Algorithmic Trading and High Frequency Trading

Hao Zhou

February 8, 2017

A thesis submitted in fulfilment of the requirements for the degree of Doctor of Philosophy

Centre for Applied Financial Studies

UniSA Business School
University of South Australia
Copyright

Under the Copyright Act 1968, this thesis must be used only under the normal conditions of scholarly fair dealings
for the purpose of research, criticism or review. In particular no results or conclusions should be extracted from
it, nor should it be copied or closely paraphrased in whole or in part without the written consent of the author.
Proper written acknowledgement should be made for any assistance obtained from this thesis.

I certify that I have made all reasonable efforts to secure copyright permissions for third-party content included
in this thesis and have not knowingly added copyright content to my work without the owner’s permission.

Hao Zhou
February 8, 2017

i
Declaration of Authorship

I hereby declare that this thesis contains no material which has been accepted for the award of any other degree
or diploma in any university or other institution and that, to the best of my knowledge this thesis contains no
material previously published or written by another person, except where due reference is made in the text of the
thesis.

Hao Zhou Date: February 8, 2017

ii
Abstract
This thesis provides one standalone survey essay and three empirical essays on algorithmic trading (AT) and its
effect on market qualities. The survey essay reviews the theoretical, empirical, and policy studies on algorithmic
and high frequency trading. We review the theoretical literature relating to: (1) market maker–taker dynamics,
(2) information content of trades and quotes, and (3) recently incurred or proposed market structural changes.
We aim to provide a comprehensive roadmap for future research by surveying the empirical literature with an
emphasis on how data and causal events can be identified. Our conclusion includes a brief discussion of policy
implications and suggestions for future work.
The first empirical essay investigates the role algorithmic trading on days when the absolute value of the market
return is more than 2%. We find that the abnormal return of a stock is related to the stock’s AT intensity, that
high AT intensity stocks experience less price drops (surges) on days when the market declines (increases) for
more than 2%. This result is consistent with the view that AT minimizes price pressures and mitigates transitory
pricing errors.
The second empirical essay examines algorithmic execution strategy and the effects of algorithmic trading
order imbalances. We find that, ex-ante, algorithmic traders execute their trades according to the prevailing
Volume-Weighted Average Price (VWAP), they are more likely to execute buy (sell) orders when the prevailing
VWAP moves lower (higher) compared to the prevailing stock price. This implies a contrarian strategy which may
mitigate the short-term price trends. Further analyses show that AT order imbalances have a smaller price impact
compared to non-AT order imbalances. These effects are robust on days when the absolute value of the market
return is more than 2%.
The last empirical essay considers the role of algorithmic trading in the price discovery process. We estimate a
state space framework that decomposes stock prices into permanent price series and transient pricing error via state
space frameworks. We find that algorithmic traders contribute more to the permanent price processes and less to
the transient pricing errors compared to other traders. Our results show that AT facilitates the price discovery
process by contributing to permanent price movements. Our results are robust on days when the absolute value of
the market return is more than 2%.

iii
Acknowledgements
First and foremost, I would like to express my sincere gratitude to my principle supervisor Prof Petko Kalev
for the continuous support of my PhD study and related research, for his patience, motivation, and immense
knowledge. His guidance helped during the research and the writing of this thesis. He was a true source of
inspiration.
I would also like to thank my co-supervisors. I would like thank Dr Michael Burrow for his help in initiating
the PhD candidature. I would like to thank Dr Andy Lian for his support in programming and technical aspects
of the thesis. A special thanks to Ron McIver for his insightful feedback and encouragement.
I am very grateful to University of South Australia for providing financial and facilities support needed to
conduct this study. I would like to thank all the staff members and fellow PhD friends for the stimulating
discussions and their selfless supports. I am grateful to the Australian Securities Exchange, the Securities Industry
Research Centre of Asia-Pacific, and Dr Hui Zheng for providing the novel data for this research. I would like to
also thank Ray Adams, who provides editing assistance on grammar and English expressions.
This thesis has benefited from various comments and suggestions from workshop and conference participants
at the 2012 AFBC meeting in Sydney, the 2014 FMA Europe meeting in Maastricht, the 2014 MFS meeting
in Prague, the 2014 SIRCA Young Researcher Workshop in Sydney, the University of Naples Federico II, the
Frankfurt School of Finance and Management, EMLYON Business School, La Trobe University, and Monash
University.
Last but not least, I would like to thank my parents, my parents-in-law, and my wife for supporting me
spiritually throughout the writing of this thesis and throughout my life in general.

iv
Contents
Copyright i

Declaration of Authorship ii

Abstract iii

Acknowledgements iv

List of Figures ix

List of Tables x

Abbreviations xi

1 Introduction and overview 1

1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

2 Algorithmic and High Frequency Trading: A Review of the Literature 6

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 Review and policy studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2.1 CT review papers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2.2 Review papers from broader perspectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2.3 Policy papers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.3 Theoretical studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.3.1 Market maker–taker dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.3.2 Information content . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.3.3 Recently incurred or proposed market structural changes . . . . . . . . . . . . . . . . . . . 19

v
 2.4 Empirical studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.4.1 Empirical studies by data identification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.4.1.1 Studies that use CT proxies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.4.1.2 Studies that classify CT on aggregate . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.4.1.3 Studies that classify HFT on account level . . . . . . . . . . . . . . . . . . . . . . 29
2.4.2 Causal inferences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.4.2.1 Technological upgrades . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.4.2.2 Exchange fee changes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.4.2.3 Financial transactions tax . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.4.2.4 Colocations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.5 Discussion on CT and market quality metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.5.1 Liquidity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.5.2 Price discovery and price efficiency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.5.3 Volatility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
2.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

3 Algorithmic Trading in Turbulent Markets 44

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.2 Literature review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.3 Data and research design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.3.1 Data description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.3.2 Stock and event day selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.4 AT intensity and abnormal returns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.4.1 Univariate analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.4.2 Multivariate analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.4.3 Post-event day analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3.4.4 Net effects of AT liquidity demand and supply . . . . . . . . . . . . . . . . . . . . . . . . . 65
3.4.5 Robustness tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66

vi
 3.5 AT, news announcements, and market conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
3.5.1 Matched event day versus non-event day difference-in-differences analysis . . . . . . . . . . 71
3.5.2 AT and news arrivals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
3.5.3 Causal implications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
3.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78

4 Algorithmic Execution Strategy and Order Imbalances 79

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
4.2 Data and event day selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
4.3 AT execution strategy and the VWAP metric . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
4.4 AT and non-AT order imbalances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
4.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90

5 State Space Models for AT and Price Discovery 93

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
5.2 Literature review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
5.3 Data and descriptive statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
5.4 The state space model backgrounds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
5.4.1 Model representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
5.4.2 Model estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
5.5 The empirical analysis of AT and price discovery . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
5.5.1 The state space decomposition of intraday price discovery . . . . . . . . . . . . . . . . . . 103
5.5.2 The price discovery of AT order flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
5.5.3 Estimation results and discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
5.5.4 Robustness test: AT on turbulent days . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
5.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116

6 Conclusion 117
6.1 Overview and conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117

vii
6.2 Suggestions for future research . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119

viii
List of Figures
2.1 HFT Public Information Demand Analysis via Google Trend. . . . . . . . . . . . . . . . . . . . . . 8
3.1 AT and non-AT Buy Volume by Event Days. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.2 AT and non-AT Sell Volume by Event Days. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.3 AT and non-AT Trading Volume on Up Days and Down Days. . . . . . . . . . . . . . . . . . . . . 59
4.1 The VWAP, AT, and non-AT of One Event Day and One non-Event Day. . . . . . . . . . . . . . . 82
5.1 Market Capitalisation of Individual Stocks in the ASX. . . . . . . . . . . . . . . . . . . . . . . . . 97
5.2 Intraday Price Series and Hidden Efficient Price Series in One Trading Session. . . . . . . . . . . . 106
5.3 Intraday Return of the Efficient Component vs the Transitory Component in One Trading Session. 107

ix
List of Tables
3.1 Market Up Days and Returns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.2 Market Down Days and Returns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.3 Descriptive Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.4 Event Day Market-Adjusted Return Regressions on Abnormal AT Ratios . . . . . . . . . . . . . 63
3.5 Post-Event Cumulative Abnormal Returns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
3.6 Event Day Market-Adjusted Return Regressions on Net AT Imbalances . . . . . . . . . . . . . . 67
3.7 Event Day Market-Adjusted Return Regressions on Innovation in AT Intensity . . . . . . . . . . 68
3.8 Sensitivity Test for Event Day Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
3.9 Panel Data Regressions for Abnormal Volume Ratios . . . . . . . . . . . . . . . . . . . . . . . . . 70
3.10 Propensity Score Matching Diagnostics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
3.11 Matched Event Day and non-Event Day Diff-in-Diff Test . . . . . . . . . . . . . . . . . . . . . . 73
3.12 AT and Firm Specific News Arrival . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
4.1 AT versus non-AT Variance from the VWAP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
4.2 Logit Regression for AT and the VWAP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
4.3 Correlations and Autocorrelations for Order Imbalances . . . . . . . . . . . . . . . . . . . . . . . 87
4.4 Event Day Market-Adjusted Return Regressions on Order Imbalances . . . . . . . . . . . . . . . 89
4.5 Event Day Return Regressions on Differences Between AT and non-AT Order Imbalances . . . . 91
5.1 Descriptive Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
5.2 State Space Model for Intraday Prices and Liquidity Demanding Order Flows . . . . . . . . . . . 110
5.3 State Space Model for Intraday Prices and Liquidity Supplying Order Flows . . . . . . . . . . . . 112
5.4 State Space Model for Intraday Prices and Liquidity Demanding Order Flows on Turbulent Days 114
5.5 State Space Model for Intraday Prices and Liquidity Supplying Order Flows on Turbulent Days . 115

x
Abbreviations
ARIMA Autoregressive Intergraded Moving Average
ARMA Autoregressive Moving Average
ASX Australian Securities Exchange
AT Algorithmic Trading
CAR Cumulative Abnormal Return
CFTC Commodity Futures Trading Commission
CT Computerized Trading
E-mini E-mini S&P 500 Futures
FTT Financial Transactions Tax
HFT High Frequency Trading
IV Instrumental Variable
NASDAQ National Association of Securities Dealers Automated Quotations
NYSE New York Stock Exchange
SEC Securities and Exchange Commission
SIRCA Securities Industry Research Centre of Asia-Pacific
VAR Vector Autoregression
VWAP Volume-Weighted Average Price

xi
 Chapter 1

Introduction and overview

1.1 Introduction

Financial markets are experiencing a major transition due to the advance of Algorithmic Trading (AT)
technologies. The traditional views regarding the roles and the responsibilities of market participants, the
characteristics of market qualities, and the determinants of security prices all need to be re-investigated. For
example, O’Hara (2015) argue that, in a high frequency environment, short term information and price changes
might not just be asset value-related but also motivated by investors’ order flows. Many of these changes can be
directly attributed to AT. This thesis surveys the recent literature on AT and High Frequency Trading (HFT) and
investigates the characteristics of AT in a turbulent post-global financial crisis period on the Australian Securities
Exchange (ASX).
This thesis consists of one standalone survey chapter and three empirical chapters. The survey chapter reviews
around 100 recent theoretical, empirical, and policy studies on AT and HFT. The theoretical literature review
highlights the heterogeneity in the assumptions and the modeling approaches and the resulting differences in
the conclusions. Most of the theoretical literature models computerized traders as fast traders which differ in
terms of the roles they play, the information contents they possess, and their trading environments. Therefore,
we review the theoretical papers based on three criteria: market maker–taker dynamics, level of information
content, and market structural changes. We review the empirical and policy literature with the aim of providing a
comprehensive roadmap to future researchers. Specifically, we focus on the two major empirical challenges faced
by computerized trading researchers: The data identification and the causal inferences. We first introduce the
empirical papers based on how AT or HFT is identified. We then discuss the papers in relation to their causal
techniques and exogenous events. In addition, we summarize and contrast the current empirical findings around
various market quality measures. Finally, we identify the gaps in the literature and suggest several directions for

1
future research.
The first empirical chapter investigates the effects of AT on individual stock returns during turbulent periods.
We identify 39 turbulent days defined as the days when the absolute value of the market return exceeds 2%. We
then assess the relation between AT and individual stock returns on turbulent days. We find that, on days when
the market moves up by more than 2%, stocks with high AT intensity experience less upward price pressures. On
market down days, the prices of high AT intensity stocks decline less. Our results are economically significant: On
market down days, a 10% (or a half standard deviation) increase in algorithmic selling, on average, corresponds to
a 12 basis point smaller price drop in individual stocks. The economic significance is similar on market up days.
Furthermore, in the five-day period post market down days, stocks with less AT experiences more return reversals.
Taken together, we find that AT negatively relates to individual stock price swings on turbulent days. Our results
supports the view that algorithmic traders minimize the price pressures of their transactions and mitigate transient
price pressures from the market.
The second empirical chapter examines AT execution strategy and order imbalances. We relate AT execution
strategy to the Volume-Weighted Average Price (VWAP). The VWAP is a common metric for traders to assess
their execution performances. We construct intraday VWAPs and apply a logit model to assess the choice of AT
order submission in relation to the prevailing VWAP. We find that algorithmic traders are more likely to initiate
transactions when intraday VWAP moves more favorably. Specifically, a buy (sell) order lower (higher) than the
VWAP is considered favorable. AT is more likely to initiate buy (sell) orders when the prevailing VWAP moves
lower (higher). Our result implies that, compared to other traders, algorithmic traders are more contrarian and
AT may mitigate the price trends over the short-term. We then assess AT order imbalances and its effect on stock
prices. We show that the price impact of AT order imbalances are smaller compared to non-AT order imbalances.
The effects documented in this chapter are robust on turbulent days.
The last empirical chapter analyze the role of AT in the process of price discovery. To isolate the efficient
(permanent) price discovery process from the short-term noises, We apply state space models to decompose
the observed price series into an unobserved efficient price component and a transient error component. The
efficient price component, modeled as a martingale, captures information arrivals relevant to the permanent price.
The transient error component, modeled as an autoregressive process, represents the short-term noises such as

2
temporary liquidity shocks and market microstructure noises. We assess the contribution of AT to both the efficient
price discovery process and transient pricing errors. We find that AT facilitates the efficient price discovery process
by trading more in the direction of efficient price changes and less in the direction of transient pricing errors.

1.2 Contributions

This thesis contributes to the literature in several ways. First, we provide the most comprehensive survey of
the recent and emerging AT and HFT literature, to the best of our knowledge. Biais and Woolley (2011) and
Jones (2013) provide early reviews on the topic of HFT. Goldstein, Kumar, and Graves (2014) describe the history
of automated trading over the last decade and survey several AT and HFT papers. SEC (2014) summarizes
the empirical HFT literature from a policy making perspective. O’Hara (2015) and Stiglitz (2014) discuss the
broader perspectives of market microstructure research and social welfare economics respectively. We add to this
burgeoning literature by reviewing the recently published and working papers, and contribute in terms of both
depth and breadth. Specifically, we aim to provide a complete picture on the topic of computerized trading by
reviewing the empirical and theoretical literature on AT and HFT. We also decompose the theoretical modeling
approaches and offer a deeper understanding of the diversity of the theoretical conclusions.
Second, we highlight the effects of AT over time intervals relevant to non-algorithmic investors. Regulatory
agencies have been expressing concerns about the implications of AT to non-AT, which has longer trading horizons
than AT. Whether the recent changes of the equity market structure due to AT are detrimental or not for the rest
of the investment community, as pointed out in a note by the U.S. regulator, is an important issue that deserves
further investigation (SEC, 2010). However, most of the AT and HFT studies have focused on ultra-short-term
effects, ranging from milliseconds to minutes.1 For instance, Hendershott and Riordan (2013) relate AT to intraday
liquidity measures such as the bid–ask spread and order book depth. Hasbrouck and Saar (2013) propose a
framework for identifying HFT and assess its intraday effects. Brogaard, Hendershott, and Riordan (2014) assess
the impact of HFT on market qualities on a second-by-second basis and, more specifically, report the effect of
HFT in the 20 seconds around public announcements. It is relevant and intuitive to analyze automated trading at
ultra-high frequencies since many proprietary trading strategies emphasize the exploitation of small and fleeting

1
A subset of AT, HFT is generally distinguished from AT by its clear emphasis on trading speed.

3
opportunities in the market. There are reasons to suggest the longer-term implications of AT. Some computerized
traders follow an extension of traditional trading strategies, such as value, momentum, and pairs trading. These
strategies often involve holding positions over days and longer horizons. Moreover, according to the ASX (2010),
execution algorithms make up the majority of AT.2 These execution algorithms are services provided to buy-side
clients to minimize the price impact of trading and thus the intention to trade is expressed by human traders.
Therefore, the inference of these trades could be studied over horizons longer than a few minutes. We show the
association between AT and price fluctuations on the turbulent days and five days immediately after the turbulent
days.
Third, we provide empirical evidence of the relationship between AT and daily price fluctuations during the
most turbulent trading days of the global financial crisis. To the best of our knowledge, our study is the first to
investigate the impact of AT on daily stock returns. Hendershott, Jones, and Menkveld (2011) show that AT
reduces the price impact of trades over the next 5 to 30 minutes. We differ in research design by focusing on
turbulent periods and longer-term effects. Our findings suggest that AT is negatively related to stock price swings
on days with extreme price movements and the stock return reversal over the subsequent days.
Fourth, we contribute to the literature by empirically showing the difference between AT and non-AT order
imbalances. A rich literature exists on the impact of order imbalances on stock markets (e.g., Hasbrouck and Seppi,
2001; Chordia, Roll, and Subrahmanyam, 2002; Chordia and Subrahmanyam, 2004). We extend this literature by
providing evidence of heterogeneity based on order imbalances from different investor groups. We find that nonAT
order imbalances are more persistent compared to AT order imbalances; however, AT order imbalances have
significantly less impact on stock returns. After controlling for trade size and the total level of trading activity, we
find that, ceteris paribus, the impact of order imbalances by nonAT is 50% larger than by AT.
In addition, we investigate the execution strategy of AT that could affect stock price fluctuations. A number of
studies have suggested that AT and HFT could follow VWAP strategies to optimize the timing of their trades (e.g.

2
Algorithms can be separated into execution algorithms and situational algorithms (ASX, 2010). Execution algorithms (or agency
algorithms) seek to reduce the costs of executing large orders by minimizing the market impact of trades. In contrast, situational
algorithms (or proprietary algorithms) profit by monitoring and analyzing market data and news. HFT is a subset of situational
algorithms.

4
Domowitz and Yegerman, 2005; Hendershott et al., 2011; Easley, Lopez de Prado, and O’Hara, 2012). Carrion
(2013) uses end-of-day VWAP metrics to show that, ex post, HFT times the market successfully. We differ from
Carrion (2013) by using the intraday dynamic VWAP, which continuously updates throughout a trading session.
We show that, ex ante, the execution decisions of algorithms are highly sensitive to the prevailing VWAP at the
time of the trade.
Finally, provided that AT is more widely adopted and incorporates longer-term trading strategies compared to
HFT. Our results on AT price discovery compliments the results of Brogaard, Hendershott, and Riordan (2014)
on HFT price discovery. High frequency traders rely on expensive low latency technologies and compete to be
a fraction of a second faster than the others (Biais, Foucault, and Moinas, 2015). Therefore, HFT is mostly
adopted by a small number of highly sophisticated proprietary traders and electronic market makers, whereas AT
technology has been widely adopted by buy-side funds and brokerages due to its lower cost and less reliance on
infrastructure.

5
 Chapter 2

Algorithmic and High Frequency Trading: A

Review of the Literature

Chapter Summary

Algorithmic trading is a specialized trading activity in which quotes and trades are computer
generated to follow certain strategies. As a subset of algorithmic trading, high frequency trading
is generally distinguished from algorithmic trading by its clear emphasis on trading speed. Our
paper reviews the theoretical, empirical, and policy studies on algorithmic and high frequency
trading. We review the theoretical literature relating to: (1) market maker–taker dynamics, (2)
information content of trades and quotes, and (3) recently incurred or proposed market structural
changes. We aim to provide a comprehensive roadmap for future research by surveying the
empirical literature with an emphasis on how data and causal events can be identified. Our
conclusion includes a brief discussion of policy implications and suggestions for future work.

6
2.1 Introduction

The past decade has seen a tremendous proliferation of AT and HFT in financial markets.3 The growth in
automated trading technology has restructured the way we trade securities now. Computer systems have replaced
trading floors and fast moving machines have taken over the role of intermediation by human market makers.
However, these developments are not without controversy. Many unprecedented events, such as the “flash crash”
and the “Knight Capital glitch”, have exposed the vulnerability of the new trading ecosystem.4 Large scale and
negative media coverage have brought these phenomena into the public eye. Academics are keen to understand
this new phenomena. To the best of our knowledge, there have been around 100 papers on the topics of AT and
HFT written over the last few years.
Figure (2.1) illustrates the public attention on HFT via a Google trend analysis. Over the long term, Google
search frequency has elevated substantially post the global financial crisis. The “flash crash”, its related media
coverage, and regulatory releases are correlated with several initial spikes of search frequencies. Public attention
peaked around the release of the popular book: “Flash Boys” (Lewis, 2014), which generated significant scepticism
over HFT.
We aim to provide a comprehensive review of the recent and emerging AT and HFT literature. Biais and
Woolley (2011) and Jones (2013) provide early reviews on the topic of HFT. Biais and Woolley (2011) focus on
the policy implications of HFT and propose several principles of HFT regulation. Jones (2013) emphasizes the
diversity of HFT strategies and the impact of HFT on market quality, particularly market fragility issues related to
the “flash crash” and similar incidents. Goldstein, Kumar, and Graves (2014) describe the evolution of automated
trading over the past decade and survey several AT and HFT studies. SEC (2014) examines the empirical HFT

3
The percentage of AT volume has increased from about non-existent in 2003 to about 70% in 2007 for euro, dollar, and yen
currency exchange trading on Electronic Broking Services (Chaboud, Chiquoine, Hjalmarsson, and Vega, 2014). ASIC (2013) estimates
that HFT accounts for 27% of Australian equity market turnover in 2012.
4
On May 6, 2010, US stock market indices dropped by more than 5%, only to recover 30 minutes later. This incident, known as
the “flash crash”, is widely discussed in the literature. See, e.g., Biais and Woolley (2011); Kirilenko, Kyle, Samadi, and Tuzun (2017);
Jones (2013); Goldstein, Kumar, and Graves (2014); and SEC (2014). On October 16, 2013, Knight Capital’s order routing system
glitch generated more than 4 million trades in 45 minutes when processing 212 small retail orders (SEC, 2013).

7
 Figure 2.1: HFT Public Information Demand Analysis via Google Trend.
This figure shows the time series of normalized Google search frequency of the key words ‘HFT’ and ‘High Frequency
Trading’ between 2007 and 2016. The blue line is global Google search frequency normalized to a maximum of 100. The
red circles denote several crucial times when HFT has entered the public sphere.

literature and focuses on US stock market to an audience of regulators, practitioners, and academics. O’Hara
(2015) considers the market microstructure changes in light of computerized trading technology. The author
discusses how the trading world has changed and how market microstructure research should adapt. Stiglitz (2014)
provides a theoretical discussion on the broader social welfare effects of HFT and other recent financial innovations.
We add to this burgeoning literature by reviewing the previous work, and contribute in terms of both depth and
breadth. In particular, we review empirical HFT studies in conjunction with AT and theoretical literature. We
also emphasise on the diversity of the modeling approaches given that the theoretical literature features a variety
of different computerized trading strategies, our goal is to compare, contrast and assimilate the literature. Another
objective of our paper is to provide a roadmap for future research. Therefore, we focus on the two main challenges
in the empirical literature: identifying AT/HFT trades and establishing causal relation between AT/HFT and
market quality metrics.

8
 Trading algorithms can generally be divided into execution algorithms and situational algorithms (ASX,
2010). Execution algorithms, also called agency algorithms, seek to reduce the costs of executing large orders by
minimizing the market impact of trades and they account for the bulk of AT. In contrast, situational algorithms,
also known as proprietary algorithms, profit by monitoring and analyzing market data and news announcements.
HFT is a subset of AT that predominantly utilizes situational algorithms.5 SEC (2014) lists four broad categories
of HFT strategies. Passive market making generally involves the submission of limit orders that provide liquidity
in the marketplace. Directional strategies aim to profit from direction change in security prices via long or short
positions. High frequency arbitrageurs scan the related products, often in different markets, to identify arbitrage
opportunities due to temporary pricing misalignments and rely on their low latency trading technology to be the
first to seize the opportunity. Structural strategies seek to exploit structural vulnerabilities in markets or in other
market participants. For instance, traders with fast trading technologies may be able to profit by trading against
stale orders submitted by slower traders. These strategies are discussed in the empirical, and to a larger extent, in
the theoretical literature.
We use the term “Computerized Trading (CT)” when the subject matter concerns both AT and HFT throughout
this paper in order to improve the precision of our discussions. Since AT encompasses HFT, when one refers to
AT, one technically also refers to HFT. However, simply using the term “AT” when referring to an effect found in
both AT and HFT studies would be inaccurate, since the reader may assume that the effect is only found in AT
studies. Therefore, we specify the term “CT” to avoid the redundant phrase “both AT and HFT” or the imprecise
term “AT” when referring to both trading groups.6
In this paper, we review the theories on CT regarding three aspects: market maker–taker dynamics, the level
of information content, and market structural changes. Under mild assumptions, the theoretical studies model
specific CT strategies. As a result, there exists a large diversity in modeling approaches and hence a variety of

5
Although there is no clear definition, HFT generally have the following attributes: (1) acting in a proprietary capacity and
executing large amount of trades; (2) using high-speed and sophisticated computer programs; (3) holding positions for a very short
time; (4) submitting numerous orders and cancelling them almost instantly; (5) ending the trading day with almost no inventory; (6)
minimizing latencies using colocation services (SEC, 2010).
6
The introduction of “CT” allows us to leave the exclusivity function of the terms “AT” and “HFT” intact. Nevertheless, we use
“AT” or “HFT” whenever possible to highlight the significant differences between AT and HFT.

