has been also
documented by Cai et al. (2004) who investigate the intraday patterns on the LSE.
It is due to the open of the US markets, as these six stocks are listed at both the
134
LSE and the NYSE. US markets opening at 2:30 p.m. GMT positively impact
on the trading activities of these stocks in terms of trading volume and trading
frequency.
Intraday patterns on trade size for the six UK stocks are shown in Figures
5.22 & 5.23. There are four stocks (BARC, BATS, BP and HSBC) exhibiting
a U-shaped intraday pattern on trade size. GSK and VOD have different shapes
of intraday pattern with the other stocks: while trade size is very large at market
close, it is smallest before market close.
135
5000 8000
4500
7000
4000
6000
3500
5000
3000
Trading Volume
Trading Volume
2500 4000
2000
3000
1500
2000
1000
1000
500
0 0
9:30 10:30 11:30 12:30 13:30 14:30 15:30 9:30 10:30 11:30 12:30 13:30 14:30 15:30
Time Time
(a) MDS (b) CHDX

4
x 10 x 10
5
6 12
5 10
4 8
Trading Volume
Trading Volume
3 6
2 4
1 2
0 0
9:30 10:30 11:30 12:30 13:30 14:30 15:30 9:30 10:30 11:30 12:30 13:30 14:30 15:30
Time Time
(c) DAR (d) F

5 6
x 10 x 10
4 2
1.8
3.5
1.6
3 1.4
1.2
Trading Volume
Trading Volume
2.5
2
0.8
1.5 0.6
0.4
1
0.2
0.5 0
9:30 10:30 11:30 12:30 13:30 14:30 15:30 9:30 10:30 11:30 12:30 13:30 14:30 15:30
Time Time
(e) GOOG (f) XOM
Figure 5.15: Intraday Trading Volume in Share (US stocks)
136
4 4
x 10 x 10
4.5 4
4 3.5
3.5
3
3
Trade Volume (dollars)

2.5
2.5
2
2
1.5
1.5
1
1
0.5 0.5
0 0
9:30 10:30 11:30 12:30 13:30 14:30 15:30 9:30 10:30 11:30 12:30 13:30 14:30 15:30
Time Time
(a) MDS (b) CHDX

5 6
x 10 x 10
3 2.5
2.5
2
2
1.5
1.5
0.5
0.5
0 0
9:30 10:30 11:30 12:30 13:30 14:30 15:30 9:30 10:30 11:30 12:30 13:30 14:30 15:30
Time Time
(c) DAR (d) F

7 7
x 10 x 10
14 15
12
10 10
6 5
2 0
9:30 10:30 11:30 12:30 13:30 14:30 15:30 9:30 10:30 11:30 12:30 13:30 14:30 15:30
Time Time
(e) GOOG (f) XOM
Figure 5.16: Intraday Trading Volume in Value (US stocks)
137
x 10
6 Trading Volume in Share 4
x 10 Trading Volume in Share
2.2 14
2
12
1.8
1.6 10
Trading Volume (Share)

1.4
8
1.2
1 6
0.8
4
0.6
0.4 2
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(a) BARC (b) BATS
x 10
6 Trading Volume in Share x 10
5 Trading Volume in Share
3 4
3.5
2.5
3
2
2.5
2
1.5
1.5
0.5 0.5
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(c) BP (d) GSK
5
x 10 Trading Volume in Share x 10
6 Trading Volume in Share
14 4
3.5
12
3
10
2.5
6
1.5
4
1
2 0.5
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(e) HSBC (f) VOD
Figure 5.17: Intraday Trading Volume in Share (UK stocks)
138
x 10
8 Trading Volume in Value 8 Trading Volume in Value
7 x 10
3
6
2.5
5
Trading Volume (Value)

2
1.5
1
2
1 0.5
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(a) BARC (b) BATS
x 10
8 Trading Volume in Value 8
x 10 Trading Volume in Value
12 4.5
11
4
10
3.5
9
3
8
7 2.5
6
2
5
1.5
4
1
3
2 0.5
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(c) BP (d) GSK
x 10
8 Trading Volume in Value 8 Trading Volume in Value
8 x 10
6
5.5
7
6
4.5
4
5
3.5
4
3
3 2.5
2
2
1.5
1 1
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(e) HSBC (f) VOD
Figure 5.18: Intraday Trading Volume in Value (UK stocks)
139
14 20
18
12
16
10
14
Number of Trades
Number of Trades
8
12
10
6
8
4
2
4
0 2
9:30 10:30 11:30 12:30 13:30 14:30 15:30 9:30 10:30 11:30 12:30 13:30 14:30 15:30
Time Time
(a) MDS (b) CHDX

140 350
120
300
100
250
Number of Trades
Number of Trades
80
200
60
150
40
100
20
0 50
9:30 10:30 11:30 12:30 13:30 14:30 15:30 9:30 10:30 11:30 12:30 13:30 14:30 15:30
Time Time
(c) DAR (d) F

2000 2400
2200
1800
2000
1600
1800
1400
Number of Trades
Number of Trades
1600
1200 1400
1200
1000
1000
800
800
600
600
400 400
9:30 10:30 11:30 12:30 13:30 14:30 15:30 9:30 10:30 11:30 12:30 13:30 14:30 15:30
Time Time
(e) GOOG (f) XOM
Figure 5.19: Intraday Trade Frequency (US stocks)
140
Market Order Frequency Market Order Frequency
500 200
450
400
150
Market Order Frequency

350
300
100
250
200
150 50
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(a) BARC (b) BATS

900 220
800 200
180
700
160
600
140
500
120
400
100
300
80
200 60
100 40
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(c) BP (d) GSK

500 350
450
300
400
250
350
300 200
250
150
200
100
150
100 50
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(e) HSBC (f) VOD
Figure 5.20: Intraday Trade Frequency (UK stocks)
141
400 450
400
350
350
300
300
Avergae Trade Size
Avergae Trade Size

250
250
200
200
150
150
100
100
50 50
9:30 10:30 11:30 12:30 13:30 14:30 15:30 9:30 10:30 11:30 12:30 13:30 14:30 15:30
Time Time
(a) MDS (b) CHDX

500 6000
450 5500
5000
400
4500
350
Avergae Trade Size
Avergae Trade Size
4000
300 3500
3000
250
2500
200
2000
150
1500
100 1000
9:30 10:30 11:30 12:30 13:30 14:30 15:30 9:30 10:30 11:30 12:30 13:30 14:30 15:30
Time Time
(c) DAR (d) F

200 900
190
800
180
700
170
Avergae Trade Size
Avergae Trade Size
600
160
150
500
140
400
130
300
120
110 200
9:30 10:30 11:30 12:30 13:30 14:30 15:30 9:30 10:30 11:30 12:30 13:30 14:30 15:30
Time Time
(e) GOOG (f) XOM
Figure 5.21: Intraday Trade Size (US stocks)
142
Trading Size in Share Trading Size in Share
4800 700
4600
4400
650
Trading Size (Share)

4200
4000
3800
600
3600
3400
3200 550
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(a) BARC (b) BATS

Trading Size in Share Trading Size in Share
4000 1700
1650
3800
1600
1550
3600
1500
3400 1450
1400
3200
1350
1300
3000
1250
2800 1200
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(c) BP (d) GSK
Trading Size in Share 4

x 10 Trading Size in Share
3400 1.25
3200 1.2
3000 1.15
1.1
2800
2600 1.05
2400 1
2200 0.95
2000 0.9
1800 0.85
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(e) HSBC (f) VOD
Figure 5.22: Intraday Trade Size in Share (UK stocks)
143
x 10
6 Trading Size in Value x 10
6 Trading Size in Value
1.55 1.55
1.5
1.5
1.45
1.45
1.4
Trading Size (Value)

1.35
1.4
1.3
1.35
1.25
1.2
1.3
1.15
1.25
1.1
1.05 1.2
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(a) BARC (b) BATS
x 10
1.9 2.1
1.8
2
1.7
1.9
1.6
1.8
1.5
1.7
1.4
1.6
1.3
1.2 1.5
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(c) BP (d) GSK
x 10
2.2 1.9
2.1
1.8
2
1.9
1.7
1.8
1.7 1.6
1.6
1.5
1.5
1.4
1.4
1.3
1.2 1.3
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(e) HSBC (f) VOD
Figure 5.23: Intraday Trade Size in Value (UK stocks)
144
5.2.6 Intraday Price Impact
Intraday Price Impact for the US stocks
Intraday patterns on price impact for the six US stocks are demonstrated in Figure
5.24. There are four stocks (DAR, F, GOOG and XOM) exhibiting a reverse S-
shaped intraday pattern on price impact: it is highest at market open and lowest at
market close. For MDS and CHDX, the price impact in the morning is generally
higher than the price impact in the afternoon.
Intraday Price Impact for the UK stocks
Intraday patterns on price impact for the UK stocks are shown in Figure 5.25. All
of the six stocks have a reverse S-shaped intraday pattern on price impact: it is
highest at market open, gradually decreases over the trading day, and is lowest at
market close. This finding is inconsistent with the U-shaped intraday pattern on
price impact observed in the HKSE by Chan (2000). The result indicates that the
market liquidity of these six stocks is increasing over the trading day, being lowest
at market open and highest at market close. It suggests that time-of-the-day is an
important consideration when constructing price impact functions and designing
trade execution strategies, which however is ignored in the literature on optimal
trading strategy (see the review in Chapter Two).
In contrast to the observation on the HKSE in Chan (2000), price impact does
not rebound before the market close as shown in Figures 5.24 & 5.25. This obser-
vation cannot be explained by private information as in Chan (2000) who asserts
that larger trades reveal more private information leading to larger price impact.
As shown in Figures 5.21, 5.22 and 5.23, the average trade size is very large at the
market close, negatively associated with low price impact in Figures 5.24 & 5.25.
It is interesting to find that price impact (Figure 5.25) has a negative associ-
145
ation with near depth (Figures 5.12 & B.7). While near depth is relatively lower
during the trading day, price impact is relatively higher; while near depth is rela-
tively higher during the trading day, price impact is relatively lower. The finding
confirms the conclusions in Hopman (2007) that price impact is mainly deter-
mined by the the order book.
146
x 10
−4 Price Impact / Trade Size (share) −5 Price Impact / Trade Size (share)
x 10
14
13
12
3 11
Price Impact Ratio
Price Impact Ratio

10
2 8
1 5
9:30 10:30 11:30 12:30 13:30 14:30 15:30 9:30 10:30 11:30 12:30 13:30 14:30 15:30
Time Time
(a) MDS (b) CHDX
x 10
−5 Price Impact / Trade Size (share) x 10
−6 Price Impact / Trade Size (share)
5 5
4.5
4.5
4
4
3.5
Price Impact Ratio
Price Impact Ratio
3.5
3
2.5
2.5
2
2
1.5
1.5 1
9:30 10:30 11:30 12:30 13:30 14:30 15:30 9:30 10:30 11:30 12:30 13:30 14:30 15:30
Time Time
(c) DAR (d) F
x 10
4.5 5
4
4.5
3.5
4
Price Impact Ratio
Price Impact Ratio
3.5
2.5
3
2
2.5
1.5
1 2
9:30 10:30 11:30 12:30 13:30 14:30 15:30 9:30 10:30 11:30 12:30 13:30 14:30 15:30
Time Time
(e) GOOG (f) XOM
Figure 5.24: Intraday Price Impact Ratio (US stocks)
147
−7
x 10 Price Impact / Trade Size (share) −6
x 10 Price Impact / Trade Size (share)
5.5 1.6
5
1.4
4.5
4 1.2
Price Impact Ratio
Price Impact Ratio

3.5
1
3
2.5 0.8
2
0.6
1.5
1 0.4
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(a) BARC (b) BATS

x 10 8
5
4.5
7
3.5 6
Price Impact Ratio
Price Impact Ratio
3
5
2.5
2 4
1.5
3
0.5 2
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(c) BP (d) GSK
x 10
−7 Price Impact / Trade Size (share) −7 Price Impact / Trade Size (share)
7 x 10
5
4.5
6
4
5
3.5
Price Impact Ratio
Price Impact Ratio
4
3
3 2.5
2
2
1.5
1
1
0 0.5
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(e) HSBC (f) VOD
Figure 5.25: Intraday Price Impact Ratio (UK stocks)
148
5.3 Summary
This chapter presented an empirical investigation of intraday behaviours of a num-
ber of interesting variables using the data drawn from the NYSE-Euronext TAQ
database and the LSE ROB database. Six stocks from the US markets and six
stocks from the UK markets were selected to perform this task.
This study contributes to the literature by documenting the up-to-date intraday
behaviours of a number of liquidity metrics. The findings in this study are of inter-
est to academics and trading desks. A better understanding of intraday behaviours
of market liquidity could help theorist in developing comprehensive models. It
also emphasises the importance of microstructure tick by tick data, which could
be applied to aggregation techniques, with implications beyond the microstructure
concerns. The results of this study are summarised as follows.
• Consistent with prior literature, this study documents U-shaped patterns on

price volatility for both the US and the UK stocks, consistent with the find-
ings in prior literature (shown in Table 2.2).
• Eleven out of twelve stocks exhibit a reverse S-shaped intraday pattern on

bid-ask spread, consistent with results in Chan et al. (1995), Henker & Wang
(2006) and Werner & Kleidon (1996).
• The observed intraday patterns on market depths in this study are incon-
sistent with the inverted U-shaped pattern demonstrated by Ahn & Cheung
(1999), Lee et al. (1993), Li et al. (2005) and Vo (2007).
• It is interesting to find that the near market depths of the UK stocks have
shown an S-shaped intraday pattern. Moreover, it is found that there is a
negative association between the near depth and bid-ask spread, confirming
149
the relationship identified by Ahn & Cheung (1999), Lee et al. (1993), Li
et al. (2005) and Vo (2007).
• Trading volume of the 12 stocks examined in this study demonstrates a

U-shaped intraday pattern, confirming the findings of most of the studies
shown in Table 2.4. Trading frequency also has a U-shaped intraday pattern
for all of the 12 stocks. Ten out of twelve stocks exhibit a U-shaped intraday
pattern on trade size.
• Ten out of twelve stocks demonstrate a reverse S-shaped intraday pattern

on price impact. It is interesting to find that price impact has a negative
association with near depth. This indicates that market is relatively more
liquid when near depth is relatively higher and price impact is relatively
lower, and vice versa. The results provide support for the conclusion in
Hopman (2007) that price impact is mainly determined by the limit order
book.
A natural question to ask which is worth further research is, what subpop-
ulation of the market participants is constituting the limit order book and what
proportion of the market participants is contributing to the trading volume and
market depth. To answer this question, further information on the identities of
market participants is needed.
150
Chapter 6
An Investigation of Price Impact

and Agent Intelligence
Parts of this chapter have appeared in:

Cui, W. and Brabazon, A. (2012). An Agent-based Modelling Approach to
Study Price Impact. Proceedings of the 2012 IEEE Conference on Computational
Intelligence for Financial Engineering & Economics, pp. 1-8, New York City,
USA. (Cui & Brabazon, 2012a)
It is well known that a buying transaction moves the stock price up and a
selling transaction drives it down. Price impact is the price change caused by
the transaction. In most limit order markets, the order book is transparent to all
market participants. A few studies have argued that the degree of price impact
is affected by the agent intelligence due to the transparency of limit order books.
As suggested by Farmer et al. (2004), the concavity of price impact is due to the
agent’s selective liquidity taking, namely agents condition the size of their trades
on liquidity, executing large orders when liquidity is high and small orders when
liquidity is low. Similar arguments are presented in Hopman (2007) and Weber &
151
Rosenow (2006).
This chapter examines whether the price impact of market orders is affected
by the intelligence of the agents who place market orders. The objective of this
chapter is to investigate whether the available market information is exploited by
“intelligent” agents when placing their market orders, thus affecting the magni-
tude of the influences of their trades on the asset’s price. Novelly, an agent-based
modelling is adopted to accomplish this task.
This chapter is organised as follows: a review of related work on agent-based
modelling of limit order markets is conducted in Section 6.1; a zero-intelligence
based model of limit order market is presented in Section 6.2; a description of
experimental setup is given in Section 6.3; results and discussions are provided in
Section 6.4, followed by a summary of chapter in Section 6.5.
6.1 Related Work

The development of zero-intelligence models of the continuous double auction
markets has a long history. One of the earliest models is built by Bollerslev &
Domowitz (1993), who extend the work of Gode & Sunder (1993) by taking into
account an order book in order to investigate the effects of order book on market
performance. Challet & Stinchcombe (2001) simulate a market which models the
Island ECN’s order books, while Bouchaud et al. (2002) model the Paris Bourse
(now known as Euronext Paris) market. In a series of studies, Zovko & Farmer
(2002), Iori et al. (2003), Smith et al. (2003), Farmer et al. (2005b) and Mike
& Farmer (2008) simulate a zero-intelligence model of double auction markets
based on random order placements to study order flows and market dynamics,
and their results demonstrate that the trading mechanism is more important than
agent behaviours in determining the statistical properties of the market. In their
152
model, the arrivals of the order flows composed of market orders, limit orders and
order cancelations are modeled as Poisson processes. Order sizes are produced
from a half-normal distribution.
Another study which simulates a limit order market in the spirit of the ZI idea
is conducted by Daniel (2006). In this model, there are 100 ZI traders endowed
with the same cash and shares. The arrival of each trade is modeled as a Poisson
process. At each arrival time, the trader chooses to buy or sell with probability
1/2, and decides to submit a market order, or a limit order, or cancels a previous
submitted order with probabilities πm , πl and 1−πm −πl respectively. This model
distinguishes two types of limit orders according to order aggressiveness. Traders
place a limit order uniformly distributed inside the bid-ask spread with probability
πin , or power-law distributed away from the spread with probability 1 − πin . The
volumes of incoming limit orders are produced from a log-normal distribution,
and the sizes of market orders are the same as those of the best counterpart orders.
Their model reproduces the principal stylized facts exhibited by real markets.
The most recent relevant work is the study by Huang et al. (2012). They build
an agent-based stock market model consisting of ZI agents to mimic the Taiwan
Stock Market (TWSE). Unlike the continuous matching in most limit order mar-
kets, order matching on the TWSE is organized every 25 seconds. In their model
at each simulated time corresponding 0.01 second of real time, there are five possi-
ble events happening in the artificial market: limit order submission, market order
submission, order cancelation, order matching and no activity. The probabilities
of these events are estimated using real data from TWSE, which are then used for
simulating the artificial market. Like Smith et al. (2003), they use a stochastic
order size which is generated randomly from a half-normal distribution. In the
study, they compare the liquidity costs in the real market to the simulated market,
which is measured as the difference of the virtual payment at the disclosed price
153
when the order is entering the market and the actual transaction payment. The re-
sult shows that the liquidity cost generated by the simulation data are higher than
those for the TWSE data which is possibly caused by their overestimated market
order size.
Many zero-intelligence models have been devoted to investigating market be-
haviours and market structure. A couple of studies have investigated the stylised
facts with zero-intelligence markets. These studies, including Ladley & Schenk
(2009), Bartolozzi (2010) and Palit et al. (2012), show that some stylised facts can
be obtained from the simulated markets with interacting zero-intelligence agents,
indicating that many stylised facts of limit order markets do not rely on individual
strategic behaviours, but are derived from the interaction of the market mechanism
and non-strategic agents. In addition, Duffy (2006) analyses bubbles and crashes
with a zero-intelligence market and show that very little trading strategy is is re-
quired to reproduce them. LiCalzi & Pellizzari (2007) investigate the performance
of four different types of market protocols. Aloud & Tsang (2011) show that the
intraday and interweek seasonality of FX market trading activity emerge with a
simple agent-based models of zero-intelligence traders.
6.2 A Zero-intelligence based Model of Limit Order