9
predictions. In other words, the conclusions reached in a theoretical study often depend on which computerized
traders groups are targeted by the authors. The modeling approaches of theoretical papers differ in the following
aspects. First, CT can be modeled as liquidity providers or liquidity consumers (e.g., Foucault, Hombert, and
Roşu 2016; Jovanovic and Menkveld 2016b). Second, the literature also models CT with different levels of private
information, i.e., informed about future order flows (Aı̈t-Sahalia and Saglam, 2014), or uninformed (Cvitanic and
Kirilenko, 2010). Third, several studies also link market structural changes related to CT, such as the introduction
of maker–taker pricing schemes (Foucault, Kadan, and Kandel, 2013), or increased degree of market fragmentation
(Biais, Foucault, and Moinas, 2015). To accommodate these three characteristics, our survey focuses on how the
models are constructed and what type of CT is discussed. We survey the theoretical literature based on whether
the targeted computerized traders are liquidity suppliers, liquidity consumers, or a mixture of both. Next, our
paper discusses the different levels of private information assumed in the literature. Last, we survey the literature
that relates CT to recently incurred or proposed market structural and policy changes.
We then survey the empirical studies on CT. The challenges faced by the empirical literature can be categorized
into two aspects. First, empirical researchers have to either proxy for CT or use datasets that flag computerized
trades. However, proxies for CT provide limited level of details for researchers to target specific questions. Moreover,
data identifying CT are usually hard to obtain and unavailable through public databases. Second, like most
other market microstructure topics, the causation between CT and market quality metrics is difficult to establish.
In the absence of direct theoretical predictions, empirical researchers can rely on “natural experiments” that
cause exogenous shocks to AT or HFT activities. In the current literature, instrumental variable regressions are
constructed based on carefully selected market events under the assumption that these events are uncorrelated
with the market quality metrics. However, some instrumental variables are not strictly exogenous. For example,
colocation is the most widely used instrument for AT and HFT. Brogaard, Hagströmer, Nordén, and Riordan
(2015) find that not only do high frequency traders employ colocation services, but many other types of traders
also use these services. To assist future research, we survey the empirical literature outlining how CT data can be
identified and introduce causal identification instruments employed in CT studies. We then summarize the effects
of CT on market quality metrics.
Although we aim to exhaust the literature on CT, it is beyond the scope of our paper to assign all existing

10
studies to each aspect that we discuss. Instead, we introduce recent CT studies based on the aforementioned three
theoretical and two empirical aspects. It is intended that this systemized approach would be more useful than
listing the contributions of individual studies in detail. Section 2.2 discusses existing literature review and policy
papers. In Section 2.3, we consider the theoretical literature with respect to three aspects of CT, namely market
maker–taker dynamics, the level of information content, and recently incurred or proposed market structural
changes. In Section 2.4, we explain the data identification and causal techniques in empirical studies. Section 2.5
examines the effects of CT on market quality and Section 2.6 provides a conclusion and suggestions for future
research.

2.2 Review and policy studies

In this section, we discuss the existing survey and policy literature on CT according to the following criteria:
First, several papers discuss the recent trading issues in the perspective of the computerized traders. The CT
literature forms a basis for their discussion in terms of various related market phenomena. Second, few studies
discuss the recent changes in market structure and its implication in social welfare, but these papers take the
broader perspectives of market designers and policy makers. Last, regulatory agencies and exchange operators
have released a number of policy papers addressing the growing popularity of CT.

2.2.1 CT review papers

Biais and Woolley (2011) provide an early review of HFT with the aim of contributing to policy debate
while drawing evidence from several theoretical and empirical studies. Specifically, the authors argue that, under
laissez-faire conditions, further development in HFT may trigger an arms’ race to minimize latency or impose
systemic risk in the market. The authors then propose potential regulatory interventions such as monitoring and
taxing HFT, imposing minimum latency, regulatory oversight, and capital requirements. Jones (2013) reviews
theoretical and empirical research on HFT and liquidity. Based on the available empirical research on HFT, the
author argues that HFT has significantly improved market liquidity and lowered trading cost for all investors. At
the same time, Jones warns that HFT may not be beneficial during extremely volatile periods such as the “flash
crash”. Goldstein, Kumar, and Graves (2014) survey a large body of literature on CT and describe its evolution as

11
well as the recent suspected decline of HFT. In addition, the authors discuss the effects of rapid computerized
traders on market fairness and market stability, giving a detailed list of market glitches during the years 2010–2013.
Goldstein, Kumar, and Graves provide additional information by quoting substantially from several reputable news
agencies (e.g., Reuters, the Wall Street Journal, and the New York Times). Menkveld (2016) provides in-depth
survey of the theory by grouping theoretical studies into seven subtopics relating to the speed advantage of HFT.
The author then discuss the empirical findings relevant to these subtopics.

2.2.2 Review papers from broader perspectives

Kirilenko and Lo (2013) survey the history and roots of AT and HFT since the beginning of quantitative
finance in the 1950s. The authors provide in-depth discussions of the extreme market events such as the “quant
meltdown”, the “flash crash”, and the “Knight Capital glitch”. Kirilenko and Lo point out that these incidents
are facilitated by the growth of AT and propose financial regulation principles to strengthen the robustness of
financial markets. Specifically, the authors recommend four AT regulation design principles: engineering based
systems, heavy safeguards, high levels of transparency, and neutral platform.
Stiglitz (2014) discusses whether recent financial innovations including CT lead to welfare improvements. On
the topic of HFT, the author argues that the acceleration price discovery do not translate into social welfare gains.
Furthermore, citing the seminal insights from Grossman and Stiglitz (1980), Stiglitz asserts that high frequency
traders are able to extract information from trades at the expense of other traders who have spent resources to
acquire information from the real economy.7 Consequently, high frequency traders dissuade others from acquiring
costly new information and thus do not contribute to better price discovery. The author also argues that HFT
can lead to less liquidity since the informational advantage of high frequency traders deters other traders from
submitting and keeping limit orders in the markets.
O’Hara (2015) describes the new market microstructure and how it is shaped by high frequency technologies.
Specifically, the author challenges the traditional definition of private information and argues that short-term

7
Grossman and Stiglitz (1980) show that prices cannot reflect all information since information is costly to acquire. Stiglitz argues
that high frequency traders extract “information rents” from that would otherwise benefit other traders who had invested in acquiring
information.

12
information in the high frequency world is not necessarily motivated by the fundamental value of assets. Instead,
informed trading is multidimensional in that trading opportunities can be motivated by traders anticipating each
other’s (or even their own) order flows. O’Hara then proposes several important market microstructure research
agendas such as market linkage and market fairness.
Kauffman, Hu, and Ma (2015) review the technological, institutional, and regulatory forces that contributed to
the development of HFT. The authors then assess the extent to which HFT has emerged in the Asian regional
financial markets. Kauffman, Hu, and Ma find that, overall, the implementation of HFT is slower in Asian markets
compared to those of the U.S. and Europe. Furthermore, the authors do not observe significant effects of HFT on
Asian markets. However, Japanese and Australian markets have the technological and institutional support for
further HFT growth.

2.2.3 Policy papers

Regulators and exchange operators have launched numerous investigations on the development and effects of CT.
SEC (2010) evaluates the U.S. equity market structure in light of HFT. The agency lists several characteristics of
HFT including the execution of large amounts of trades in a proprietary capacity, using high-speed and sophisticated
computer programs, holding short-term positions, submitting and cancelling numerous orders, maintaining a low
end-of-day inventory position, and the tendency to use colocation services. ASX (2010) assesses the impact of AT
and changes in market structure on the ASX. The report also raises concerns about balancing the interests of
short-term algorithmic traders with those of other investors. ASIC (2013) examines the impact of dark trading and
HFT on Australian financial markets. The agency find that HFT accounts for 27% of the market and some of the
common negative perceptions about HFT seem to be overstated. Specifically, the order-to-trade ratios have only
experienced moderate increase and only 1.2% of high frequency traders hold positions for an average of 2 minutes
or less. ESMA (2014) estimates the level of HFT activity in European markets. HFT activities are measured
in two ways: The direct approach identifies HFT on an institutional level based on the institutions’ business
models. The indirect approach identifies HFT according to trading behavior such as lifetime order-to-trade ratio.
The paper finds that there are significant differences in the amount of estimated HFT between the direct and
indirect approach. Finally, SEC (2014) provides an extensive survey of HFT empirical literature and discusses

13
some limitations of the literature such as the challenges of collecting data and the difficulties in distinguishing
different HFT strategies.

2.3 Theoretical studies

The theories on CT are new and emerging. Due to the diversity of CT strategies, theoretical models generally
do not attempt to capture the overarching impact of AT or HFT on security markets. Instead, the literature focuses
on incorporating one or few salient characteristics of CT into standard trading games. The most prevalent and
generally agreed characteristic is that computer algorithms are faster than human traders. However, beyond the
notion that computerized traders are fast traders, the literature has not reached a consensus on what characteristics
of CT should be modeled. Therefore, the generated predictions, which are based on different assumptions, can
often be distinct and sometimes contradictory. The differences in the modeling approaches can be summarized in
three ways: First, computerized traders are modeled as liquidity providers, liquidity consumers, or a mixture of
both. Second, computerized traders can be uninformed, partially informed, or fully informed, so their information
content varies across papers. Third, some papers assess the market design changes in the presence of computerized
traders. In this section, we survey the growing theoretical literature in order of the three CT aspects: market
maker–taker dynamics, the level of information content, and recently incurred or proposed market structural
changes.8 It is important to note that the theoretical literature mainly discusses HFT. While some studies explicitly
model HFT (e.g., Jovanovic and Menkveld, 2016b; Jarrow and Protter, 2015), other papers implicitly target HFT
by modeling “fast traders”(see, among others, Biais, Foucault, and Moinas, 2015; Hoffmann, 2014).

2.3.1 Market maker–taker dynamics

One of the aspects of contrast is whether CT provides or supplies liquidity. CT can specialize anywhere on
the continuum between pure market makers and pure market takers. An array of literature models CT as either
speculators that only consume liquidity, market makers that try to minimize adverse selection risks, or a mixture
of both.
8
Menkveld (2016) employs a similar approach that groups the theoretical HFT literature into seven categories. For each category,
the author also discusses several related empirical papers.

14
 HFT can employ pure liquidity consuming strategies. The theoretical model by Foucault, Hombert, and Roşu
(2016) studies a “fast trader” who closely resembles the directional high frequency speculators. The authors extend
Kyle’s (kyle1985continuous) model by incorporating heterogeneity in the speed of speculators reaction to new
information. The model features a fast or slow informed trader and a competitive dealer who are both continuously
trading a risky asset. The informed trader faces short and long run price movements that could have opposite
signs depending on the latest news signal and the informed trader’s forecast of the asset’s long-term payoff.9 If the
latest news signal for the risky asset is negative and the risky asset is overvalued in the long-term, the asset price
will move in a “U” shape trajectory. A slow trader, due to insufficient computing power or reaction speed, can
only execute a buy trade according to the long run price projection. However, a fast trader who possesses superior
reaction speed can quickly carry out round-trip trades and profit from the short-term dip as well as the long-term
rise. The model predicts that the net positions of fast traders are more volatile and related to short term price in
comparison to those of slow traders. When news becomes more informative, fast traders increase their amount of
trades and liquidity improves.
High frequency traders can be modeled as market makers. Jovanovic and Menkveld (2016b) model high
frequency traders as the machine middlemen between a human buyer and a human seller. In their model, the seller
and the buyer enter the market sequentially due to their opportunity costs. Without the machine, the seller would
enter the market first and post a quote. The buyer would enter later and decide whether to buy at the quoted
price. The machine is assumed to have no opportunity cost and thus can be present during both periods. After
the introduction of the machine, the seller can use the market order against the machine’s quote. The machine is
specified to possess “hard” information about the risky asset that is unavailable to human traders. Jovanovic and
Menkveld (2016b) argue that the machine middleman is more useful if its “hard” information (information that
is easily accessible and quantifiable by computers) accounts for a larger fraction of the risky asset value. Their
empirical analysis shows that the entry of HFT reduces adverse selection cost by 23% and increases the amount of

9
We note that the theoretical definition of “news” often means news from the latest price change, the price change is not necessarily
caused by a news announcement. We draw a distinction between “news feed/annoucement” in the general sense and the theoretical
“news” throughout our paper. We reserve the term “news” for the news in a theoretical context, while we use “news announcement” or
“news feed” when we discuss the news in general. We thank the reviewer for pointing out this important distinction and other highly
constructive suggestions.

15
trades by 17%. The authors also advocate a double auction mechanism instead of the prevalent limit order book
market structure.
Fast traders can be liquidity suppliers and consumers at the same time. Hoffmann (2014) extends Foucault
(1999) and features fast and slow traders who can submit limit orders and market orders. A risky asset is traded
by sequentially arrived fast and slow traders over an infinite time horizon. Slow traders can choose to either submit
a limit order or a market order, then leave before the next trader arrives. Since fast traders can better monitor
the market and react to new information, they can revise their orders before the next trader arrives if the next
trader is a slow trader. It is implied in the model that slow traders can endogenously react to the speed advantage
of fast traders. Hoffmann predicts that, overall, fast traders reduce price inefficiencies and create more trades.
However, slow traders are worse off in equilibrium due to their reduced bargaining power. When the cost of fast
trading technology is considered similar to that described in Biais, Foucault, and Moinas (2015), the investment
in fast technology is likely to result in a social welfare loss. Hoffmann recommends policies that can reduce the
bargaining power of fast traders such as randomized speed bumps.
Menkveld and Zoican (2017) study the interaction between market makers and speculators when both trader
groups employ HFT technologies. The authors incorporate high frequency market makers, high frequency
speculators (or bandits), and liquidity traders into a trading game on a single risky asset. The game is repeated
on a fixed time interval which represents the latency of the exchange. High frequency speculators and market
makers only submit or cancel orders at the beginning or the ending of each time interval. During the time interval,
both HFT groups are not allowed to access the market but the market makers can decide whether to pay the
monitoring cost. Liquidity traders might arrive to consume the quote posted by market makers, while news could
occur to change the price of the asset. These two events happen independently with a probability proportional
to the length of the time interval. The model predicts that as the time interval becomes shorter (the latency
diminishes) the high frequency market maker encounters speculators more often. Therefore, the spread increases
due to a higher adverse selection cost.
In their “white” paper, the Securities and Exchange Commission (SEC) expressed concerns about the possibility
of proprietary “front running” algorithms (p. 54, SEC, 2010). To address this concern about front running practice
of HFT, Cartea and Penalva (2012) assume high frequency traders as pure surplus extractors who stand between

16
liquidity traders and market makers. The model is a three-stage trading game, in which two liquidity traders seek
to trade with the market makers to absorb their liquidity shock in Period 1 and Period 2 respectively. However,
liquidity traders’ orders cannot reach the market maker before being “front run” by machine traders and vice
versa. Consequently, both liquidity traders and the market makers take a “haircut” by the machine traders.
Cartea and Penalva (2012) find that HFT extracts trading profits from liquidity traders and causes a greater
price impact on liquidity trades. The market maker suffers increased trading costs due to HFT extractions but
gains expected returns through higher liquidity discounts. Baldauf and Mollner (2015) model the interaction
between anticipatory high frequency traders, an analyst, and liquidity investors on multiple exchanges. The analyst
collect private information at a cost and submit orders to profit from the costly information while high frequency
traders anticipate the analyst’s orders. In equilibrium, high frequency traders profit by free-riding on the analyst’s
information. The authors argue that the anticipatory behavior of HFT discourages information acquisition and
the prices become less informative.

2.3.2 Information content

An alternative aspect of contrast is the degree of private information possessed by computerized traders.
Traders with high frequency technology can be assumed to possess private information about future order flow
from other traders. Aı̈t-Sahalia and Saglam (2014) construct a trading game in which high frequency traders are
liquidity suppliers posting limit orders and other traders demand liquidity by posting market orders. In their
model, high frequency market makers periodically receive signals about the direction of future order flow from
slow traders. The authors find that high frequency market makers can optimally adjust their limit orders based
on observed frequency and quality of the order flow signals. Aı̈t-Sahalia and Saglam (2014) then expand their
model to include both high frequency and low frequency market makers. High frequency traders exert adverse
selection risk on slower market makers and low frequency market makers respond by posting wider bid–ask spreads.
Overall, Aı̈t-Sahalia and Saglam (2014) find that better HFT technology can improve profits, liquidity provision,
and generate a higher order cancellation rates in normal times. However, the liquidity provision of HFT would
deteriorate in light of increased price volatility.
High frequency traders can acquire information and become informed about price impact by submitting

17
“exploratory” trades. Clark-Joseph (2014) updates the study of specialist’s price experimentation behavior by
Leach and Madhavan (1992) and Leach and Madhavan (1993) in the context of modern limit order book markets
with high frequency technologies.10 Specifically, Clark-Joseph (2014) constructs an order-driven market to analyze
the motivation for exploratory trades. The model is a two period (t1 and t2 ) trading game that features a high
frequency trader, the aggressive order flows, and an uncertain liquidity state. There are two possible liquidity
states: good or bad state. The liquidity state does not change between t1 and t2 . The high frequency trader can
test the prevailing liquidity state by submitting a small but costly exploratory trade. The aggressive order flow
arrives at t2 , the high frequency trader is informed about this order flow and has a chance to trade ahead of it. At
t1 , the high frequency trader faces the trade-off between submitting a costly exploratory trade at t1 to acquire
information about the liquidity state and front–running the predicted order flow at t2 without knowledge of the
liquidity state. If the high frequency trader chooses the latter, then high frequency trades would not be profitable
if the liquidity state is bad. In other words, the high frequency trader can choose to submit a small exploratory
trade at a minor cost, in order to avoid potential large costs from unfavorable order executions due to the lack of
liquidity in the market. The empirical part of Clark-Joseph (2014) tests and validates the existence of exploratory
HFT.
Li (2015) studies the competition among high frequency traders who possess persistent private information
about a risky asset’s value. Extending Kyle (1985) and Holden and Subrahmanyam (1992), Li models a risky asset
that is traded by multiple informed high frequency traders, liquidity traders, and a risk neutral market maker. The
asset is traded over discrete periods. The informed traders observe private information about the value of the asset
and submit orders accordingly, whereas the liquidity traders’ order flows are normally distributed. The designated
market maker sets prices at each period based on the order flows of informed traders and liquidity traders. The
author finds that, when the time interval between trades reduces, the order size of both informed traders and
liquidity traders decreases. In contrast to Holden and Subrahmanyam (1992), Li finds that the competition

10
Leach and Madhavan (1992) and Leach and Madhavan (1993) propose theoretical models in which market making specialists can
adjust their price quotes to induce informed order flow and expedite the price formation process. The model by Clark-Joseph (2014)
extends this idea by assuming that high frequency traders would submit small marketable transactions to gain information about the
liquidity state of the market.

18
between informed traders does not eliminate the rent from private information at equilibrium. Overall, the result
provides an explanation for the persistent profitability of HFT, despite the increasing competition between high
frequency traders.
Some high frequency traders are assumed to be uninformed. Cvitanic and Kirilenko (2010) analyze the price
formation process in a limit order book with and without an uninformed machine trader. The authors assume
that the machine trader is extremely fast in its order submissions and cancellations. The machine trader employs
a “sniping” strategy to “pick-off” stale orders in the limit order book. By following this strategy, the machine
repeatedly submits orders to the limit order book and immediately cancels any orders that are not executed. This
behavior is empirically documented by Hasbrouck and Saar (2013). Cvitanic and Kirilenko (2010) find that, even
in the absence of new information, the average transaction price is likely to change following the machine trader’s
entrance. The presence of a machine trader changes the human-to-human transaction price distribution into a
mixture of human-to-human and machine-to-human transaction price distributions.
Jarrow and Protter (2015) investigate high frequency traders’ informational advantages in an arbitrage-free
environment. The authors model a continuous trading economy with ordinary traders, sophisticated traders, and
high frequency traders. The high frequency traders are allowed to trade continuously due to their speed advantage,
whereas ordinary traders and sophisticated traders are assumed to trade in discrete times. The authors find that,
under the arbitrage-free condition, HFT can profit by executing market orders at limit order prices. HFT can also
“front run” slower market orders by submitting limit orders. In a similar environment, Jarrow and Protter (2012)
show that HFT can increase volatility and create stock price deviations from their fundamental values.

2.3.3 Recently incurred or proposed market structural changes

Recent changes in market structure and design often contribute to the proliferation of AT and HFT.11 For
example, the degree of market fragmentation is increasing in equity markets across the world. The cost of
monitoring order books and executing trades across markets have substantially increased, despite the generally

11
The popularity of AT and HFT have also significantly altered the market structure, see O’Hara (2015) for a detailed survey on
the recent and predicted market microstructure changes brought by the development of high frequency technologies.

19
beneficial effects of increased competition across trading venues.12 This gives rise to computer algorithms that
are designed to efficiently trade under conditions of market fragmentation. Brokers employ execution algorithms
to seek optimal prices for their clients, whereas proprietary traders rely on low latency technologies to identify
and trade on any arbitrage opportunities before others. Due to competition among trading venues, lower trading
costs can induce traders to post more limit orders (Colliard and Foucault, 2012). CT benefits more from limit
orders due to their higher monitoring capacities. Biais, Foucault, and Moinas (2015) capture the advantage
of CT in fragmented markets. The authors extend Grossman and Stiglitz (1980) and endogenize the trading
decisions of financial institutions on the trade-off between the gains from trade and the costs of adverse selection
and information acquisition. In their model, financial institutions seek to profit by trading a risky asset based
on their private valuation of the asset. However, the asset is traded in multiple venues and only a fraction of
those venues are liquid enough to have attractive quotes. Institutions are then at risk of not finding an attractive
quote due to their limited monitoring capacities. Financial institutions can invest in fast trading technologies to
search across venues and guarantee trading against an attractive quote. Biais, Foucault, and Moinas find that fast
traders generate a negative externality upon slow traders. In equilibrium, institutions overinvest in fast trading
technologies which results in social welfare loss. The authors also argue that, due to the benefit of fast trading
technologies in facilitating quote searches, a complete ban on fast trading is not advised but Pigovian taxes on
investment in fast trading technology are recommended.
An additional effect of increased market fragmentation is a heightened competition between exchanges. To
increase liquidity and attract investors, many exchanges have introduced market maker–taker price schemes, which
are two-level pricings that offer lower transaction fees to liquidity providers compared to liquidity consumers.
Foucault, Kadan, and Kandel (2013) provide an explanation for this scheme. The authors model market makers as
specialized high frequency liquidity providers. Market takers are assumed to behave like buy-side investors or their
brokers employing execution algorithms to break a large order into multiple small trades. In this model, a new
quote submitted by market makers creates a trading opportunity for market takers as they can trade against this

12
For instance, Shkilko, Van Ness, and Van Ness (2012) study the competition between National Association of Securities Dealers
Automated Quotations (NASDAQ) and three major electronic communication networks. The authors find that the probability of
executions on all four venues have increased due to the competition among exchanges.

20
quote. Similarly, a new trade is a profit opportunity for market makers as a quoting opportunity arises in the
absence of a consumed quote. A maker (taker) is more likely to quote (trade) when his/her monitoring intensity
increases. Therefore, trading and quoting reinforces each other, thus makers and takers exert positive externalities
on each other. The trading rate is then determined by the aggregate monitoring intensity and trading cost on
both sides. If the aggregate monitoring intensity of makers decreases compared to those of takers, the quoting
intensity would decrease as makers struggle to find quoting opportunities. The trading rate would then decrease,
since lower quoting intensity generates fewer opportunities for takers and vice versa. Exchanges could then charge
a lower fee on market making via maker–taker pricing.13

2.4 Empirical studies

Despite the exponential growth of empirical AT and HFT studies, the research on CT still faces two major
challenges. First, the data of CT are difficult to define and identify. AT and HFT are technological tools which
have the purpose to assist trading rather than distinct strategies or confined trader groups. Therefore, algorithmic
traders, and to a larger extent high frequency traders, are not clearly defined. For instance, Brogaard, Hagströmer,
Nordén, and Riordan (2015) analyze two different definitions of HFT and one widely used HFT event and find that
these three measures have substantial differences in many cases. Furthermore, many of the data used in AT and
HFT studies are not publicly available. Second, the empirical literature is placing an increasing emphasis on causal
inferences. Techniques such as instrumental variables and difference-in-difference regressions, although not new,
have received increased attention in the literature. CT studies generally relate trading activities to market quality
metrics. In a typical regression setup, a significant coefficient can arise because CT causes changes in market
quality metrics. However, the same result can be observed if CT simply reacts to changes in the market conditions.
Therefore, it is important to identify the direction of the causal relations between CT and market quality metrics.
Given the challenges in AT and HFT empirical literature, we first survey recent studies based on how CT is
identified or proxied and then discuss the events used as causal instruments and other causal establishing techniques.
We argue that, at least in the current state of the literature, AT and HFT studies cannot be discussed in isolation

13
Brolley and Malinova (2012) find that a decrease in maker fee increases liquidity but decreases market participation by investors.
However, it is uncertain whether maker–taker fee structure would receive wider adoption (see, e.g., Lam, 2015; ASIC, 2014).

21
from each other. First, proxies of AT and HFT could be correlated. The proxies introduced by Hendershott,
Jones, and Menkveld (2011) and Hasbrouck and Saar (2013) both measure, in different ways, the intensity of order
submissions, revisions, cancellations, and trades. Second, HFT makes up a significant portion of AT and there are
many similar results found on the aggregate effect of AT and HFT. Hendershott and Riordan (2013) and Carrion
(2013) find that AT and HFT, respectively, provide liquidity when it is expensive and consume liquidity when it is
cheap. Both studies also incorporate information from the futures markets into the spot markets (see, among
others, Hendershott and Riordan, 2013; Zhang, 2013).

2.4.1 Empirical studies by data identification

2.4.1.1 Studies that use CT proxies

Compared to human trading, CT generally creates larger amounts of message traffic, in terms of order
submissions, revisions, and cancellations. For instance, Hasbrouck and Saar (2013) find that high frequency traders
tend to generate “fleeting orders” to the market by repeatedly submitting limit orders only to cancel them in
the next milliseconds. SEC (2010) and SEC (2014) also list “bursts of order cancellation and modification” as a
trait for identifying HFT. Therefore, CT is likely more prevalent when there are large amounts of quoting and
trading activities compared to dollar value exchanges. Researchers exploit this trait and proxy CT as the amount
of market activities scaled by dollar volume traded. The use of proxies enables researchers to leverage publicly
available data to conduct large scale studies. However, the drawback is clear: these proxies are a “catch all” type
of measure that allows the observation only of aggregate CT effects with noise. Microstructure intricacies such as
trade imbalance and order aggressiveness cannot be uncovered via proxies.
The most widely used proxy is message traffic introduced by Hendershott, Jones, and Menkveld (2011):

HJM DollarV olumei,t

ATi,t =− , (1)
M essagesi,t
HJM
where ATi,t is the message traffic proxy for AT and messagesi,t is the electronic message traffic sent to the
exchange.14 Electronic message traffic includes quote submissions, cancellations, and trades. Boehmer, Fong, and

14
To differentiate each proxy, we added the author’s surname initials “HJM” (and “AAFH”, “HF” in Equation (2), Equation (3)

22
Wu (2015) apply this proxy in 42 exchanges internationally over the period between 2001 and 2011. The authors
find AT increases liquidity and price discovery, but causes higher volatility. Boehmer, Fong, and Wu (2013) use a
similar dataset to describe the relation between AT and the ability of firms to raise new capital. The authors find
that greater AT intensity is associated with declines in equity capital over the next year. Frino, Mollica, Monaco,
HJM
and Palumbo (2017) use ATi,t to proxy AT around earnings announcements and find that liquidity is more
resilient during periods when the level of AT activities is high.
Using a similar AT proxy, Skjeltorp, Sojli, and Tham (2015) study the effect of AT on asset prices of U.S. listed
stocks during the period 1999–2012. The authors find that the returns of low AT stocks are higher compared to
those of high AT stocks. Scholtus, van Dijk, and Frijns (2014) use the amount message traffic, the number of orders
submitted and canceled shortly after, and the number of order executions as AT proxies around macroeconomic
news announcements. Their study highlights the importance of speed during news announcements, and shows that
an additional latency of 300ms can cause an annualized loss of 1.94%. Scholtus, van Dijk, and Frijns also find
that AT increases trading volume and market depth at the best quotes, but also increases volatility and causes a
decline in the overall market depth.
Lepone and Sacco (2013) employs order-to-trade ratio as an AT proxy to investigate the impact of the exogenous
implementation of a message traffic tax by the regulators in Chi-X Canada. Order-to-trade ratio is measured as
the amount of raw message traffic over the number of trades. The authors find that the trading cost increase due
to the message traffic tax coincides with the deterioration of liquidity.
Harris (2015) propose cancel-to-trade ratio as an alternative proxy to message traffic:

AAF H Cancellationsi,t
ATi,t = , (2)
T radesi,t
AAF H
where ATi,t is the order cancellation proxy for AT. Cancellationsi,t is the number of cancellations for
orders that are within the first 10 levels of the limit order book. T radesi,t is the number of transactions within the
same time interval t. The authors argue that the cancel-to-trade ratio is a superior AT proxy to order-to-trade
(Hendershott, Jones, and Menkveld, 2011) because it is less correlated with HFT according to a Hausman test. The

respectively) as superscripts. These superscripts do not exist in original studies.

23
authors then construct a simultaneous equations system for market integrity violations (measured by closing price
manipulation and information leakage during public announcements), market efficiency (measured by effective
spread), and AT. (Hendershott, Jones, and Menkveld, 2011) find that AT reduces the incidence of closing price
manipulation and information leakage. AT also improves market efficiency by reducing effective spreads.
HFT can be proxied as the intensity of “strategic runs”. Hasbrouck and Saar (2013) show that high frequency
traders conduct strategic runs when they quickly submit a series of linked quotations, cancellations, and executions.
The authors proxy HFT as the time-weighted number of strategic runs:

HS T imeof Runsi,t
HF Ti,t = , (3)
T imeIntervali,t
HS
where HF Ti,t is the strategic runs proxy for HFT T imeof Runsi,t is the amount of time during which strategic
runs are occurring in stock i and time interval t. T imeIntervali,t is the amount of time in interval t. For example,
if we observed two strategic runs during a 10-minute interval t in stock i, and these two runs lasted for 1 minute
HS
and 2 minutes respectively, then HF Ti,t equals 0.3 (3/10). Hasbrouck and Saar (2013) compare this proxy with
the HFT data identified by NASDAQ (Brogaard, Hendershott, and Riordan, 2014). The authors find that the
proxy is highly correlated with the observed HFT activities. Hasbrouck and Saar then apply this metric to analyze
market quality of NASDAQ stocks from October, 2007 to June, 2008. The authors show that HFT is positively
associated with liquidity and volatility measures.

2.4.1.2 Studies that classify CT on aggregate

The majority of the recent studies use data that directly flag each trade or quote as CT or human trading.
These data are usually provided by exchanges to academics on a non-disclosure agreement. The exchanges identify
CT by leveraging their specialized knowledge about the account-level trading and quoting activities by their
market participants. The advantage of these data is that they allow researchers to perform more detailed analysis
in intraday and high frequency context. The drawback is that the effect of CT is assessed on aggregate. In
other words, it is difficult to analyze the heterogeneity of CT beyond a simple separation of liquidity-taking and
liquidity-providing trades (SEC, 2014). In this section, we introduce CT studies by the datasets used.

24
 AT datasets: Hendershott and Riordan (2013) use an AT dataset provided by the Deutsche Boerse. The
dataset contains the quotes and trades submitted by algorithms in the top 30 stocks of the German stock market.
The time period is January 1 to January 18, 2008. The authors analyze the order submission strategies of AT and
non-AT by order types. Compared to non-AT, AT is more likely to initiate trades when spreads are small and
more likely to provide liquidity when spreads are wide. Hendershott and Riordan then further validate the finding
via a probit regression in event-time (trade-by-trade). The dependent variable is equal to 1 if the trade is initiated
by AT (or supplied/cancelled by AT in later variations). The regressors are market conditions and the recent DAX
futures returns. The authors confirm the earlier results: algorithms are more likely to trade when it is cheaper
to trade. Furthermore, algorithmic traders are more likely to buy (sell) if the recent futures returns are positive
(negative). These results imply that AT monitors the market better than non-AT, since algorithmic traders submit
quotes and trades based on the spot market conditions as well as the returns in the corresponding futures markets.
Chaboud, Chiquoine, Hjalmarsson, and Vega (2014) employ quotes and trades data provided by Electronic
Broking Services. The paper analyzes the effect of AT on price efficiency of the triangular exchanges rate of
EUR-JPY-USD between September 2003 to December 2007. High frequency Vector Autoregression (VAR) is
constructed to assess the effects of AT on triangular arbitrage opportunities and autocorrelation in returns. The
authors find that AT reduces both triangular arbitrage opportunities and autocorrelation in return series. More
interestingly, the decreases in arbitrage opportunities are attributed to algorithmic trades that consume liquidity
whereas autocorrelation reductions are caused by algorithmic trades that supply liquidity. The result implies that
algorithmic traders engage in arbitrage strategies; they scan the market for temporary price misalignments in
related financial products and profit by initiating trades, thus AT improves price efficiency by speeding up price
discovery.
Frino, Prodromou, Wang, Westerholm, and Zheng (2017) use a transaction level dataset provided by the
Australian Securities Exchange and assess the information content of AT by linking it to corporate earnings
announcements between October 2008 to October 2009. The authors use VAR to model the series of return, AT
volume imbalance, and non-AT volume imbalance around earnings announcements. The paper finds that non-AT
volume imbalance leads AT volume imbalance prior to corporate earnings announcements but the lead direction
reverses immediately after the announcements. AT also reacts faster to public information arrivals since their

25
trades are more profitable after earnings announcements. This dataset is also used in Zhou, Kalev, and Lian
(2016). The authors investigate the effect of AT during the volatile periods around the global financial crisis. The
study finds that on volatile days, when the absolute value of the market return exceeds 2%, AT is negatively
related to stock price fluctuations. The paper also finds that order imbalances from AT have a smaller effect on
stock returns compared to those from non-AT. Last, Zhou, Kalev, and Lian show that AT tends to follow intraday
volume weighted average price metrics as suggested by Domowitz and Yegerman (2005).15
Using the same dataset, Duong, Kalev, and Sun (2016) study earnings announcements and the subsequent
information transmission among related stocks. The announcing stocks are matched with their rivals within the
same industry. The authors apply a VAR framework to examine the AT order flows between the announcing
stocks and the rival stocks. Duong, Kalev, and Sun show that, during the earnings announcements period, AT
incorporates information in other stocks that compete with the announcing stock. Their results imply that AT
synchronizes public information among related stocks.

HFT datasets: Many studies use a proprietary dataset provided by NASDAQ. This dataset is first acquired by
Terrence Hendershott and Ryan Riordan.16 The sample of stocks are stratified based on their market capitalization.
The stocks are evenly assigned into large-cap, mid-cap, and small-cap groups with 40 stocks in each group. Each
market cap group contains 20 New York Stock Exchange (NYSE) listed stocks and 20 NASDAQ listed stocks.
NASDAQ manually identified 26 HFT firms based on its knowledge of the firms and their trading activities, such
as daily net positions, order-to-trade ratio, and order durations. The time period covers 2008 and 2009.17 This
dataset allows accurate identification of HFT firms without relying on proxies or quantitative classifications. The
limitation of this dataset is that not all high frequency traders are identified, since this classification technique
is only applied at the firm level. Other traders with high frequency technologies that conduct their trading in
conjunction with activities in the same firm are not included. For example, a proprietary trading desk that is
part of a large non-HFT institution is not identified since the trades from the proprietary trading desk cannot be

15
The volume weighted average price is commonly used by the practitioners to measure the execution quality of their trades. See
Madhavan (2002) for a comprehensive survey.
16
See Brogaard, Hendershott, and Riordan (2014) for a detailed description of the NASDAQ dataset.
17
SEC (2014) provides the most comprehensive survey of the empirical literature that uses the NASDAQ dataset.

26
separately identified. A similar situation can arise if a small HFT firm routes its orders through a non-HFT firm.
Brogaard, Hendershott, and Riordan (2014) use the NASDAQ dataset to study the relation between HFT and
the price formation process, applying state space methods to decompose the stock price series into a permanent
component and a transitory component. The permanent component reflects the underlying “efficient” stock price.
Changes in the permanent prices indicate the new information arrivals. The transitory component captures pricing
errors or liquidity shocks that are uncorrelated with the “efficient” price process. Brogaard, Hendershott, and
Riordan then estimate the effects of HFT and non-HFT on these two components via the Kalman filter and
smoother, and maximum likelihood. The authors find that both HFT and non-HFT initiate trades in the direction
of permanent price movements and in the opposite direction of transitory price movements. The effects of HFT are
significantly larger than those of non-HFT. This implies that high frequency traders initiate trades to incorporate
more information into the market at the same time as reducing the transitory pricing error. Brogaard, Hendershott,
and Riordan then analyze the passive trades by HFT and non-HFT. The authors find that passive trades by
HFT, to a larger extend compared to those by non-HFT, are negatively related to permanent price changes and
positively related to transitory pricing errors. This effect is consistent with the notion that passive high frequency
traders act as market makers in the modern financial markets.
Using a similar dataset, Carrion (2013) studies how high frequency traders change their trading behavior based
on market conditions and other traders. The author first compares HFT-executed prices with the end-of-day
VWAP metrics. VWAP metrics are widely used by practitioners to evaluate their trade performance. Carrion finds
that HFT time the market better by beating the VWAP. Furthermore, similar to AT effects found by Hendershott
and Riordan (2013), HFT provides liquidity when it is cheap and consumes liquidity when it is expensive. Carrion
(2013) then tests the relation between HFT and price efficiency in terms of order imbalances and price delays. The
order imbalance results are similar to the AT findings in Zhou, Kalev, and Lian (2016); Carrion (2013) finds that
the predictive power of order imbalances on stock returns is weaker on high HFT participation days. The study
also shows a significant reduction in price delays on high HFT participation days. Overall, these results indicate
that HFT is positively associated with stock price efficiencies.
Hirschey (2016) also uses the NASDAQ dataset to study HFT order anticipation strategy. If HFT follows the
order anticipation algorithm that aims to trade ahead of large non-HFT orders, the cost of trading for non-HFT

27
will increase because the favorable quotes are already taken by HFT. Moreover, anticipatory HFT would increase
the indirect cost of trading in the form of a higher price impact.18 Emphasizing the initiated trades. Hirschey
(2016) first sorts the stocks into deciles based on net marketable buying by high frequency traders. Net marketable
buying, also known as order flow or volume imbalance in other studies, is simply the buy-trades less sell-trades
initiated by a certain trader group. Non-HFT net marketable buying is then sorted by HFT net marketable buying
deciles. The results indicate that HFT net marketable buying leads non-HFT net marketable buying by up to
5 minutes, at least in the top and bottom deciles of HFT net marketable buying. VAR is then constructed to
estimate the lead-lag relationship among stock returns, HFT net marketable buying, and non-HFT net marketable
buying. The analysis confirms that HFT leads non-HFT in order flows.
Several other studies use the NASDAQ dataset. Gerig (2015) studies how HFT synchronizes security prices
among related stocks. Synchronization occurs when price changes in one security are reflected in other related
securities. The author finds that prices synchronized much faster in 2010, compared to 2005 and 2000, and the
increased synchronization speed is largely contributed by HFT. However, the author also notes that there is
potential for HFT to “synchronize” pricing errors more than real news.19 Gao and Mizrach (2011) examine HFT
during large scale asset purchases by the Federal Reserve. The authors find that HFT is less likely to provide
liquidity, measured by the frequency of HFT quotes within the bid–ask spread. HFT is also more likely to initiate
trades against non-HFT firms during large scale purchases. Brogaard, Carrion, Moyaert, Riordan, Shkilko, and
Sokolov (2016) investigate HFT around price jumps, classified as 99.9% percentile of price changes during 10-second
intervals, and show that its liquidity provision increases. The authors suggest that HFT can provide liquidity
during price jumps due to its quicker information processing abilities.
Tong (2015) combines the NASDAQ dataset with a proprietary dataset of equity transactions by institutional
investors provided by Ancerno LTD. The execution shortfalls of 204 institutions are studied between 2008 and
2009. Execution shortfall is the percentage difference between the executed price and a prevailing benchmark price
when the order is submitted to the broker. The author finds that HFT is positively associated with the trading
cost of traditional institutions. Moreover, institutions with better historical trading-desk performance are less

18
A similar situation is incorporated into the theoretical model by Cartea and Penalva (2012).
19
Moosa and Ramiah (2015) find that the short-term profitability of HFT is likely to be overstated.

28
affected by HFT. Zhang (2013) follows the intuition in Jovanovic and Menkveld (2016b) and investigates HFT’s
reaction to “hard” and “soft” information. Hard information is easily accessible and quantifiable by computers,
whereas soft information is qualitative and hard for computers to interpret. The author proxies hard information
by extreme price movements (shocks) in the E-mini S&P 500 Futures (E-mini) futures contracts and the VIX
index. News announcements are used to proxy soft information. Zhang finds that HFT dominates non-HFT in
reacting to shocks in E-mini futures and the VIX index, and non-HFT reacts better to news announcements.
O’Hara, Yao, and Ye (2014) investigate the odd-lot trades by HFT and non-HFT. Odd-lot trades are trades
that are small and not reported to the consolidated tape. The authors show that although high frequency traders
are more likely to use odd-lot trades, slow traders’ odd-lot trades contain more information compared to high
frequency odd-lot trades. This result supports the premise of execution algorithms, whereby institutions use
algorithms to split up their large liquidity orders. Johnson, Van Ness, and Van Ness (2016) study odd-lot trades
with the NASDAQ dataset and a more recent dataset in 2013. The authors confirm that non-HFT odd-lot trades
contribute more to price discovery processes. Furthermore, odd-lot trades originated from a larger order contain
less information than other trades.
O’Hara, Saar, and Zhong (2015) employ order-level NYSE DLE (Display Book Data Log Extractor) dataset
that covers all trades and quotes between May and June 2012. The dataset categorizes each trading account into
institutions, individuals, quantitative traders, and high frequency market makers. The authors find that HFT
market makers benefit from a larger relative tick size. In a large relative tick size environment, high frequency
market makers can reduce the frequency of their limit order revisions, and high frequency market makers are
more likely to improve prices by undercutting existing limit orders. O’Hara, Saar, and Zhong suggests that a
“one-size-fits-all” tick size regulation is not optimal. The authors recommend larger tick sizes for less liquid stocks.

2.4.1.3 Studies that classify HFT on account level

Many empirical studies quantitatively classify traders as high frequency traders or slow traders based on their
trading characteristics. SEC (2010) identifies several common characteristics of HFT which can be identified via
careful examination of trader accounts: (1) HFT primarily holds positions over short-term; (2) HFT generates
large amounts of order submissions, revisions, and cancellations; (3) high frequency traders aim to close out

29
the day with no inventory; (4) traders with high frequency technologies tend to employ colocation services to
minimize latencies. A range of identification techniques are employed based on these principles. The advantage of
quantitative identification is that it provides even more detailed data. Once HFT accounts have been identified,
researchers can trace all market activities (if available) generated by these accounts, including HFT quotations,
cancellations as well as the entry of new high frequency traders. These data also allow a dichotomy of aggressive
and passive high frequency traders based on their order aggressiveness.20 Moreover, these data enable enquiries
into the “arms race” between different HFT accounts which are a primary concern discussed in policy documents
and theoretical papers. However, this identification method may not be applicable to many public data, since
public data usually do not flag activities based on individual accounts (or each proprietary firm).

Quantitative identification of HFT: Kirilenko, Kyle, Samadi, and Tuzun (2017) are the first to adopt
quantitative identification. The authors apply account level identification on the E-mini trading and trader
accounts data around the “flash crash” from May 3, 2010 to May 6, 2010. Kirilenko, Kyle, Samadi, and Tuzun
categorize trading accounts into high frequency traders, market makers, fundamental buyers, fundamental sellers,
opportunistic traders, and small traders. The 16 (out of a total of 15,422 trading accounts) HFT accounts are
identified based on three criteria:

1. Daily volume: the number of trades is more than 9 in at least one of the three days before the “flash crash”.
2. End-of-day position: on the days when the accounts have 10 or more trades, the end-of-day net positions are
no more than 5% of the daily volume.
3. Intraday position: the square root of the sum of squared deviations of each minute’s net holding from the
end-of-day net positions does not exceed 1.5% of its total trading volume for the day.

The authors find that a non-HFT large sell order triggered the “flash crash”. During the initial decline on May
6, 2010, high frequency traders provided liquidity to fundamental sellers. However, as the price declined further
and high frequency traders’ positions kept accumulating, they started to rewind their positions by consuming
liquidity and thus exacerbated the downward price pressure. Their results suggest that high frequency traders did
20
We use “aggressive” and “passive” as neutral terms to distinguish whether the traders predominantly initiate trades by submitting
market orders or facilitate trades by submitting limit orders.

30
not cause the “flash crash” but that their reactions contributed to the extreme selling pressure during the episode.
The “flash crash” has been extensively discussed in the literature, see Jones (2013) for a detailed description of
the incident and an extensive review of the related literature. A joint investigation between Commodity Futures
Trading Commission (CFTC) and SEC finds that the “flash crash” is not caused by “fat fingers” or another
singular entity; it is rather the result of a severe liquidity mismatch relating to a number of market stability issues
(CFTC/SEC, 2010). Furthermore, the “flash crash” is not an isolated incident, Goldstein, Kumar, and Graves
(2014) provide a list of other similar incidents and numerous related news feeds.

Aggressive versus passive HFT: Baron, Brogaard, Hagströmer, and Kirilenko (2016) adopt a similar but
more inclusive classification method on the E-mini data from August 2010 to August 2012. For example, instead
of criteria (3) in Kirilenko, Kyle, Samadi, and Tuzun (2017), Baron, Brogaard, Hagströmer, and Kirilenko require
HFT accounts to have a median net position range (maximum position less minimum position) of less than 10% of
the total volume traded on that day. Furthermore, the authors distinguish the identified HFT accounts based on
how often they initiate trades:21

1. Aggressive high frequency traders are HFT accounts that initiate at least 60% of their trades.
2. Mixed high frequency traders are those that initiate between 20% and 60% of their trades.
3. Passive high frequency traders are those that initiate less than 20% of their trades.

Baron, Brogaard, Hagströmer, and Kirilenko find that aggressive high frequency traders are more profitable
than passive ones. The profits of aggressive HFT are concentrated towards a small number of firms. Overall, HFT
firms outperform non-HFT firms and generate a median Sharpe ratio of 4.5 in August 2010. Baron, Brogaard,
Hagströmer, and Kirilenko (2016) further proxy the speed of high frequency traders by how fast they switch from
a passive trade to an aggressive trade. Their results suggest that speed is positively associated with profitability of
HFT. Last, HFT new entrants are more likely to underperform and exit the market. This result implies that the
competition among HFT firms is high.

21
Many market makers employ HFT technologies to reduce the cost of monitoring the market, for example Menkveld (2013) finds
that a single HFT market maker accounted for 64.4 % trade participation in the new trading venue, Chi-X Europe, in 2007.

31
 Benos and Sagade (2016) further explore the heterogeneity in HFT. The authors study a dataset that contains
account level information of transactions on UK markets between September 1, 2012 and December 31, 2012.
A classification of 26 HFT firms on account level is generated with the assistance of the UK Financial Conduct
Authority. The HFT firms are then segregated into three subcategories: aggressive HFT, neutral HFT, and
passive HFT. Benos and Sagade then apply the VAR frameworks to time-series of the stock prices, the order flow
initiated by the three groups of high frequency traders, and the order flow initiated by other traders to study the
price discovery characteristics of HFT firms in terms of their aggressiveness. Aggressive HFT firms dominate the
price discovery process followed by neutral HFT firms and market making HFT firms. The authors’ event-time
(trade-by-trade) analysis confirms their calendar-time results and shows that aggressive HFT contributes more
to the permanent price changes in proportion to noise. The results suggest that the aggressive high frequency
traders are most informed about short-term price innovations, while the passive ones use market orders primarily
to re-balance their inventory.

2.4.2 Causal inferences

A common issue in CT studies is establishing causality. CT studies aim to address the effects of CT on
market quality metrics using econometric tests. However, the results can be correlations instead of causations. For
example, CT is found, on aggregate, to improve liquidity in terms of bid–ask spreads and market depth (see, e.g.,
Hendershott, Jones, and Menkveld, 2011; Hasbrouck and Saar, 2013). At the same time, Hendershott and Riordan
(2013) and Carrion (2013) find that algorithmic and high frequency traders execute their transactions based on
the prevailing market liquidity conditions. Therefore, the two-way causation between CT and market conditions
should be disentangled. In this section, we review the techniques and the exogenous events used in the literature
to establish the causal relation between CT and market quality measures.

2.4.2.1 Technological upgrades

Instrumental Variable (IV) analysis is the most commonly used causal identification technique in the empirical
literature on CT. IV analysis requires an exogenous shock to the variable of interest (AT or HFT). Since CT
is facilitated by technological advances, many studies use trading system upgrades and the introduction of fast

32
trading services as an instrument for AT.
The first causal event for AT in the literature is a technology upgrade to the NYSE trading system called
Autoquote. With the Autoquote upgrade, the market displays a large liquidity quote (generally more than 15,000
shares) alongside the best bid and offer (Hendershott, Jones, and Menkveld, 2011). This upgrade allows AT to
better monitor the market and execute trades accordingly. Autoquote is not introduced simultaneously for all
NYSE stocks. The upgrade is phased in gradually in stock batches. This allows researchers to isolate market-wide
time effects from the effects of increased AT. Hendershott, Jones, and Menkveld (2011) use Autoquote as an
instrument to analyze the causal effects of AT on NYSE stocks around the period of Autoquote implementation
(December 2002 to July 2003). The authors find that AT improves liquidity by narrowing spreads and reducing
adverse selection costs. AT reduces the price impact of trades and improves the informativeness of quotes by
reducing the fraction of price discovery related to trades.
Goldstein, Kwan, and Philip (2016) investigate the effect of HFT on the transaction costs of non-HFT around a
technological update introduced by the Australian Securities Exchange in April 2012. The new technology, known
as ASX ITCH, enables access to ultra-low latency market information that attracts latency-sensitive traders. Their
study exploits exogenous increase in the effectiveness of HFT to establish causal inference. Goldstein, Kwan, and
Philip conduct difference-in-difference tests for limit order transaction costs over the three months before and three
months after the introduction of ASX ITCH. The authors find that HFT increases the limit order transaction
costs of non-HFT and that HFT is more successful in trading ahead of limit orders of non-HFT.

2.4.2.2 Exchange fee changes

Another possible instrumental variable is related to transaction and quotation fee changes. Many algorithmic
strategies rely heavily on generating large amounts of message traffic. For instance, execution algorithms break
up large orders into small packets of trades to minimize their price impact and transaction costs, whereas high
frequency market makers submit large amounts of limit order quotations, revisions, and cancellations to minimize
their adverse selection costs. It is conceivable that a fee change would have a larger effect on CT in comparison to
human trading.