Market
The model simulates a limit order market. The trading mechanism follows the
price/time priority described in Section 1.2.1 of Chapter 1. Traders are allowed
to submit orders at any time, and cancel their limit orders which are not executed.
The model extends the model of Smith et al. (2003) by distinguishing different
limit orders according to order aggressiveness, considering log-normal distributed
order size and taking into account power-law distributed limit order placements in
154
order to come as close as possible to a realistic order flow.
6.2.1 Specification of Model
In the model, there are two groups of agents: buy agents and sell agents. All the
buy (sell) limit/market orders are placed by the buy (sell) agent. It is assumed that
each hypothetical time in the artificial market corresponds to one millisecond in
real market. At each time in the artificial market, a buy agent or a sell agent is
chosen with probability 1/2. The chosen agent at each time can perform one of
the four actions as below to fulfill his investment object:
• do nothing,
• submit a market order,
• submit a limit order, or
• cancel an outstanding limit order.1
With probability λo , the agent will do nothing at all; with probability λm , the agent
will submit a market order to the market; with probability λl , the agent will submit
a limit order to the market; with probability λc , the agent will cancel a limit order
which he previously placed and is not executed. The sum of these probabilities
has to be one (λo + λm + λl + λc = 1). Without loss of generality, it is assumed
that the agent always cancels the oldest limit order which he previously placed.
If the agent wants to submit a limit order, he can choose from four types of
limit order according to the order’s aggressiveness. These four types of limit
orders are:
• a crossing limit order, which causes an immediate execution,

1
An outstanding limit order means that the limit order is still listed on the order book
155
• an inside-spread limit order, which is placed inside the bid-ask spread,
• a spread limit order, which is placed at the best bid (ask) price if it is a buy
(sell), and
• an off-spread limit order, which is placed inside the order book with a less
attractive price than the best quote.
With probability λcrs , the agent chooses a crossing limit order; with probability
λinspr , the agent uses an inside-spread limit order; with probability λspr , the agent
adopts a spread limit order; with probability λof f spr , the agent places an off-spread
limit order. The sum of these probabilities is one (λcrs + λinspr + λspr + λof f spr =
1). It is assumed that crossing buy (sell) limit orders are always placed at the best
ask (bid) price.2 If the size of the crossing buy (sell) limit order exceeds the depth
at the best ask (bid) quote, the unexecuted part of the buy (sell) limit order will
be placed at the best ask (bid) quote. The inside-spread limit orders are uniformly
placed between the best bid quote and the best ask quote. The placement of off-
spread limit order follows a power-law distribution with exponent βof f −p ,3 and
various order sizes are allowed in the model, which are generated from log-normal
distributions.4
6.2.2 Algorithm of Model
An algorithm describing the artificial market simulation process is shown in Al-

gorithm 2.
2
This assumption is made based on the observation of LSE market that crossing limit orders
are rarely placed far away from the best quote on the opposite side of the market.
3
Empirical work (Zovko & Farmer, 2002; Potters & Bouchaud, 2003) finds that the relative
price of the off-spread limit order follows a power-law distribution. The relative price is the dif-
ference between the limit price of the order and the best quote.
4
Empirical work (Maslov & Millis, 2001) finds that order size is roughly distributed like a
log-normal distribution with a power law tail.
156
Algorithm 2 Artificial Limit Order Market
Generate a random number from a uniform distribution (0, 1) and determine the agent type according to the
probability 0.5.
switch Agent do
case Buyer
Generate a random number from a uniform distribution (0, 1) and determine the action type
according to the probabilities λo , λm , λl and λc .
switch Action do
case Do nothing
Do nothing at all.
endsw
case Submit a market order
Execute a market order with the order size generated from a log-normal distribution with
parameters µmo and σmo
endsw
case Submit a limit order
Generate a random number from a uniform distribution (0, 1) and determine the limit
order type according to probabilities λcrs , λinspr , λspr and λof f spr .
switch LimitOrder do
case Crossing Limit Order
Submit a limit order at the best ask price with the order size generated from a
log-normal distribution with parameters µcrs and σcrs .
Execute the crossing limit order. If it is not fully executed, place the rest of the
limit order at the best ask price on the order book.
endsw
case Inside-spread Limit Order
Generate a random value denoted as Pinspr from a uniform distribution
(BestBid, BestAsk).
Place a limit order at the price Pinspr on the order book with the order size
generated from a log-normal distribution with parameters µinspr and σinspr .
endsw
case Spread Limit Order
Place a limit order at the best bid price on the order book with the order size
generated from a log-normal distribution with parameters µspr and σspr .
endsw
case Off-spread Limit Order
Generate a random value denoted as RPof f spr from a power-law distribution
with exponent βof f −p .
Place a limit order at the price (BestBid − RPof f spr ) on the order book
with the order size generated from a log-normal distribution with parameters
µof f spr and σof f spr .
endsw
endsw
endsw
case Cancel an outstanding limit order
Cancels the oldest outstanding limit order that the buyer previously submitted.
endsw
endsw
endsw
case Seller
(similar to the case of buyer)
endsw
endsw
157
6.3 Experimental Setup
The setup of the experiment is explained in this section.
6.3.1 Parameters of Model
In order to make the model more realistic, the 14 parameters (shown in Table 6.2)
of the model are estimated using real data from London Stock Exchange (LSE).
One important reason why the LSE market is studied is that it provides data that
contain details of every order and every trade which enables us to calculate the
parameters of the model. This section provides a brief introduction to the LSE
data and describes how to estimate the parameters of the model using the data.
6.3.2 Parameter Estimation
For this study, the stock Barclays Capital is chosen, which is one of the most fre-
quently traded stocks in LSE. The sample data covers details of all orders and
trades from 1th April to 30th September 2010. Only the continuous trading ses-
sion is considered. The records before 8:00:00 and after 16:30:00 are ignored.
Table 6.1 shows the descriptive statistics of the trades and orders data for Bar-
clays Capital.
The data covers 126 trading days. During the continuous trading session, there
are 30,600,000 milliseconds in each trading day. The numbers of market orders,
limit orders and order cancelations5 and the probabilities of these events on each
day are calculated. As it is assumed that only one event occurs at each millisecond
in the model, the events occurring at the rest of the trading period are ‘do noth-
ing’ events. The probability parameters for the four events (do nothing λo , limit
5
We take order deletion and order expiration in ROB data as the same, both are counted as
order cancelation events.
158
Table 6.1: Descriptive Statistics of ROB Data at Sep. 6, 2010 for Barclays Capital
Descriptive Statistics
Average midquote price: 323.04
Average bid-ask spread: 0.11
Average trade size (pounds): 2,110,606
Average trade size (shares): 6,537
Number of trades: 4,329
Number of order cancelations: 68,783
Number of market orders: 3,972
Number of limit orders: 142,368
order λl , cancelation λc and market order λm ,) of the model are estimated as the
average daily probabilities from the real data. Similar calculations are performed
for the probabilities of the limit order types on each trading day using LSE data.
The probability parameters for limit order types (λcrs , λinspr , λspr and λof f spr ) in
the model are estimated as the average daily probabilities over the whole period
weighted by the total number of limit orders on each trading day.
6.3.3 Simulation Setup
In the model, each order size is drawn from a log-normal distribution of the func-
tional form:
exp(µ + σ ∗ rnorm )
where µ and σ are parameters, rnorm is a random number drawn from N(0,1), and
exp is an exponential function of the natural number. Each relative limit price is
drawn from the power-law function as:
xmin × (1 − r)− 1−β

1
159
where r is a random number uniformly generated from (0,1), xmin and β are the
parameters which need to be estimated. The data from the whole period is used
to estimate these parameters for all kinds of orders using the method of maximum
likelihood.6 In order to better fit the power-law distribution, extreme values xt
whose probabilities satisfy P (x > xt ) < 0.01 or P (x < xt ) < 0.01 are excluded.
All the estimated parameters are shown in Table 6.2.
The model is simulated using Matlab. In the model, there are always at least
three levels on each side of the order book in order to prevent the order book from
being empty. The artificial market runs for 34,200,000 hypothetical milliseconds
(corresponding to 9 and a half hours) in each simulation. In some cases, the ar-
tificial market may be unstable initially due to the stochastic order flows. Thus,
the first 3,600,000 milliseconds (corresponding to 1 hour) are used to warm up the
market. Only the later 30,600,000 milliseconds which correspond to the contin-
uous trading session in one trading day are used. The artificial market is run for
30 artificial trading days. Like the LSE, the system has a data recording function
which records the details of every trade and every order.
6
The matlab code for estimating the power-law distributions is developed by
the researchers at Santa Fe Institute, which can be downloaded from the website
http://tuvalu.santafe.edu/aaronc/powerlaws/.
160
Table 6.2: Parameters of Artificial Limit Order Market Simulation
Market Settings Values

Initial mid-quote price 300
Initial bid-ask spread 0.5
Tick size 0.01
Agent Type Probabilities
Buy or sell 0.5
Event Type Probabilities
Do nothing λo = 0.9847
Submit a market order λm = 0.0003
Submit a limit order λl = 0.0077
Cancel a limit order λc = 0.0073
Limit Order Type Probabilities
Crossing limit order λcrs = 0.0032
Inside-spread limit order λinspr = 0.0978
Spread limit order λspr = 0.1726
Off-spread limit order λof f spr = 0.7264
Order Size Type Parameters of Log-normal Distribution
Market order size µmo = 7.5663 σmo = 1.3355
Crossing limit order size µcrs = 8.4701 σcrs = 1.1982
Inside-spread limit order size µinspr = 7.8709 σinspr = 0.9799
Spread limit order size µspr = 7.8929 σspr = 0.8571
Off-spread limit order size µof f spr = 8.2166 σof f spr = 0.9545
Limit Price Type Parameters of Power-law Distribution
Off-spread relative limit price xminof f −p = 0.05 βof f −p = 1.7248
6.4 Results and Discussions

The results of this chapter are presented and discussed in this section.
161
6.4.1 Results of Probabilities of Events and Limit Order Types
After 30 simulation runs of the artificial market, the numbers of events, orders and
probabilities were recorded. Table 6.3 shows the means and standard deviations
of these over the 30 simulated trading days. An observation from the table is that
the events and orders probabilities in the simulated market are very close to those
in the LSE market (shown in Table 6.2).
Table 6.3: Events and Limit Order Types in Artificial Market
Probabilities Daily Occurrences

Order Sign
buy 0.5000 ± 0.0001 15, 300, 854 ± 2, 517
sell 0.5000 ± 0.0001 15, 299, 145 ± 2, 517
Event Type
do nothing 0.9847 ± 0.0000 30, 132, 401 ± 722
submit a market order 0.0003 ± 0.0000 9, 399 ± 104
submit a limit order 0.0076 ± 0.0001 233, 785 ± 483
cancel a limit order 0.0073 ± 0.0000 224, 413 ± 472
Limit Order Type
crossing limit order 0.0032 ± 0.0001 736 ± 29
inside-spread limit order 0.0977 ± 0.0006 22, 829 ± 152
spread limit order 0.1728 ± 0.0010 40, 408 ± 232
off-spread limit order 0.7262 ± 0.0011 169, 810 ± 444
6.4.2 Summary Statistics of Order Sizes and Relative Limit

Prices
The order sizes of different order types in the LSE market and those from the
simulated market were compared. Table 6.4 shows some summary statistics of
order sizes and relative limit prices for the six-month period (126 trading days) in
the LSE data and for the 30-artificial-day period in the simulated data respectively.
One can observe that the average order sizes for all order types and the average
162
relative limit prices in the simulated market are very close to those in the LSE.
163
Table 6.4: Summary Statistics of Order Sizes and Relative Limit Prices
Min Max Mean S.D. Observations
LSE Data 1 789,367 3,916 6,558 1,184,045
Market Order Size
Simulated Data 2 786,637 4,721 10,481 281,970
LSE Data 3 12,500,000 10,562 46,111 93,200
Crossing Limit Order Size
Simulated Data 36 543,378 9,676 16,716 22,080
LSE Data 1 2,476,197 3,902 4,761 2,881,792
Inside-spread Limit Order Size
164
Simulated Data 16 351,554 4,237 5,384 684,870
LSE Data 1 3,398,413 3,723 4,782 5,087,395
Spread Limit Order Size
Simulated Data 33 183,278 3,867 4,028 1,212,240
LSE Data 1 3,666,101 5,966 11,910 21,412,078
Off-spread Limit Order Size
Simulated Data 16 564,201 5,839 7,123 5,094,300
LSE Data 0.05 100.00 1.29 2.76 21,412,078
Off-spread Limit Order’s Relative Price
Simulated Data 0.05 99.99 0.93 4.72 5,107,980
6.4.3 Results of Price Impact vs. Trade Size
This study adopts the method used in Lillo et al. (2002) and Lillo et al. (2003)
to measure the price impact of a market order. Letting the logarithm of midquote
price be p, the price impact caused by a market order is calculated as
∆p = paf ter − pbef ore
where pbef ore is the price just before the market order arrived at the market and
paf ter is the price immediately after the order is executed. The methods used to
measure the trade size vary in previous literature. Lillo et al. (2003) measure it as
traded value in dollars divided by the stock value while Hopman (2007) measure it
as the number of shares in the order divided by the number of shares outstanding.
In this study, the trade size is measured as the shares of the market order divided
by the total trading volume in each trading day.
The average behavior of price impact is investigated by dividing the data based
on trade size into 10 bins and compute the average price impact for the data in each
bin. Figures 6.1 and 6.2 depict the relationship between the price impact and trade
sizes in the LSE and the simulated market respectively.
From Figures 6.1, one finds that the relationship between the price impact and
trade size is nonlinear and concave. Larger trades do not always have higher price
impact than smaller trades. This finding is consistent with the conclusion in pre-
vious literature. There are three explanations for this concavity in the literature.
The first explanation is that large-order trades have private information about the
market price. The second explanation is that large-order traders are patient traders
who wait for periods of high liquidity (Farmer et al., 2004). The third explana-
tion is that the order book has deeper liquidities away from the best bid or ask
(Bouchaud et al., 2002).
165
x 10
−4 price impact vs. trade size in LSE market
3.5
2.5
price impact
1.5
0.5
1 2 3 4 5 6 7 8 9 10
normalized trade size −3
x 10
Figure 6.1: Price Impact vs. Trade Size in the LSE market
166
x 10
−3 price impact vs. trade size in simulated market
8
5
price impact
0
1 2 3 4 5 6 7 8 9 10
normalized trade size −3
x 10
Figure 6.2: Price Impact vs. Trade Size in the simulated market
167
By comparing the price impacts in the two markets, one can find that generally
the price impact in simulated market is larger than that in the LSE market. The
price impacts for the smallest trade sizes are very close in the two markets, but
the price impacts for larger trade sizes (ranging from 2 ∗ 10−3 to 10 ∗ 10−3 ) in
simulated market are bigger than those in the LSE market.
There are two possible reasons to explain the difference on price impact be-
tween the simulated market and the real market. One possible reason is due to the
difference between the price gaps of limit order book in these two markets. If the
limit price gaps in the simulated market are larger than those in the real market, it
is possible that the price impact in the simulated market is larger than that in real
market. However, this reason can be excluded by the fact that the average relative
limit price in simulated market is smaller than that in the LSE market (shown in
Table 6.4).
The other possible reason is that traders in the LSE market execute their orders
intelligently, for example, by observing the market conditions. The intelligent
agents in real market condition their order sizes on market liquidity. Farmer et al.
(2004) and Hopman (2007) both find that large orders are traded when the market
liquidity is deep and small orders are traded when the market liquidity is shallow.
However, the trader agents in the simulated market are ‘zero intelligent’ who do
not care about the market liquidity when trading their orders. Therefore, the price
impacts caused by trading large orders in the simulated market are higher than the
price impacts in the LSE market, which is shown by the results in Figures 6.1 &
6.2. This indicates that agent intelligence plays an important role in determining
the magnitude of price impact, and suggests that agent intelligence is needed when
simulating an artificial market which tries to replicate the relationship between
price impact and market order size.
168
6.5 Summary
Many studies have shown that the price impact of a trade is a concave function of
its trade size, which is often called the “concavity” of price impact. Recently, a
number of studies (Farmer et al., 2004; Hopman, 2007; Weber & Rosenow, 2006)
argue that the concavity is due to selective liquidity taking: traders submit large
orders when the market is more liquidity, and small orders when the market is
less liquid. In an earlier study, Admati & Pfleiderer (1988) also argue that both
liquidity and informed traders prefer to trade when the market is thick, that is when
their trading has little effect on prices. This chapter investigates whether agent
intelligence plays an important role in determining the magnitude of price impact.
An agent-based modeling approach is adopted to perform this task. The advantage
of this approach is that it is easy to analyse one specific factor by isolating it in the
simulation. An zero-intelligence based artificial market was simulated in order to
examine the price impact caused by the zero-intelligence agents.
This chapter provides evidence supporting that agent intelligence is a critical
factor in determining the magnitude of price impact. It contributes to the liter-
ature on price impact. The finding is of interest to academics and practitioners.
It suggests that agent intelligence is an important factor when modelling price
impact. This chapter also contributes to the literature on agent-based modelling
of financial markets. It indicates that agent intelligence is a necessary condition
when simulating an artificial market where replicating realistic price impact is a
concern.
169
Chapter 7
Dynamic Trade Execution with

Limit and Market Orders
Much of the content of this chapter has been in various publications:
• Cui, W., Brabazon, A. and O’Neill, M. (2011). Adaptive Trade Execution:

A Grammatical Evolution Approach. International Journal of Financial
Markets and Derivatives. Vol. 2, No. 1/2, pp. 4-31. (Cui et al., 2011)

and Trade Execution. Natural Computing in Computational Finance (vol-
ume III), pp. 51-69, Springer-Verlag, Berlin. (Cui et al., 2010b)
• Cui, W., Brabazon, A. and O’Neill, M. (2010). Evolving Efficient Limit Or-
der Strategy Using Grammatical Evolution. Proceedings of the 2010 IEEE
Congress on Evolutionary Computation, pp. 1-6, Barcelona, Spain. (Cui
et al., 2010d)
• Cui, W., Brabazon, A. and O’Neill, M. (2010). Evolving Dynamic Trade