33
 Malinova, Park, and Riordan (2013) draw their causal inference between AT and market quality via a regulatory
fee change. On April 1, 2012, the Investment Industry Regulatory Organization of Canada, a regulatory body,
started to charge its dealers an IT cost-recovery fee based on the message traffic generated by trading. The
dealers passed on these fees to its high activity clients, including computerized traders. The authors apply the
first stage IV regression on the fee change and the level of AT. Malinova, Park, and Riordan find that the fee
hike exogenously decreased message traffic generated by AT. The authors then show that liquidity deteriorates in
the form of increased spreads and adverse selection costs, due to the exogenous decrease in AT message traffic
following the fee change.

2.4.2.3 Financial transactions tax

Many computerized algorithms rely on submitting large amounts of small trades and quotes to minimize
transaction costs or profit from fleeting opportunities. Therefore, Financial Transactions Tax (FTT) on trades
and especially quotes would disproportionably affect HFT. Several studies recommend implementing FTT to
internalize the negative externalities of HFT (see e.g., Jones 2013 and Goldstein, Kumar, and Graves 2014).
However, the efficacy of FTT is widely supported by empirical evidences. Colliard and Hoffmann (2017) analyze the
implementation of a French FTT in August, 2012 and find that market quality deteriorates after the introduction
of the FTT.Lepone and Sacco (2013) investigate the impact of a message traffic tax implementation in Chi-X
Canada in April, 2012. A message traffic can be a quote submission, revision, or cancellation. The authors find
that the trading cost increase due to the message traffic tax coincides with the deterioration of liquidity.

2.4.2.4 Colocations

The trading opportunities identified by algorithms are generally small and fleeting. To trade on these
opportunities requires nearly instant connection speed with the exchange. A tiny delay may result in a non-
execution if another order arrived a millisecond earlier. For instance, Goldstein, Kumar, and Graves (2014) argue
that HFT strategies rely on being the first to seize profit opportunities. O’Hara (2015) asserts that high frequency
traders can turn reaction speed via colocation and other technologies into informational advantage. High frequency
market makers also need to minimize the delay to minimize the risk of their stale order being picked off due to new

34
information arrivals. The desire to be the first to trade motivates exchanges and computerized traders to reduce
latencies across various trading venues. Laughlin, Aguirre, and Grundfest (2014) document a three millisecond
reduction of intermarket latencies between Chicago and New York from April 2010 to August 2012. The authors
attribute the drop in delays to the increased presence of CT, alongside other technological infrastructure upgrades.
However, the delay caused by the physical distance between the location of the exchange and the location of the
trader cannot be eliminated by faster computers. For example, Garvey and Wu (2010) analyze the execution
quality of traders in relation to their geographical locations. The authors find that traders with closer proximity to
New York City areas experience faster order executions. To target the demand for closer proximity to the exchange
servers, exchanges around the world have introduced colocation services that enable market participants to locate
their machines next to the exchange server. These services aim to minimize the travel time of signals between the
exchange and the computers. The level of CT is expected to increase after computerized traders employ colocation
services.
Boehmer, Fong, and Wu (2015) use colocation introduction as an exogenous shock that increases the intensity
of AT, and compare its market quality effects before and after the colocation event in multiple countries to establish
a causal link between AT and market quality measures.22 The authors find that AT improves liquidity and price
discovery, however, AT is associated with greater volatility. Moreover, on days when AT is positively related to
volatility, AT is also negatively associated with liquidity. Based on this finding, the authors argue that algorithmic
activities on volatile days are not desirable since they also reduce liquidity. Boehmer, Fong, and Wu highlight
the cross-sectional differences in algorithmic traders’ effects stating that the benefits of AT are stronger in larger
stocks, whereas AT actually reduces liquidity in the smaller stocks.
Aggarwal and Thomas (2014) study the introduction of colocation services on the Indian stock market in
January 2010. Stocks that experience a large increase in AT after colocation are matched with other stocks based
on market quality metrics such as market capitalisation, stock price, turnover, number of trades, and floating stocks.
The authors then apply difference-in-difference methods to compare the difference of market quality measures
between the matched groups. The pre-colocation period is January 2009 to December 2009 and the post-colocation
period is July 2012 to August 2013. Aggarwal and Thomas find that AT reduces liquidity costs, order imbalance,

22
See Aitken, Cumming, and Zhan (2015) for the global starting dates of colocation services.

35
and price volatility. Using similar colocation events, Boehmer and Shankar (2014) investigate the effect of AT on
the comovement of order flow, prices, and market conditions. The authors find that AT decreases commonalities
in order flow, return, liquidity, and volatility due to the more intense competition among algorithmic traders after
the introduction of colocation.

2.5 Discussion on CT and market quality metrics

2.5.1 Liquidity

The studies on CT generally agree that CT, on aggregate, improves liquidity. For example, the theoretical
model by Roşu (2016a) predicts that fast traders generate a large trading volume, and market liquidity increases
as the number of fast traders rises. Empirically, Hendershott, Jones, and Menkveld (2011) find that AT improves
liquidity by narrowing spreads and reducing adverse selections. Bershova and Rakhlin (2013) analyze the effect of
buy-side HFT on Japanese and UK markets. Their results indicate that HFT is negatively associated with bid–ask
spreads. However, there are different results in the literature when the heterogeneity of high frequency traders
is considered. Aggressive high frequency traders primarily initiate trades to profit from short-term information,
so aggressive high frequency traders could be associated with a reduction in liquidity. For example, Foucault,
Kozhan, and Tham (2017) argue that the opportunistic arbitrage by HFT can impair liquidity since it exposes
liquidity suppliers to the risk of being adversely selected. Passive/market making high frequency traders possess
superior reaction speeds and can monitor the market cheaply. Therefore, passive HFT can outperform traditional
intermediaries and improve liquidity provision. Jarnecic and Snape (2014) find that HFT market makers submit
limit orders closer to the best quotes, and HFT limit order strategy improves bid–ask spread.
Despite the overall positive liquidity effect of CT, the literature also discusses how traditional liquidity measures
such as bid–ask spread and market depth might be problematic due to the “quote flickering” characteristic of
HFT. “Quote flickering” is an order submission strategy in which high frequency traders quickly submit limit
orders and cancel any unexecuted orders within milliseconds.23 If the market is experiencing a large amount

23
Some papers also call this “fleeting orders” or “strategic runs”. Hasbrouck and Saar (2009) show that this phenomenon is
relatively recent. Cartea, Payne, Penalva, and Tapia (2016) find pervasive quote flickering episodes on NASDAQ in March 2007–2015.
Table 2 of Hasbrouck and Saar (2013) provides an empirical example of “quote flickering”. Blocher, Cooper, Seddon, and Van Vliet

36
of quote flickering, then the calculated bid–ask spread and market depth at the best quotes would vary greatly.
Consequently, the liquidity metrics could be inflated since the flickering quotes generally cannot be utilized by
market orders. Baruch and Glosten (2013) provide the theoretical basis for quote flickering and argue that limit
order traders manage their adverse selection risk by rapidly canceling and resubmitting their quotes. Hasbrouck
and Saar (2013) use quote flickering to construct proxies for HFT. Kang and Shin (2012) find that flickering quotes
reduce the informational content of the limit order book. Van Ness, Van Ness, and Watson (2015) find that higher
order cancellation rates are detrimental to market liquidity.
In addition, CT may decrease market stability during difficult times. The theoretical model by Cespa and
Vives (2016) shows that the fragmentation of liquidity supply induced by HFT may cause traders to consume
more liquidity when the cost of supplying liquidity increases. This behavior can generate market instability.
Boehmer, Fong, and Wu (2015) argue that the benefit of AT liquidity provision reduces in volatile markets.
Hagströmer, Nordén, and Zhang (2014) analyze order aggressiveness of high frequency market makers, high
frequency speculators, and slow traders. The authors show that high frequency traders supply more liquidity when
the bid–ask spread is wide. Korajczyk and Murphy (2016) investigate HFT liquidity provision in difficult times,
and find that HFT liquidity provision is reduced in the presence of large, directional institutional trades. high
frequency traders also provide less liquidity after they incur losses. Finally, Han, Khapko, and Kyle (2014) model
the order submission and cancellation behavior of high frequency market makers. The authors argue that, since
high frequency market makers impose adverse selection on slower market makers, high frequency market makers
may drive out low frequency market makers, and liquidity would consequently deteriorate.

2.5.2 Price discovery and price efficiency

Due to the higher speed and cheaper monitoring, CT is able to quickly incorporate information into the prices
and accelerate the price discovery process. A wide range of studies provide evidence that CT improves price
discovery and reduces price inefficiencies. As described earlier in Section 2.3.2, Foucault, Hombert, and Roşu

(2016) find that high frequency traders create clusters of order cancellations by competing to get to the front of the limit order queue.
Egginton, Van Ness, and Van Ness (2016) study the intense episodic spikes in quoting activity and find that these spikes are associated
with lower liquidity, higher trading costs, and increased short-term volatility.

37
(2016) argue that fast traders can incorporate information by trading on the short-term news and long-term price
forecasts whereas slow traders can only trade according to long-term price movements. Roşu (2016b) shows that
informed trading improves liquidity and has no effect on the impact of orders. Boehmer, Fong, and Wu (2015)
analyze a wider range of markets and find that AT improves informational efficiency across international markets.
Chaboud, Chiquoine, Hjalmarsson, and Vega (2014) show that algorithmic traders improve price efficiency: AT
market orders reduce arbitrage opportunities and their passive orders reduce autocorrelations in high frequency
returns. Brogaard, Hendershott, and Riordan (2014) find that high frequency traders facilitate the price discovery
process by trading in the direction of permanent price changes and in the opposite direction of transitory pricing
errors. In addition, the literature finds that, compared to passive HFT, aggressive HFT dominates the price
discovery process. For instance, Benos and Sagade (2016) separate HFT into aggressive HFT, mixed HFT, and
passive HFT. The authors show that aggressive HFT market orders are more informed than those from mixed
and passive HFT. Brogaard, Garriott, and Pomeranets (2016) investigate the effects of competition between high
frequency traders. The authors find that price efficiency improves after the entry of aggressive high frequency
traders.
Furthermore, price discovery is increasingly caused by updates to limit order quotations without actual
transactions. Hendershott, Jones, and Menkveld (2011) show that trade-correlated information declines when
AT increases around the implementation of Autoquote around 2003. Chordia, Green, and Kottimukkalur (2016)
investigate the role of AT around macroeconomic announcements and find that order flow becomes less informative
over time as quotes respond to news directly rather than indirectly through trading. The authors also show that
there are no significant incremental profits when algorithmic traders can access macroeconomic news two seconds
earlier via Reuters’ premium news services. However, the costs of price efficiency may outweigh its benefits. Stiglitz
(2014) questions whether the improvements in price discovery translate into gains in social welfare. Chakrabarty,
Jain, Shkilko, and Sokolov (2015) investigate the effects of SEC’s ban on unfiltered connections to exchange
servers, which slowed down the speed of a significant fraction of market participants. The authors find that both
transaction costs and price efficiency have declined after the market slow down. Moreover, the reduction in trading
cost is substantially larger than the increase in pricing errors.
The literature also uncovers several sources of the information advantage of computerized traders. First,

38
computerized traders execute their orders on the spot markets based on price changes in related indices, markets,
and derivatives. For example, the theoretical model by Jovanovic and Menkveld (2016b) described earlier in
Section 2.3.2 assumes that machine traders possess “hard” information. Hard information such as prices of market
indices is easy for computers to process. Zhang (2013) shows that HFT reacts strongly to hard information proxied
by shocks in E-mini prices. Hendershott and Riordan (2013) also find that algorithmic traders can monitor prices
from futures markets and execute trades on spot markets based on futures price movements. Ito and Yamada
(2015) suggest that HFT contributes to the spillover effect between NASDAQ and Forex markets by incorporating
common information about Forex rates. Second, CT can profit from price discrepancies in related assets. Budish,
Cramton, and Shim (2015) point out that arbitrage opportunities arise due to the breakdowns in the correlations
between related assets over ultra-short time intervals (i.e. 100 milliseconds). Chaboud, Chiquoine, Hjalmarsson,
and Vega (2014) provide evidence that liquidity consuming AT is negatively associated with the frequency of
arbitrage opportunities in JPY-USD-EUR triangular currency exchange prices. Gerig (2015) suggests that HFT
improves the speed of synchronization among related stocks. Alampieski and Lepone (2012) find that HFT trades
more UK-US cross-listed stocks in the UK market when the US market opens. Last, CT can quickly adapt to
public information arrivals. Frino, Prodromou, Wang, Westerholm, and Zheng (2017) demonstrate that although
algorithmic traders are uninformed before corporate earnings announcements, they time trades better immediately
after the earnings announcements. Jiang, Lo, and Valente (2014) show that HFT improves price efficiency and
reduces liquidity during macroeconomic news announcements. Alampieski and Lepone (2011) find that HFT
liquidity consuming and liquidity providing activities increase around macroeconomic announcements.
In addition, some studies also show that small trades initiated by HFT can be uninformed. Clark-Joseph
(2014) argues that high frequency traders submit small exploratory trades to test the liquidity state of the market
and decide whether to execute their larger trades. O’Hara, Yao, and Ye (2014) provide supporting evidence that
small odd-lot trades by high frequency traders contain less information in comparison to those by other investors.
Johnson, Van Ness, and Van Ness (2016) note that odd-lot trades originated from larger orders (i.e. generated by
execution algorithms) contain less information in comparison to other trades.

39
2.5.3 Volatility

The effect of CT on short-term volatility is more divided in the literature than that on liquidity and price
discovery. On one hand, Aggarwal and Thomas (2014) estimate volatility in terms of realized volatility and high-low
midquote range over 5-minute intervals. The authors find that AT reduces volatility following the colocation
event. Hasbrouck and Saar (2013) show that HFT reduces volatility, measured as a high-low midquote range over
10-minute intervals. Hagströmer and Nordén (2013) estimate the realized volatility over 1-minute to 15-minute
intervals and separate HFT into aggressive and passive HFT. The authors find that high frequency market makers
reduce volatility. On the other hand, Boehmer, Fong, and Wu (2015) calculate a wide range of volatility metrics
over 10-minute, 30-minute, and daily time intervals. Their results suggest that AT increases volatility. Bershova
and Rakhlin (2013) measure high-low price ranges over 5-minute, 10-minute, and 30-minute intervals and show
that buy-side HFT is positively associated with volatility.
Several intricacies are noted in the CT literature on volatility. First, the “flash crash” and similar events have
raised concerns about the role of CT during high volatility periods. Biais and Woolley (2011) argue that HFT may
impose systemic risk to the stability of the markets. Jones (2013) describes in detail the impact of the “flash crash”
on the E-mini prices and several individual stocks, and surveys relevant academic and policy papers. Goldstein,
Kumar, and Graves (2014) compile an extensive list of market glitches between 2010 and 2013 associated with
the rapid responses of computerized traders. Kirilenko, Kyle, Samadi, and Tuzun (2017) is the first empirical
study to investigate the causes of the “flash crash”. The authors find that HFT, while not triggering the “flash
crash”, exacerbated the price declines. Bernales (2014) provides a dynamic equilibrium model with AT in a limit
order market, arguing that AT prefers volatile assets since it can generate larger profits during periods of high
market volatility. Zhou, Kalev, and Lian (2016) study AT on volatile days defined as days when the market rises
or declines by more than 2%. The authors find that AT is associated with smaller price fluctuations in stock prices
on volatile days. Gao and Mizrach (2016) associate HFT with market breakdowns. Market breakdown days are
defined as days when the stock falls by more than 10% of the 09:35 price and subsequently reverts back to within
2.5% of the 09:35 price at the end of the day. The authors show that HFT is positively associated with liquidity
and volatility.
Second, volatility can arise when new information is incorporated into prices. It is, therefore, important to

40
consider the effects of CT on both information related and noise related volatility. While Brogaard, Hendershott,
and Riordan (2014) do not associate HFT with traditional volatility measures, the authors show that, compared
to non-HFT, HFT initiated order flow is more positively associated with the permanent (information) component
of the price changes. Moreover, the HFT initiated order flow is more negatively associated with the transitory
(noise) component of the price changes.
In addition, CT, especially HFT, may have different effects on volatilities over different intervals. Hasbrouck
(2015) applies wavelet transform on the time-series of bid and ask quotes in US markets. The author then extracts
volatilities and variance ratios from the wavelets over different time intervals. The results indicate that during the
period of 2001-2011, although volatility of bid and ask quotes measured in ultra-short horizons (i.e. 50 to 100
milliseconds) does not exhibit strong trend, the variance ratios between ultra-short horizons and longer horizons
(i.e. 27 minutes) have increased. Breckenfelder (2013) estimates realized volatility over a wide range of time
intervals from a few minutes to a day. Breckenfelder finds that HFT competition increases intraday volatility but
has no effect on interday volatility.

2.6 Conclusion

AT and its subset, HFT, have experienced formidable growths over the past decade. In this paper, we survey
the academic and policy papers on AT and HFT. First, we discuss review papers specifically on CT, review
papers on broader topics that discusses CT, and policy papers. We point out the differences in terms of their
perspectives and contributions. Second, We survey the theoretical literature on HFT and fast traders. Since fast
trading technology can be employed by a variety of investors in the market ecosystem, theoretical studies can
target various types of computerized traders. Consequently, predictions by the theoretical models depend on what
types of computerized traders are considered. To synthesize these theoretical predictions, we focus on gauging how
HFT is modeled. We review the theoretical literature in relation to market maker–taker dynamics, information
content of fast traders, recently incurred market structural changes, and proposed market changes. Third, we
review the empirical literature with an emphasis on how to identify the proxies and data used as well as the
techniques employed in establishing the causal links between CT and market quality. We hope that this review
provides a roadmap for future research, particularly in overcoming difficulties of identifying CT and establishing

41
its causal relations with market quality measures. Last, we summarize and discuss the impact of CT on market
quality. Given the different effects documented in the literature, we point out several issues related to measuring
and interpreting the observed market quality effects.
We note that the results discussed in this paper are subject to a few caveats. Since the literature on CT is
relatively new and emerging, little consensus has been reached on the effects of CT. In addition, many papers
discussed are working papers, which are subject to changes as they go through the peer review processes. A more
significant challenge is that the empirical studies on CT are limited by data constraints. Specifically, various
proxies and datasets mainly focus on the aggregated effects of CT. Account level data are confronted by privacy
issues. These data restrictions contribute to the gap between the negative public perceptions and the mixed
results in the literature. Due to the extensive media coverage and popular books featuring anecdotal stories,
regulatory agencies around the globe are under pressure to make substantial policy changes.24 We advise against
indiscriminate restrictions on CT such as global speed bumps and transaction fee hikes.25 Instead, we encourage
regulators to collaborate with academics by providing anonymously identified account level data, in order to
promote rigorous, robust, and independent scholarly research.
Finally, we suggest some directions for future research. First, further research is needed to determine the
optimal market design in the new trading environment. This line of inquiry could focus on: what is the socially
optimal transaction speed;26 and what is the best transaction fee structure. Second, with the proliferation of
account level HFT data and detailed market microstructure prices and quotes, future research could focus on the
finer details in terms of both HFT and microstructure metrics. For instance: how to distinguish HFT beyond our
current binary aggressive/passive dichotomy; how to identify HFT based on the heterogeneity of HFT strategies;
and how to link the accelerated price discovery processes to volatility metrics in different time intervals. Third,
the source of the private information of computerized traders is not clear. O’Hara (2015) notes that, in a high
frequency world, private information (and the associated price changes) over very short time intervals might

24
For instance, Lewis (2014) claims that the market is “rigged”, however, no computerized trader is cited. Mary Jo White, the
chairwoman of the SEC, says that “the markets are not rigged” (see The Committe on Financial Services, 2014, p. 12).
25
See Malinova, Park, and Riordan (2013) for the adverse effect of regulatory fee increases.
26
Several theoretical studies discuss the optimal trading frequency, see, among others, Budish, Cramton, and Shim (2015); Foucault,
Kadan, and Kandel (2013); Guo (2015); Bongaerts and Van Achter (2016).

42
not just be asset value-related but also investor’s order-related. Therefore, the boundary between “informed”
and “uninformed” traders could become blurry. Future studies on the source, heterogeneity, and the implications
of computerized traders’ information content are warranted. In addition, market fairness should attract more
attention due to recent changes in the trading ecosystem. HFT can potentially impose adverse effects on other
market participants and more market fairness metrics are needed to capture these effects. Last, researchers could
exploit the frequent policy and market structure changes because the important changes, such as order priorities
and minimum order resting time, can potentially have large adverse effects. In addition, many changes provide
clean natural experiment opportunities that can enable researchers to disentangle causations from correlations.

43
 Chapter 3

Algorithmic Trading in Turbulent Markets

Chapter Summary

We investigate the role algorithmic trading on days when the absolute value of the market return
is more than 2%. We find that the abnormal return of a stock is related to the stock’s AT
intensity, that high AT intensity stocks experience less price drops (surges) on days when the
market declines (increases) for more than 2%. This result is consistent with the view that AT
minimizes price pressures and mitigates transitory pricing errors.

44
3.1 Introduction

Technological developments over the past decade have substantially increased the use of computer algorithms
by equity investors. In light of extreme market events such as the “flash crash”, academics, market regulators, and
finance practitioners are keen to understand the role of AT in turbulent markets.27 Does AT contribute to further
price decline in individual stocks when there is a large drop in the overall market? The answer is not clearly
answered by the literature. On one hand, algorithmic traders act as “messengers” by transmitting price movements
among derivatives, indices, and related stock prices (See, e.g. Jovanovic and Menkveld, 2016b; Hendershott and
Riordan, 2013; Zhang, 2013; Chaboud, Chiquoine, Hjalmarsson, and Vega, 2014). Therefore, AT would exert
price pressure in individual stocks by reacting to the large drop in market indices. On the other hand, computer
algorithms can minimize price pressure, reduce the price impact of trades, and mitigate pricing errors (See, e.g.
Hendershott, Jones, and Menkveld, 2011; Malinova, Park, and Riordan, 2013; Brogaard, Hendershott, and Riordan,
2014). One could argue that AT would mitigate the downward price pressure from overall market drop. The
answer to the question of whether AT contribute to extreme price movements is important to the stability of the
new trading ecosystem. Building on the research framework of Dennis and Strickland (2002), we address this
question by analyzing the relation between AT and the returns of stocks on days when there is a large movement
in market indices defined as days when the absolute value of the market return exceeds 2%. Specifically, we aim to
answer the following questions: Does AT contribute to the stock price decline (increase) when there is a large
market drop (gain)? If so, what properties of AT are the cause?
We employ a novel dataset in which the trades are flagged based on whether they are generated by computer
algorithms. We focus on the turbulent days on the ASX from October 27, 2008, till October 23, 2009. Our
sample has the following advantages. First, we are able to identify AT buy and sell trades without relying on
proxies such as message traffic. Second, in their seminal study, Hendershott, Jones, and Menkveld (2011) find that,
overall, AT plays a beneficial role in terms of liquidity and price discovery in rising markets. At the same time,
the authors warn that investigations of the characteristics of AT “in turbulent or declining markets” (p. 31) are

27
On May 6, 2010, U.S. stock market indices and related securities experienced a sharp price drop of more than 5%, only to recover
in the course of about 30 minutes. See Kirilenko, Kyle, Samadi, and Tuzun (2017) for a detailed analysis.

45
equally important. We study the effects of AT during the most turbulent times in decades, since our sample period
is at the peak of the financial crisis and immediately after the collapse of Lehman Brothers in September 2008.
Third, Biais, Foucault, and Moinas (2015) find that fast traders are better off than slow traders in fragmented
markets. To highlight the effects of AT in the absence of market fragmentation, we analyze AT prior to the entry
of Chi-X in October 2011 when the ASX was the only exchange operator in Australia.
We find that, controlling for size, risk, liquidity, and information, stocks sold by AT experience less downward
price pressure than those sold by non-AT when there is a broad market decline of more than 2%. We obtain
similar results for algorithmic buy trades when the market is up by more than 2%. Our results are economically
significant. For example, we find that a 10% (or a half standard deviation) increase in AT selling, on average,
corresponds to a 12 basis point increase in returns for individual stocks in bear markets. Furthermore, stocks that
have low levels of AT experiences significant return reversals following market decline days. Specifically, stocks
traded less by AT tend to recover from their large turbulent day price drops during the subsequent trading days.
Our findings on post-event day return reversals imply that non-AT overreacts to the overall market pressure by
pushing stock prices beyond their fundamental values. Overall, our evidence suggests that, compared to non-AT,
AT does not contribute to price swings among individual stocks in turbulent markets.
We isolate the effect of AT on event days from the endogenous effect between AT and stock movements
by applying a propensity score matching algorithm. We match event day observations (the treatment group)
with non-event day observations (the control group) based on return, volatility, market cap, and liquidity. The
matching effectively eliminates any meaningful difference between the treatment group and the control group in
terms of these market condition covariates. We apply the matching algorithm to market up and market down
days separately. This matching procedure gives us a representative control group that has market characteristics
similar to those on event days. We then use a difference-in-differences method to distinguish the effects of AT on
event days in comparison to a sample of stock-days with similar market conditions. We show that the negative
association between AT and price fluctuation only exists in the treatment group despite the similarity between the
treatment and the control groups. Furthermore, we show that this relation is unchanged in light of firm-specific
news arrivals. In summary, we show that the relation between AT and market fluctuations is not driven by AT
reacting to an individual stock’s return, volatility, liquidity, size, or public information arrivals.

46
 AT is associated with fewer price swings on stressful days; however, the mechanics of this relation are beyond
the scope of the current chapter. While the results in this chapter imply that AT actively counteracts pricing
errors, it is also possible that algorithms aim to minimize transaction costs and stay away from stocks experiencing
the largest price swings.28 However, it is sufficiently clear that AT does not exacerbate extreme price movements.
The rest of the chapter is organized as follows. Section 3.2 briefly discuss the related literature on AT and
HFT. Section 3.3 describes our data. Section 3.4 presents our findings on AT in turbulent markets. Section 3.5
discusses the association between AT, market conditions, and news. Finally, Section 3.6 concludes the chapter.

3.2 Literature review

Our paper relates to the rapidly growing literature of AT/HFT. Several authors have surveyed this topic:
Goldstein, Kumar, and Graves (2014) describe the evolution of automated trading over the past decade and
survey several AT and HFT studies; SEC (2014) examines the empirical HFT literature and focuses on the US
stock market for an audience of regulators, practitioners, and academics; O’Hara (2015) considers the market
microstructure changes in light of computerized trading technology. The author discusses how the trading world has
changed and how market microstructure research should adapt. The theoretical literature focuses on incorporating
the “fast” nature of computerized trading into trading games. Foucault, Hombert, and Roşu (2016) argue that fast
traders can incorporate information by trading on the short-term news and long-term price forecasts whereas slow
traders can only trade according to long-term price movements. The theoretical model of Jovanovic and Menkveld
(2016b) considers high frequency traders as the machine middlemen and finds that HFT incorporates information
that is easily accessible and quantifiable by computers such as index values. Foucault, Kadan, and Kandel (2013)
study maker–taker price schemes in high frequency environments. The authors find that charging market makers
lower fees encourages trades and increases liquidity. Roşu (2016a) predicts that fast traders generate a large
amount of trading volume, and market liquidity increases as the number of fast traders rises. Biais and Woolley
(2011) argue that HFT may impose systemic risk to the stability of the markets.
The empirical literature studies the impact of AT/HFT on market quality. Overall, AT and HFT improve

28
In support of our interpretation, Chapter 5 in this thesis show that AT trade less in the direction of transient pricing errors
compared to other traders.

47
liquidity (see, e.g. Hendershott, Jones, and Menkveld, 2011; Hasbrouck and Saar, 2013; Brogaard, Hagströmer,
Nordén, and Riordan, 2015). However, AT and HFT may retreat from liquidity provision during difficult times
(for instance, Hagströmer, Nordén, and Zhang, 2014; Boehmer, Fong, and Wu, 2015; Korajczyk and Murphy,
2016). AT and HFT contribute to informational efficiency (Chaboud, Chiquoine, Hjalmarsson, and Vega, 2014;
Brogaard, Hendershott, and Riordan, 2014; among others). Their information is acquired from derivatives and
indices (Hendershott and Riordan, 2013; Zhang, 2013; Ito and Yamada, 2015); related assets (Budish, Cramton,
and Shim, 2015; Gerig, 2015); and public announcements (Jiang, Lo, and Valente, 2014; Frino, Prodromou, Wang,
Westerholm, and Zheng, 2017). Boehmer, Fong, and Wu (2015) and Breckenfelder (2013) find that AT/HFT is
positively associated with volatility whereas Aggarwal and Thomas (2014) and Hasbrouck and Saar (2013) observes
a reduction in volatility when AT/HFT increases. Although it is beyond the scope of our study to measure and
model volatility, our research design allows us to draw inferences on price swings in a stock by investigating whether
the stock experiences a larger price surge (decline) when the overall market rises (drops) by more than 2%.

3.3 Data and research design

3.3.1 Data description

We employ a novel AT dataset provided by the ASX. This dataset contains all equity transactions on the
ASX between October 27, 2008, and October 23, 2009. Each trade reports the company code, trade price, trade
volume, buy/sell indicator, time stamp to the nearest millisecond, and a special indicator for both sides of the
transaction showing whether the trade was initiated by a computer or a human. Algorithmic trades are identified
based on their digital imprints on the ASX. Specifically, trades automatically generated by computers are assigned
terminal IDs different from human trades in the exchange. This classification does not completely eliminate the
possibility that human traders would submit their orders through the computer based system gateway, and vice
versa. Therefore, the classification in this paper is a proxy for human and computer based trading.29 Similar to
Hendershott and Riordan (2013), we merge the AT dataset with order-level data provided by Securities Industry

29
This dataset is also used by Frino, Prodromou, Wang, Westerholm, and Zheng (2017). For more details, see Frino, Prodromou,
Wang, Westerholm, and Zheng (2017).

48
Research Centre of Asia-Pacific (SIRCA). The SIRCA data enable the accurate identification of buy/sell trades.
This combination allows us to identify whether a trade was initiated by a buyer or a seller, as well as whether
the trade was algorithm driven. The return of the All Ordinaries Index (All Ords) was acquired from Thompson
Reuters’ tick history. The sample period is approximately one calendar year, covering the Australian stock market
around the peak of the global financial crisis. The market turbulence during our sample period allows us to study
AT in relation to extreme market movements.
Besides the ability to differentiate AT from non-AT transactions, the Australian data provide additional benefits.
First, while many previous studies rely on Lee and Ready’s (1991) algorithm, we use the “true” classification of buys
and sells from order-level data provided by SIRCA.30 Ellis, Michaely, and O’Hara (2000) and Chakrabarty, Moulton,
and Shkilko (2012) find the accuracy of Lee and Ready’s algorithm to be 81.05% and 69%, respectively. In the ASX,
Aitken and Frino (1996) show the accuracy to be 74%. Therefore, it would be beneficial to improve the accuracy
of trade direction classification. Additionally, Dennis and Strickland (2002) use quarterly sampled corporate
filing data to identify the participation rate of each investor group. We utilize real-time transaction-level data,
which enables us to better analyze the time-series of investor participation and control for potential autoregressive
properties.

3.3.2 Stock and event day selection

To ensure that there are sufficient AT and non-AT volumes in our sample to generate robust results, we limit
the sample stocks to those that were present throughout the sample period. We further delete stocks that were
traded on fewer than 200 days over the 252 trading days in our sample. Our final sample contains 384 stocks.
Researchers apply a wide range of techniques to identify the turbulent trading periods. Brogaard, Hendershott,
and Riordan (2014) select the top 10% of stock-days sorted by the volatility in the permanent (efficient) component
of the intraday price changes. Brogaard, Carrion, Moyaert, Riordan, Shkilko, and Sokolov (2016) detect extreme

30
Our data include every order in the limit order book and each order is given an unique ID. Therefore, each order’s time stamp
can be dynamically updated upon submission, revision, cancellation, and execution. In a limit order market, a trade occurs when an
ask-side order has a lower or equal price compared to a bid-side order. The trade initiator can therefore be identified by comparing
the time stamps of the orders on both sides. If the time stamp of the ask-side (bid-side) is earlier, then the trade resulted from a
liquidity-demanding buy (sell) order hitting the ask (bid).

49
price jumps as the top 0.1% of the 10-second trading intervals sorted by absolute midquote changes. Shkilko,
Van Ness, and Van Ness (2012) identify episodes of intraday downward price pressures as stock-days when the
individual stock prices fall for more than two standard deviations of historical intraday returns and then rebound.
Our approach differs from these methods based on two objectives. First, we assess algorithmic buys (sells) in
individual stocks in relation to the upward (downward) pressure from the market. Therefore, we separately identify
extreme market up days and down days. Second, we analyze the longer-term effects of AT and investigate the
stock prices on the event days and five days after the event days.
Similar to Dennis and Strickland (2002), we define turbulent days as days when the absolute values of the
returns on the market are greater than 2%.31 There are 19 market up days and 20 market down days. We use
the All Ords as our proxy for market returns. The All Ords contains the top 500 Australian ordinary stocks and
amounts to over 95% of the value of all stocks listed in the ASX. The final sample contains 9,896 stock trading
days across 384 stocks. Table (3.1) and (3.2) report the event days, number of stocks for each event day, and
returns of the All Ords.
The ASX is a highly concentrated market, large movements in the market index could be caused by a few of
the largest firms. Consequently, the selected days may contain days when the index change does not represent a
price shift among a wide range of stocks. To avoid this potential bias in our event day selection, we calculate
the percentages of firms with positive returns, zero returns, and negative returns. Furthermore, we calculate
the ratios of stocks with positive returns (negative returns) to those with negative returns (positive returns) for
positive (negative) market return days. Table (3.1) presents the market up days, the mean percentage of positive
return stocks is 72.40%, with a maximum of 81.47% on July 14, 2009, and a minimum of 57.45% on January
27, 2009. The ratio of stocks with positive returns to stocks with negative returns indicates that there are, on
average, 2.75 times more stocks with positive returns than those with negative returns over our sample period.
Table (3.2) presents the market down days, the findings for market down days are stronger, with a maximum of
92.68% negative return stocks on January 15, 2009. Overall, the results imply that individual stock returns are
overwhelmingly positive (negative) on market up (down) days.

31
As a robustness test, we select the same amount of turbulent days based on intraday realized variances of the market index, the
results are qualitatively and quantitatively similar.

50
 Table 3.1: Market Up Days and Returns
This table contains the dates, market returns, and number of stocks in the sample and the proportion of stocks that have
positive, zero, and negative returns on days when the the return of the market portfolio rises by more than 2%. The
sample period is from 27 October 2008 to 23 October 2009. Stocks are included on each event day if they were traded by
both ATers and non-ATers on the day. The market portfolio is defined as the Australian All Ordinaries index. In this
table, percent positive is the percentage of stocks with returns greater than zero, percent zero is the percentage of stocks
with returns equal to zero, percent negative is the percentage of stocks with returns less than zero, and ratio is the ratio of
percent positive (negative) to percent negative (positive) on market up (down) days. There are 19 up days and 20 down
days in our sample.
Market Number Percent Percent Percent
Return
Date of Stocks Positive Zero Negative Ratio
(%)
05-Nov-08 2.82 277 72.92 5.78 21.30 3.42
25-Nov-08 5.51 253 73.52 7.91 18.58 3.96
28-Nov-08 4.10 244 73.77 8.61 17.62 4.19
08-Dec-08 3.69 221 71.04 4.98 23.98 2.96
15-Dec-08 2.41 224 72.32 4.46 23.21 3.12
27-Jan-09 2.79 235 57.45 8.94 33.62 1.71
13-Mar-09 3.27 242 80.99 4.96 14.05 5.76
17-Mar-09 2.91 252 75.00 5.16 19.84 3.78
23-Mar-09 2.29 241 64.73 8.30 26.97 2.40
02-Apr-09 2.69 269 76.58 5.58 17.84 4.29
14-Apr-09 2.22 274 75.55 5.84 18.61 4.06
30-Apr-09 2.26 265 76.98 5.28 17.74 4.34
04-May-09 2.89 266 73.31 5.64 21.05 3.48
19-May-09 2.12 274 67.52 6.93 25.55 2.64
10-Jun-09 2.10 277 67.51 10.11 22.38 3.02
14-Jul-09 3.23 259 81.47 8.11 10.42 7.81
13-Aug-09 2.74 340 73.75 5.01 21.24 3.47
16-Sep-09 2.32 334 71.56 8.68 19.76 3.62
07-Oct-09 2.15 330 69.70 8.48 21.82 3.19

51
 Table 3.2: Market Down Days and Returns
This table contains the dates, market returns, and number of stocks in the sample and the proportion of stocks that have
positive, zero, and negative returns on days when the the return of the market portfolio drops by more than 2%. The
sample period is from 27 October 2008 to 23 October 2009. Stocks are included on each event day if they were traded by
both ATers and non-ATers on the day. The market portfolio is defined as the Australian All Ordinaries index. In this
table, percent positive is the percentage of stocks with returns greater than zero, percent zero is the percentage of stocks
with returns equal to zero, percent negative is the percentage of stocks with returns less than zero, and ratio is the ratio of
percent negative to percent positive. There are 20 down days in our sample.
Market Number Percent Percent Percent
Return
Date of Stocks Positive Zero Negative Ratio
(%)
06-Nov-08 −4.22 233 13.30 3.00 83.69 6.29
07-Nov-08 −2.43 236 34.75 5.51 59.75 1.72
11-Nov-08 −3.40 224 15.18 2.68 82.14 5.41
13-Nov-08 −5.44 220 8.64 5.00 86.36 10.00
17-Nov-08 −2.32 225 22.67 6.22 71.11 3.14
18-Nov-08 −3.47 238 15.13 5.88 78.99 5.22
20-Nov-08 −4.32 266 14.29 5.64 80.08 5.61
26-Nov-08 −2.68 225 23.56 8.44 68.00 2.89
02-Dec-08 −4.02 217 13.82 5.53 80.65 5.83
12-Dec-08 −2.31 211 29.38 5.21 65.40 2.23
08-Jan-09 −2.27 220 20.91 4.55 74.55 3.57
15-Jan-09 −4.07 205 2.93 4.39 92.68 31.67
20-Jan-09 −3.00 223 15.25 4.04 80.72 5.29
23-Jan-09 −3.83 216 13.43 6.48 80.09 5.97
02-Mar-09 −2.82 225 22.22 7.11 70.67 3.18
08-Apr-09 −2.22 255 20.78 4.71 74.51 3.58
21-Apr-09 −2.40 265 19.25 2.64 78.11 4.06
14-May-09 −3.43 277 13.72 2.17 84.12 6.13
23-Jun-09 −3.01 308 11.04 6.17 82.79 7.50
02-Oct-09 −2.04 331 9.37 4.23 86.40 9.23

52
 Although the numbers of stocks included for each event day are not identical, the sample size for each event
day is sufficiently large. The minimum is 205 on January 15, 2009, and the maximum is 340 on both August 14,
2009, and September 17, 2009. The distribution of market up days is relatively spread out throughout the year.
For market down days, however, there is a cluster of event days in November 2008.32

3.4 AT intensity and abnormal returns

In this section, we investigate whether the intensities of AT is associated with the returns in individual stocks
when the overall market experiences large surges or declines. Therefore, the cross-sectional distribution of individual
stock returns will be a function of AT intensity. We first analyze the univariate properties of AT and then report
the multivariate regression results.

3.4.1 Univariate analysis

Figure (3.1) (Figure (3.2)) presents AT and non-AT buy (sell) volume statistics for all days, market up days,
and market down days. The volume statistics are calculated as the mean of the daily trading volume across all
stocks in our sample.
Figure (3.1) and Figure (3.2) indicate that, on market up and down days, the fraction of AT liquidity-demanding
and liquidity-supplying trades compared to non-AT trades does not change significantly. The findings support
our premise that AT does not drastically change their trading behavior compared to non-AT in light of extreme
market movements.
Table (3.3) reports descriptive statistics. We measure trading activity by volume traded and the number of
transactions. We then separate buy-initiated trades from sell-initiated trades to identify additional information
from the trade direction. Panel A presents the cross-sectional averages of volume, the number of transactions, and
various AT volume (number of trades) ratios measured by the AT-initiated volume (number of trades) over the
total volume (number of trades). The daily statistics are reported for all 252 trading days, including 19 up days
and 20 down days. In line with Chordia and Subrahmanyam (2004), the number of buy trades is slightly higher
than the number of sell trades, with means of 118 and 107, respectively. AT trade size is much smaller than that

32
This turbulent period is attributed to the collapse of Lehman Brothers and the peak of the financial crisis (Longstaff, 2010).

53
 Figure 3.1: AT and non-AT Buy Volume by Event Days.
These figures contain AT and non-AT buy volume statistics by event days from 27 October 2008 to 23 October 2009.
The up (down) days are defined as the days when the market returns exceed 2% (-2%). Trader group ‘AT vs non-AT’
denotes the group of trades that are initiated by AT, and non-AT is on the passive side. Other trader groups are defined
analogously. Volume is presented in millions of shares.

54
 Figure 3.2: AT and non-AT Sell Volume by Event Days.
These figures contain AT and non-AT sell volume statistics by event days from 27 October 2008 to 23 October 2009.
The up (down) days are defined as the days when the market returns exceed 2% (-2%). Trader group ‘AT vs non-AT’
denotes the group of trades that are initiated by AT, and non-AT is on the passive side. Other trader groups are defined
analogously. Volume is presented in millions of shares.

55
of non-AT. This finding is consistent with Hendershott and Riordan (2013), in that algorithmic traders break their
orders into smaller packets to achieve the best prices.
The average daily volume differences between regular days and event days can be interpreted based on the
first two rows of Panel A in Table (3.3). Consistent with the notion that large market returns should correlate
with elevated trading volume, the average buy volume on market up days is 13.44% higher than those on all
days. On market down days, the proportion of sell volume to overall volume increased to 54.81% compared to
49.28% on regular days. This increase implies that traders are responding to the market downward pressure by
disproportionately selling stocks. However, the overall trading volume decreased by 20.57%. This abnormal decline
in volume is likely a result of the highly turbulent sample period: several of the market down days are in November
2008, at the height of the Global Financial Crisis first originated in the U.S. markets. It is an unusual time period
for the U.S. and the global economy. Australian investors were reasonably uncertain about the severity of impact
on Australian companies.
Panel B of Table (3.3) presents the cross-sectional means of individual stock time-series correlations and
autocorrelations between all AT trade, AT buy, and AT sell ratios, measured by the number of trades and volume.
The corresponding ratios measured by the number of trades and volume are highly correlated, with correlations of
0.672, 0.666, and 0.680 for all AT trades, AT buys, and AT sells, respectively. The correlations between AT buys
and AT sells measured by the number of trades and volume are 0.093 and 0.110, respectively. Panel C contains
the cross-sectional average autocorrelations of AT ratios measured by volume and the number of trades. The
autocorrelation of the ratio of AT trades to all trades is substantially high; the first-lag autocorrelation is 0.211.
The autocorrelations of AT buys and AT sells are smaller but also significant: 0.167 and 0.176, respectively. The
autocorrelations decay at a moderate speed.
For the majority of this chapter, we focus on AT liquidity-demanding trades for the following reasons.33 First,
during our sample period, buy-side execution algorithms account for the majority of AT (ASX, 2010). large
institutions employ execution algorithms to minimize the price impact of pre-existing large orders (Menkveld,
2016). Unlike many HFT firms, these buy-side institutions generally do not specialize in liquidity provision.34 The

33
We use Section 3.4.4 to explore the net effects of AT liquidity demand and supply.
34
Menkveld (2013) characterizes the trading strategy of one large HFT firm in the European markets. The author finds that over

56
 Table 3.3: Descriptive Statistics
This table contains summary statistics for daily AT ratios between 27 October 2008 and 23 October 2009. The sample
comprises 384 stocks. The event days are defined as the days when the absolute values of the market returns exceed 2%.
Panel A presents the means and standard deviations of trading volume and the AT volume ratios. Buy (sell) volume is the
mean of the total daily buy (sell) volume across all stocks in our sample. The AT volume ratio is defined as the daily
ratio between the AT volume and the overall volume. Other ratios are defined analogously. Panels B and C present the
cross-sectional means of the individual stock time-series correlations and autocorrelations. There are 252 trading days, 19
up days, and 20 down days.
Panel A: Descriptive Statistics
All Days Up Days Down Days
Mean Std. Dev. Mean Std. Dev. Mean Std. Dev.
Buy Volume (,000,000) 424 145 481 141 300 74
Sell Volume (,000,000) 412 121 393 99 364 89
No. of Buy Trades (,000) 118 25 133 29 115 28
No. of Sell Trades (,000) 107 21 101 15 108 20
AT Volume Ratio (%) 68.25 17.26 69.97 16.76 71.56 15.89
AT Buy Volume Ratio (%) 68.22 18.27 69.74 17.46 69.64 16.38
AT Sell Volume Ratio (%) 67.24 19.01 67.21 18.91 69.86 17.25
AT No. of Trades Ratio (%) 80.84 11.75 81.40 11.33 81.75 10.47
AT No. of Buy Trades Ratio (%) 79.56 12.46 80.67 11.53 80.08 10.85
AT No. of Sell Trades Ratio (%) 79.32 12.84 79.06 12.68 80.43 11.62
Panel B: Correlations
AT Buy AT Sell AT No. of AT No. of AT No. of
Volume Volume Trades Buy Trades Sell Trades
Ratio Ratio Ratio Ratio Ratio
AT Volume Ratio 0.622 0.678 0.672 0.440 0.484
AT Buy Volume Ratio 0.093 0.435 0.666 0.090
AT Sell Volume Ratio 0.441 0.090 0.680
AT No. of Trades Ratio 0.649 0.643
AT No. of Buy Trades Ratio 0.110
Panel C: Autocorrelations
AT Volume Ratios AT No. of Trades Ratios
lag All Trades Buy Trades Sell Trades All Trades Buy Trades Sell Trades
1 0.211 0.167 0.176 0.254 0.211 0.204
2 0.149 0.118 0.109 0.191 0.150 0.142
3 0.116 0.085 0.082 0.159 0.117 0.113
4 0.100 0.075 0.071 0.140 0.102 0.093
5 0.087 0.059 0.056 0.127 0.086 0.085

57
major sell-side electronic market makers have yet to enter the ASX. For example, two major global electronic
market makers, Global Electronic Trading Company (GETCO) and Virtu Financial, entered the ASX in May and
August 2011 respectively (ASX 2011a and ASX 2011b). Second, our data identify AT on a trade-by-trade basis
instead of order-by-order basis, the coarseness of our data limits us from fully interpreting the liquidity provision
results. For example, an execution of a passive AT sell order can be viewed as AT providing liquidity to inpatient
buyers. Alternatively, it is possible that the passive sell order is stale and is being picked off by other traders.
With order level data, the intentions and implications of limit order activities can be better teased out. We leave
the issues of automated market making for future studies with more granulated datasets.
Similar to Dennis and Strickland (2002), our main variable of interest is the level of AT activity in proportion
to total trading activity. We define rAT as the ratio of the AT volume divided by the total trading volume:

AT volumei,t
rATi,t = , (4)
T otalvolumei,t

where AT volumei,t is AT initiated volume for stock i on day t. T otalvolumei,t is the total volume for stock i
on day t. rAT buy and rAT sell are defined analogously for AT buy and sell volume. Throughout the paper, we
focus on AT liquidity-demanding trades instead of AT liquidity-supplying trades for the following reasons.35 First,
buy-side execution algorithms, which primarily initiates trades constitute the majority during our sample period
(ASX, 2010). On the other hand, sell-side algorithms such as HFT market makers, do not account for a significant
fraction during our sample period. Second, our robustness tests show that liquidity-supplying trade contains
insignificant information. Third, based on the results in Figure (3.1), AT does not supply more liquidity than
the amount of liquidity it consumes. Our measures are aggregated daily. The daily frequency optimally reflects
the richness of transaction data and the relevance of our results to the longer-term investors. Buy-side agency
execution algorithms remain the majority among automated algorithms during our sample period (ASX, 2010).
These algorithms tend to split large orders into smaller trades. The number of trade measures would be biased,
since the algorithms would split a large order into multiple trades. Therefore, we use volume ratios instead of the

75% of its trades are liquidity providing trades.

35
Similar to Brogaard, Hendershott, and Riordan (2014), we classify a trade as AT liquidity-demanding (supplying) if AT is on the
aggressive (passive) side of the trade.

58
number of trades.
We assess the univariate relation between rAT and the stock returns on turbulent days. Figure (3.3) presents
the association between rAT and individual stock returns on market up days and down days.

Figure 3.3: AT and non-AT Trading Volume on Up Days and Down Days.
These figures illustrate the univariate relation between AT and event day stock returns. The up (down) days are defined
as the days when the market returns exceed 2% (-2%). The top (bottom) panel is for up (down) days.

Consistent with the notion that AT minimizes price pressure and reduce the price impact of trades, we find
that low rAT stocks experience less upward (downward) pressure when the market goes up (down) by more than
2% (Hendershott, Jones, and Menkveld, 2011 and Malinova, Park, and Riordan, 2013).