Execution Strategies Using Grammatical Evolution. Proceedings of the
170
2010 International Conference on Applications of Evolutionary Computa-
tion, Lecture Notes in Computer Science 6025, pp. 192-201. (nominated as
best paper) (Cui et al., 2010c)
• Cui, W., Brabazon, A. and O’Neill, M. (2009). Efficient Trade Execution

Using a Genetic Algorithm in an Order Book Based Artificial Stock Market.
Proceedings of the 11th Annual Conference Companion on Genetic and
Evolutionary Computation Conference, pp. 2023-2028. (Cui et al., 2009a)
Trading cost is a significant concern for financial investors as it directly im-

pacts on investment returns. As shown by Perold (1988), a hypothetical portfolio
without loss caused by execution costs outperforms the actual portfolio by almost
20% per year during the period from 1965 to 1986. In another study, Wermers
(2002) find that mutual funds outperform the market by 1.3 percent per year be-
fore accounting for transaction costs, but underperform the market by 0.3 percent
after accounting for expenses and transaction costs while investigating mutual
fund performance. Trading cost is composed of broker commission, fee, bid-ask
spread, price impact and opportunity cost. Empirical studies (Chan & Lakon-
ishok, 1995, 1997; Collins, 1991; Kraus & Stoll, 1972) have identified that the
dominant component is price impact. Price impact is variable and depends on
market liquidity. Many studies have shown that a small-quantity market order can
cause significant impact on an asset’s price when volume at price levels on the
order book is sparse at the time the order is submitted to the market.
Controlling price impact is an important issue in financial markets. It is espe-
cially important when trading large-quantity orders as their impact on the market
can be significant. In electronic limit order markets, this trading process is imple-
mented by computers automatically according to predefined trading algorithms. In
recent years, there is a growing literature on trading algorithms that are designed
171
to minimise the price impact of trading a large order. As reviewed in Section
2.5.1 of Chapter 2, one of the key assumptions in these models is that only market
orders are adopted.
This chapter aims to shed light on the effect of order choice on the price impact
of trading a large order. To the best of my knowledge, this is the first study which
specifically looks at this issue in the context of designing optimal trade execution
strategies. This chapter finds that order choice is an important determinant of the
design of optimal trading strategies. Using limit orders can help to minimise the
price impact of trading large quantities of orders. As the important task of provid-
ing liquidity in limit order markets is assigned to the complex trading interactions
enabled by the emergence and disclosure of the limit order book, the state of the
limit order book is extremely important for practitioners who use it to optimise
their trading strategies. In this chapter, this consideration is also accounted in the
design of execution strategies.
This chapter is organised as follows. A review of related work on trade exe-
cution and the genetic algorithm is given in Section 7.1; experimental design is
described in Section 7.2; experimental setup is explained in Section 7.3; empiri-
cal results are documented in Section 7.4; a summary of this chapter is given in
Section 7.5.
7.1 Related Work

Despite the importance of optimising trade execution, there has been relatively
little attention paid in the literature to the application of evolutionary methods for
this task. One notable exception is Lim & Coggins (2005b) who applied a Evo-
lutionary Algorithm to evolve a trade execution strategy using order book data
from a fully electronic limit order market, the Australian Stock Exchange (ASX).
172
In their study, a large order is to be completed within one trading day. The or-
der is divided into ten child orders which are submitted to the market at regular
intervals of half an hour. The relative sizes of these child orders are determined
according to share volume trading patterns, which typically follow a U-shaped
pattern with increased volumes trading at the open and close. The child orders
are placed into the market as limit orders at the best available price and the Evo-
lutionary Algorithm is used to find the optimal lifetime that a limit order would
remain on the order book (if it had not already been executed) before it was au-
tomatically ticked over the spread to close out the trade. The fitness function was
the Volume Weighted Average Price (VWAP) performance of that strategy rela-
tive to the benchmark daily VWAP. Each strategy was trained on three months’
worth of transaction-level data using a market simulator. The results were tested
out of sample on three highly liquid stocks and tested separately for sell side and
buy side. The in sample and out of sample performances of the evolved strategies
were better than those of pure market order strategies and pure limit order strate-
gies. In Lim & Coggins (2005b), the strategies were tested using historical data.
An agent-based artificial market is used instead in the present study.
7.2 Designs of Experiments

In order to investigate whether order choice is an important factor in determining
price impact when trading a large order, three experiments are conducted in this
chapter where three different combined strategies adopting different types of or-
ders and benchmark strategies using one type of orders are designed. The three
combined strategies are named Combined-I Strategy, Combined-II Strategy, and
Combined-III Strategy.
In the first experiment, the aim is to examine whether order choice between
173
various types of limit orders with different levels of aggressiveness plays an im-
portant role in determining the price impact when trading a large order. Empir-
ical evidence suggests that the state of the order book affects order placement
strategies. This effect is taken into account in the second and the third exper-
iments, where the trader is allowed to use market information to determine his
order choice. In the second experiment, the trader can place a limit order at three
levels of aggressiveness, aggressive limit order, moderate limit order, and passive
limit order, which are explained below. The aim of the second experiment is to
examine whether the choice between these three types of limit orders influences
the price impact of trading a large order. In the third experiment where the trader
can use a combination of market order and moderate limit order, the objective is
to investigate whether the choice between market order and limit order affects the
price impact of trading a large order.
• Aggressive limit order is a limit order whose limit price is one tick better
than the best price in the order book.
• Moderate limit order is a limit order which is placed at the best price in the
order book.
• Passive limit order is a limit order whose limit price is one tick worse than
the best price in the order book.
In these three experiments, the performances of the evolved strategies are com-
pared with those of simple execution strategies which only use one type of order.
In the first experiment, the evolved strategy is compared with a simple market
order strategy which only adopts market orders. In the second experiment, the
evolved strategy is compared with a simple aggressive limit order strategy, a sim-
ple moderate limit order strategy and a simple passive limit order strategy which
only use aggressive limit orders, moderate limit orders and passive limit orders
174
respectively. In the third experiment, the combined strategy is compared with
simple market order strategy and simple limit order strategy which only employs
market orders and moderate limit orders respectively.
Performance Evaluation
The standard industry metric for measuring trade execution performance is
the VWAP measure, short for Volume Weighted Average Price. The VWAP price
as a quality of execution measurement was first developed by Berkowitz et al.
(1988), who argued that ‘a market impact measurement system requires a bench-
mark price that is an un-biased estimate of prices that could be achieved in any
relevant trading period by any randomly selected trader’ and then defined VWAP
as an appropriate benchmark that satisfied this criteria.
The VWAP is calculated as the ratio of the value traded and the volume traded
within a specified time horizon (Berkowitz et al., 1988)
∑
(V olume ∗ P rice)
V W AP = ∑
(V olume)
where V olume represents the number of shares in each trade and P rice represents
its corresponding traded price. An example is shown in Table 7.1.
In order to evaluate the performance of a trade execution strategy, its VWAP
is compared against the VWAP of the overall market. The rationale here is that
performance of a trade execution strategy is considered good if the VWAP of
the strategy is more favorable than the VWAP of the market within the trading
period and poor if the VWAP of the strategy is less favorable than the VWAP of
the market within the trading period. For example, if the VWAP of a buy strategy
(V W AP strategy ) is lower than the market VWAP (V W AP market ), it is considered
a good trade execution strategy. Conversely, if the V W AP strategy is higher than
the V W AP market , it is considered a poor trade execution strategy. Although this
175
Table 7.1: VWAP Calculation of A Sample Buy Strategy
Submission Traded
Shares Value
Time Price
Child Order 1: t0 400 ∗ 50.15 = 20,060
600 ∗ 50.16 = 30,096
Child Order 2: t1 (t0 + ∆t) 1,000 ∗ 50.40 = 50,400
Child Order 3: t2 (t0 + 2∆t) 200 ∗ 50.34 = 10,068
800 ∗ 50.36 = 40,288
Child Order 4: t3 (t0 + 3∆t) 1,000 ∗ 50.39 = 50,390
Child Order 5: t4 (t0 + 4∆t) 1,000 ∗ 50.68 = 50,680
Child Order 6: t5 (t0 + 5∆t) 1,000 ∗ 51.10 = 51,100
Child Order 7: t6 (t0 + 6∆t) 1,000 ∗ 50.87 = 50,870
Child Order 8: t7 (t0 + 7∆t) 700 ∗ 50.98 = 35,686
300 ∗ 51.00 = 15,300
Child Order 9: t8 (t0 + 8∆t) 1,000 ∗ 50.39 = 50,390
Child Order 10: t9 (t0 + 9∆t) 1,000 ∗ 50.26 = 50,260
Total: 10,000 505,588
VWAP = 505, 588/10, 000 = 50.5588
is a simple metric, it largely filters out the effects of volatility, which composes
market impact and price momentum during the trading period (Almgren, 2008).
The performance evaluation functions are as follows (which were used by Lim
& Coggins (2005b)):


 104 ∗(V W AP strategy −V W AP market )

 (BuyStrategy)

 V W AP market
V W AP Ratio =





 10 ∗(V W AP market −V W AP strategy )
4
V W AP market
(SellStrategy)
where V W AP market is the average execution price which takes into account all
the trades over the day excluding the strategy’s trades. This corrects for bias,
especially if the order is a large fraction of the daily volume (Lim & Coggins,
176
2005b). For both buy and sell strategies, the smaller the VWAP Ratio, the better
the strategy is. If the combined strategies outperform the simple execution strate-
gies in terms of minimising price impact, it can be concluded that order choice is
an important determinant of trade execution strategies.
The design of an execution strategy can be considered of consisting of two
stages. The first stage is to divide a big block of shares into multiple small orders,
and the second stage is to determine the parameters of each small order, including
order type (limit/market order), submission time, limit price of limit order and
lifetime (the time length when a limit order appears in the order book before it
is cancelled or changed). How to divide a large trade depends on the order size
and trading time. The following three sections describe the designs of execution
strategies in these three experiments respectively.
7.2.1 Design of Experiment One
In this experiment, the large-quantity order is equally divided into 30 small orders,
and submit each small order into the market at regular intervals over one trading
day. In the design of the combined execution strategy, limit orders with varied
levels of aggressiveness are allowed. A genetic algorithm is adopted as the learn-
ing method of the trader who learns where to place his limit order in the order
book and when to cross his limit order over the bid-ask spread if it has not been
fully executed. The representation of the trading strategy in the genetic algorithm
method is shown in Figure 7.1.
The order aggressiveness is distinguished by the variable ci which represents
the relative price of order i. The following examples are used to illustrate how
to use the variable ci to determine the level of aggressiveness of each order. For
example, at the time of order submission, the best bid and best ask in the market
are 49.00 and 51.00 respectively and then the bid-ask spread is 2.00, assuming
177
1 ĂĂ 30 1 ĂĂ 30
1 30
1̢ 30
Figure 7.1: Representation of Order’s Relative Price and Lifetime
that it is a buy strategy:
• ci < 2.00 means that order i will be placed at 49.00 + ci , and its lifetime bi
means that at the end of its lifetime any unfilled part (if there is) of order i
will get crossed over the bid-ask spread for trading completion;
• ci > 2.00 or ci = 2.00 means that order i is a marketable limit order which
will be executed immediately at price 51.00.
7.2.2 Design of Experiment Two
In this experiment, a large-quantity order is equally divided into ten small orders
and submitted to the market at regular intervals over the trading day. Three types
of limit orders with varied levels of aggressiveness are considered: aggressive
limit order, moderate limit order and passive limit order. Two execution strategies
with two different settings of limit order’s lifetime are devised. One takes a short
lifetime, while the other takes a long lifetime. These two lifetime settings are
explained as follows. If the limit order is not fully executed at the end of its
lifetime, it will get crossed over the bid-ask spread for trading completion.
• In the short-lifetime strategy, each limit order can survive until the end of
its interval period.
• In the long-lifetime strategy, each limit order can survive until the end of
the trading day.
178
As prior literature indicates that the state of the limit order book affects or-
der placement strategies, both this experiment and the next experiment take into
account this effect. A range of possible explanatory variables, also known as in-
formation indicators, have been suggested in the literature. Six variables, which
are commonly employed in recent literature of market microstructure, are adopted
in the experiments. The explanations of these six variables are shown in Table 7.2.
Table 7.2: Definitions of Market Variables
Variables Definitions
BidDepth Number of shares at the best bid
AskDepth Number of shares at the best ask
Total number of shares at the best five ask prices divided by
RelativeDepth
total number of shares at the best five bid and ask prices
Spread Difference between the best bid price and best ask price
Volatility Standard deviation of the most recent 20 mid-quotes
Number of positive price changes within the past ten
PriceChange minutes divided by the total number of quotes submitted
within the past ten minutes
In these experiments, a trader’s order choice is determined by a number of

rules which are composed of these six information indicators. A grammatical
evolution algorithm is used as the learning method of the trader in the experiments
in order to learn the rules which are used to determine his order choice. The
grammar used in the grammatical evolution algorithm is defined in Figure 7.2.
In this grammar (Figure 7.2), two additional variables P ercOf T radedV olume
and P ercOf P astT ime (used in experiment two) are included, which represent
the percentage of the traded volume accounting for the total volume V shares and
the percentage of the past time accounting for the whole trading period at the ob-
served time respectively. The six indicators, BidDepth, AskDepth, RelativeDepth,
Spread, V olatility, and P riceChange are observed at the time of order sub-
mission. In Figure 7.2, AvgBidDepth, AvgV olatility, andAvgP riceChange,
179
represents the average bid depth of the market, AvgAskDepth represents the av-
erage ask depth of the market, AvgRelativeDepth represents the average rela-
tive depth of the market, AvgSpread represents the average spread of the market,
AvgV olatility represents the average volatility of the market and AvgP riceChange
represents the average price change of the market.
<lc> ::= if (<stamt>)
class = "AggressiveLimitOrder"
else {
if (<stamt>)
class = "ModerateLimitOrder"
else
class = "PassiveLimitOrder"
}
<stamt> ::= (<stamt><op><stamt>)|<cond1>|<cond2>|
<cond3>|<cond4>|<cond5>|<cond6>|
<cond7>|<cond8>
<op> ::= and
<cond1> ::= (BidDepth<comp>AvgBidDepth)
<cond2> ::= (AskDepth<comp>AvgAskDepth)
<cond3> ::= (RelativeDepth<comp>AvgRelativeDepth)
<cond4> ::= (Spread<comp>AvgSpread)
<cond5> ::= (Volatility<comp>AvgVolatility)
<cond6> ::= (PriceChange<comp>AvgPriceChange)
<cond7> ::= (PercOfTradedVolume<comp><threshold>)
<cond8> ::= (PercOfPastTime<comp><threshold>)
<comp> ::= <less>|<more>|<lessE>|<moreE>
<less> ::= "<"
<more> ::= ">"
<lessE> ::= "<="
<moreE> ::= ">="
<threshold> ::= 0.1|0.2|0.3|0.4|0.5|0.6|0.7|0.8|0.9
Figure 7.2: Grammars in Experiment Two
An example of an trade execution strategy evolved using the grammar in Fig-

ure 7.2 is illustrated in Figure 7.3. Using this example strategy, if P ercOf T radedV olume <
0.5andP ercOf P astT ime >= 0.8 is satisfied, an aggressive limit order is sub-
180
mitted. Otherwise, a moderate limit order is submitted if V olatility <= AvgV olatility,
or a passive limit order is submitted if V olatility > AvgV olatility.
if (PercOfTradedVolume<0.5 and PercOfPastTime>=0.8)
AggressiveLimitOrder
else{
if (Volatility<=AvgVolatility)
ModerateLimitOrder
else
PassiveLimitOrder
}
Figure 7.3: An Example of Trade Execution Strategy in Experiment Two
7.2.3 Design of Experiment Three
In this experiment, a large-quantity order is equally divided into ten small orders
and submitted to the market at regular intervals over the trading day. Both market
orders and moderate limit orders are allowed in the design of execution strategy.
Any unfilled orders at the end of trading day will get crossed over the bid-ask
spread for trading completion. As in experiment two, a grammatical evolution
algorithm is used as the learning method for the trader in order to learn the rules
on how to choose between market orders and moderate limit orders based on the
six information indicators of the limit order book. The grammar adopted in the
grammatical evolution algorithm is shown in Figure 7.4.
An example of an trade execution strategy evolved using the grammar in Fig-
ure 7.4 is illustrated in Figure 7.5. Using this example strategy, if (RelativeDepth >
AvgRelativeDepth) is true, a market order is submitted, otherwise, a moderate
limit order is submitted.
181
<lc> ::= if (<stamt>)
class = "MarketOrder"
else
class = "ModerateLimitOrder"
<stamt> ::= <cond1><op><cond2><op><cond3><op>
<cond4><op><cond5><op><cond6>
<op> ::= and | or
<cond1> ::= (BidDepth>AvgBidDepth) is <boolean>
<cond2> ::= (AskDepth>AvgAskDepth) is <boolean>
<cond3> ::= (RelativeDepth>AvgRelativeDepth) is
<boolean>
<cond4> ::= (Spread>AvgSpread) is <boolean>
<cond5> ::= (Volatility>AvgVolatility) is <boolean>
<cond6> ::= (PriceChange>AvgPriceChange) is <boolean>
<boolean> ::= True | False
Figure 7.4: Grammars in Experiment Three
if (RelativeDepth>AvgRelativeDepth) is True
MarketOrder
else
ModerateLimitOrder
Figure 7.5: An Example of Trade Execution Strategy in Experiment Three
182
7.3 Artificial Market Simulators
The execution strategies are tested in an artificial market simulator. Critically, in
the experiments the actions of the evolved execution strategies employed by the
simulated trader impacts on the state of the order book facing all the other agents
in the stock market and therefore impacts on their actions. In turn, the actions of
those agents impact on the order book facing the trader and therefore on the utility
of her execution strategy. In other words, the training and testing environment is
dynamic and allows examination of the issues of market impact and opportunity
cost. on trade execution. The use of an artificial stock market environment to
test the utility of execution strategies is a novel contribution of this chapter, and
opens up the door to a wide range of future work in this domain. Different mar-
ket simulators are used to test those three experiments. The purpose of this is to
examine whether the execution strategies keep robust in different market environ-
ments. Specifically, simulator I is adopted to test the strategies in experiment one,
while simulator II is employed to test the strategies in experiments two and three.
7.3.1 Simulator I
In the artificial market simulator I, the agents are equally likely to generate a
buy order or a sell order. The order can be a market order, a limit order, or a
cancellation of a previous limit order. When an agent is active, she can try to
issue a cancelation order with probability λc (oldest orders are canceled first), a
market order with probability λm , a limit order with probability λl = 1 − λc − λm .
In the model, the limit order price is uniformly distributed inside the spread with
probability λin , and power-law distributed outside the spread with probability 1 −
λin . Limit order price ranges are illustrated in Figure 7.6.
In this model, order generation is modeled as a Poisson process, which means
183
Price range of incoming limit buy orders
Current Best Bid Price

Price increases
Current Best Ask Price
Price range of incoming limit sell orders
Figure 7.6: Place Limit Price
that the time between orders follows an exponential distribution. As each incom-
ing buy (sell) market order arrives, the market agent will match it with the best ask
(bid) limit order stored in the order book. If this market order is fully filled by the
first limit order, the unfilled part will be matched to the next best ask (bid) limit
order until it is fully filled. As each incoming limit order arrives, the market agent
will store it in the order book according to price and time priority. As each incom-
ing cancelation order arrives, the market agent will delete the relevant limit order
in the order book. In the simulation, a Swarm platform in JAVA (Swarm, 2009) is
adopted, which is one of the most popular agent-based modeling platforms. The
algorithm used in the simulation I is described in Algorithm 3.
In order to ensure that the order flows generated by the artificial market are
economically plausible, all the parameters in the model are derived from previous
literature (Chakraborti et al., 2011; Farmer et al., 2005b,c,a, 2006; Mike & Farmer,
2008; Toth et al., 2009). The parameters used in the simulation are presented in
Table 7.3.
184
Algorithm 3 Artificial Market Simulator I
Generate t from an exponential distribution with parameter τ ;

currentt ime = currentt ime + t.
Generate Psign from Bernoulli distribution, and determine the agent type with probability 0.5.
switch Agent Type do
case Buyer
Generate Ptype from a uniform distribution, and determine the order type according to the
probabilities λm , λl and λc .
switch Order Type do
case Market order
Submit a market order.
endsw
case Limit order
Generate Plo from a uniform distribution, and determine the limit order type with
probabilities PinSpread and Pof f Spread .
switch Limit Order Type do
case Spread limit order
Submit a limit order whose limit price is drawn from a uniform distribution of
values between the best bid and best ask on the order book.
endsw
case Off-spread limit order
with exponent α.
Place a limit order at the price (BestBid − RPof f spr ) on the order book.
endsw
endsw
endsw
case Cancelation order
Cancel the oldest outstanding limit order.
endsw
endsw
endsw
case Seller
endsw
endsw
Table 7.3: Initial Parameters for Order Book based ASM I
Explanation Value
Initial Price price0 = 100
Tick Price δ = 0.01
Probability of Cancellation Order λc = 0.07
Probability of Market Order λm = 0.33
Probability of Limit Order λl = 0.60
Probability of Limit Order in Spread λin = 0.35
Probability of Limit Order Out of Spread λout = 0.65
Limit Price Tail Index 1 + α = 1.3
Order Size (µ, σ) ∼ (4.5, 0.8) ∗ 100 shares
Waiting Time τ = 6, 90
185
7.3.2 Simulator II
As in simulator I, an agent in simulator II is equally likely to generate a buy

order or a sell order at each time step, and this order can be a market order, or
a limit order, or a cancellation of a previous limit order, with probabilities λm ,
λl , and λc respectively. The sum of these probabilities is one (λm + λl + λc =
1). In simulator II, for a limit buy (sell) order, it has a probability of λinSpread
falling inside the bid-ask spread, a probability of λatBest falling at the best bid
(ask) price, and a probability of λinBook falling off the best bid (ask) price in the
book, (λinSpread +λatBest +λinBook = 1). The limit price inside the spread follows
a uniform distribution. The limit price off the best bid (ask) price follows a power
law distribution with the exponent of (1 + µ1 ). The log order size of a market
order follows a power law distribution with the exponent of (1 + µ2 ), while the
log order size of a limit order follows a power law distribution with the exponent
of (1 + µ3 ). The algorithm used in the simulation II is described in Algorithm 4.
In order to ensure that the order flows generated by the artificial market are
plausible, all the parameters in the model are derived from previous literature
(Chakraborti et al., 2011; Farmer et al., 2005a,b; Mike & Farmer, 2008; Toth
et al., 2009). The parameters used in the simulator II are presented in Table 7.4.
186
Algorithm 4 Artificial Market Simulator II
Generate a random number from a uniform distribution (0, 1) and determine the agent type according to the
probability 0.5.
switch Agent do
case Buyer
Generate a random number from a uniform distribution (0, 1) and determine the order type according
to the probabilities λm , λl and λc .
switch Order Type do
case Market order
Execute a market order with the order size generated from a power-law distribution with
exponent µ2
endsw
case Limit order
Generate a random number from a uniform distribution (0, 1) and determine the limit
order type according to probabilities λinSpread , λatBest , andλinBook .
switch Limit Order Type do
case Inside-spread Limit Order
Generate a random value denoted as Pinspr from a uniform distribution
(BestBid, BestAsk).
Place a limit order at the price Pinspr on the order book with the order size
generated from a power-law distribution with exponent µ3 .
endsw
case Spread Limit Order
Place a limit order at the best bid price on the order book with the order size
generated from a power-law distribution with exponent µ3 .
endsw
case Off-spread Limit Order
with exponent µ1 .
Place a limit order at the price (BestBid − RPof f spr ) on the order book
with the order size generated from a power-law distribution with exponent µ3 .
endsw
endsw
endsw
case Cancelation order
Cancels the oldest outstanding limit order that the buyer previously submitted.
endsw
endsw
endsw
case Seller
endsw
endsw
187
Table 7.4: Initial Parameters for Artificial Limit Order Market II
Explanation Value
Initial Price price0 = 50
Tick Price δ = 0.01
Probability of Order Cancellation λc = 0.34
Probability of Market Order λm = 0.16
Probability of Limit Order λl = 0.50
Probability of Limit Order in Spread λinSpread = 0.32
Probability of Limit Order at Best Quote λatBest = 0.33
Probability of Limit Order off the Best Quote λinBook = 0.35
Limit Price Power Law Exponent 1 + µ1 = 2.5
Market Order Size Power Law Exponent 1 + µ2 = 2.7
Limit Order Size Power Law Exponent 1 + µ3 = 2.1
188
7.4 Results and Discussions
Results of the three experiments are presented and discussed as below.
7.4.1 Results of Experiment One
The Combined-I strategy was evolved using the parameters presented in Table 7.5,
and then was evaluated in the artificial market for 30 simulation runs. Its VWAP
ratio is shown in Table 7.6. The VWAP Ratio reveals the difference between the
volume weighted execution price of Combined-I orders and the average traded
price of all orders during the whole simulation time. The better strategies have
smaller VWAP ratios.
Table 7.5: Parameters for Genetic Algorithm in Experiment One
Parameter Value
Population size 30
Maximum number of generation 100
Generation gap 0.8
Crossover rate 0.75
Mutation rate 0.05
Selection method Stochastic Universal Sampling
Crossover method Single-Point
The VWAP ratio of simple market order strategy (SM), which was evaluated
in the artificial market for 30 simulation runs, is also presented in Table 7.6. In
order to analyze the Combined-I strategy, the execution types of the 30 orders
are also recorded in the experiment. The three execution types are explained as
follows.
1. MLO. It represents marketable limit order. If the order is placed at or be-

yond the best price on the opposite side of the order book, and gets executed
immediately at its limit price, it is denoted as MLO.
189
2. LO. It represents limit order. If the order is executed against market orders
with opposite order sign some time after its submission, it is denoted as LO.
3. Crossed. If the order is not executed in the required time period but gets
executed at the end of the period for trading completeness, it is denoted as
Crossed.
Table 7.6: Results of Experiment One.
SM Strategy Combined-I Strategy

VWAP Ratio VWAP Ratio TradedOrderType
(10−3 ) (10−3 ) MLO LO Crossed
Buy Order 5.9748 0.5146 11 8 11
Sell Order 3.2378 -1.8244 13 7 10
Note: SM Strategy means Simple Market Order Strategy.
MLO is short for Marketable Limit Order.
LO is short for Limit Order.
Crossed represents Crossed Limit Order.
From Table 7.6, the Combined-I strategy outperforms the SM strategy signif-
icantly, both in-sample and out-of-sample, which is consistent with the results in
Lim & Coggins (2005b). Table 7.6 shows that the Combined-I strategy, which
has more orders executed via LOs, has a smaller VWAP ratio, meaning better per-
formance. All the Combined-I strategies with negative VWAP ratios have more
orders executed in the way of LO than those executed in the two other ways, ex-
cept the best out-of-sample strategy in Table 7.6. Also, Combined-I strategies
have achieved better VWAP than that of the whole simulation time for buy and
sell in in-sample test, which is showed by the negative values of VWAP ratios.
This is more significant for the sell order.
190
7.4.2 Results of Experiment Two
The Combined-II strategy was evolved using the parameters presented in Table
7.7, and then was evaluated in the artificial market for 240 simulation runs. Its
VWAP ratio is shown in Tables 7.8 & 7.9. The VWAP Ratio reveals the differ-
ence between the volume weighted execution price of Combined-II orders and the
average traded price of all orders during the whole simulation time. The better
strategies have smaller VWAP ratios.
Table 7.7: Parameters for Grammatical Evolution in Experiment Two
Parameter Value
Population size 100
Maximum number of generation 40
Generation gap 0.8
Crossover rate 0.9
Mutation rate 0.01
Selection method Roulette
The results (all out of sample) of buy strategies and sell strategies are provided
in Tables 7.8 & 7.9. The “S-T” represents short-term lifetime and the “L-T” repre-
sents long-term lifetime. The “Mean” is the average VWAP ratio of each strategy
over the 240 days, and “S.D.” represents the standard deviation of the average
(daily) VWAP ratio. P-values for the null hypothesis H1 : meanSA ≤ meanGE ,
H2 : meanSM ≤ meanGE , H3 : meanSP ≤ meanGE are also shown in the table,
to indicate the degree of statistical significance of the performance improvement
of Combined-II strategies over the two simple strategies. The figures show that
the null hypotheses are rejected at the ≤ 0.01 level.
Based on the results, Combined-II strategies notably outperform the three
benchmark strategies, a simple aggressive limit order strategy (SA), a simple mod-
est limit order strategy (SM) and a simple passive limit order strategy (SP). The
191
performances of these four strategies can be described as follows.
P erf ormanceCombined−II > P erf ormanceSP
> P erf ormanceSA > P erf ormanceSM
Combined-II strategies perform the best, while SM strategies perform the

worst. The negative VWAP ratios show that the Combined-II strategies achieve
better execution prices than the average execution price of the market. The values
of standard deviation show that the performance of Combined-II strategies is more
stable than most of the other three kinds of strategies over the tested trading days,
and the performance of SM strategies is more volatile than the other three kinds
of strategies. Comparing the performance of the strategies for buy and sell orders,
it can be observed that the performances of sell strategies are better than those of
buy strategies in most cases, which means that trading costs of buys are higher
than those of sells. This asymmetry is consistent with previous empirical find-
ings, for example, see Kraus & Stoll (1972), Keim & Madhavan (1997), Bikker
et al. (2007). And it is observed that L-T strategies all perform better than S-T
strategies, which indicate that strategies with longer lifetime can achieve better
execution prices than those with short lifetime.
192
Table 7.8: Results of Experiment Two (Buy Orders)
SA Strategy SM Strategy SP Strategy Combined-II Strategy

Mean (S.D.) Mean (S.D.) Mean (S.D.) Mean (S.D.) H1 H2 H3
S-T 14.7 (1.74%) 66.25 (2.53%) 13.19 (1.76%) -2.35 (1.52%) 0.01 0.00 0.01
L-T 5.14 (1.69%) 60.7 (2.03%) 9.37 (1.48%) -5.86 (1.3%) 0.01 0.00 0.00
193
Note: SA Strategy represents Simple Aggressive limit order Strategy.
SM Strategy represents Simple Moderate limit order Strategy.
SP Strategy represents Simple Passive limit order Strategy.
S.D. is short for Standard Deviation.
S-T is short for Short-Term lifetime.
L-T is short for Long-Term lifetime.
Table 7.9: Results of Experiment Two (Sell Orders)
SA Strategy SM Strategy SP Strategy Combined-II Strategy

Mean (S.D.) Mean (S.D.) Mean (S.D.) Mean (S.D.) H1 H2 H3
S-T 24.7 (1.86%) 43.6 (2.18%) 6.78 (2.03%) -8.15 (1.87%) 0.00 0.00 0.01
L-T 5.28 (1.56%) 37.23 (2.27%) 4.91 (1.39%) -12.07 (1.53%) 0.00 0.00 0.01
194
Note: SA Strategy represents Simple Aggressive limit order Strategy.
SM Strategy represents Simple Moderate limit order Strategy.
SP Strategy represents Simple Passive limit order Strategy.
S-T is short for Short-Term lifetime.
L-T is short for Long-Term lifetime.
7.4.3 Results of Experiment Three
The Combined-III strategy was evolved using the parameters presented in Table
7.10, and then was evaluated in the artificial market for 240 simulation runs. Its
VWAP ratio is shown in Table 7.11. The VWAP Ratio reveals the difference
between the volume weighted execution price of Combined-III orders and the
average traded price of all orders during the whole simulation time. The better
strategies have smaller VWAP ratios.
Table 7.10: Parameters for Grammatical Evolution in Experiment Three
Parameter Value
Population size 100
Maximum number of generations 40
Generation gap 0.8
Crossover rate 0.9
Mutation rate 0.01
Selection method Roulette
The results (all out of sample) of buy strategies and sell strategies are pro-
vided in Table 7.11. The “Mean” is the average VWAP ratio of each strategy over
the 240 days, and “S.D.” represents the standard deviation of the average (daily)
VWAP ratio. P-values for the null hypothesis H1 : meanSM ≤ meanCombined−III
and H2 : meanSL ≤ meanCombined−III are also shown in the table, to indicate the
degree of statistical significance of the performance improvement of Combined-III
strategies over the two simple strategies. The figures show that the null hypotheses
are rejected at the ≤ 0.01 level.
Based on the results, Combined-III strategies notably outperform the two bench-
mark strategies, a simple market order strategy (SM) and a simple limit order strat-
egy (SL). The negative VWAP ratios of -1.42 and -23.21 show that the GE evolved
strategies achieve better execution prices than the average execution price of the
195
Table 7.11: Results of Experiment Three
SM Strategy SL Strategy Combined-III Strategy

Mean (S.D.) Mean (S.D.) Mean (S.D.) H1 H2
Buy Order 69.64 (0.42%) 42.54 (1.45%) -1.42 (0.49%) 0.00 0.01
Sell Order 68.73 (0.36%) 13.81 (1.59%) -23.21 (0.48%) 0.00 0.01
Note: SM Strategy represents Simple Market order Strategy.

SL Strategy represents Simple Limit order Strategy.
market. The small standard deviations of 0.49 and 0.48 indicate that Combined-III
strategy is robust over the tested trading days. The performance of SL strategies
seems more volatile than those of SM strategies and Combined-III strategies, and
the performance of SM strategies is more stable than those of the other two kinds
of strategies. Comparing the performance of the strategies for buy and sell orders,
it can be observed that the performances of sell strategies are better than those of
buy strategies, which means that trading costs of buys are higher than those of
sells. This asymmetry is consistent with previous empirical findings.
7.4.4 Discussions
The results in the three experiments all show that the combined strategies out-
perform the simple strategies. This indicates that order choice affects the perfor-
mance of trade execution strategies, that is the price impact of trading large-orders,
and suggests that order choice is an important decision in large-order trade execu-
tion.
In the above experiments, both buy and sell strategies are tested. The results
of these experiments all show that sell strategies generally perform better than
buy strategies, confirming the asymmetric price impacts of institutional trades
found in previous literature (Bikker et al., 2007; Chan & Lakonishok, 1993, 1995;
196
Gemmill, 1996; Holthausen et al., 1987, 1990; Keim & Madhavan, 1995b, 1996,
1997; Kraus & Stoll, 1972).
Several explanations appear in the literature to account for this asymmetry
phenomenon. Chan & Lakonishok (1993) and Keim & Madhavan (1996) argue
that sells are more often liquidity-motivated rather than information-based than
buys. Buys create new long-term positions and thus imply a preference to hold
a particular stock. Saar (2001) provides a different explanation for the price im-
pact asymmetry. He builds a theoretical model which demonstrates how the price
impact asymmetry can arise. The main implication of the model is that the his-
tory of price performance of a stock affects the degree of asymmetry: the longer
the run of price appreciations, the less positive the difference in permanent price
impact between buys and sells. When the price run-up is long enough, sells may
even have higher price impact than buys. Another explanation for the price impact
asymmetry is given by Chiyachantana et al. (2004). They find that the asymmetry
depends on particular market conditions: price effects of buyer-initiated trades are
greater in bull markets (as in 1997-1998) whereas those of seller-initiated trades
are larger in bear markets (as in 2001). Boscuk & Lasfer (2005) take a different
view and show that the type of investor and the combination of the size of the
trade and the trader’s resulting level of ownership are the major determinants of
price impact asymmetry at the London Stock Exchange.
7.5 Summary
Order choice problem and large-order trading problem have been studied sepa-
rately in the market microstructure literature. However, large order trading with
various order types has not been investigated. This chapter sheds light on large-
order traders’ use of limit orders. It examined whether order choice affects the
197
price impact of trading a large order. Different combinations of order types were
investigated. Agent-based artificial markets were simulated in order to evaluate
the trading strategies. One advantage of this performance evaluation approach is
that the simulated market can capture the dynamic effect of trading while histori-
cal data can not.
The results show that the combined-order-type strategies perform better than
the simple-order-type strategies in terms of price impact. This confirms the finding
in Keim & Madhavan (1997) that the investment style affects trading costs. This
study also demonstrates that buys are more expensive than sells, which is consis-
tent with the findings in Bikker et al. (2007); Chan & Lakonishok (1993, 1995);
Gemmill (1996); Holthausen et al. (1987, 1990); Keim & Madhavan (1995b,
1996, 1997); Kraus & Stoll (1972).
This chapter compliments the prior literature on trade execution strategies,
providing evidence that order choice plays an important role in determining the
price impact when trading a large order. The findings in this chapter are of interest
to academics. When modelling the behaviours of large-order traders, order choice
as an import factor should be taken into account. Moreover, the results provide
useful insights for institutional traders who are obligated to liquidate or acquire
a big position, suggesting that large-order execution strategies adopting multiple
order types can perform better than those using single order type.
In addition, this chapter contributes to the literature on agent-based modellings
of financial markets. It makes the first attempt to use agent-based modelling to
evaluate the performance of trade execution strategies. The artificial market model
developed in this study can be extended to examine the effect of high-frequency
trading on market quality.
198
Part III
Conclusions
199
Chapter 8
Conclusions and Future Work
This chapter begins with a brief summary of the thesis and the contributions it has
made. The limitations of this thesis are highlighted in Section 8.2. Suggestions
for future work complete the chapter in Section 8.3.
8.1 Thesis Summary