59
3.4.2 Multivariate analysis

In the regression analyses, we define the intensity of AT, iAT , as the value of rAT on the event day less the
mean of rAT over the past five days:36

t−1
1 X
iATt = rATt − rATi . (5)
5 i=t−5

The decision to use abnormal values as opposed to raw values is based on the autoregressive properties reported
in Table (3.3). Specifically, we find that AT and non-AT consistently prefer certain stocks to others. Applying
the raw values on the event day for the cross-section of stocks would incorporate the information about these
preferences. The lag length is determined by autocorrelation analysis, wherein the lag length is determined by the
Akaike information criterion.
To assess how the trading activities of different investor groups correlate to individual stock returns in a
turbulent market, we model the market-adjusted return on each event day as a function of the AT intensity and
control variables. The most efficient estimation method for our panel data would be a pooled ordinary least squares
estimator. However, possible cross-sectional correlations in the error terms could be a problem. To mitigate this
issue, we follow Dennis and Strickland (2002) and use Fama–MacBeth (1973) regression on each event day:

ari = α + β1 iATi + β2 sizei + β3 turnoveri + β4 idiovari + β5 betai + i , (6)

where ari is the market-adjusted return for stock i on the event day, iATi is the intensity of AT, sizei is the
logarithm of the market value of stock i five days prior to the event day, turnoveri is the ratio of the daily volume
over the number of shares outstanding on the event day, idiovari is the idiosyncratic variance of the market model
residual of stock i on days [−125, −5], and betai is the beta of stock i for days [−125, −5]. In Panel A of Table
(3.4),
In Panel B of Table (3.4), iATi is further segregated into iAT buyi and iAT selli , corresponding to AT buy

36
We obtain similar regression results using rAT .

60
intensity and AT sell intensity, respectively:

ari = α + β1 iAT buyi + β2 iAT selli + β3 sizei + β4 turnoveri + β5 idiovari + β6 betai + i . (7)

We include beta as an independent variable, which is calculated based on historical returns over the past year.
The magnitude of beta is directly associated with market-adjusted returns and volatility. There are two reasons to
include turnover in our regression. First, previous studies have established the link between stock trades and stock
price changes (for a detailed survey, see Karpoff, 1987). Although our main variables capture AT/non-AT trading
effects, we include turnover to account for overall liquidity effects. Second, the theoretical model of Foucault,
Kadan, and Kandel (2013) predicts a strong association between AT and trading rates. Empirically, AT is reported
to follow a liquidity-driven strategy (Hendershott and Riordan, 2013). If AT is correlated with liquidity in our
sample, omitting turnover would likely force our main variables to become proxies for liquidity effects. Therefore,
we include turnover to ensure that the estimated relation between iAT and returns is robust to pricing and proxy
effects.
As for turnover, we include size in the regression to account for its possible association with return and iAT .37
Moreover, the All Ords, a value-weighted index, places more weight on larger stocks. Therefore, including size
could alleviate potential biases of returns toward larger stocks. We also include idiosyncratic variance in our
regression. Dierkens (1991) suggests using idiosyncratic volatility as a measure of informational effects. If AT
has an informational advantage, as argued by Biais, Foucault, and Moinas (2015), it will be correlated with
idiosyncratic variance.
Our main variable of interest is the intensity of AT in individual stocks in light of large market movements. We
measure the association between AT and market-adjusted returns for each stock. When the market suffers from
a price decline of more than 2%, further decline in a given stock represented by a decrease in market-adjusted
return would indicate its higher downward price pressure. If we find more AT in stocks that have a smaller
market-adjusted return, then AT would cause price fluctuations by exerting further downward pressure on individual
stocks. Alternatively, if AT levels are positively associated with market-adjusted returns on market decline days,

37
Banz (1981) finds size to be a significant factor of stock returns.

61
then AT would be beneficial to stock price stability. To emphasize the importance of trade direction, we then
segregate trading volume into buy and sell volumes. We expect the buy (sell) volume to be more relevant than the
sell (buy) volume on market up (down) days.
Table (3.4) presents the results. Panel A reports the estimations with unsigned AT intensity. The signs of
the coefficients for the AT intensity (iAT) are as expected: negative for market up days and positive for market
down days. The coefficients suggest that non-AT creates price pressure in the direction of market movement.
However, the AT intensity (iAT) is marginally insignificant (p-value of 0.113) on market up days and insignificant
(p-value of 0.181) on market down days. To further disentangle the informativeness in trade signals, we report the
estimations for the AT buy intensity (iAT buy) and the AT sell intensity (iAT sell) in Panel B. On market up days,
the AT buy intensity is significant and the AT sell intensity is highly insignificant, whereas the distribution of
significance reverts on market down days. This result suggests that the predictive power of the buy (sell) volume
on up (down) days is diluted by not assigning trade direction in Panel A. Taken together, the AT buy intensity
is negatively correlated with the market-adjusted return on up days, whereas the AT sell intensity is positively
correlated with the market-adjusted return on down days. This finding supports the notion that stocks with less
AT buying (selling) would incur a greater upward (downward) price swing on up (down) days. As a result, stocks
with higher levels of AT would reduce volatility on event days.
Economically, the association between AT intensity and market-adjusted returns are substantial. The economic
significance is most pronounced when we segregate the trades into buy- and sell-initiated trades. The coefficient
of the AT buy intensity on market up days is 1.61, which implies a decrease of 16 basis points in the predicted
abnormal returns for a 10% (or 0.53 standard deviation) increase in the AT buy intensity. Likewise, the coefficient
of the AT sell intensity on market down days indicates that a 10% (or 0.49 standard deviation) increase in AT
selling, on average, corresponds to a 12 basis point increase in abnormal returns. The coefficients of the control
variables in Panels A and B of Table (3.4) are similar in magnitude and significance. As expected, turnover is
positively (negatively) related to market-adjusted returns on market up (down) days. However, the association on
market down days is not significant, implying that market up days are more likely to be liquidity driven compared
to market down days. Size and idiosyncratic variance are not significantly related to market-adjusted returns.

62
 Table 3.4: Event Day Market-Adjusted Return Regressions on Abnormal AT Ratios
This table presents coefficient estimates from Fama–MacBeth regressions using the following model:

ari = α + β1 iATi + β2 sizei + β3 turnoveri + β4 idiovari + β5 betai + i ,

where ari is the market-adjusted abnormal return for stock i on the event day. The event days are defined as the days
when the absolute values of market returns exceed 2%. In Panel A, iATi is the abnormal volume ratio between the AT
volume and the overall volume on the event day less the mean volume ratio over the past five days. The term sizei is the
logarithm of the market value of stock i five days prior to the event day and turnoveri is the ratio of the daily volume over
the number of shares outstanding on the event day. The variable idiovari is the idiosyncratic variance of the market model
residual of stock i on days [−125, −5] and betai is the beta of stock i for days [−125, −5]. The event days are segregated
into 19 up days and 20 down days. In Panel B, iATi is further segregated into iAT buyi and iAT selli corresponding to the
abnormal buy volume ratio and the abnormal sell volume ratio, respectively:

ari = α + β1 iAT buyi + β2 iAT selli + β3 sizei + β4 turnoveri + β5 idiovari + β6 betai + i .

The control variables are identical to those in Panel A. The coefficients for iATi , iAT buyi , iAT selli , and betai are multiplied
by 100 in both panels. The coefficients for sizei are multiplied by 1,000. The p-values are reported from a t-test of the
mean coefficient being different from zero.
Up Days Down Days
Mean p-value Min Max Mean p-value Min Max
Panel A: Aggregated AT Ratio
iAT −1.01 0.113 −5.22 3.98 0.73 0.181 −2.83 6.47
beta 1.86 0.000 −0.12 4.74 −1.93 0.000 −4.66 0.29
turnover 0.60 0.001 −0.72 1.94 −0.36 0.239 −3.54 2.57
size −0.19 0.838 −5.10 10.49 0.09 0.889 −4.42 8.40
idiovar 0.42 0.380 −3.61 4.53 −0.58 0.316 −3.71 5.22
Panel B: Segregated Buy/Sell Ratio
iATbuy −1.61 0.005 −6.54 1.43 −0.31 0.579 −3.61 6.18
iATsell −0.13 0.764 −3.52 2.80 1.19 0.011 −2.58 5.05
beta 1.86 0.000 −0.11 4.80 −1.95 0.000 −4.58 0.52
turnover 0.59 0.002 −0.91 1.98 −0.33 0.272 −3.68 2.66
size −0.24 0.803 −5.35 10.46 0.13 0.848 −4.13 8.41
idiovar 0.39 0.427 −3.69 4.57 −0.61 0.300 −3.63 5.20

63
3.4.3 Post-event day analysis

The empirical findings on event days indicate that the absolute value of individual stock returns with lower AT
intensity exceeds that of returns with higher AT intensity. In this section, we show the post-event return differences
between stocks with high and low AT intensity. The return difference on the event day could be explained by
non-AT reactions to information and driving prices to their fundamental values. If this is the case, we should
observe no return reversal during the period immediately after the event day for stocks with lower AT intensity
compared to those with higher AT intensity. If, however, there are significant return reversals among stocks with
lower AT intensity, then non-AT increases price fluctuations and causes prices to deviate from their fundamental
values.
The time span of our data dictates that longer-term analysis is not feasible; nevertheless, we provide a post-event
cumulative return analysis over the five days immediately after each event day. We sum post-event market-adjusted
returns as Cumulative Abnormal Returns (CARs) for each stock and partition them into quartiles based on their
event day AT intensity. We then calculate the mean difference between CARs of higher and lower AT intensity.
The intuition is that if the return effects on event days are temporary, we will observe significantly higher CARs in
low AT stocks, compared to high AT stocks, immediately after market down days. Alternatively, if the return
effects are fundamental on event days, we will observe insignificant differences in post-event CARs.
Table (3.5) contains the results of the post-event CAR differences. In Panel A, the first (third) row contains
the mean CAR difference between the top and bottom 50% (25%) based on AT activities. The second and fourth
rows report the p-values corresponding to a test of a null hypothesis that the CARs from high/low AT quartiles
are identical. On market down days, the post-event CAR difference is significantly negative. This finding implies
significant return reversals in stocks of low AT intensity.

64
 Table 3.5: Post-Event Cumulative Abnormal Returns
This table presents the results of post-event CAR analysis for stocks ranked by AT activity quartiles. The event days
are defined as the days when the absolute values of market returns exceed 2%. In Panel A, the CAR for stock i is the
market-adjusted return over five days after each event day. The CARs for individual stocks are partitioned into quartiles
based on the AT buy (sell) volume market share on each up (down) day. The mean CAR difference between high- and
low-AT stocks are presented. In Panel B, the CAR for stock i is calculated as the five-day post-event return for stock i less
the mean five-day returns for all stocks in the same beta quartile as stock i on the event day. The p-values in parentheses
correspond to a test of a null hypothesis that the CARs from high/low-AT quartiles have identical means.
Up Days Down Days
5 Days 5 Days (Indep) 5 Days 5 Days (Indep)
Panel A: Post Event CAR Partitioned by AT Activities
Top Less Bottom Half (%) −0.201 −0.249 −0.690 −1.000
(0.595) (0.446) (0.027) (0.017)
Top Less Bottom Quartile (%) −0.095 0.348 −0.775 −1.380
(0.817) (0.607) (0.084) (0.024)
Panel B: Post Event CAR Partitioned by AT Activities (Robustness)
Top Less Bottom Half (%) −0.337 −0.227 −0.775 −1.080
(0.219) (0.480) (0.011) (0.009)
Top Less Bottom Quartile (%) −0.090 −0.180 −0.832 −1.420
(0.820) (0.700) (0.058) (0.018)

3.4.4 Net effects of AT liquidity demand and supply

In this section, we assess the net effect of AT liquidity demand and supply on turbulent days. Brogaard,
Carrion, Moyaert, Riordan, Shkilko, and Sokolov (2016) use an activity metric that nets liquidity supply and
demand from HFT. We apply a similar metric for AT. To measure AT demand and supply, we use AT volume
imbalances of liquidity demanding and supplying trades respectively:

dATi,t = dAT buyi,t − dAT selli,t

sATi,t = sAT buyi,t − sAT selli,t , (8)

where dATi,t (sATi,t ) is the volume imbalance of liquidity demanding (supplying) AT for stock i on day t.
dAT buyi,t (dAT selli,t ) is the buy (sell) initiated trading volume where the initiator is AT. sAT buyi,t (sAT selli,t )
is the buy (sell) initiated trading volume where the seller (buyer) is an algorithmic trader. The difference between
liquidity demand and liquidity supply volume imbalances is our main metric:

65
 netATi,t = dATi,t − sATi,t . (9)

A positive net AT imbalance, netATi,t , indicates net trading activity of AT in the direction of the positive
returns. I.e., A positive net AT imbalance results from either AT demands more liquidity by initiating more
buy trades than sell trades (positive dATi,t ) or AT supplies more liquidity to buy-initiated trades compared to
sell-initiated trades (negative sATi,t ).
On market up (down) days, most storks experience positive (negative) returns. A positive (negative) net AT
imbalance shows that AT facilitates trades in the direction of the upward (downward) price movements. Similar to
Equation (6), we apply Fama–MacBeth (1973) regression on event days:

ari = α + β1 netATi + β2 sizei + β3 turnoveri + β4 idiovari + β5 betai + i , (10)

where ari is the abnormal return for stock i. The main variable of interest is netATi , which measures the
net effect of AT liquidity demand and supply. Table (3.6) presents the association between abnormal return of
individual stocks on event days and the net AT imbalance. On both market up and down days, the coefficients of
netAT are negatively significant. These results imply that AT liquidity demand and supply activities are inversely
associated with upward (downward) price swings in individual stocks on market up (down) days.

3.4.5 Robustness tests

We perform several tests to assess the robustness of our results. We first test an alternative construction of the
AT intensity, iAT . Madhavan, Richardson, and Roomans (1997) argue that innovation in the order flow is more
indicative for security prices if the order flow is correlated. Innovation is modeled as the “surprise” component:
the raw value less the expected value from the previous period. We follow a methodology similar to that of
Brogaard, Hendershott, and Riordan (2014) and model the innovation in rAT , innAT , as the residual of a five-lag
autoregressive model. The number of lags is determined by the Akaike information criterion. The innovations in
the AT buys (innAT buy) and AT sells (innAT sell) are obtained analogously. Table (3.7) presents the results.
The re-estimation results in Table (3.7) are quantitatively and qualitatively similar to the initial results in

66
 Table 3.6: Event Day Market-Adjusted Return Regressions on Net AT Imbalances
This table presents coefficient estimates from Fama–MacBeth regressions using the following model:

ari = α + β1 netATi + β2 sizei + β3 turnoveri + β4 idiovari + β5 betai + i ,

where ari is the market-adjusted abnormal return for stock i on event days. The event days are defined as the days when
the absolute values of market returns exceed 2%. netATi is the net effect of AT liquidity demand and supply defined in
Equation (8) and (9). sizei is the logarithm of the market value of stock i five days prior to the event day and turnoveri
is the ratio of the daily volume over the number of shares outstanding on the event day. The variable idiovari is the
idiosyncratic variance of the market model residual of stock i on days [−125, −5] and betai is the beta of stock i for days
[−125, −5]. The event days are segregated into 19 up days and 20 down days. The coefficients for netATi are scaled by
10,000,000. The coefficients of betai (sizei ) are multiplied by 100 (1000). The p-values are reported from t-tests of the
mean coefficient being different from zero.
Up Days Down Days
mean p-value min max mean p-value min max
netAT -0.859 0.001 -2.194 2.015 -0.832 0.000 -2.559 0.523
beta 1.792 0.000 0.001 0.041 -1.921 0.000 -0.050 0.003
turnover 0.552 0.001 -0.689 1.873 -0.477 0.027 -3.266 0.582
size 0.142 0.884 -0.006 0.010 0.303 0.654 -0.006 0.009
idiovar 0.481 0.298 -3.722 3.801 -0.316 0.567 -3.402 5.519

Table (3.4): innAT is statistically insignificant when we do not distinguish buy-initiated from sell-initiated trades
and innAT buy (innAT sell) is significantly related to abnormal returns on market up (down) days, with similar
coefficients. The re-estimation confirms that our finding is robust with regard to the measurement of AT.
We also provide a sensitivity analysis of the event day selection criteria. In Table (3.8), we relax the threshold
of event day selection criteria from an absolute market return of more than 2% to a range between 1.5% and 2.5%.
Furthermore, if the market moves during the day but reverses toward the initial value at closing, the daily market
return would not capture these days. Therefore, we replace the daily market return with the daily high (low) to
open price return to capture the extreme price increase (decline) in the intraday. This measure allows us to include
stocks that rose (fell) more than 2% in the intraday but reversed below 2% at the end of the day.
We then delete all consecutive event days to eliminate the effect of interday fluctuations. In addition, we
further limit our sample of stocks to those that have at least ten trades initiated by both AT and non-AT in Table
(3.9). Finally, the result is also robust to day-of-the-week effects, firm fixed effects, and double clustering of the
standard error in stocks and days.
All robustness results in Tables (3.8) and (3.9) are quantitatively and qualitatively similar to those estimated
using the initial specifications described in Section 3.4.2.

67
 Table 3.7: Event Day Market-Adjusted Return Regressions on Innovation in AT Intensity
This table presents coefficient estimates from Fama–MacBeth regressions using the following model:

ari = α + β1 innATi + β2 sizei + β3 turnoveri + β4 idiovari + β5 betai + i ,

where ari is the market-adjusted abnormal return for stock i on the event day. The event days are defined as the days
when the absolute values of market returns exceed 2%. In Panel A, innATi is volume ratio innovation obtained as the
residual of an autoregressive model with five lags applied to the individual stock volume ratios. The term sizei is the
logarithm of the market value of stock i five days prior to the event day and turnoveri is the ratio of the daily volume
over the number of shares outstanding on the event day. The variable idiovari is the idiosyncratic variance of the market
model residual of stock i on days [−125, −5] and betai is the beta of stock i for days [−125, −5]. The event days are
segregated into 19 up days and 20 down days. In Panel B, innATi is further segregated into innAT buyi and innAT selli ,
corresponding to the abnormal buy volume ratio and the abnormal sell volume ratio, respectively:

ari = α + β1 innAT buyi + β2 innAT selli + β3 sizei + β4 turnoveri + β5 idiovari + β6 betai + i .

The coefficients for innATi , innAT buyi , innAT selli , and betai are multiplied by 100 in both panels. The coefficients for
sizei are multiplied by 1,000. The p-values are reported from a t-test of the mean coefficient being different from zero.
Up Days Down Days
Mean p-value Min Max Mean p-value Min Max
Panel A: Aggregated AT Ratio
innAT −0.93 0.160 −5.43 4.30 0.48 0.378 −3.22 5.58
beta 1.86 0.000 −0.15 4.75 −1.94 0.000 −4.66 0.32
turnover 0.60 0.001 −0.71 1.94 −0.36 0.235 −3.54 2.56
size −0.15 0.871 −4.87 10.42 0.09 0.891 −4.52 8.42
idiovar 0.42 0.391 −3.74 4.54 −0.56 0.338 −3.65 5.28
Panel B: Segregated Buy/Sell Ratio
innATbuy −1.65 0.005 −6.48 1.58 −0.49 0.398 −4.06 5.52
innATsell 0.10 0.820 −2.98 3.57 0.94 0.059 −3.44 5.17
beta 1.87 0.000 −0.12 4.80 −1.95 0.000 −4.64 0.59
turnover 0.59 0.002 −0.85 2.02 −0.35 0.255 −3.71 2.62
size −0.11 0.902 −4.91 10.49 0.11 0.871 −4.34 8.38
idiovar 0.36 0.462 −3.90 4.52 −0.58 0.319 −3.50 5.29

68
 Table 3.8: Sensitivity Test for Event Day Selection
This table presents coefficient estimates from Fama–MacBeth regressions using the following model:

ari = α + β1 iAT buyi + β2 iAT selli + β3 sizei + β4 turnoveri + β5 idiovari + β6 betai + i .

Sensitivity analysis of the event day selection method is presented based on the absolute value of market returns exceeding
the range from 1.5% to 2.5%. In the last column, we estimate the event days when the absolute difference between intraday
high/low prices exceed 2%. The variable ari is the market-adjusted abnormal return for stock i on the event day; iAT buyi
(iAT selli ) is the abnormal volume ratio between the AT buy (sell) volume and the overall buy (sell) volume on the event
day less the mean volume ratio over the past five days; sizei is the logarithm of the market value of stock i five days prior
to the event day; turnoveri is the ratio of the daily volume to the number of shares outstanding on the event day; idiovari
is the idiosyncratic variance of the market model residual of stock i on days [−125, −5]; and betai is the beta of stock i for
days [−125, −5]. The coefficients for iAT buyi , iAT selli , and betai are multiplied by 100 in both panels. The coefficients
for sizei are multiplied by 1,000. The p-values are reported in parentheses from a t-test of the mean coefficient being
different from zero.
1.50% 1.75% 2% 2.25% 2.50% high/low
No. of Up Days 32 25 19 14 10 23
No. of Down Days 29 24 20 18 13 31
Panel A: Market Up Days
iATbuy −1.199 −1.528 −1.606 −1.834 −1.592 −1.510
(0.012) (0.008) (0.005) (0.016) (0.101) (0.002)
iATsell −0.202 −0.104 −0.128 −0.137 0.064 0.041
(0.494) (0.772) (0.764) (0.775) (0.923) (0.920)
beta 1.527 1.664 1.865 1.901 1.705 1.632
(0.000) (0.000) (0.000) (0.000) (0.003) (0.000)
turnover 0.475 0.539 0.587 0.642 0.611 0.461
(0.000) (0.001) (0.002) (0.008) (0.042) (0.007)
size −0.800 −1.320 −0.236 0.105 0.746 0.328
(0.313) (0.142) (0.803) (0.929) (0.646) (0.721)
idiovar 0.163 0.194 0.387 0.334 0.565 0.645
(0.663) (0.645) (0.427) (0.601) (0.406) (0.897)
Panel B: Market Down Days
iATbuy −0.485 −0.475 −0.313 −0.416 −0.854 −0.301
(0.246) (0.345) (0.579) (0.532) (0.246) (0.495)
iATsell 0.829 0.790 1.194 1.117 1.260 0.827
(0.021) (0.057) (0.011) (0.027) (0.036) (0.021)
beta −1.719 −1.796 −1.953 −2.004 −2.126 1.447
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
turnover −0.315 −0.426 0.330 −0.368 −0.394 −0.434
(0.239) (0.175) (0.272) (0.277) (0.396) (0.124)
size −0.884 0.047 0.130 −0.012 0.375 0.398
(0.218) (0.937) (0.848) (0.988) (0.701) (0.532)
idiovar −0.663 −0.531 −0.610 −0.526 −1.205 0.417
(0.155) (0.298) (0.300) (0.411) (0.084) (0.451)

69
 Table 3.9: Panel Data Regressions for Abnormal Volume Ratios
This table presents coefficient estimates from panel data regressions using the following model:

ari,t = α + β1 iAT buyi,t + β2 iAT selli,t + β3 sizei,t + β4 turnoveri,t + β5 idiovari,t + β6 betai,t + i,t .

The variable ari,t is the market-adjusted abnormal return for stock i on the event day; iAT buyi,t (iAT selli,t ) is the
abnormal volume ratio between the AT buy (sell) volume and the overall buy (sell) volume on the event day less the mean
volume ratio over the past five days; sizei,t is the logarithm of the market value of stock i five days prior to the event
day; turnoveri,t is the ratio of the daily volume to the number of shares outstanding on the event day; idiovari,t is the
idiosyncratic variance of the market model residual of stock i on days [t − 125, t − 5]; and betai,t is the beta of stock i for
days [t − 125, t − 5].
In the first (second) column, we included stocks that has at least 1 (10) AT and non-AT initiated trades on each event day.
In the third column, we exclude consecutive event days. In the last column, we include weekday dummies to control for day
of the week effect. The coefficients for iAT buyi , iAT selli , and betai are multiplied by 100 in both panels. The coefficients for
sizei are multiplied by 1,000. The standard errors are clustered by stock and day. The p-values are reported in parentheses.
All 10 Non-consec Day of
Trades Days the Week
Panel A: Market Up Days
iATbuy −1.400 −1.270 −1.580 −1.400
(0.014) (0.045) (0.013) (0.014)
iATsell −0.050 0.520 −0.240 −0.090
(0.893) (0.266) (0.586) ((0.818))
beta 1.500 1.400 1.450 1.500
(0.000) (0.000) (0.000) (0.000)
turnover 0.546 0.000 0.572 0.537
(0.268) (0.314) (0.271) ((0.269))
size −0.200 −1.500 −0.100 −0.200
(0.810) (0.102) (0.935) (0.843)
idiovar 1.503 2.031 1.633 1.478
(0.003) (0.023) (0.002) (0.002)
adjusted R2 0.052 0.058 0.058 0.057
Panel B: Market Down Days
iATbuy −0.280 −0.698 −0.080 −0.290
(0.577) (0.203) (0.898) (0.555)
iATsell 1.270 1.636 0.980 1.280
(0.010) (0.004) (0.068) (0.010)
beta −1.790 −1.901 −1.670 −1.800
(0.000) (0.000) (0.000) (0.000)
turnover −0.080 −0.005 −0.080 −0.080
(0.462) (0.691) (0.439) (0.458)
size −0.100 0.472 0.500 −0.100
(0.909) (0.574) (0.627) (0.937)
idiovar −1.380 −1.384 −1.547 −1.333
(0.011) (0.013) (0.012) (0.015)
adjusted R2 0.050 0.058 0.053 0.052

70
3.5 AT, news announcements, and market conditions

While we observe economically large and robust results for AT intensity and price fluctuations, the results
could be driven by the reactions of AT or non-AT to relevant market conditions and information rather than AT
reducing price fluctuations. In this section, we explore plausible mechanisms through which market conditions and
information might influence AT and non-AT.

3.5.1 Matched event day versus non-event day difference-in-differences analysis

Algorithmic traders may react to the return, liquidity, size, and volatility of a particular stock and adjust their
execution accordingly. This may alter our inference about AT intensity and stock price swings. For example,
computerized traders may choose to buy more stocks that have lower returns and not as many higher-return
stocks. Consequently, on market up days, we observe a negative association between AT buy intensity and
upward price swings. We aim to address this issue using a difference-in-differences regression between event and
non-event days. We separate our sample into event days as our treatment group and non-event days as our control
group. However, we cannot directly compare the treatment and control groups since the market conditions in
terms of return, liquidity, and volatility could be significantly different. To avoid traders with AT technologies
reacting to these differences, we match the treatment group with a subset of the control group based on market
conditions. Specifically, we acquire propensity scores based on a logit regression on a set of covariates. The
dependent variable is a dummy indicating whether the observation is on an event day. The independent variables
are the market-adjusted return, size, turnover, beta, and idiosyncratic variance. We then match each observation
in the treatment group with two observations in the control group that have similar market characteristics.38 Table
(3.10) presents the market condition differences before and after the match.
We perform the matching analysis for market up days and down days separately, with the results presented
in Panels A and B of Table (3.10), respectively. The left-hand columns show the difference prior to the match
and the right-hand column show the difference after the match. All differences in market conditions becomes

38
We apply a 1:2 match because we have a large control group compared to the treatment group. We also performed a 1:1 match
as a robustness check, the results of which are available upon request. The implications are unchanged.