This thesis empirically investigated the effect of trading on price, that is price
impact. Understanding price impact is a crucial task in financial markets. In order
to get a better understanding of this effect, the following questions were asked:
1. Does price impact on the LSE and the NYSE exhibit an intraday pattern?
2. Does agent intelligence affect the magnitude of price impact?
3. Does order choice affect the price impact of trading large orders?
Empirical analysis on intraday behaviours of market liquidity and price impact

were carried out in Chapter 5. The analysed data was drawn from the NYSE-
Euronext TAQ database and the LSE ROB database. Before using the data for
200
analysis, several steps were implemented to preprocess the data in Chapter 4. The
entries in the TAQ data recorded during non-normal trading periods were deleted,
as well as the entries with non-positive values. Identified outliers in the TAQ
data were also removed. An algorithm was developed for inferring the hidden
information contained in the ROB data. Then the limit order book for each trading
day was rebuilt using the ROB data. With the clean data, six stocks selected from
the US markets and six stocks selected from the UK markets were analysed. The
results on intraday patterns generally confirm those found in the previous studies
as reviewed in Chapter 2. In particular, intraday price impact is found to be reverse
S-shaped for US and UK stocks.
The study in Chapter 6 addressed the second research question. As motivated
by the argument in Farmer et al. (2004) that traders condition their order sizes
based on market liquidity, this study investigated whether agent intelligence af-
fects the magnitude of price impact. An agent-based modelling approach was
adopted, as this method facilitates the isolation of a relevant factor in the market
simulation. An artificial limit order market composed of zero-intelligence agents
was simulated. It captured the salient features of electronic limit order markets
including continuous trading, a visible book of orders, price-time priority rules,
instantaneous trade reporting rules, order cancellation capabilities, and both limit
order and market order functionality. In addition, the artificial market was cali-
brated using the LSE ROB data. The results showed that the price impact observed
in the artificial market was generally larger than that observed in the real market,
indicating that agent intelligence is an important determinant of the magnitude of
price impact.
Price impact is an significant component of trading cost, and thus controlling
price impact is important for improving execution performance. The study in
Chapter 7 examined whether order choice affects the performance of large-order
201
execution. Several combined strategies, which used various types of orders, were
developed and their performances was then evaluated using price impact. The
results were benchmarked against simple strategies which adopted a single or-
der type. The results show that the combined strategies outperformed the simple
strategies, indicating that order choice is important in controlling price impact for
large-order execution.
Arising from this thesis, these series of experiments help our understanding
of price impact. It identifies the important intraday behaviours of price impact,
the crucial role of agent intelligence in determining the magnitude of price im-
pact, and the significance of order choice in determining the price impact of trad-
ing large orders. These findings have further implications for research on market
microstructure, for example, understanding how securities markets function and
developing comprehensive models of market microstructure. Moreover, they are
also of interest to institutional traders and trading desks.
8.1.1 Contributions
Through the gathering together of the literature on market microstructure and the
literature on agent-based financial markets and the subsequent implementation of
a series of experiments in the domain, a number of key insights are generated.
This section reviews the contributions of this thesis as follows.
• Up-to-date review of relevant literature
Chapters 2 & 3 provided an up-to-date survey of relevant literature on mar-

ket microstructure, price impact, and agent-based financial markets. Chap-
ter 2 initially gave an overview of the literature of market microstructure,
and then specifically reviewed several streams of market microstructure re-
search relevant to this thesis, including studies on the relationship between
202
price impact and trade size, studies on large price impact, studies of intraday
phenomena on price volatility, bid-ask spread, trading volume and market
depth, studies on optimal trading strategies for a single asset, and studies on
order submission strategies. This chapter uncovered several research gaps,
which led to the experimental studies of this thesis. Chapter 3 reviewed
various applications of agent-based modelling to financial markets, and re-
vealed that there had been little attempt to use agent-based modelling for
investigating price impact.
• Analysis of intraday behaviours of price impact in UK and US markets
As some studies (Ahn & Cheung, 1999; Lee et al., 1993; Li et al., 2005; Vo,
2007) have documented an intraday pattern on market depth, price impact
may exhibit an intraday pattern because price impact is affected by market
liquidity. This thesis analysed intraday behaviours of price impact, using
the recent data drawn from the NYSE-Euronext TAQ database and the LSE
ROB database. It for the first time documented a reverse S-shaped intraday
pattern on price impact in UK and US markets. Moreover, it found that
price impact had a negative association with near market depth. The result
suggests that the information conveyed by price impact is consistent with
near market depth as indicators of intraday market liquidity.
• Up-to-date evidence on intraday behaviours of market liquidity in UK

and US markets
The previous studies on intraday phenomena are dated. This thesis anal-
ysed intraday behaviours of market liquidity, using recent data drawn from
the NYSE-Euronext TAQ database and the LSE ROB database, and pro-
vided up-to-date evidence on intraday behaviours of price volatility, bid-ask
spread, market depth, trading volume, trading frequency and trade size in
203
UK and US markets. Unlike previous studies on intraday phenomena which
only examined the market depth at the best level of the limit order book, this
thesis for the first time analyses the intraday behaviour of market depth for
up to ten levels of the limit order book in the UK market.
• Investigation of the role of agent intelligence in determining the magni-

tude of price impact
Prior studies on price impact typically found that price impact is a concave
function of trade size. Some of them (e.g. Farmer et al. (2004); Hopman
(2007); Weber & Rosenow (2006)) argued that this concavity was due to
selective liquidity taking, namely agents conditioning order size on market
liquidity. However, little study was conducted to investigate this. This thesis
for the first time investigated whether agent intelligence plays an important
role in determining the magnitude of price impact. The results support the
hypothesis that agent intelligence plays an important role in determining the
magnitude of price impact.
• Examination of the effect of order choice on the price impact when

trading a large order
Controlling price impact is an important issue in financial markets. Many

studies have been devoted to examining this problem. However, these stud-
ies only adopted market orders, and no studies were conducted to investigate
the effect of order choice on the price impact of trading a large order. This
thesis for the first time analysed this. The results demonstrate that order
choice plays an important role in determining the price impact of trading a
large order.
• Use of ABM for investigating price impact
204
Agent-based modelling have been applied to economics and finance for a
long time, but there has been little attempt to use ABM for investigating
price impact. This thesis makes the first attempt to investigate price impact
using ABM in its two experimental studies. It laid the framework for this
use, and the model developed here can be extended for other applications,
like investigation of the market effect of high-frequency trading.
8.2 Limitations of this Thesis

The primary focus of this thesis is on the investigation of price impact, and it
makes useful contributions to our understanding of price impact. However, one
thesis can not cover all aspects of this topic. There are a number of areas worthy
of further research that fall outside the scope of this work.
Price impact can be caused by market orders, limit orders and upstairs or-
ders. When examining the intraday behaviours of price impact, the focus of the
present work is on the price impact caused by market orders. A few studies have
investigated the price impact of other trading activities. For example, Eisler et al.
(2012); Hautsch & Huang (2012) examine the price impact caused by limit orders,
and Gabaix et al. (2006); Keim & Madhavan (1996) investigate the price impact
of upstairs orders.
A number of studies have shown that price impact decays over time. In the
current work, concentration is placed on the immediate effect of trading on prices.
Some studies have been devoted to modelling the resilience of price impact. For
example, Gatheral et al. (2011) model that price impact decays exponentially over
time.
Large-order execution can occur at several trading venues. In this thesis, at-
tention is limited to trading at one trading venue. Many traders are willing to
205
route their orders to the cheapest trading place in order to reduce trading costs. A
number of studies are focusing on this issue. For example, Foucault & Menkveld
(2008) empirically examines smart order routing systems, and Rawal (2010) dis-
cusses the importance of intelligent decision-making in order routings.
8.3 Future Work

Building upon this research, there are numerous avenues for future work. A few
of them are outlined as follows.
The study in Chapter 5 shows that market volume has a U-shaped intraday
pattern and near market depth has an S-shaped intraday pattern. It would be in-
teresting to determine who contributes to trading volume and market depths, that
is the constituents of the market participants at different times of the trading day.
Understanding whether institutional investors or individual investors dominate the
market when trading volume is high or when depth is high would help us better
understand the underlying drivers of the intraday patterns.
The price impact studied in Chapter 5 is limited to the price effect of market
orders. The price impact of limit orders is not explored here. A recent study by
Hautsch & Huang (2012) quantifies the price effect of submitting a limit order in a
limit order market and find that limit orders have significant impact. However, the
intraday behaviour of this impact is unknown. It would be interesting to analyse
the intraday patterns of price impact of limit orders using the method in Hautsch
& Huang (2012).
The study in Chapter 7 can be extended to an investigation of large-order trad-
ing in multiple markets. A number of artificial markets can be simultaneously
simulated with an agent-based modelling approach. The choice of trading venue
will be decided on for each order at the same time with other decisions like choice
206
of order type and order size. This investigation would shed light on how liquidity
is supplied and demanded across multiple linked markets.
207
Appendix A
Explanation of the Chosen Time

Period in Chapter 5
One of the aims of this thesis is to investigate whether price impact exhibits an
intraday pattern and whether this intraday pattern is consistent across different
stocks in US and UK markets. Therefore, the stocks are chosen from different
industries and with different market capitalisations. The chosen time period is
a normal period. During these time periods, there were no disturbances in the
markets. The chosen length of the time period is enough to generate intraday
patterns. Figures A.1 & A.2 illustrate the intraday curves of price volatility using
one-day data and one-month day separately.
208
−3
x 10
7
5
Absolute Log Return
0
9:30 10:30 11:30 12:30 13:30 14:30 15:30
Time
Figure A.1: Intraday Price Volatility (One-day data)
209
−3
x 10
9
6
Absolute Log Return
1
9:30 10:30 11:30 12:30 13:30 14:30 15:30
Time
Figure A.2: Intraday Price Volatility (One-month data)
210
Appendix B
More Results of Chapter 5
This appendix includes some results from experiment two in Chapter 5. This the-
sis measures market depth in share and in value as well as time-weighted market
depth for UK stocks. Chapter 5 includes the figures showing intraday pattern on
market depth in share. This appendix presents the intraday patterns on market
depth in value and time-weighted market depth. From Figures B.1-B.10, it can be
observed that the intraday patterns on market depth in value and on time-weighted
market depths are almost identical to the intraday patterns on market depth in share
as shown in Chapter 5.
211
Time−weighted Depth in Share Time−weighted Depth in Share
11000 3000
10500
2800
10000
2600
TW Depth (Share)
TW Depth (Share)
9500
2400
9000
2200
8500
2000
8000
7500 1800
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(a) BARC (b) BATS

Time−weighted Depth in Share Time−weighted Depth in Share
11000 14000
13000
10500
12000
10000 11000
TW Depth (Share)
TW Depth (Share)
10000
9500
9000
9000 8000
7000
8500
6000
8000 5000
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(c) BP (d) GSK

x 10
4 Time−weighted Depth in Share
Time−weighted Depth in Share 11
9000
10
8500
8000
TW Depth (Share)
TW Depth (Share)
7500
7000
6
6500
5
6000 4
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(e) HSBC (f) VOD
Figure B.1: Intraday Time-weighted Depth in Share (UK stocks)
212
6
x 10 Depth in Value 6
x 10 Depth in Value
3.2 6.5
3.1
6
3
2.9 5.5
Depth (Value)
Depth (Value)
2.8
5
2.7
2.6 4.5
2.5
4
2.4
2.3 3.5
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(a) BARC (b) BATS

6
x 10 Depth in Value
4.6 1.6
1.5
4.4
1.4
1.3
4.2
1.2
Depth (Value)
Depth (Value)
4 1.1
3.8
0.9
0.8
3.6
0.7
3.4 0.6
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(c) BP (d) GSK
6
x 10 Depth in Value
5.4 1.5
1.4
5.2
1.3
5
1.2
4.8
1.1
Depth (Value)
Depth (Value)
4.6 1
0.9
4.4
0.8
4.2
0.7
4
0.6
3.8 0.5
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(e) HSBC (f) VOD
Figure B.2: Intraday Depth in Value (UK stocks)
213
6
x 10 Time−weighted Depth in Value 6
x 10 Time−weighted Depth in Value
3.6 7
3.4 6.5
3.2 6
TW Depth (Value)
TW Depth (Value)
3 5.5
2.8 5
2.6 4.5
2.4 4
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(a) BARC (b) BATS

6
5 1.6
1.5
4.8
1.4
1.3
4.6
TW Depth (Value)
TW Depth (Value)
1.2
4.4 1.1
4.2
0.9
0.8
4
0.7
3.8 0.6
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(c) BP (d) GSK
6
6 1.6
5.8 1.5
1.4
5.6
1.3
5.4
TW Depth (Value)
TW Depth (Value)
1.2
5.2
1.1
5
1
4.8
0.9
4.6
0.8
4.4 0.7
4.2 0.6
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(e) HSBC (f) VOD
Figure B.3: Intraday Time-weighted Depth in Value (UK stocks)
214
6
x 10 Bid Depth in Value x 10
6 Bid Depth in Value
3.1 6.5
3
6
2.9
2.8 5.5
Bid Depth (Value)
Bid Depth (Value)

2.7
5
2.6
2.5 4.5
2.4
4
2.3
2.2 3.5
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(a) BARC (b) BATS

6
x 10 Bid Depth in Value x 10
7 Bid Depth in Value
4.6 1.5
1.4
4.4
1.3
1.2
4.2
Bid Depth (Value)
Bid Depth (Value)
1.1
4 1
0.9
3.8
0.8
0.7
3.6
0.6
3.4 0.5
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(c) BP (d) GSK
6
x 10 Bid Depth in Value 6
x 10 Bid Depth in Value
5.4 14
13
5.2
12
5
11
4.8
Bid Depth (Value)
Bid Depth (Value)
10
4.6
9
4.4
8
4.2
7
4 6
3.8 5
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(e) HSBC (f) VOD
Figure B.4: Intraday Bid Depth in Value (UK stocks)
215
6
x 10 Ask Depth in Value 6
x 10 Ask Depth in Value
3.2 7
3.1
6.5
6
2.9
Ask Depth (Value)
Ask Depth (Value)

5.5
2.8
2.7
5
2.6
4.5
2.5
4
2.4
2.3 3.5
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(a) BARC (b) BATS

6
4.6 1.5
1.4
4.4
1.3
4.2 1.2
Ask Depth (Value)
Ask Depth (Value)
1.1
4
1
3.8 0.9
0.8
3.6
0.7
3.4 0.6
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(c) BP (d) GSK
6
5.4 1.5
1.4
5.2
1.3
5
1.2
4.8
Ask Depth (Value)
Ask Depth (Value)
1.1
4.6 1
0.9
4.4
0.8
4.2
0.7
4
0.6
3.8 0.5
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(e) HSBC (f) VOD
Figure B.5: Intraday Ask Depth in Value (UK stocks)
216
x 10
5 Time−weighted Near Depth in Share 4
x 10 Time−weighted Near Depth in Share
1.6 7
6.5
1.5
1.4
5.5
TW Near Depth (Share)

5
1.3
4.5
1.2
4
3.5
1.1
1
2.5
0.9 2
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(a) BARC (b) BATS
x 10
1.8 3
1.7
2.5
1.6
1.5
2
1.4
1.3
1.5
1.2
1.1
1
0.9 0.5
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(c) BP (d) GSK

x 10 2.2
1.35
1.3 2
1.25 1.8
1.2
1.6
1.15
1.4
1.1
1.2
1.05
1
1
0.8
0.95
0.9 0.6
0.85 0.4
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(e) HSBC (f) VOD
Figure B.6: Intraday Time-weighted Near Depth in Share (UK stocks)
217
7
x 10 Near Depth in Value 8
x 10 Near Depth in Value
5 1.6
1.4
4.5
1.2
Near Depth (Value)
Near Depth (Value)

4
3.5
0.8
3
0.6
2.5 0.4
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(a) BARC (b) BATS

7
8 3.5
7.5
3
2.5
6.5
Near Depth (Value)
Near Depth (Value)
6 2
5.5
1.5
1
4.5
4 0.5
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(c) BP (d) GSK
7
9 3
8.5
2.5
8
7.5
Near Depth (Value)
Near Depth (Value)
1.5
6.5
6
1
5.5
5 0.5
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(e) HSBC (f) VOD
Figure B.7: Intraday Near Depth in Value (UK stocks)
218
7
x 10 Time−weighted Near Depth in Value 8
x 10 Time−weighted Near Depth in Value
5 1.6
1.4
4.5
1.2
TW Near Depth (Value)

4
3.5
0.8
3
0.6
2.5 0.4
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(a) BARC (b) BATS

7
8.5 3.5
8
3
7.5
7 2.5
6.5
2
6
5.5 1.5
5
1
4.5
4 0.5
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(c) BP (d) GSK
7
9 3
8.5
2.5
8
2
7.5
7
1.5
6.5
5.5 0.5
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(e) HSBC (f) VOD
Figure B.8: Intraday Time-weighted Near Depth in Value (UK stocks)
219
7
x 10 Bid Near Depth in Value 8
x 10 Bid Near Depth in Value
5 1.6
1.4
4.5
1.2
Bid Near Depth (Value)

4
3.5
0.8
3
0.6
2.5 0.4
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(a) BARC (b) BATS

7
8.5 3.5
8
3
7.5
7 2.5
6.5
2
6
5.5 1.5
5
1
4.5
4 0.5
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(c) BP (d) GSK
7
9 3
8.5
2.5
8
7.5
2
1.5
6.5
6
1
5.5
5 0.5
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(e) HSBC (f) VOD
Figure B.9: Intraday Near Bid Depth in Value (UK stocks)
220
7
x 10 Ask Near Depth in Value 8
x 10 Ask Near Depth in Value
5 1.6
1.4
4.5
1.2
Ask Near Depth (Value)

4
3.5
0.8
3
0.6
2.5 0.4
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(a) BARC (b) BATS

7
8 3.5
7.5
3
2.5
6.5
6 2
5.5
1.5
1
4.5
4 0.5
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(c) BP (d) GSK
7
9 3
8.5
2.5
8
7.5
2
1.5
6.5
6
1
5.5
5 0.5
8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 8:00 9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00
Time Time
(e) HSBC (f) VOD
Figure B.10: Intraday Near Ask Depth in Value (UK stocks)
221
Bibliography
Abhyankar, A., Ghosh, D., Levin, E. and Limmack, R. (1997). ‘Bid-ask spreads,
trading volume and volatility: Intra-day evidence from the London Stock Ex-
change’, Journal of Business Finance and Accounting, Vol. 24, No. 3, pp. 343–
362.
Admati, A. and Pfleiderer, P. (1988). ‘A theory of intraday patterns: volume and

price variability’, Review of Financial Studies, Vol. 1, No. 1, pp. 3–40.
Ahn, H., Bae, K. and Chan, K. (2001). ‘Limit Orders, Depth, and Volatility:
Evidence from the Stock Exchange of Hong Kong’, Journal of Finance, Vol. 56,
No. 2, pp. 767–788.
Ahn, H. and Cheung, Y. (1999). ‘The intraday patterns of the spread and depth in
a market without market makers: The Stock Exchange of Hong Kong’, Pacific-
Basin Finance Journal, Vol. 7, No. 5, pp. 539–556.
Aitken, M., Almeida, N., Harris, F. and McInish, T. (2007). ‘Liquidity supply in
electronic markets’, Journal of Financial Markets, Vol. 10, No. 2, pp. 144–168.
Aitken, M., Berkman, H. and Mak, D. (2001). ‘The use of undisclosed limit orders
on the Australian Stock Exchange’, Journal of Banking and Finance, Vol. 25,
No. 8, pp. 1589–1603.
222
Aitken, M., Kua, A., Brown, P., Watter, T. and Izan, H. (1995). ‘An Intraday Anal-
ysis of the Probability of Trading on the ASX at the Asking Price’, Australian
Journal of Management, Vol. 20, No. 2, pp. 115–154.
Al-Suhaibani, M. and Kryzanowski, L. (2000). ‘An exploratory analysis of the

order book, and order flow and execution on the Saudi stock market’, Journal
of Banking and Finance, Vol. 24, No. 5, pp. 1323–1357.
Alfonsi, A., Fruth, A. and Schied, A. (2010). ‘Optimal execution strategies in

limit order books with general shape functions’, Quantitative Finance, Vol. 10,
No. 2, pp. 143–157.
Alfonsi, A. and Schied, A. (2010). ‘Optimal Trade Execution and Absence of

Price Manipulations in Limit Order Book Models’, SIAM J. on Financial Math-
ematics, Vol. 1, No. 3, pp. 490–522.
Allen, F. and Karjalainen, R. (1999). ‘Using genetic algorithms to find technical

trading rules’, Journal of Financial Economics, Vol. 51, No. 3, pp. 245–271.
Almgren, R. (2003). ‘Optimal execution with nonlinear impact functions and

trading-enhanced risk’, Applied Mathematical Finance, Vol. 10, No. 1, pp. 1–
18.
Almgren, R. (2008). ‘Execution costs’, in R. Cont (ed.), Encyclopedia of Quanti-

tative Finance, pp. 1–5, Wiley, UK.
Almgren, R. and Chriss, N. (1999). ‘Value under liquidation’, Journal of Risk,

Vol. 12, No. 12, pp. 61–63.
Almgren, R. and Chriss, N. (2000). ‘Optimal execution of portfolio transactions’,

Journal of Risk, Vol. 3, No. 2, pp. 5–39.
223
Aloud, M. and Tsang, E. (2011). ‘Modelling the trading behaviour in high-
frequency markets’, in Proceedings of the 3rd Computer Science and Electronic
Engineering Conference, pp. 7–12.
Anand, A., Chakravarty, S. and Martell, T. (2005). ‘Empirical evidence on the

evolution of liquidity: Choice of market versus limit orders by informed and
uninformed traders’, Journal of Financial Markets, Vol. 8, No. 3, pp. 289–309.
Arifovic, J. (1995). ‘Genetic algorithms and inflationary economies’, Journal of

Monetary Economics, Vol. 36, No. 1, pp. 219–243.
Arifovic, J. (1996). ‘The Behavior of the Exchange Rate in the Genetic Algorithm
and Experimental Economies’, Journal of Political Economy, Vol. 104, No. 3,
pp. 510–541.
Arthur, W., Holland, J., LeBaron, B., Palmer, R. and Tayler, P. (1997). ‘Asset pric-
ing under endogenous expectations in an artificial stock market’, in W. Arthur,
S. Durlauf, & D. Lane (eds.), The Economy as an Evolving Complex System II,
Addison-Wesley.
Atkins, A. and Basu, S. (1995). Journal of Business Finance and Accounting,

Vol. 22, No. 6, pp. 789–809.
Bae, K., Jang, H. and Park, K. (2003). ‘Traders’ choice between limit and market
orders: evidence from NYSE stocks’, Journal of Financial Markets, Vol. 6,
No. 4, pp. 517–538.
Bagehot, W. (1971). ‘The only game in town’, Financial Analysts Journal,

Vol. 27, No. 2, pp. 12–14.
Bak, P., Paczuski, M. and Shubik, M. (1997). Physica A: Statistical Mechanics

and its Applications, Vol. 246, No. 3-4, pp. 430–453.
224
Barclay, M. and Warner, J. (1993). ‘Stealth trading and volatility : Which trades
move prices?’, Journal of Financial Economics, Vol. 34, No. 3, pp. 281–305.
Barndorff-Nielsen, O., Hansen, P., Lunde, A. and Shephard, N. (2009). ‘Realized

kernels in practice: trades and quotes’, Econometrics Journal, Vol. 12, No. 3,
pp. 1–32.
Bartolozzi, M. (2010). ‘A Multi Agent Model for the Limit Order Book Dynam-
ics’, European Physical Journal B, Vol. 78, No. 2, pp. 265–273.
Beber, A. and Caglio, C. (2005). ‘Order Submission Strategies and

Information: Empirical Evidence from the NYSE’. Available at:
http://ideas.repec.org/p/fam/rpseri/rp146.html.
Becker, G. (1962). ‘Irrational Behavior and Economic Theory’, Journal of Politi-

cal Economy, Vol. 70, No. 1, pp. 1–13.
Berkowitz, S., Logue, D. and Noser, E. (1988). ‘The total cost of transactions on
the NYSE’, Journal of Finance, Vol. 43, No. 1, pp. 97–112.
Bertsimas, D. and Lo, A. (1998). ‘Optimal control of execution costs’, Journal of

Financial Markets, Vol. 1, No. 1, pp. 1–50.
Bessembinder, H. (2003). ‘Issues in assessing trade execution costs’, Journal of

Financial Markets, Vol. 6, No. 3, pp. 233–257.
Bessembinder, H., Panayides, M. and Venkataraman, K. (2009). ‘Hidden liq-

uidity: An analysis of order exposure strategies in electronic stock markets’,
Journal of Financial Economics, Vol. 94, No. 3, pp. 361–383.
Biais, B., Hillion, P. and Spatt, C. (1995). ‘An empirical analysis of the limit order
book and the order flow in the Paris Bourse’, Journal of Finance, Vol. 50, No. 3,
pp. 1655–1689.
225
Bikker, J., Spierdijk, L. and Sluis, P. (2007). ‘Market Impact Costs of Institutional
Equity Trades’, Journal of International Money and Finance, Vol. 1, No. 27, pp.
974–1000.
Black, F. (1971). ‘Toward a Fully Automated Stock Exchange, Part I’, Financial
Analysts Journal, Vol. 27, No. 4, pp. 28–35.
Bloomfield, R., O’Hara, M. and Saar, G. (2005). ‘The ”make or take” decision
in an electronic market: Evidence on the evolution of liquidity’, Journal of
Financial Economics, Vol. 75, No. 2, pp. 165–199.
Blume, M. and Goldstein, M. (1997). ‘Quotes, Order Flow, and Price Discovery’,
Journal of Finance, Vol. 52, No. 1, pp. 221–244.
Bollerslev, T. and Domowitz, I. (1993). ‘Some effects of restricting the electronic

order book in an automated trade execution system’, in D. Friedman & J. Rust
(eds.), The Double Auction Market: Institutions, Theories, and Evidence, pp.
221–252, Perseus Publishing.
Bongiovanni, S., Borkovec, M. and Sinclair, R. (2006). ‘Let’s Play Hide-and-

Seek: The Location and Size of Undisclosed Limit Order Volume’, Journal of
Trading, Vol. 1, No. 3, pp. 34–46.
Boscuk, A. and Lasfer, M. (2005). ‘The information content of institutional trades

on the London stock exchange’, Journal of Financial and Quantitative Analy-
sis, Vol. 40, No. 3, pp. 621–644.
Bouchaud, J., Farmer, J. and Lillo, F. (2009). ‘How markets slowly digest changes
in supply and demand’, in T. Hens & K. Schenk-Hoppe (eds.), Handbook of
Financial Markets: Dynamics and Evolution, pp. 57–160, North Holland, US.
226
Bouchaud, J., Gefen, Y., Potters, M. and Wyart, M. (2004). ‘Fluctuations and
response in financial markets: The subtle nature of ‘random’ price changes’,
Quantitative Finance, Vol. 4, No. 2, pp. 57–62.
Bouchaud, J., Mezard, M. and Potters, M. (2002). ‘Statistical properties of stock

order books: empirical results and models’, Quantitative Finance, Vol. 2, No. 4,
pp. 251–256.
Brabazon, A. and O’Neill, M. (2006). Biologically Inspired Algorithms for Fi-

nancial Modeling, Spring-Verlag, Berlin.
Brabazon, A., O’Neill, M. and Dempsey, I. (2008). ‘An Introduction to Evolu-

tionary Computation in Finance’, IEEE Computational Intelligence Magazine,
Vol. 3, No. 4, pp. 42–55.
Brock, W. and Kleidon, A. (1992). ‘Periodic market closure and trading volume:
A model of intraday bids and asks’, Journal of Economic Dynamics and Con-
trol, Vol. 16, No. 3, pp. 451–489.
Brownlees, C. and Gallo, G. (2006). ‘Financial Econometric Analysis at Ultra-

high Frequency: Data Handling Concerns’, Computational Statistics and Data
Analysis, Vol. 51, No. 4, pp. 2232–2245.
Cai, C., Hudson, R. and Keasey, K. (2004). ‘Intraday bid-ask spreads, trading
volume and volatility: recent empirical evidence from the London Stock Ex-
change’, Journal of Business Finance and Accounting, Vol. 31, No. 5, pp. 647–
676.
Cao, C., Hansch, O. and Wang, X. (2008). ‘Order Placement Strategies In A Pure
Limit Order Book Market’, Journal of Financial Research, Vol. 26, No. 2, pp.
113–140.
227
Cao, C., Hansch, O. and Wang, X. (2009). ‘The information content of an open
limit order book’, Journal of Futures Markets, Vol. 29, No. 1, pp. 16–41.
CFA (2009). ‘Market Microstructure: The Impact of Fragmentation under the

Markets in Financial Instruments Directive’, CFA Institute Publications, Vol.
2009, No. 13, pp. 1–60.
Chaboud, A., Chiquoine, B., Hjalmarsson, E. and Vega, C. (2009). ‘Rise of the
machines: algorithmic trading in the foreign exchange market’. Working paper,
Board of Governors of the Federal Reserve System (US).
Chakraborti, A., Toke, I., Patriarca, M. and Abergel, F. (2011). ‘Econophysics

review: I. Empirical facts’, Quantitative Finance, Vol. 11, No. 7, pp. 991–1012.
Chakravarty, S. and Holden, C. (1995). ‘An integrated model of market and limit
orders’, Journal of Financial Intermediation, Vol. 4, No. 3, pp. 213–241.
Challet, D. and Stinchcombe, R. (2001). ‘Analyzing and modeling 1+ 1d markets’,

Physica A, Vol. 300, No. 2, pp. 285–299.
Chan, K., Christie, W. and Schultz, P. (1995). ‘Market structure and the intra-
day pattern of bid-ask spreads for NASDAQ securities’, Journal of Business,
Vol. 68, No. 1, pp. 35–60.
Chan, L. and Lakonishok, J. (1995). ‘The behavior of stock prices around institu-
tional trades’, Journal of Finance, Vol. 50, No. 5, pp. 1147–1174.
Chan, L. and Lakonishok, J. (1997). ‘Institutional Equity Trading Costs: NYSE

versus NASDAQ’, Journal of Finance, Vol. 52, No. 2, pp. 713–735.
Chan, L. K. and Lakonishok, J. (1993). ‘Institutional trades and intraday stock

price behavior’, Journal of Financial Economics, Vol. 33, No. 3, pp. 173–199.
228
Chan, N., LeBaron, B., Lo, A. and Poggio, T. (2001). ‘Agent-based models of
financial markets: A comparison with experimental markets’. Working paper,
MIT.
Chan, Y. (2000). ‘The price impact of trading on the stock exchange of Hong
Kong’, Journal of Financial Markets, Vol. 3, No. 1, pp. 1–16.
Chavez-Demoulin, V. and McGill, J. (2012). ‘High-frequency financial data mod-

eling using Hawkes processes’, Journal of Banking & Finance, Vol. 36, No. 12,
pp. 3415 – 3426.
Chen, S. (2007). ‘Computationally intelligent agents in economics and finance’,

Information Sciences, Vol. 177, No. 5, pp. 1153–1168.
Chen, S. (2012). ‘Varieties of agents in agent-based computational economics: A

historical and an interdisciplinary perspective’, Journal of Economic Dynamics
and Control, Vol. 36, No. 1, pp. 1–25.
Chen, S., Chang, C. and Du, Y. (2012). ‘Agent-based economic models and econo-
metrics’, Knowledge Engineering Review, Vol. 27, No. 2, pp. 187–219.
Chen, S. and Yeh, C. (1997). ‘Toward a computable approach to the efficient mar-
ket hypothesis: An application of genetic programming’, Journal of Economic
Dynamics and Control, Vol. 21, No. 5, pp. 1043–1063.
Chen, S. and Yeh, C. (2002). ‘On the emergent properties of artificial stock mar-
kets: the efficient market hypothesis and the rational expectations hypothesis’,
Journal of Economic Behavior & Organization, Vol. 49, No. 2, pp. 217–239.
Cheung, Y., Ho., R., Pope, P. and Draper, P. (1994). ‘Intraday stock return volatil-
ity: The Hong Kong evidence’, Pacific-Basin Finance Journal, Vol. 2, No. 1,
pp. 261–276.
229
Chiarella, C., Dieci, R. and He, X. (2009a). ‘Heterogeneity, Market Mechanisms,
and Asset Price Dynamics’, in Handbook of Financial Markets: Dynamics and
Evolution, pp. 277–344, North Holland, Amsterdam.
Chiarella, C. and Iori, G. (2002). ‘A simulation analysis of the microstructure of

double auction markets’, Quantitative Finance, Vol. 2, No. 2, pp. 246–253.
Chiarella, C., Iori, G. and Perello, J. (2009b). ‘The impact of heterogeneous

trading rules on the limit order book and order flows’, Journal of Economic
Dynamics and Control, Vol. 33, No. 3, pp. 525–537.
Chiyachantana, C., Jain, P., Jiang, C. and Wood, R. (2004). ‘International Evi-
dence on Institutional Trading Behavior and Price Impact’, Journal of Finance,
Vol. 59, No. 2, pp. 869–898.
Chung, K., Van Ness, B. and Van Ness, R. (1999). ‘Limit orders and the bid-ask
spread’, Journal of Financial Economics, Vol. 53, No. 3, pp. 255–287.
Collins, B.M.and Fabozzi, F. (1991). ‘A Methodology for Measuring Transaction

Costs’, Financial Analysts Journal, Vol. 47, No. 2, pp. 27–36.
Consiglio, A. (ed.) (2007). Artificial Markets Modeling: Methods and Applica-

tions, Springer.
Cont, R. (2001). ‘Empirical properties of asset returns: stylized facts and statisti-
cal issues’, Quantitative Finance, Vol. 1, No. 2, pp. 223–236.
Copeland, T. and Galai, D. (1983). ‘Information Effects on the Bid-Ask Spread’,

Coppejans, M., Domowitz, I. and Madhavan, A. (2003). ‘Dynamics of Liquidity

in an Electronic Limit Order Book Market’. Working paper, Duke University.
230
Cristelli, M., Pietronero, L. and Zaccaria, A. (2011). ‘Critical Overview of Agent-
Based Models for Economics’. Working paper, arXiv. org.
Cui, W. and Brabazon, A. (2011). ‘Ultra High-frequency Financial Data:

Analysing NYSE Euronext TAQ Data’, in Proceedings of the 24nd Irish Ac-
counting and Finance Association Conference, pp. 1–38.
Cui, W. and Brabazon, A. (2012a). ‘An Agent-based Modelling Approach to

Study Price Impact’, in Proceedings of the 2012 IEEE Conference on Compu-
tational Intelligence for Financial Engineering & Economics, pp. 1–8, IEEE
Press.
Cui, W. and Brabazon, A. (2012b). ‘An Analysis of Liquidity in London Stock

Exchange’, in Proceedings of the 25th Irish Accounting and Finance Associa-
tion Conference, pp. 1–45.
Cui, W., Brabazon, A. and O’Neill, M. (2009a). ‘Efficient Trade Execution Us-
ing a Genetic Algorithm in an Order Book Based Artificial Stock Market’, in
Proceedings of the 11th Annual Conference Companion on Genetic and Evolu-
tionary Computation Conference, pp. 2023–2028, ACM Press, New York, NY,
USA.
Cui, W., Brabazon, A. and O’Neill, M. (2009b). ‘Evolutionary Computation in

Trade Execution’, in Proceedings of the 22nd Irish Accounting and Finance
Association Conference, pp. 1–10, Dublin, Ireland.
Cui, W., Brabazon, A. and O’Neill, M. (2010a). ‘An Analysis of Price Impact’,
in Proceedings of the 3rd International Conference of European Research Con-
sortium on Informatics and Mathematics, London, UK.
Cui, W., Brabazon, A. and O’Neill, M. (2010b). ‘Evolutionary Computation and
231
Trade Execution’, in A. Brabazon & M. O’Neill (eds.), Natural Computing in
Computational Finance (volume III), pp. 51–69, Springer-Verlag, Berlin.
Cui, W., Brabazon, A. and O’Neill, M. (2010c). ‘Evolving Dynamic Trade Ex-
ecution Strategies Using Grammatical Evolution’, in Proceedings of the 2010
International Conference on Applications of Evolutionary Computation, Lec-
ture Notes in Computer Science 6025, pp. 191–200, Springer-Verlag, Berlin.
Cui, W., Brabazon, A. and O’Neill, M. (2010d). ‘Evolving Efficient Limit Or-
der Strategy Using Grammatical Evolution’, in Proceedings of the 2010 IEEE
Congress On Evolutionary Computation, pp. 1–6, IEEE Press.
Cui, W., Brabazon, A. and O’Neill, M. (2011). ‘Dynamic Trade Execution: A

Grammatical Evolution Approach’, International Journal of Financial Markets
and Derivatives, Vol. 2, No. 1/2, pp. 4–31.
Daniel, G. (2006). Asynchronous simulations of a limit order book, PhD thesis,

University of Manchester, UK.
Daniels, M., Farmer, D., Gillemot, L., Iori, G. and Smith, E. (2003). ‘Quantitative
Model of Price Diffusion and Market Friction Based on Trading as a Mecha-
nistic Random Process’, Physical Review Letters, Vol. 90, No. 10, pp. 1–4.
Darley, V. and Outkin, A. (2007). A Nasdaq Market Simulation: Insights on a