71
 Table 3.10: Propensity Score Matching Diagnostics
This table shows the diagnostics for propensity score matching between event days and non-event days. The sample is
separated into up (down) days as the treatment group and non-event days as the control group. Each stock-day in the
treatment group is matched with a stock-day in the control group based on their propensity scores. The propensity score
is acquired from the logit regression for the treatment and control group. The dependent variable is a dummy that takes
1 if the observation is on a event day and 0 otherwise. The set of covariates are presented in the first column: ari,t is
the market-adjusted abnormal return for stock i on day t; sizei,t is the logarithm of the market value of stock i five days
prior to day t; turnoveri,t is the ratio of the daily volume to the number of shares outstanding on day t; idiovari,t is the
idiosyncratic variance of the market model residual of stock i on days [t − 125, t − 5]; and betai,t is the beta of stock i for
days [t − 125, t − 5]. Panel A (B) presents the diagnostics for market up (down) days. Column 2-4 (5-7) presents pairwise
comparison of the covariates and their difference before (after) propensity matching. Ar and Turnover are in percentages.
The p-values for the differences in each covariates between treatment and control groups are reported in parentheses.
Panel A: Market Up Days
Pre-Match Post-Match
Control Treatment Difference Control Treatment Difference
ar 0.168 0.122 0.046 0.098 0.122 -0.024
(0.562) (0.798)
beta 0.865 0.976 -0.111 0.980 0.976 0.004
(0.000) (0.697)
turnover 0.256 0.353 -0.097 0.313 0.353 -0.040
(0.000) (0.717)
size 19.933 20.315 -0.382 20.337 20.315 0.022
(0.000) (0.432)
idiovar 0.231 0.240 -0.009 0.237 0.240 -0.003
(0.014) (0.434)
No. of observations 79,488 5,076 84,564 10,152 5,076 15,228
Panel B: Market Down Days
Pre-Match Post-Match
Control Treatment Difference Control Treatment Difference
ar 0.168 -0.104 0.272 -0.119 -0.104 -0.015
(0.000) (0.858)
beta 0.865 0.978 -0.113 0.995 0.978 0.017
(0.000) (0.213)
turnover 0.256 0.302 -0.046 0.314 0.302 0.012
(0.000) (0.500)
size 19.933 20.448 -0.515 20.420 20.448 -0.028
(0.000) (0.335)
idiovar 0.231 0.228 0.003 0.232 0.228 0.004
(0.391) (0.340)
No. of observations 79,488 4,820 84,308 9,640 4,820 14,460

72
 Table 3.11: Matched Event Day and non-Event Day Diff-in-Diff Test
This table shows the results for the difference-in-difference test on market up days and down days using following model:

ari,t = α + β1 iAT buyi,t + β2 iAT selli,t + β3 upt + β4 upt · iAT buyi,t + β5 upt · iAT selli,t
+β6 sizei,t + β7 turnoveri,t + β8 idiovari,t + β9 betai,t + i,t ,

over a sample in which each event day is matched with two non-event days based on the criteria described in Table (3.10).
ari,t is the market-adjusted abnormal return for stock i on the event day; iAT buyi,t (iAT selli,t ) is the abnormal volume
ratio between the AT buy (sell) volume and the overall buy (sell) volume on the event day less the mean volume ratio over
the past five days; sizei,t is the logarithm of the market value of stock i five days prior to the event day; turnoveri,t is the
ratio of the daily volume to the number of shares outstanding on the event day; idiovari,t is the idiosyncratic variance of
the market model residual of stock i on days [t − 125, t − 5]; and betai,t is the beta of stock i for days [t − 125, t − 5]; upt
is a dummy variable that takes 1 if day t is a market up day. Column 2 and 3 presents the results for matched market up
days. In column 3 and 4, we repeat the analysis on market down days. The coefficients for iAT buyi , iAT selli , and betai
are multiplied by 100 in both panels. The coefficients for sizei are multiplied by 1,000. In order to control for correlations
in cross-section and time-series, we cluster the standard errors by stock and day. The p-values are reported in parentheses.
Up Days Down Days
Coefficient p-value Coefficient p-value
iATbuy 0.490 (0.197) -0.350 (0.126)
iATsell -0.420 (0.197) 0.050 (0.836)
up 0.050 (0.816)
iATbuy*up -1.950 (0.005)
iATsell*up 0.340 (0.488)
down -0.010 (0.949)
iATbuy*down 0.030 (0.956)
iATsell*down 1.150 (0.021)
beta 0.420 (0.005) -0.350 (0.009)
turnover 0.306 (0.251) 0.064 (0.440)
size -0.200 (0.592) 0.400 (0.324)
idiovar 0.762 (0.026) -0.487 (0.062)
adjusted R2 0.013 0.005

insignificant after the propensity matching algorithm is applied.

We then run a difference-in-differences regression on the matched sample to show whether there is a differential
effect of AT intensity on event days compared to non-event days, given that the observations on event and non-event
days have similar market characteristics. Specifically, we interact the AT intensity variables with event day dummy
variables to assess the marginal effect of AT on price fluctuations in light of overall market pressure. The standard
errors are two-way clustered to control for persistent effects in the stock and time dimensions (Thompson, 2011).
Table (3.11) presents the coefficient estimates.
On market up days, the AT intensity variables (iAT buyi,t and iAT selli,t ) are not significant, whereas the

73
interaction between the market up dummy (upt ) and AT buy intensity (iAT buyi,t ) is negatively significant. These
two results imply that, despite the treatment and control groups having similar market characteristics, the
association between AT intensity and abnormal return only exists when there is overall market pressure. The
results and implication on market down days are similar. The adjusted R -squared is artificially low because we
include two observations in the control group for each one in the treatment group.39
While beyond the scope of this chapter, it is worth mentioning that the effects of AT differ during episodes of
market-wide pressure in our treatment group compared to idiosyncratic shocks in our control group. Brogaard,
Carrion, Moyaert, Riordan, Shkilko, and Sokolov (2016) find that high frequency traders on average switch to
liquidity demanding when more than one stock simultaneously experiences extreme price movements. Our evidence
suggests that the liquidity demanding AT is associated with individual stock returns during market-wide shocks.
Overall, the matching algorithm eliminates any meaningful differences in market condition measures between
event day and non-event day observations. The difference-in-differences regression highlights the contrast of AT
effects between event and non-event days. More importantly, the analyses provide confidence that the association
between AT and price fluctuation is exogenous to the market condition measures in individual stocks.

3.5.2 AT and news arrivals

While we observe an economically large effect between AT and stock price fluctuation on event days, it is
possible that this effect is information driven. For instance, Hendershott and Riordan (2013) suggest that, due to
the cheaper monitoring costs of AT, AT may quickly incorporate relevant market information. Frino, Prodromou,
Wang, Westerholm, and Zheng (2017) find that non-AT volume imbalance leads AT volume imbalance prior to
corporate earnings announcements but AT can quickly adjust trades immediately after the announcements. To
the best of our knowledge, the literature does not have a consensus on whether AT is more informed compared to
non-AT. However, we need to consider the possibility that one party is more informed than the other.40 If non-AT
is more informed compared to AT, then non-AT may incorporate price-relevant information during announcement

39
The adjusted R-squared is between 5% and 6% in the previous sections when we include only event days.
40
The literature has not decided whether AT is more informed compared to non-AT. This question, however, is beyond the scope of
the current paper.

74
periods. Therefore, the observed association between AT and stock price fluctuation may be a manifestation of
non-AT incorporating information. To assess the effect of public information arrivals, we include firm-specific and
price-sensitive announcements as dummy variables and interact with our main AT variables. An announcement is
included if it occurred during the period between the closing of the last trading day and the closing of the current
day. Our sample includes 293 announcements on market up days and 291 announcements on market down days. If
event day returns are driven by AT or non-AT reacting to public information arrivals, we would observe significant
coefficients for the interaction variable.
Table (3.12) presents panel regression results for public information arrivals. To control for correlations in the
stock and time dimensions, the standard errors are clustered by stock and date (Thompson, 2011).
The interaction variables are not significant in Table (3.12), which implies that the correlation between AT
and price fluctuation observed in Table (3.4) is not driven by traders reacting to information. Furthermore, all
announcement stock days in the third column of up and down days are excluded. We observe qualitatively and
quantitatively similar results compared to Table (3.4). Overall, the results suggest that the association between
AT and information is not driven by trader groups reacting to public information arrivals.

3.5.3 Causal implications

While we have excluded several plausible alternatives to the premise of AT reducing price fluctuation on
turbulent days, it would be ideal to establish causal implications directly via instrumental variables that are
exogenous to price fluctuation on turbulent days. For example, an instrument for AT is the introduction of the
“autoquote” on the New York Stock Exchange.41 This technological improvement enables faster dissemination
of information for AT and is introduced in batches of stocks. Another popular instrument is the co-location
services provided by the exchanges.42 This service enables specialized low-latency driven traders to achieve
near-instantaneous execution speeds for a substantial fee. For example, Boehmer, Fong, and Wu (2013) use
co-location events as an instrument to assess the impact of AT on various market quality metrics. However,
autoquotes and similar technological events are not available in our sample. In addition, co-location services

41
See Hendershott, Jones, and Menkveld (2011) for the application and details.
42
See Aitken, Cumming, and Zhan (2015) for a comprehensive survey on the starting dates of co-location services internationally.

75
 Table 3.12: AT and Firm Specific News Arrival
This table presents coefficient estimates from panel data regressions using the following model:

ari,t = α + β1 iAT buyi,t + β2 iAT selli,t + β3 sizei,t + β4 turnoveri,t + β5 idiovari,t + β6 betai,t + i,t .

The variable ari,t is the market-adjusted abnormal return for stock i on the event day; iAT buyi,t (iAT selli,t ) is the
abnormal volume ratio between the AT buy (sell) volume and the overall buy (sell) volume on the event day less the mean
volume ratio over the past five days; sizei,t is the logarithm of the market value of stock i five days prior to the event
day; turnoveri,t is the ratio of the daily volume to the number of shares outstanding on the event day; idiovari,t is the
idiosyncratic variance of the market model residual of stock i on days [t − 125, t − 5]; and betai,t is the beta of stock i for
days [t − 125, t − 5].
In the second column of up days and down days, we proxy public information arrival by including price sen-
sitive firm specific announcements as dummy variables and interact with our main AT variables. In the third
column of up days and down days, stocks that have firm specific announcements during the event days are
excluded. The coefficients for iAT buyi , iAT selli , and betai are multiplied by 100 in both panels. The co-
efficients for sizei are multiplied by 1,000. The p-values are reported in parentheses, which are calculated
based on standard errors clustered by stock and day to control for correlations in cross-section and time-series.
Up Days Down Days
Without With Exclude Without With Exclude
Interaction Interaction News Days Interaction Interaction News Days
iATbuy -1.400 -1.370 -1.410 -0.280 -0.280 -0.280
(0.012) (0.004) (0.004) (0.577) (0.589) (0.589)
iATsell -0.050 -0.070 -0.070 1.270 0.850 0.860
(0.841) (0.865) (0.867) (0.010) (0.032) (0.032)
iATbuy*news -1.030 -0.970
(0.742) (0.717)
iATsell*news 0.350 5.450
(0.877) (0.182)
news 1.330 0.320
(0.093) (0.586)
beta 1.500 1.490 1.570 -1.790 -1.790 -1.800
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
turnover 0.545 0.533 0.405 -0.080 -0.081 -0.035
(0.268) (0.264) (0.248) (0.462) (0.454) (0.618)
size -0.200 -0.200 0.500 -0.100 -0.100 0.300
(0.810) (0.814) (0.536) (0.909) (0.917) (0.755)
idiovar 1.503 1.484 1.160 -1.380 -1.364 -1.360
(0.003) (0.003) (0.012) (0.011) (0.013) (0.018)
adjusted R2 0.052 0.054 0.053 0.050 0.052 0.057

76
were only available on the ASX at the start of our sample in November 2008, which eliminates the possibility
of a pre–post implementation comparison. Besides autoquotes and co-location, the literature uses many other
instruments. For instance, Hasbrouck and Saar (2013) use other stocks’ HFT activity as an instrument of the target
stock’s HFT activity in an attempt to circumvent the endogeneity between the target stock’s market quality and
HFT activity. Aitken, Cumming, and Zhan (2015) use “permanent change” in trade size as one of the instruments.
The permanent change event is defined as the first of four continuously declining months in average market trading
size or the biggest single drop from the previous month. Aggarwal and Thomas (2014) compare market quality
measures from January 2009 to December 2009 against those from July 2012 to August 2013. These two periods
are chosen to be around the co-location event in January 2010. These methods are not suitable for our analysis.
We further explored other technological upgrades and trading rule amendments and were unable to find suitable
instruments for the two-stage least squares analysis.
Our approach is to exclude the most probable alternatives. First, we match the observations on the event days,
as the treatment group, with those that have similar market characteristics on non-event days, as the control
group. We then use a difference-in-differences approach to show that the effects are observed only in the treatment
group, despite the control group not being meaningfully different in terms of market conditions. Second, we show
that the effect is not driven by algorithmic traders’ reactions to firm-specific news arrivals. Third, the effects are
not likely to be caused by situational algorithms, since the significant effects of AT buys (sells) were observed
only on market up (down) days. Specifically, unlike execution algorithms, which have a pre-existing intention
to buy or sell, situational algorithms tend to be indifferent in terms of profiting from buy trades or sell trades.
For example, many HFT algorithms frequently switch between buy and sell trades and maintain a low level of
inventory ((Brogaard, Garriott, and Pomeranets, 2016)). If situational algorithms buy more stocks with lower
returns to exploit the cross-sectional return differences on market up days, these algorithms are likely to also sell
less (more) of the same (other) stocks. As a result, we would observe significant results in both buy and sell trades
on event days. Overall, the evidence suggests that the association between AT intensity and price fluctuation
is not driven by algorithmic traders’ reactions to market conditions in terms of return, liquidity, volatility, and
firm-specific information arrivals.

77
3.6 Conclusions

In this paper we examine the impact of AT during the most turbulent trading days on the Australian Securities
Exchange from October 2008 till October 2009. Building on Dennis and Strickland (2002), we define turbulent
days as days when the absolute values of the market returns are greater than 2%. We identify 39 event days of
extreme market movements during the sample period - there are 19 trading days when the market gains at least
2% and 20 trading days when the market drops by at least 2%. For those turbulent days we analyze how individual
stock returns correlate with the level of AT intensity, which is measured as the abnormal ratio of trading volume
of AT to the total volume traded.
We show that the level of AT intensity in individual stocks is significantly related to abnormal returns in
turbulent markets. In particular, stocks with lower levels of AT intensity experience greater price swings when the
absolute return of the market exceeds 2%. Moreover, these effects are economically very large. We find that with
an increase of 10% (or half standard deviation) in AT buying on average, there is a decrease of 16 basis points in
market adjusted returns for individual stocks when the market goes up by 2% or more. To isolate the effects of AT
on market adjusted returns, We perform a difference-in-difference method on the event day observations with a
propensity score matched sample of non-event day observations. We show that the effects of AT exists only on
turbulent days, despite the fact that the matched non-event day sample is similar in terms of stock size, return,
liquidity, and volatility.
Our study is subject to a few caveats. First, while the sample period covers the most turbulent time in decades,
the sample size is limited to a time span over one year. Second, although we find beneficial effects of AT during
turbulent times based on AT’ characteristics, we do not support the premise that AT is harmless in other periods
or other aspects of the financial market. In other words, while we observe the majority of AT to consist of benign
execution algorithms, we do not exclude the possibility of market disruption caused by manipulative situational
algorithms or high-frequency algorithms. Overall, our findings have important policy implications. Since AT, as a
whole, plays a beneficial role in turbulent markets, we advise against indiscriminate restrictions on AT. Future
research could focus on heterogeneity among AT groups, identify nefarious algorithms, and explore the effects of
AT in an international, multi-year context.

78
 Chapter 4

Algorithmic Execution Strategy and Order Im-

balances

Chapter Summary

We examine algorithmic execution strategy and the effects of algorithmic trading order imbalances.
We find that, ex-ante, algorithmic traders execute their trades according to the prevailing volume-
weighted average prices, they are more likely to execute buy (sell) orders when the prevailing
volume-weighted average price moves lower (higher) compared to the prevailing stock price. This
implies a contrarian strategy which may mitigate the short-term price trends. Further analyses
show that AT order imbalances have a smaller price impact compared to non-AT order imbalances.
These effects are robust on days when the absolute value of the market return is more than 2%.

79
4.1 Introduction

In this chapter, we analyze the execution strategy and order imbalances of AT. Hendershott, Jones, and
Menkveld (2011) suggest that execution algorithms may track the VWAP metric to reduce execution costs. The
VWAP is the average price of each transaction over a certain time horizon (typically one day) weighted by the
volume of each trade. The executed price of each trade can then be compared with the VWAP to evaluate the
trade’s execution performance.43 By monitoring the VWAP, algorithmic traders are more likely to trade when the
relative position between the VWAP and the stock price becomes more favorable. A favorable position is one in
which the price moves downward (upward) relative to the VWAP for buy (sell) trades. AT execution strategy
would effectively be contrarian by design. Specifically, in the initial surge of a large market-wide drop, stock prices
would unfavorably deviate downward compared to the VWAPs and human traders would be more likely to sell.
After the initial surge, the VWAPs would catch up with prices as more trades are executed and algorithmic traders
would be more likely to trade based on this favorable VWAP movement. Therefore, algorithmic traders would
be less likely to herd with human traders. As a result, AT would smooth out the liquidity demand and would
not contribute to further price declines. After controlling for market conditions, we find that this is exactly the
case. In particular, we estimate logit models of AT execution and show that traders with AT technologies are
more likely to buy (sell) when the difference between the stock price and the intraday VWAP at the time becomes
smaller (larger) on market decline days, market rise days, and other days in our sample. This finding is consistent
with the notion that AT does not exacerbate the price pressure from the overall market.
We then investigate the effects of AT order imbalances on stock prices. Trades initiated by AT and non-AT may
exert different levels of price pressure. To explore this hypothesis, we extend Chordia and Subrahmanyam (2004)
and model market-adjusted returns as a function of order imbalances from AT and non-AT. After controlling
for trade size and the total level of trading activity, we find that, on average, the impact of order imbalances
by non-AT is 58% larger than that by AT on event days. This finding highlights the heterogeneity in the price
impacts of order imbalances from different investor groups.
The rest of the chapter is organized as follows. Section 4.3 presents the VWAP-tracking algorithms and their

43
The VWAP was first proposed by Berkowitz, Logue, and Noser (1988). See Madhavan (2002) for a comprehensive survey.

80
effects on the characteristics of AT in turbulent markets. Section 4.4 reports our findings on order imbalances.
Section 4.5 provides the conclusion to the chapter.

4.2 Data and event day selection

The dataset used in the current chapter is similar to the dataset used in Chapter 3.4. Briefly speaking, we
combine Australian equity market data provided by SIRCA with a novel dataset that classifies the traders on
both sides of each stock transaction into algorithmic and non-algorithmic traders. The sample period is between
October 27, 2008, and October 23, 2009. We focus on the stocks that are present throughout the sample period
and have more than 200 active trading days out of the 252 trading days during the sample period.
To assess the characteristics of AT when the market is volatile, we define turbulent days as days when the
absolute values of the returns on the market are greater than 2% (Dennis and Strickland, 2002). During our
sample period, there are 19 (20) days when the market surges (declines) for more than 2%. Our final sample
contains 9,896 stock-days across 384 stocks and 252 trading days. Refer to Section 3.3 for a detailed discussion on
the data and the filtering process.

4.3 AT execution strategy and the VWAP metric

In this section, we investigate the possible strategies that AT could employ. Easley, Lopez de Prado, and
O’Hara (2012) suggest that execution algorithms track the VWAP metric to reduce execution costs. The VWAP
is the average price of each transaction over a certain time horizon (typically one day) weighted by the volume
of each trade. The executed price of each trade can then be compared with the VWAP to evaluate the trade’s
execution performance. A buy (sell) trade is considered favorable if the transacted price is lower (higher) compared
to the VWAP. If algorithms closely monitor the intraday VWAP, then AT would act as counter-trend traders.
Although VWAP-tracking algorithms do not trade less than usual, they would optimize their execution timing
based on the VWAP-to-price relation. Therefore, AT would smooth out the liquidity demand and would not
contribute to further price swings.
By way of illustration, Figure (4.1) highlights the prices of trades initiated by AT and non-AT and the intraday
dynamics of the VWAP on an event day (the top graph) and a non-event day (the bottom graph). In terms of

81
 Figure 4.1: The VWAP, AT, and non-AT of One Event Day and One non-Event Day.
These figures illustrate the intraday dynamics of the VWAP and AT and non-AT prices for one event day (November 6,
2008) and one non-event day (January 21, 2009) for stocks ALS and PMV, respectively.

trading volume, both algorithmic traders and human traders initiated similar amounts of transactions. However,
algorithmic traders initiated trades when the price was much closer to the VWAP, compared to human traders.
This finding implies that a substantial portion of AT monitored the VWAP, whereas non-AT did not track the
VWAP as much.

82
 We reset the VWAP benchmark on a daily basis for the following reasons. First, Berkowitz, Logue, and Noser
(1988), who initially proposed the VWAP as a measure of transactions cost, calculate the VWAPs on a daily
basis. Second, Carrion (2013) applies end-of-day VWAP metrics that resets every day to assess how HFT times
the market. Third, we have contacted several day traders and funds managers about their application of VWAP
metric. These practitioners confirm that the daily VWAP is the a commonly used metric to assess day traders’
performance.
We start the analysis by testing whether algorithmic traders traded closer to the VWAP compared to human
traders. Table (4.1) shows the univariate results for the relation between the intraday VWAP and trades initiated
by AT and non-AT. Panel A reports the average variance from the VWAP (vari,t ), defined as the squared difference
between the transacted price and the prevailing VWAP weighted by the prevailing VWAP at the time of the
transaction, for both AT and non-AT. On event and non-event days, vari,t is significantly smaller for AT than for
non-AT. Algorithms therefore trade significantly closer to the VWAP compared to human traders. This finding
suggests that algorithmic traders employ a VWAP-tracking strategy. During turbulent markets, algorithmic traders
do not alter their strategy and therefore do not exacerbate price fluctuations.
In Panel B of Table (4.1) , we explore whether algorithmic traders buy (sell) lower (higher) than the VWAP
compared to human traders. Specifically, variance from the VWAP is replaced with deviation from the VWAP
(devii,t ), which measures the signed difference between the prevailing VWAP and the transaction price weighted
by the VWAP. Deviation from the VWAP is interacted with an indicator variable (Ii,t ) that equals one (-1) if the
trade is buyer (seller) initiated. The term Ii,t · devii,t is positive when a buy (sell) trade occurs in which the stock
price is below (above) the prevailing VWAP. Therefore, Ii,t · devii,t can gauge whether and by how much each
trade can beat the VWAP metric. Although the average value of the measure is negative for AT and non-AT,
algorithms beat human traders in regard to the intraday VWAP metric.

83
 Table 4.1: AT versus non-AT Variance from the VWAP
This table presents the variance from the VWAP from AT and non-AT. The VWAP for stock i at time t is defined as
Pt
j=1 voli,j pricei,j
vwapi,t = Pt ,
j=1 voli,j

where voli,j and pricei,j are the volume and price of the trade at time j (j = 1, 2, . . . , t − 1, t) for stock i, respectively.
vwapi,t resets every trading day. I.e., the prevailing VWAP at time t is the volume weighted average of all transactions
from the beginning of the day up to the current trade at time t.
We then calculate the variance of each trade relative to the prevailing VWAP at the time (without incorporating the trade
at the time):
pricei,t − vwapi,t−1 2
vari,t = ( ) .
vwapi,t−1
Panel A contains the variance from the VWAP (vari,t ) per trade for AT and non-AT on all days, non-event days, and event
days. In Panel B, variance from the VWAP is replaced by deviation from the VWAP (devii,t ) with a buy/sell indicator
(Ii,t ) that is one (or −1) if the trade is initiated by a buyer (or seller):

vwapi,t−1 − pricei,t
devii,t = .
vwapi,t−1

The coefficients in Panel A (B) are multiplied by 10,000 (100). The p-values in parentheses correspond to a test of a null
hypothesis that the variances of AT and non-AT trades have same means.
All Nonevent Up Down
Days Days Days Days
Panel A: Variance from the VWAP (vari,t )
AT 0.168 0.158 0.197 0.238
non-AT 0.327 0.332 0.317 0.389
AT less non-AT −0.159 −0.174 −0.120 −0.151
(0.000) (0.000) (0.000) (0.000)
Panel B: Deviation from VWAP with Buy/Sell Indicator (Ii,t · devii,t )
AT −0.144 −0.138 −0.183 −0.161
non-AT −0.178 −0.177 −0.192 −0.176
AT less non-AT 0.034 0.039 0.009 0.015
(0.000) (0.000) (0.004) (0.000)

84
 Table 4.2: Logit Regression for AT and the VWAP
This table presents the coefficient estimates from logit regression using the following model:

ati,t = α + β1 idevii,t−1 · Ii,t−1 + β2 spreadi,t−1 + β3 sizei,t−1 + β4 depthi,t−1 + β5 volai,t−1 + β5 voli,t−1 + i,t , (11)

where ati,t equals one if the trade of stock i at time t is AT initiated and zero otherwise; idevii,t−1 is the innovation of
deviation from the VWAP (devii,t−1 , defined in Table (4.1)), and is measured as the residual of an AR(8) autoregression
multiplied by 100; Ii,t−1 is an indicator variable which is equal to one (−1) for a buy (sell) trade; spreadi,t−1 is the quoted bid–
ask spread divided by 10; sizei,t−1 is the dollar volume of the trade divided by 100,000; depthi,t−1 is the market depth at the
best bid and ask in 10 million dollars; and volai,t−1 and voli,t−1 are the lagged volatility and lagged volume, respectively, in
the 15 minutes before the trade. Lagged volatility is measured as the absolute value of the stock return over the interval and
lagged volume is the total volume in 10 million shares over the interval. The odds ratios are calculated based on the regression
coefficient. The odds ratios and marginal effects are omitted for the control variables and are available upon request. Time
of day dummies for each half-hour of the trading day are included but not reported. The p-values are included in parentheses.
All Nonevent Up Down
Days Days Days Days
idevi*I 0.279 0.292 0.299 0.145
−odds ratio 1.322 1.339 1.349 1.156
−marginal effect 2.613 2.547 2.654 1.305
−p-value (0.000) (0.000) (0.000) (0.000)
spread 0.024 0.024 0.025 3.972
−p-value (0.000) (0.000) (0.005) (0.346)
size −0.696 −0.688 −0.705 −0.771
−p-value (0.000) (0.000) (0.000) (0.000)
depth −0.144 −0.150 −0.104 0.594
−p-value (0.161) (0.146) (0.496) (0.503)
vola −2.076 −1.993 −1.546 −3.471
−p-value (0.108) (0.135) (0.379) (0.036)
vol −0.410 −0.361 −0.705 −1.312
−p-value (0.124) (0.195) (0.000) (0.002)

To more formally establish the statistical association between the intraday VWAP and the strategy of AT, we
estimate logit regressions for algorithm-initiated trades. Following Hendershott and Riordan (2013), we control
for market conditions by including the bid–ask spread, trade size, market depth, lagged volatility, and lagged
volume. Also included but do not reported are the time-of-day dummies for each half-hour period. To control for
correlations across stocks and across time, standard errors are double clustered in the cross-section and time-series
(Thompson, 2011).
The key variable in Table (4.2) is idevii,t−1 · Ii,t−1 , where Ii,t−1 is an indicator variable that equals one (-1) if
the trade is buyer (seller) initiated. As for Brogaard, Hendershott, and Riordan (2014), idevii,t−1 is the deviation
innovation obtained as the residual of an eight-lag autoregressive model on the time-series of devii,t−1 . The

85
number of lags is determined by the Akaike information criterion. The term idevii,t−1 · Ii,t−1 measures whether the
price–VWAP relation becomes favorable. A positive idevii,t−1 · Ii,t−1 indicates that the stock price has decreased
(increased) compared to the VWAP just before a trader buys (sells) stock i at time t. The marginal effect of
idevii,t−1 · Ii,t−1 on all days show that AT is 2.613% more likely to execute buy (sell) trades when the price
decreases (increases) by 1% compared to the VWAP. The marginal effects on up days and non-event days are
similar to that on all days. AT is less sensitive to changes in the VWAP on market down days. Overall, the
logit regression presented in Table (4.2) show that AT is more likely to initiate buy (sell) trades when the price
decreases (increases) compared to the VWAP.