Major Market from the Science of Complex Adaptive Systems, World Scientific
Publishing Company.
Darwin, C. (1859). On The Origin of Species: By Means of Natural Selection or

the Preservation of Favoured Races in the Struggle for Life (reprinted 1985),
Penguin Books, London.
232
Decovny, S. (2008). ‘Asian-Pacific gears up for algorithmic trading’, Market View,
Vol. 1, No. 1, p. 13.
Dempsey, I., O’Neill, M. and Brabazon, A. (2009). Foundations in Grammatical

Evolution for Dynamic Environment, Springer-Verlag, Berlin.
Dempster, M. and Jones, C. (2001). ‘A real-time adaptive trading system using

Genetic Programming’, Quantitative Finance, Vol. 1, No. 2, pp. 397–413.
Duffy, J. (2006). ‘Agent-Based Models and Human Subject Experiments’, in

L. Tesfatsion & K. L. Judd (eds.), Handbook of Computational Economics,
Vol. 2, Elsevier.
Dufour, A. and Engle, R. (2000). ‘Time and the price impact of a trade’, Journal
of Finance, Vol. 55, No. 6, pp. 2467–2498.
Duong, H., Kalev, P. and Krishnamurti (2009). ‘Order aggressiveness of institu-

tional and individual investors’, Pacific-Basin Finance Journal, Vol. 1, No. 4,
pp. 1–14.
Durbin, M. (2010). All about high-frequency trading, McGraw-Hill.
Dworman, G., Kimbrough, S. and Laing, J. (1996). ‘On automated discovery

of models using genetic programming:bargaining in a three-agent coalitions
game’, Journal of Management Information Systems, Vol. 12, No. 3, pp. 97–
125.
Easley, D. and O’Hara, M. (1992). ‘Time and the process of security price adjust-
ment’, Journal of Finance, Vol. 47, No. 2, pp. 577–605.
Eisler, Z., Bouchaud, J. and Kockelkoren, J. (2012). ‘The price impact of or-
der book events: market orders, limit orders and cancellations’, Quantitative
Finance, Vol. 12, No. 9, pp. 1395–1419.
233
Ellul, A., Holden, C., Jain, P. and Jennings, R. (2007). ‘Order dynamics: recent
evidence from the NYSE’, Journal of Empirical Finance, Vol. 14, No. 3, pp.
636–661.
Engle, R. and Patton, A. (2004). ‘Impacts of trades in an error-correction model

of quote prices’, Journal of Financial Markets, Vol. 7, No. 1, pp. 1–25.
Esser, A. and Monch, B. (2007). ‘The navigation of an iceberg: The optimal use
of hidden orders’, Finance Research Letters, Vol. 4, No. 1, pp. 68–81.
Evans, M. and Lyons, R. (2002). ‘Order Flow and Exchange Rate Dynamics’,
Journal of Political Economy, Vol. 110, No. 1, pp. 170–180.
Farmer, D., Gillemot, L., Iori, G., Krishnamurthy, S., Smith, E. and Daniels, M.
(2005a). ‘A random order placement model of price formation in the continuous
double auction’, in L. Blume & S. Durlauf (eds.), The Economy as an Evolving
Complex System III, Oxford University Press.
Farmer, J. and Foley, D. (2009). ‘The economy needs agent-based modelling’,

Nature, Vol. 460, pp. 685–686.
Farmer, J., Gerig, A., Lillo, F. and Mike, S. (2006). ‘Market efficiency and the
long-memory of supply and demand: Is price impact variable and permanent or
fixed and temporary?’, Quantitative Finance, Vol. 6, No. 2, pp. 107–112.
Farmer, J. and Lillo, F. (2004). ‘On the origin of power-law tails in price fluctua-
tions’, Quantitative Finance, Vol. 4, No. 1, pp. 7–10.
Farmer, J. and Zamani, N. (2007). ‘Mechanical vs. informational components of

price impact’, European Physical Journal B., Vol. 55, No. 1, pp. 1899–2000.
Farmer, J. D., Gillemot, L., Lillo, F., Mike, S. and Sen, A. (2004). ‘What really
causes large price changes?’, Quantitative Finance, Vol. 4, No. 4, pp. 383–397.
234
Farmer, J. D., Patelli, P. and Zovko, I. (2005b). ‘The predictive power of zero intel-
ligence models in financial markets’, in Proceedings of the National Academy
of Sciences of the United States of America, Vol. 102, pp. 2254–2259.
Farmer, J. D., Patelli, P. and Zovko, I. (2005c). ‘Supplementary Material for

‘The Predictive Power of Zero Intelligence in Financial Markets”. Available at:
www.santafe.edu/jdf/papers/zerosuppl.pdf.
Feldman, T. and Friedman, D. (2010). ‘Human and Artificial Agents in a Crash-

Prone Financial Market’, Computational Economics, Vol. 36, No. 3, pp. 201–
229.
FinancialTimes (2011a). ‘Daily Market Share: equities (source

Thompson Reuters Equity Market Share Reporter)’. Available at:
http://www.ft.com/intl/tradingroom (accessed 3 February 2011).
FinancialTimes (2011b). ‘Nasdaq OMX lures trades to platform’. Available at:

http://www.ft.com/tradingroom (accessed 30 January 2011).
Foster, F. and Viswanathan, S. (1993). ‘Variations in Trading Volume, Return

Volatility, and Trading Costs: Evidence on Recent Price Formation Models’,
Foucault, T. (1999). ‘Order flow composition and trading costs in a dynamic limit
order market’, Journal of Financial Markets, Vol. 2, No. 2, pp. 99–134.
Foucault, T., Kaden, O. and Kandel, E. (2005). ‘The limit order book as a market
for liquidity’, Review of Financial Studies, Vol. 18, No. 1, pp. 1171–1217.
Foucault, T. and Menkveld, A. (2008). ‘Competition for Order Flow and Smart
Order Routing Systems’, Journal of Finance, Vol. 63, No. 1, pp. 119–158.
235
French, K. and Roll, R. (1986). ‘Stock return variances: the arrival of information
and the reaction of traders’, Journal of Financial Economics, Vol. 17, No. 1,
pp. 5–26.
Fruth, A., Schoeneborn, T. and Urusov, M. (2011). ‘Optimal trade execution and
price manipulation in order books with time-varying liquidity’. Available at:
http://arxiv.org/abs/1109.2631.
Gabaix, X., Gopikrishnan, P., Plerou, V. and Stanley, H. (2003). ‘A theory of

power law distributions in financial market fluctuations’, Nature, Vol. 423, pp.
267–270.
Gabaix, X., Gopikrishnan, P., Plerou, V. and Stanley, H. (2006). ‘Institutional

Investors and Stock Market Volatility’, Quarterly Journal of Economics, Vol.
121, No. 2, pp. 461–504.
Gatheral, J. (2010). ‘No-Dynamic-Arbitrage and Market Impact’, Quantitative

Finance, Vol. 10, No. 7, pp. 749–759.
Gatheral, J., Schied, A. and Slynko, A. (2011). ‘Exponential Resilience and Decay
of Market Impact’, in F. Abergel, B. Chakrabarti, A. Chakraborti, & M. Mitra
(eds.), Econophysics of Order-driven Markets, Spring-Verlag, Italy.
Gemmill, G. (1996). ‘Transparency and Liquidity: A Study of Block Trades

on the London Stock Exchange under Different Publication Rules’, Journal of
Finance, Vol. 51, No. 5, pp. 1765–1790.
Gillemot, L., Farmer, J. and Lillo, F. (2005). ‘There’s more to volatility than
volume’. Available at: http://arxiv.org/abs/physics/0510007.
Glosten, L. (1994). ‘Is the Electronic Open Limit Order Book Inevitable?’, Jour-
nal of Finance, Vol. 49, No. 4, pp. 1127–1161.
236
Glosten, L. and Milgrom, P. (1985). ‘Bid, ask and transaction prices in a special-
ist market with heterogeneously informed traders’, Journal of Financial Eco-
nomics, Vol. 14, No. 1, pp. 71–100.
Gode, D. and Sunder, S. (1993). ‘Allocative efficiency of markets with zero-

intelligence traders’, Journal of Political Economy, Vol. 101, No. 1, pp. 119–
137.
Goettler, R., Parlour, C. and Rajan, U. (2005). ‘Equilibrium in a dynamic limit

order book’, Journal of Finance, Vol. 60, No. 2, pp. 2149–2192.
Goldstein, M., Irvine, P., Kandel, E. and Wiener, Z. (2009). ‘Brokerage Commis-
sions and Institutional Trading Patterns’, Review of Financial Studies, Vol. 22,
No. 12, pp. 5175–5212.
Gould, M., Porter, M., Williams, S., McDonald, M., Fenn, D. and Howison, S.
(2011). ‘Limit order books’. Working paper, arXiv. org.
Grant, J. (2012). ‘China’s Citic in algorithmic trading deal’. Available at:

http://www.ft.com/intl (accessed 14 May 2012).
Griffiths, M., Smith, B., Turnbull, D. and White, R. (2000). ‘The costs and de-
terminants of order aggressiveness’, Journal of Financial Economics, Vol. 56,
No. 1, pp. 65–88.
Hall, A. and Hautsch, N. (2006). ‘Order aggressiveness and order book dynamics’,
Empirical Economics, Vol. 30, No. 1, pp. 973–1005.
Hall, A. and Hautsch, N. (2007). ‘Modelling the buy and sell intensity in a limit
order book market’, Journal of Financial Markets, Vol. 10, No. 3, pp. 249–286.
Handa, P. and Schwartz, R. (1996). ‘Limit Order Trading’, Journal of Finance,

Vol. 51, No. 5, pp. 1835–1861.
237
Handa, P., Schwartz, R. and Tiwari, A. (2003). ‘Quote setting and price formation
in an order-driven market’, Journal of Financial Markets, Vol. 6, No. 1, pp.
461–489.
Harris, L. (1986). ‘A transaction data study of weekly and intraday patterns in

stock returns’, Journal of Financial Economics, Vol. 16, No. 2, pp. 99–117.
Harris, L. (1998). ‘Optimal Dynamic Order Submission Strategies in Some

Stylized Trading Problems’, Financial Markets, Institutions and Instruments,
Vol. 7, No. 2, pp. 1–76.
Harris, L. and Hasbrouck, J. (1996). ‘Market vs. Limit Orders - The SuperDOT
Evidence On Order Submission Strategy’, Journal of Financial and Quantita-
tive Analysis, Vol. 31, No. 2, pp. 213–231.
Hasbrouck, J. (1991). ‘Measuring the information content of stock trades’, Jour-

nal of Finance, Vol. 46, No. 2, pp. 178–208.
Hasbrouck, J. (1999). ‘The dynamics of discrete bid and ask quotes’, The Journal
of finance, Vol. 54, No. 6, pp. 2109–2142.
Hasbrouck, J. and Saar, G. (2002). ‘Limit Orders and Volatility in a Hybrid Mar-
ket: The Island ECN’. Available at: http://ssrn.com/abstract310940.
Hasbrouck, J., Sofianos, G. and Sosebee, D. (1993). ‘New York

Stock Exchange System and Trading Procedures’. Available at:
http://citeseerx.ist.psu.edu/viewdoc/summarydoi10.1.1.132.5638.
Hausman, J., Lo, A. and MacKinlay, A. (1992). ‘An ordered probit analysis of
transaction stock prices’, Journal of Financial Economics, Vol. 31, No. 3, pp.
319–379.
238
Hautsch, N. and Huang, R. (2012). ‘The market impact of a limit order’, Journal
of Economic Dynamics and Control, Vol. 36, No. 4, pp. 501 – 522.
Hendershott, T., Jones, C. and Menkveld, A. (2011). ‘Does Algorithmic Trading

Improve Liquidity?’, Journal of Finance, Vol. 66, No. 1, pp. 1–33.
Hendershott, T. and Riordan, R. (2012). ‘Algorithmic Trading and the Market for
Liquidity’, Journal of Financial and Quantitative Analysis (forthcoming).
Henker, T. and Wang, J. (2006). ‘On the importance of timing specifications in

market microstructure research’, Journal of Financial Markets, Vol. 9, No. 2,
pp. 162–179.
Ho, Y. and Cheung, Y. (1991). ‘Behaviour of intra-daily stock return on an Asian

emerging market - Hong Kong’, Applied Economics, Vol. 23, No. 5, pp. 957–
966.
Holthausen, R., Leftwich, R. and Mayers, D. (1987). ‘The effect of large block
transactions on security prices: A cross-sectional analysis’, Journal of Finan-
cial Economics, Vol. 19, No. 2, pp. 237–267.
Holthausen, R., Leftwich, R. and Mayers, D. (1990). ‘Large-block transactions,

the speed of response, and temporary and permanent stock-price effects’, Jour-
nal of Financial Economics, Vol. 26, No. 1, pp. 71–95.
Hommes, C. and Wagener, F. (2009). ‘Complex Evolutionary Systems in Behav-

ioral Finance’, in T. Hens & K. Schenk (eds.), Handbook of Financial Markets:
Dynamics and Evolution, North Holland.
Hopman, C. (2007). ‘Do supply and demand drive stock prices?’, Quantitative
Finance, Vol. 7, No. 1, pp. 37–53.
239
Hora, M. (2006). ‘Tactical Liquidity Trading and Intraday Volume’. Available at:
http://ssrn.com/abstract=931667.
Huang, Y., Chen, S., Hung, M. and Yu, T. (2012). ‘Liquidity cost of market orders
in the Taiwan Stock Market: A study based on an order-driven agent-based
artificial stock market’, International Review of Financial Analysis, Vol. 23,
pp. 72–80.
Hussain, S. (2011). ‘The Intraday Behaviour of Bid-Ask Spreads, Trading Vol-

ume and Return Volatility: Evidence from DAX30’, International Journal of
Economics and Finance, Vol. 3, No. 1, pp. 23–34.
Iori, G., Daniels, M., Farmer, J., Gillemot, L., Krishnamurthy, S. and Smith, E.
(2003). ‘An analysis of price impact function in order-driven markets’, Physica
A, Vol. 324, No. 1-2, pp. 146–151.
Izumi, K., Toriumi, F. and Matsui, H. (2009). ‘Evaluation of automated-trading

strategies using an artificial market’, Neurocomputing, Vol. 72, No. 1, pp. 3469–
3476.
Jain, P. (2003). ‘Institutional design and liquidity at stock exchanges around the
world’.
Jain, P. and Gun-Ho, J. (1988). ‘The Dependence between Hourly Prices and Trad-
ing Volume’, Journal of Financial and Quantitative Analysis, Vol. 23, No. 3,
pp. 269–283.
Joulin, A., Lefevre, A., Grunberg, D. and Bouchaud, J. (2008). ‘Stock

price jumps: news and volume play a minor role’. Available at:
http://arxiv.org/abs/0803.1769.
240
Kaniel, R. and Liu, H. (2006). ‘So What Orders Do Informed Traders Use?’,
Journal of Business, Vol. 79, No. 4, pp. 1867–1913.
Keim, D. and Madhavan, A. (1995a). ‘Anatomy of the trading process: Em-

pirical evidence on the behavior of institutional traders’, Journal of Financial
Economics, Vol. 37, No. 1, pp. 371–398.
Keim, D. and Madhavan, A. (1995b). ‘Execution Costs and Investment Perfor-

mance: An Empirical Analysis of Institutional Equity Trades’. Available at:
http://ideas.repec.org/p/fth/pennfi/09-95.html.
Keim, D. and Madhavan, A. (1996). ‘The upstairs market for large-block transac-
tions: analysis and measurement of price effects’, Review of Financial Studies,
Vol. 9, No. 1, pp. 1–36.
Keim, D. and Madhavan, A. (1997). ‘Transactions costs and investment style:

an inter-exchange analysis of institutional equity trades’, Journal of Financial
Economics, Vol. 46, No. 3, pp. 265–292.
Kharroubi, I. and Pham, H. (2010). ‘Optimal Portfolio Liquidation with Execution

Cost and Risk’, SIAM Journal on Financial Mathematics, Vol. 1, No. 1, pp.
897–931.
Kim, G. and Markowitz, H. (1989). ‘Investment rules, margin, and market volatil-
ity’, Journal of Portfolio Management, Vol. 16, No. 1, pp. 45–52.
King, L. (2011). ‘London Stock Exchange price data fail-

ures emerged immediately at Millenium launch’. Available at:
http://www.computerworlduk.com/news/itbusiness/3261816/.
Koksal, B. (2012). ‘An analysis of intraday patterns and liquidity on the Istanbul
stock exchange’. Available at: http://mpra.ub.unimuenchen.de/35968/.
241
Koza, J. (2010). ‘Human-competitive results produced by genetic programming’,
Genetic Programming and Evolvable Machines, Vol. 11, No. 3, pp. 251–284.
Kozhan, R. and Salmon, M. (2011). ‘The information content of a limit order

book: The case of an FX market’, Journal of Financial Markets, Vol. 15, No. 1,
pp. 1–28.
Kraus, A. and Stoll, H. (1972). ‘Price impacts of block trading on the New York
Stock Exchange’, Journal of Finance, Vol. 27, No. 3, pp. 569–588.
Kyle, A. (1985). ‘Continuous auctions and insider trading’, Econometrica,

Vol. 53, No. 3, pp. 1315–1335.
Ladley, D. and Schenk, K. (2009). ‘Do stylised facts of order book markets need
strategic behaviour’, Journal of Economics Dynamics & Control, Vol. 33, No. 3,
pp. 817–831.
Large, J. (2007). ‘Measuring the resiliency of an electronic limit order book’,

Journal of Financial Markets, Vol. 10, No. 1, pp. 1–25.
LeBaron, B. (2000). ‘Agent-based computational finance: suggested readings and

early research’, Journal of Economic Dynamics and Control, Vol. 24, No. 1, pp.
679–702.
LeBaron, B. (2001). ‘A Builder’s Guide to Agent Based Financial Markets’,

Quantitative Finance, Vol. 1, No. 2, pp. 254–261.
LeBaron, B. (2006). ‘Agent-based Computational Finance’, in L. Tesfatsion &

K. Judd (eds.), Handbook of Computational Economics, Volume 2: Agent-
based Computational Economics, pp. 134–151, Elsevier.
242
Lee, C., Mucklow, B. and Ready, M. (1993). ‘Spreads, depths, and the impact
of earnings information: An intraday analysis’, Review of Financial Studies,
Vol. 6, No. 2, pp. 345–375.
Lee, C. and Ready, M. (1991). ‘Inferring Trade Direction from Intraday Data’,
Lee, Y., Fok, R. and Liu, Y. (2001). ‘Explaining intraday pattern of trading volume
from the order flow data’, Journal of Business Finance and Accounting, Vol. 28,
No. 3, pp. 199–230.
Lehmann, B. and Modest, D. (1994). ‘Trading and Liquidity on the Tokyo Stock
Exchange: A Bird’s Eye View’, Journal of Finance, Vol. 49, No. 3, pp. 951–
984.
Li, Y., Van Ness, B. and Van Ness, R. (2005). ‘Daily and Intraday Patterns in
Spread and Depth for Limit Orders and Specialists’, Quaterly journal of busi-
ness and economics, Vol. 44, No. 3, pp. 3–14.
LiCalzi, M. and Pellizzari, P. (2003). ‘Fundamentalists clashing over the book: a

study of order-driven stock markets’, Quantitative Finance, Vol. 3, No. 6, pp.
470–480.
LiCalzi, M. and Pellizzari, P. (2007). ‘Simple market protocols for efficient risk
sharing’, Journal of Economic Dynamics and Control, Vol. 31, No. 11, pp.
3568–3590.
Lillo, F., Farmer, J. and Mantegna, R. (2002). ‘Single Curve Collapse of the
Price Impact Function for the New York Stock Exchange’. Available at:
http://arxiv.org/abs/cond-mat/0207428.
243
Lillo, F., Farmer, J. and Mantegna, R. (2003). ‘Master curve for price impact
function’, Nature, Vol. 421, No. 3, pp. 129–130.
Lillo, F., Mike, S. and Farmer, J. (2005). ‘Theory for long memory in supply and
demand’, Physical Review E, Vol. 71, No. 6, pp. 1–11.
Lim, M. and Coggins, R. (2005a). ‘The immediate price impact of trades on the
Australian Stock Exchange’, Quantitative Finance, Vol. 5, No. 4, pp. 365–377.
Lim, M. and Coggins, R. (2005b). ‘Optimal trade execution: an evolutionary

approach’, in Proceedings of the IEEE Congress on Evolutionary Computation,
Vol. 2, pp. 1045–1052.
Liu, W. (2009). ‘Monitoring and limit order submission risks’, Journal of Finan-
cial Markets, Vol. 12, No. 1, pp. 107–141.
Lo, A., MacKinlay, A. and Zhang, J. (2002). ‘Econometric models of limit-order

executions’, Journal of Financial Economics, Vol. 65, No. 1, pp. 31–71.
Lo, I. and Sapp, S. (2010). ‘Order Aggressiveness and Quantity: How Are They
Determined in a Limit Order Market?’, Journal of International Financial Mar-
kets, Institutions and Money, Vol. 20, No. 3, pp. 223–243.
Lockwood, L. and Linn, S. (1990). ‘An Examination of Stock Market Return

Volatility During Overnight and Intraday Periods’, Journal of Finance, Vol. 45,
No. 2, pp. 591–601.
Lux, T. (1998). ‘The socio-economic dynamics of speculative markets: interacting

agents, chaos, and the fat tails of return distributions’, Journal of Economic
Behavior & Organization, Vol. 33, No. 2, pp. 143–165.
244
Lux, T. (2009). ‘Stochastic Behavioral Asset Pricing Models and the Stylized
Facts’, in Handbook of Financial Markets: Dynamics and Evolution, North
Holland.
Madhavan, A. (1992). ‘Trading Mechanisms in Securities Markets’, Journal of

Finance, Vol. 47, No. 2, pp. 607–641.
Madhavan, A. (2000). ‘Market Microstructure: A Survey’, Journal of Financial

Markets, Vol. 3, No. 1, pp. 205–258.
Madhavan, A., Richardson, M. and Roomans, M. (1997). ‘Why do security prices

change? A transaction-level analysis of NYSE stocks’, Review of Financial
Studies, Vol. 10, No. 4, pp. 1035–1064.
Maslov, S. (2000). ‘Simple model of a limit order-driven market’, Physica A, Vol.

278, No. 3, pp. 571–578.
Maslov, S. and Millis, M. (2001). ‘Price fluctuations from the order book per-
spective - empirical facts and a simple model’, Physica A, Vol. 299, No. 1, p.
234.
McInish, T. and Wood, R. (1990a). ‘An analysis of transactions data for the
Toronto Stock Exchange: Return patterns and end-of-the-day effect’, Journal
of Banking & Finance, Vol. 14, No. 2, pp. 441–458.
McInish, T. and Wood, R. (1990b). ‘A transactions data analysis of the variability

of common stock returns during 1980-1984’, Journal of Banking and Finance,
Vol. 14, No. 1, pp. 99–112.
McInish, T. and Wood, R. (1992). ‘An Analysis of Intraday Patterns in Bid-Ask

Spreads for NYSE Stocks’, Journal of Finance, Vol. 47, No. 2, pp. 753–764.
245
McKay, R., Hoai, N., Whigham, P., Shan, Y. and O’Neill, M. (2010). ‘Grammar-
based Genetic Programming: a survey’, Genetic Programming and Evolvable
Machines, Vol. 11, No. 4, pp. 365–396.
Menkhoff, L., Osler, C. and Schmeling, M. (2010). ‘Limit-order submission

strategies under asymmetric information’, Journal of Banking and Finance,
Vol. 34, No. 11, pp. 2665–2677.
Mike, S. and Farmer, J. (2008). ‘An empirical behavioral model of liquidity and
volatility’, Journal of Economic Dynamics and Control, Vol. 32, No. 1, pp.
200–234.
Miller, M. (2008). ‘Don’t let your robots grow up to be traders: Artificial intel-
ligence, human intelligence, and asset-market bubbles’, Journal of Economic
Behavior & Organization, Vol. 68, No. 1, pp. 163–166.
Mishra, S., Daigler, R. and Holowczack, R. (2012). ‘The Effect of Al-

gorithmic Trading on Liquidity in the Options Market’. Available at:
http://people.stern.nyu.edu/jhasbrou (accessed 12 Nov. 2012).
Moazeni, S., Coleman, T. and Li, Y. (2010). ‘Optimal Portfolio Execution Strate-
gies and Sensitivity to Price Impact Parameters’, SIAM Journal on Optimiza-
tion, Vol. 20, No. 3, pp. 1620–1654.
Neely, C. and Weller, P. (1999). ‘Technical trading rules in the European monetary
system’, Journal of International Money and Finance, Vol. 18, No. 1, pp. 429–
458.
Neely, C. and Weller, P. (2001). ‘Technical analysis and central bank intervention’,
Technical analysis and central bank intervention, Vol. 20, No. 7, pp. 949–970.
246
Neely, C. and Weller, P. (2003). ‘Intraday technical trading in the foreign exchange
market’, Intraday technical trading in the foreign exchange market, Vol. 22,
No. 2, pp. 223–237.
Neely, C., Weller, P. and Dittmar, R. (1997). ‘Is technical analysis in the foreign
exchange market profitable? A genetic programming approach’, Journal of
Financial and Quantitative Analysis, Vol. 32, No. 4, pp. 405–428.
NYSE (2006). TAQ 3 User’s Guide (version 1.0.9), New York Stock Exchange
Inc.
Obizhaeva, A. A. and Wang, J. (2012). ‘Optimal trading strategy and sup-

ply/demand dynamics’, Journal of Financial Markets (forthcoming).
O’Hara, M. (1995). Market Microstructure Theory, Blackwell Publishers, Cam-

bridge US.
Omura, K., Tanigawa, Y. and Uno, J. (2000). ‘Execution Probabil-

ity of Limit Orders on the Tokyo Stock Exchange’. Available at:
http://ssrn.com/abstract=252588.
O’Neill, M. and Brabazon, A. (2009). ‘Recent Patents on Genetic Programming’,

Recent Patents on Computer Science, Vol. 2, No. 1, pp. 43–49.
O’Neill, M. and Ryan, C. (2003). Grammatical Evolution: Evolutionary Auto-

matic Programming in an Arbitrary Language, Kluwer Academic Publishers,
Boston US.
Palit, I., Phelps, S. and Ng, W. (2012). ‘Can a Zero-Intelligence Plus Model Ex-
plain the Stylized Facts of Financial Time Series Data?’, in Proceedings of the
11th International Conference on Autonomous Agents and Multiagent Systems,
pp. 653–660.
247
Palmer, R., Arthur, W., Holland, J., LeBaron, B. and Tayler, P. (1994). ‘Artifi-
cial economic life: a simple model of a stockmarket’, Physica D: Nonlinear
Phenomena, Vol. 75, No. 1-3, pp. 264–274.
Parlour, C. (1998). ‘Price dynamics in limit order markets’, Review of Financial

Studies, Vol. 11, No. 4, pp. 789–816.
Parlour, C. and Seppi, D. (2008). ‘Limit Order Markets: A Survey’, in A. V.

Thakor & A. W. Boot (eds.), Handbook of Financial Intermediation and Bank-
ing, pp. 63–96, Elsevier, Amsterdam.
Pascual, R. and Verdas, D. (2009). ‘What pieces of limit order book information
matter in explaining order choice by patient and impatient traders?’, Quantita-
tive Finance, Vol. 9, No. 1, pp. 527–545.
Perold, A. (1988). ‘The Implementation Shortfall: Paper vs. Reality’, Journal of

Portfolio Management, Vol. 14, No. 3, pp. 4–9.
Peterson, M. and Sirri, E. (2002). ‘Order Submission Strategy And The Curi-
ous Case Of Marketable Limit Orders’, Journal of Financial and Quantitative
Analysis, Vol. 37, No. 2, pp. 221–241.
Peterson, M. and Sirri, E. (2003). ‘Evaluation of the biases in execution cost

estimation using trade and quote data’, Journal of Financial Markets, Vol. 6,
No. 3, pp. 259–280.
Potters, M. and Bouchaud, J. (2003). ‘More statistical properties of order books

and price impact’, Physica A: Statistical Mechanics and its Applications, Vol.
324, No. 1, pp. 133–140.
Predoiu, S., Shaikhet, G. and Shreve, S. (2011). ‘Optimal Execution in a Gen-
248
eral One-Sided Limit-Order Book’, SIAM Journal on Financial Mathematics,
Vol. 2, No. 1, pp. 183–212.
Raberto, M. and Cincotti, S. (2005). ‘Modeling and simulation of a double auc-

tion artificial financial market’, Physica A: Statistical Mechanics and its appli-
cations, Vol. 355, No. 1, pp. 34–45.
Ranaldo, A. (2004). ‘Order Aggressiveness in Limit Order Book Markets’, Jour-

nal of Financial Markets, Vol. 7, No. 1, pp. 53–74.
Rawal, D. (2010). ‘Bringing intelligent decision-making to order routing’, Journal

of Trading, Vol. 5, No. 1, pp. 30–34.
Rosu, I. (2009). ‘A Dynamic Model of the Limit Order Book’, Review of Financial
Studies, Vol. 22, No. 1, pp. 4601–4641.
Saar, G. (2001). ‘Price Impact Asymmetry of Block Trades: an institutional trad-

ing explanation’, Review of Financial Studies, Vol. 14, No. 4, pp. 1153–1181.
Samanidou, E., Zschischang, E., Stauffer, D. and Lux, T. (2007). ‘Agent-based

models of financial markets’, Reports on Progress in Physics, Vol. 70, No. 3,
pp. 409–450.
Sandas, P. (2001). ‘Adverse Selection and Competitive Market Making: Empirical

Evidence from a Limit Order Market’, Review of Financial Studies, Vol. 14,
No. 3, pp. 705–734.
Schelling, T. (1971). ‘Dynamic models of segregation’, Journal of Mathematical

Sociology, Vol. 1, No. 2, pp. 143–186.
Schmidt, A. (2010). ‘Optimal Execution in the Global FX Market’, Journal of

Trading, Vol. 5, No. 3, pp. 68–77.
249
Sefiane, S. and Benbouziane, M. (2012). ‘Portfolio Selection Using Genetic Al-
gorithm’, Journal of Applied Finance & Banking, Vol. 2, No. 4, pp. 143–154.
Seppi, D. (1997). ‘Liquidity provision with limit orders and a strategic specialist’,
Review of Financial Studies, Vol. 10, No. 2, pp. 103–150.
Simon, H. (1996). The Sciences of the Artificial, MIT Press, 3rd edn.
Smith, E., Farmer, J., Gillemot, L. and Krishnamurthy, S. (2003). ‘Statistical

theory of the continous double auction’, Quantitative Finance, Vol. 3, No. 1,
pp. 481–514.
Spencer, H. (1864). The Principles of Biology, Vol. 1, Williams and Norgate,

London and Edinburgh.
Suzuki, K., Shimokawa, T. and Misawa, T. (2009). ‘An Agent-Based Approach to

Option Pricing Anomalies’, IEEE Transactions on Evolutionary Computation,
Vol. 13, No. 1, pp. 19–32.
Swarm (2009). ‘Swarm package can be obtained from’. Available at:

http://www.swarm.org.
Tesfatsion, L. (2006). ‘Agent-based computational economics: A constructive

approach to economic theory’, in L. Tesfatsion & K. Judd (eds.), Handbook of
computational economics: agent-based computational economics, pp. 52–74,
Elsevier, Amsterdam.
Torre, N. (1997). BARRA Market Impact Model Handbook, BARRA Inc., Berke-
ley.
Toth, B., Kertesz, J. and Farmer, J. (2009). ‘Studies of the limit order book around
large price changes’, European Physical Journal B., Vol. 71, No. 1, pp. 499–
510.
250
Vergote, O. (2005). ‘How to match trades and quotes for NYSE stocks’. Available
at: http://ssrn.com/abstract808984.
Verhoeven, P., Ching, S. and Ng, H. (2004). ‘Determinants of the decision to sub-
mit market or limit orders on the ASX’, Pacific-Basin Finance Journal, Vol. 12,
No. 3, pp. 1–18.
Vo, M. (2007). ‘Limit orders and the intraday behavior of market liquidity: Ev-
idence from the Toronto stock exchange’, Global Finance Journal, Vol. 17,
No. 3, pp. 379–396.
Wald, J. and Horriga, H. (2005). ‘Optimal Limit Order Choice’, Journal of Busi-
ness, Vol. 78, No. 2, pp. 597–620.
Weber, P. and Rosenow, B. (2005). ‘Order book approach to price impact’, Quan-
titative Finance, Vol. 5, No. 1, pp. 357–364.
Weber, P. and Rosenow, B. (2006). ‘Large stock price changes: volume or liquid-
ity?’, Quantitative Finance, Vol. 6, No. 1, pp. 7–14.
Wermers, R. (2002). ‘Mutual fund performance: An empirical decomposition

into stock-picking talent, style, transactions costs, and expenses’, The Journal
of Finance, Vol. 55, No. 4, pp. 1655–1703.
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,
Vol. 9, No. 2, pp. 619–664.
Wood, R., MiInish, T. and Ord, J. (1985). ‘An Investigation of Transactions Data
for NYSE Stocks’, Journal of Finance, Vol. 40, No. 3, pp. 723–739.
Xu, Y. (2009). ‘Order aggressiveness on the ASX market’, International Journal

of Economics and Finance, Vol. 1, No. 1, pp. 51–75.
251
Yang, J. (2002). ‘The Efficiency of an Artificial Double Auction Stock Market
with Neural Learning Agents’, in S. Chen (ed.), Evolutionary Computation in
Economics and Finance, pp. 85–106, Springer-Verlag.
Zovko, I. and Farmer, J. (2002). ‘The power of patience: a behavioural regularity

in limit-order placement’, Quantitative Finance, Vol. 2, No. 5, pp. 387–392.
252

Wei Thesis PDF

Uploaded by

Copyright:

Available Formats

Wei Thesis PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Wei Thesis PDF

Uploaded by

Copyright:

Available Formats

An Empirical Investigation of Price Impact:

An Agent-based Modelling Approach

Thesis presented for the degree of Ph.D.

Understanding price impact is a fundamental task in finance. Many ex-

List of Tables xii

3 Agent-based Modelling in Finance 51

4 Datasets and Data Processing 66

5 An Analysis of Intraday Behaviours of Price Impact and Market Liq-

6 An Investigation of Price Impact and Agent Intelligence 151

7 Dynamic Trade Execution with Limit and Market Orders 170

8 Conclusions and Future Work 200

A Explanation of the Chosen Time Period in Chapter 5 208

B More Results of Chapter 5 211

2.1 Different Shapes of Intraday Patterns . . . . . . . . . . . . . . . . 30

3.1 Evolutionary Cycle . . . . . . . . . . . . . . . . . . . . . . . . . 58

4.1 Data Selection Screen for TAQ3 Extraction Tool . . . . . . . . . . 70

7.1 Representation of Order’s Relative Price and Lifetime . . . . . . . 178

A.1 Intraday Price Volatility (One-day data) . . . . . . . . . . . . . . 209

B.1 Intraday Time-weighted Depth in Share (UK stocks) . . . . . . . 212

1.1 Order Book 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

2.1 Price Impact Function of Trade Size . . . . . . . . . . . . . . . . 25

4.1 A Sample of Trade Data for Ford Drawn from 2/2/09 . . . . . . . 72

5.1 Companies Listed in the US markets. Market capitalisation and

6.1 Descriptive Statistics of ROB Data at Sep. 6, 2010 for Barclays

7.1 VWAP Calculation of A Sample Buy Strategy . . . . . . . . . . . 176

• Cui, W., Brabazon, A. and O’Neill, M. (2011). Adaptive Trade Execution:

• Cui, W., Brabazon, A. and O’Neill, M. (2010). Evolutionary Computation

• Cui, W. and Brabazon, A. (2012). An Analysis of Liquidity in London

• Cui, W. and Brabazon, A. (2012). An Agent-based Modelling Approach to

• Cui, W. and Brabazon, A. (2011). Ultra High-frequency Financial Data:

• Cui, W., Brabazon, A. and O’Neill, M. (2010). Evolving Dynamic Trade

• Cui, W., Brabazon, A. and O’Neill, M. (2009). Efficient Trade Execution

• Cui, W., Brabazon, A. and O’Neill, M. (2009). Evolutionary Computation

• Cui, W., Brabazon, A. and O’Neill, M. (2010). An Analysis of Price Im-

1.1 Limit Order Market and Trade Execution

• Examples of newly set-up limit order markets in traditional exchanges in-

1.1.2 Trade Execution

“The investment process has been described as a three-legged

A. Madhavan, J.L. Treynor and W.H. Wagner (2007)

Trade execution is the implementation of trading decisions. The costs that

1. Commission is the payment made to brokers, which varies from broker to

1.2 Research Aims

These questions have practical significance. Without an understanding of the in-

1.2.1 Objectives of Research

1. Review relevant literature on market microstructure and price impact.

2. Develop an algorithm for preprocessing TAQ data.

4. Develop an agent-based model of a limit order market.

5. Design trade execution strategies.

6. Analyse price impact.

7. Draw conclusions from the results.

1.3 Framework of Research

1.4 Contributions of Thesis

• Up-to-date review of relevant literature

This thesis presents an up-to-date survey of relevant literature on market

• Analysis of intraday behaviours of price impact in UK and US markets

• Up-to-date evidence on intraday behaviours of market liquidity in UK

• Investigation of the role of agent intelligence in determining the magni-

• Examination of the effect of order choice on the price impact when

• Use of ABM for investigating price impact