4.4 AT and non-AT order imbalances

In this section, we analyze the impact of AT order imbalances on stock prices. Our premise is that trades from
AT exert less price pressure and order imbalance from AT would have less of a price impact compared to non-AT
order imbalance. As a result, the stock price would fluctuate less during periods of high AT intensity. We model
market-adjusted returns as a function of order imbalances from AT and non-AT and explore our hypothesis.
Order imbalances are measured as the scaled and unscaled imbalances in the number of transactions and
volume. The unscaled volume imbalance is defined as the daily buy volume less the daily sell volume. The scaled
volume imbalance is defined as the unscaled volume imbalance divided by the total daily volume. We separately
measure each order imbalance metric for AT and non-AT on a daily basis. The cross-sectional averages of the
correlations between various order imbalance metrics are reported in Panel A of Table (4.3). The correlations
between scaled and unscaled order imbalances are high. For example, the correlations between the unscaled and
scaled volume imbalances for AT and non-AT are 0.705 and 0.730, respectively. The correlations between volume
imbalances and the number of trades imbalances are lower (0.427 and 0.570, respectively, for AT and non-AT). The
correlations between AT and non-AT are very small across different metrics. The largest value (in absolute terms)
for AT versus non-AT correlations is −0.071. This finding highlights the heterogeneity in the trading strategies of
AT and non-AT.
Panel B of Table (4.3) presents the cross-sectional averages of the autocorrelation of AT and non-AT order
imbalances measured by the number of trades and volume. Domowitz and Yegerman (2005) argue that a substantial

86
 Table 4.3: Correlations and Autocorrelations for Order Imbalances
This table presents summary statistics for daily AT and non-AT order imbalances between 27 October 2008 and 23 October
2009. atoib (atoibs) is the (scaled) order imbalance in the volume traded by AT, nonatoib (nonatoibs) is the (scaled) order
imbalance in the volume traded by non-AT. We then replace atoib and nonatoib with atnoib and nonatnoib, corresponding
to order imbalance measured in number of trades for AT and non-AT, respectively. The unscaled AT volume imbalance is
defined as the daily AT buy volume less the daily AT sell volume. The scaled AT volume imbalance is defined as the daily
AT buy volume less the daily AT sell volume divided by the daily total volume. Number of trades imbalances are defined
similar to volume imbalances (replace volume figures with number of trades). Panel A presents the cross-sectional means
of the individual stock time-series correlations and Panel B contains the autocorrelations.
Panel A: Correlations
nonatoib atoibs nonatoibs atnoib nonatnoib atnoibs nonatnoibs
atoib −0.071 0.705 −0.062 0.427 −0.061 0.430 −0.058
nonatoib −0.055 0.730 −0.019 0.570 −0.024 0.503
atoibs −0.057 0.401 −0.055 0.641 −0.052
nonatoibs −0.019 0.520 −0.016 0.726
atnoib −0.018 0.756 −0.020
nonatnoib −0.027 0.738
atnoibs −0.021
Panel B: Autocorrelations
Volume Imbalances No. of Trades Imbalances
lag atoib nonatoib atoibs nonatoibs atnoib nonatnoib atnoibs nonatoibs
1 0.122 0.142 0.120 0.144 0.182 0.183 0.170 0.156
2 0.059 0.069 0.066 0.074 0.097 0.094 0.107 0.085
3 0.033 0.046 0.048 0.047 0.074 0.058 0.085 0.060
4 0.017 0.030 0.023 0.030 0.043 0.042 0.059 0.042
5 0.015 0.018 0.014 0.023 0.049 0.030 0.040 0.034

87
proportion of buy-side AT uses VWAP-monitoring strategies to execute trades over time to minimize execution
costs. For all but one metric, non-AT daily autocorrelations in order imbalance are larger than those for AT. This
result suggests that AT does not seem to break down orders across days, which minimizes potential bias due to
autocorrelation.
We model individual stock returns as a function of order imbalance by AT and non-AT. Although our order
imbalance metrics have a lower autocorrelation than those of Chordia and Subrahmanyam (2004), we include four
lags of order imbalance. We also use market-adjusted returns to mitigate the cross-sectional correlation in error
terms. Similar to Equation ((6)), we estimate the Fama–MacBeth (1973) regression for each event day:

4
X 4
X
ari,t = α + β1 atoibi,t + β2 nonatoibi,t + β2+k atoibi,t−k + β6+k nonatoibi,t−k
k=1 k=1
+β11 sizei,t + β12 turnoveri,t + β13 idiovari,t + β14 betai,t + i,t , (12)

where ari,t is the abnormal return for stock i on event day t and atoibi,t is the volume imbalance from AT in stock
i on day t. The control variables are included to account for risk factors, informational effects, liquidity effects, and
the potential preferences of algorithmic traders. Panel A of Table (4.4) presents the regression results. The results
for lagged order imbalances are mostly insignificant and in line with the findings of Chordia and Subrahmanyam
(2004). Therefore, the coefficients for lagged metrics are omitted. In Panel B, we replace the volume imbalance
(atoibi,t and nonatoibi,t ) metrics in Equation (12) with a number of trades imbalance metrics for AT (atnoibi,t )
and non-AT (nonatnoibi,t ).
The core variables of interest are the order imbalance metrics from AT and from non-AT. The relation between
market-adjusted returns and imbalance from the order flow can be expected to be similar to that from the previous
literature (e.g., Chordia, Roll, and Subrahmanyam, 2002): Higher-order imbalance would create more price pressure
on the buy side and cause prices to go up. In this regression, however, our objective is to find out whether order
imbalances from AT and non-AT affect abnormal return differently. In other words, if the coefficients for AT
imbalance (atoibi,t ) are larger than the coefficients for non-AT imbalance (nonatoibi,t ), then the implication is that
AT exerts greater price pressure compared to non-AT and the market responds differently to imbalances from the
different trading groups. We expect, however, non-AT imbalances to exert greater price pressure compared to AT

88
 Table 4.4: Event Day Market-Adjusted Return Regressions on Order Imbalances
This table reports the coefficient estimates from Fama–MacBeth regressions using the following model:
4
X 4
X
ari,t = α + β1 atoibi,t + β2 nonatoibi,t + β2+k atoibi,t−k + β6+k nonatoibi,t−k
k=1 k=1
+β11 sizei,t + β12 turnoveri,t + β13 idiovari,t + β14 betai,t + i,t ,

where ari,t is the market-adjusted abnormal return for stock i on event day t. The event days are defined as the days when
the absolute values of market returns exceed 2%. In Panel A, atoibi,t is the order imbalance in the volume traded by AT,
nonatoibi,t is the order imbalance in the volume traded by non-AT, sizei,t is the logarithm of the market value of stock i
five days prior to event day t, turnoveri,t is the ratio of the daily volume over the number of shares outstanding on event
day t for stock i, idiovari,t is the idiosyncratic variance of the market model residual of stock i on days [t−125, t−5], and
betai,t is the beta of stock i for days [t−125, t−5]. The event days are segregated into 19 up days and 20 down days. The
coefficients for atoibi,t and nonatoibi,t are scaled by 100,000,000.
In Panel B, atoibi,t and nonatoibi,t are replaced by atnoibi,t and nonatnoibi,t , corresponding to order imbalance measured in
number of trades for stock i on event day t for AT and non-AT, respectively. The coefficients for atnoibi,t and nonatnoibi,t
are multiplied by 10,000. The control variables are identical to those in Panel A. The coefficients for betai,t (sizei,t ) are
multiplied by 100 (1,000). The p-values are reported from a t-test of the mean coefficient being different from zero.
Up Days Down Days
mean p-value min max mean p-value min max
Panel A: Volume Imbalance
atoib 1.43 0.001 0.12 5.22 1.28 0.000 −0.33 4.87
nonatoib 2.13 0.000 0.25 8.15 2.53 0.000 0.42 7.76
beta 1.71 0.000 −0.02 4.52 −1.79 0.000 −4.34 0.65
turnover 0.36 0.001 −0.15 1.23 −0.76 0.147 −8.16 1.36
size −0.55 0.590 −6.50 9.75 0.25 0.701 −4.25 7.99
idiovar 0.49 0.307 −3.57 4.42 −0.01 0.984 −4.72 4.78
Panel B: Number of Trades Imbalance
atnoib 0.20 0.000 −0.04 0.63 0.14 0.051 −0.39 0.74
nonatnoib 1.09 0.000 −0.48 2.44 0.62 0.006 −0.61 2.57
beta 1.75 0.000 −0.40 4.78 −1.78 0.000 −4.62 0.69
turnover 0.62 0.002 −0.35 2.23 −0.53 0.063 −3.39 1.48
size −1.44 0.181 −7.31 11.08 0.46 0.512 −5.39 6.83
idiovar 0.37 0.430 −3.88 4.56 −0.69 0.204 −4.62 4.23

89
 All the contemporaneous imbalance metrics in Table (4.4) are significant. The results for market up and
market down days are quantitatively and qualitatively similar. Therefore, we discuss the results on up and down
days together. For volume imbalance, the average coefficients of imbalances from AT and non-AT are 1.35 and
2.33, respectively, corresponding to a 72.59% difference in effects. The results from the number of trades order
imbalance regression are similar. Overall, the results from the estimation are consistent with our expectation
that the abnormal returns of an individual stock are related to the stock’s level of AT intensity. Specifically, AT
executes trades that minimize the price pressure compared to non-AT. As a result, stocks with higher AT trading
experience lower price swings on turbulent days.
To further highlight the heterogeneity effects of AT and non-AT order imbalances, we repeat the analysis
by explicitly modeling the difference between AT and non-AT order imbalances. We replace atoibi,t , nonatoibi,t ,
and their lags with dif oibi,t , oibi,t , and their corresponding lags, respectively. The key variable is the difference
between the AT and non-AT order imbalances (dif oibi,t ). A positive (negative) coefficient for this variable can be
interpreted, given the same level of overall order imbalance, as meaning that stocks that have more AT (non-AT)
order imbalance experience more price movement. Table (4.5) contains the coefficient estimates of the differences
between AT and non-AT order imbalances.
Consistent with the findings of Chordia and Subrahmanyam (2004), the overall order imbalance (oibi,t ) is
the main predictor of stock returns. The coefficients for the differences between AT and non-AT (dif oibi,t ) are
significantly negative on up and down days measured by the volume and number of trades. This result is as
expected: Controlling for market conditions and overall order imbalance, we find that stocks with larger AT order
imbalances experience fewer price swings on turbulent days. Finally, the result is robust to nonconsecutive event
day selection, event day selection based on daily high/low prices, day-of-the-week effects, and double clustering of
the standard error in stocks and days. For brevity, the results are not reported here but are available upon request.

4.5 Conclusions

In this chapter, we investigate AT execution strategy in relation to the VWAP metrics and the price impact of
AT order imbalances. We first analyze the determinants of AT executions using a logit model. Specifically, we
relate algorithmic trades to the prevailing VWAP changes and other market quality measures. The VWAP is the

90
 Table 4.5: Event Day Return Regressions on Differences Between AT and non-AT Order Imbalances
This table reports the coefficient estimates from Fama–MacBeth regressions using the following model:
4
X 4
X
ari,t = α + β1 dif oibi,t + β2 oibi,t + β2+k dif oibi,t−k + β6+k oibi,t−k
k=1 k=1
+β11 sizei,t + β12 turnoveri,t + β13 idiovari,t + β14 betai,t + i,t ,

where ari,t is the market-adjusted abnormal return for stock i on event day t. The event days are defined as the days
when the absolute values of market returns exceed 2%. In Panel A, dif oibi,t is the AT order imbalance less non-AT
order imbalance measured in volume traded, oibi,t is the overall order imbalance measured in volume traded, sizei,t is
the logarithm of the market value of stock i five days prior to event day t, turnoveri,t is the ratio of the daily volume
over the number of shares outstanding on event day t for stock i, idiovari,t is the idiosyncratic variance of the market
model residual of stock i on days [t−125, t−5], and betai,t is the beta of stock i for days [t−125, t−5]. The event days are
segregated into 19 up days and 20 down days. The coefficients for dif oibi,t and oibi,t are scaled by 100,000,000.
In Panel B, dif oibi,t and oibi,t are replaced by dif f noibi,t and noibi,t , corresponding to AT less non-AT order imbalance
and overall order imbalance measured in number of trades for stock i on event day t. The coefficients for dif f noibi,t and
noibi,t are multiplied by 10,000. The control variables are identical to those in Panel A. The coefficients for betai,t (sizei,t )
are multiplied by 100 (1,000). The reported p-values are based on testing the null hypothesis that the mean is equal to
zero.
Up Days Down Days
mean p-value min max mean p-value min max
Panel A: Volume Imbalance
difoib −0.35 0.015 −1.47 0.42 −0.62 0.016 −3.03 0.93
oib 1.78 0.000 0.25 6.69 1.91 0.000 0.47 6.31
beta 1.71 0.000 −0.02 4.52 −1.79 0.000 −4.34 0.65
turnover 0.36 0.001 −0.15 1.23 −0.76 0.147 −8.16 1.36
size −0.55 0.590 −6.50 9.75 0.25 0.701 −4.25 7.99
idiovar 0.49 0.307 −3.57 4.42 −0.01 0.984 −4.72 4.78
Panel B: Number of Trades Imbalance
diffnoib −0.44 0.000 −1.09 0.43 −0.24 0.033 −1.48 0.38
noib 0.64 0.000 −0.07 1.50 0.38 0.002 −0.47 1.23
beta 1.75 0.000 −0.40 4.78 −1.78 0.000 −4.62 0.69
turnover 0.62 0.002 −0.35 2.23 −0.53 0.063 −3.39 1.48
size −1.44 0.181 −7.31 11.08 0.46 0.512 −5.39 6.83
idiovar 0.37 0.430 −3.88 4.56 −0.69 0.204 −4.62 4.23

91
average price of each transaction over a certain time horizon weighted by the volume of each trade. We find that,
on average, algorithmic traders are more likely to submit market buy (sell) orders when the VWAP decreases
(increases) compared to the stock price. Our results imply that algorithmic traders follow VWAP tracking and
short-term contrarian strategies, consistent with the notion that AT minimizes its impact to stock prices. Based
on the prior literature and our discussion with the industry practitioners, we apply the VWAP that resets on a
daily basis. Therefore, our findings are subject to the assumption that VWAP generally resets on a daily basis.
We then investigate the reasons for the effects of AT on market adjusted returns. We extend Chordia and
Subrahmanyam (2004) by separating order imbalances in individual stocks into AT and non-AT order imbalances.
We find that AT order imbalances have smaller price impacts compared to non-AT order imbalances. This finding
implies that, consistent with Hendershott, Jones, and Menkveld (2011), the effect of AT on stock returns is likely
caused by AT exerting less price pressure.

92
 Chapter 5

State Space Models for AT and Price Discovery

Chapter Summary

We investigate the role of algorithmic trading (AT) in the price discovery process. We estimate a
state space framework that decomposes stock prices into permanent price series and transient
pricing error via state space frameworks. We find that algorithmic traders contribute more to the
permanent price processes and less to the transient pricing errors compared to other traders. Our
results show that AT facilitates the price discovery process by contributing to permanent price
movements. Our results are robust on days when the absolute value of the market return is more
than 2%.

93
5.1 Introduction

Price discovery is the process by which the actions of buyers and sellers determine the price of an asset in a
market (O’Hara, 2003). This process is one of the essential functions provided by financial markets. Asset pricing
and market efficiency relies on the speed and accuracy of price discovery processes in financial markets. The
proliferation of CT has important implications on the price discovery processes. Due to computerized traders’
higher speed and cheaper monitoring, CT is able to quickly incorporate information to the prices and accelerate
the price discovery process. However, short-term information incorporated through accelerated price discovery
process may not be motivated by the fundamental value of assets (O’Hara, 2015). In today’s high frequency world,
security prices are more suspectable to short-term noises such as liquidity shocks and inventory constraints. We
analyze the characteristics of AT in relation to the efficient component and the noise component of the price
discovery process.
We apply state space models to analyze the relation between AT, non-AT, and stock price discoveries (Brogaard,
Hendershott, and Riordan, 2014). The state space models decompose observed stock price series into unobserved
permanent price components and transient pricing errors. The permanent (or efficient) price series is modeled as
a martingale to capture information arrivals that affect the permanent value of a stock. The transient pricing
error series represents the short-lived price deviations that is not driven by fundamental value related information.
The transient price changes are assumed to be stationary with an autoregressive component. We then relate AT
and non-AT order flows to the increments of efficient and transitory components. We find that AT contributes
more to the permanent price discovery process on a per dollar volume basis compared to non-AT. For large stocks,
transient pricing errors are mitigated by AT whereas non-AT contributes to the pricing errors.
Our findings on AT compliments the results of Brogaard, Hendershott, and Riordan (2014) on HFT in several
ways. First, we provide empirical evidence on AT, which is a broader computerized trading group. While
high frequency traders consist of mainly the proprietary trading desks and specialized electronic market makers,
AT technology has been widely adopted by buy-side institutions due to its less reliance on ultra low latency
infrastructure. Second, one major advantage for computerized traders is that they can sift through the prices
in dozens of trading venues and related securities to execute on small and fleeing arbitrage opportunities (Biais,

94
Foucault, and Moinas, 2015). This advantage is most pronounced in highly fragmented markets such as the U.S.
market. By studying AT in Australian market, we are able to assess their performance on a more leveled playing
field.44
The rest of the chapter is organized as follows. Section 5.2 discusses the related literature on CT and price
discovery. Section 5.3 provides descriptive statistics. Section 5.4 briefly describes the the state space models.
Section 5.5.2 discusses the empirical decomposition of the observed price series and the estimation results. Section
5.6 concludes the chapter.

5.2 Literature review

Several studies provide evidence that CT improves price discovery and reduces price inefficiencies. Foucault,
Hombert, and Roşu (2016) argue that fast traders can incorporate information by trading on the short-term
news and long-term price forecasts whereas slow traders can only trade according to long-term price movements.
Boehmer, Fong, and Wu (2015) analyze a wider range of markets and find that AT improves informational efficiency
across international markets. Chaboud, Chiquoine, Hjalmarsson, and Vega (2014) show that AT improves price
efficiency: AT market orders reduce arbitrage opportunities and their passive orders reduce autocorrelations in
high frequency returns. Brogaard, Hendershott, and Riordan (2014) find that HFT facilitates the price discovery
process by trading in the direction of permanent price changes and in the opposite direction of transitory pricing
errors. In addition, the literature finds that, compared to passive HFT, aggressive HFT dominates the price
discovery process. For instance, Benos and Sagade (2016) separate HFT into aggressive HFT, mixed HFT, and
passive HFT. They show that market orders placed by aggressive HFT are more informed than those placed by
mixed and passive HFT. Brogaard, Garriott, and Pomeranets (2016) find that price efficiency improves after the
entry of aggressive HFT.
The literature also uncovers several sources of computerized traders’ information advantage. First, computerized
traders execute their orders on the spot markets based on price changes in related indices and derivatives. For
example, the theoretical model by Jovanovic and Menkveld (2016b) assumes that machine traders possess “hard”
information. Hard information is easy for computers to process such as prices of market indices. Zhang (2013)

44
The ASX was the only securities exchange in Australia during our sample period.

95
shows that HFT reacts strongly to hard information proxied by shocks in E-mini prices. Hendershott and Riordan
(2013) also find that AT can monitor prices from futures markets and execute trades on spot markets based on
futures price movements. Second, CT can profit from price discrepancies in related assets. Budish, Cramton,
and Shim (2015) point out that arbitrage opportunities arise due to the breakdowns in the correlations between
related assets over ultra-short time intervals (i.e. 100 milliseconds). Chaboud, Chiquoine, Hjalmarsson, and Vega
(2014) provide evidence that liquidity consuming AT is negatively associated with the frequency of arbitrage
opportunities in JPY-USD-EUR triangular currency exchange prices. Gerig (2015) suggests that HFT improves
the speed of synchronization among related stocks. Alampieski and Lepone (2012) find that aggressive HFT trade
more UK-US cross-listed stocks in the UK market when the US market opens. Last, CT can quickly adapt to
public information arrivals. Frino, Prodromou, Wang, Westerholm, and Zheng (2017) demonstrate that although
AT is uninformed before corporate earnings announcements, they time their trades better immediately after the
earnings announcements. Alampieski and Lepone (2011) find that HFT liquidity consuming and liquidity providing
activities increase around macroeconomic announcements.
In addition, some studies also show that small trades initiated by HFT can be uninformed. Clark-Joseph (2014)
argues that computerized traders submit small exploratory trades to test the liquidity state of the market and
decide whether to execute their larger trades. O’Hara, Yao, and Ye (2014) provide supporting evidence that small
odd-lot trades by HFT contain less information in comparison to those by non-HFT. Johnson, Van Ness, and
Van Ness (2016) note that trades originated from larger orders (i.e. generated by execution algorithms) contain
less information in comparison to other trades.

5.3 Data and descriptive statistics

In this chapter, we use the same raw data as those used in Chapter 3.4. I.e., we employ a novel AT dataset
provided by the ASX that classifies each stock transaction into algorithmic and non-algorithmic trades. We
compliment the AT dataset with equity market data provided by SIRCA. Our sample period is between October
27, 2008, and October 23, 2009. We filter out stocks that are not present throughout the sample period or have
less than 200 active trading days out of the 252 trading days in our sample. Figure (5.1) presents the sorted
market capitalisations of the 384 remaining stocks after our initial selection criteria.

96
 The Australian stock market is highly concentrated with the top 200 stocks account for approximately 95% of
the total market value. The highest valued stocks such as BHP Billiton and Commonwealth Bank have market
capitalisation over $100 Billion. The size of the stocks quickly diminishes pass the few largest stocks.

Figure 5.1: Market Capitalisation of Individual Stocks in the ASX.

This figure provides the market capitalisation of 384 stocks in the ASX on October 23 2009. The stocks are sorted by
market capitalisation. The X axis represents the rank of the stock whereas the Y axis is the market capitalisation in
billions of dollars.

We form large, medium, and small stock groups with 30 stocks each. The large group consists of the top 30
stocks in our sample. The medium (small) group includes the 50th to the 80th (140th to the 170th ) stocks. By
separating the stocks into size groups, we provide additional evidences in terms of cross-sectional variations in AT
effects. Table (5.1) presents the descriptive statistics of the three size groups.

97
 Table 5.1: Descriptive Statistics
This table reports the descriptive statistics of 30 large, medium, and small stock in the ASX from 27 October 2008 to 23
October 2009. The large group contains the top 30 stocks in Australia by market capitalization, the stock symbols are:
AMP, ANZ, AXA, BHP, BXB, CBA, CCL, CSL, FGL, FMG, IAG, LEI, MQG, NAB, NCM, NWS, ORG, ORI, OSH,
QBE, RIO, SGP, STO, SUN, TLS, WBC, WDC, WES, WOW, and WPL. The medium group consists of the 51st to the
80th largest stocks, their stock symbols are: AOE, BBG, BEN, BLD, BOQ, COH, CTX, DJS, DOW, DXS, EQN, FBU,
FXJ, GMG, HVN, IPL, JHX, MIG, MTS, NHC, OST, PDN, PRY, PTM, RMD, SGM, SOL, TAH, TEL, and TTS. The
small group contains the 141–170th largest stocks, the stock symbols are: ABP, AHD, AIX, AND, AWB, BKN, BPT,
BWP, CAB, CDU, CRG, CXP, CZA, ENV, ESG, EWC, FKP, GNS, GWT, IRE, KCN, LNC, MBN, MCW, MIN, OMH,
REA, SLX, SOT, and WSA. Market capitalisation is calculated based on the closing price on 23 October 2009 in billions
of dollars. Relative bid–ask spread is in percentages. Dollar volumes are scaled by 1,000,000.
Stock size group
Large Medium Small
Panel A: Market conditions
Market capitalisation 29.285 3.590 0.862
Price 18.320 6.891 2.826
Bid–ask spread 0.017 0.014 0.013
Relative bid–ask spread 0.150 0.353 0.730
Panel B: AT and non-AT dollar volumes
AT total 92.810 12.074 2.318
AT active 45.017 5.844 1.127
AT passive 47.793 6.230 1.191
non-AT 22.355 3.296 0.856
non-AT active 12.620 1.858 0.462
non-AT passive 9.735 1.438 0.394

98
5.4 The state space model backgrounds

The state space methods are applied in a variety of research disciplines such as engineering, biology, medicine,
and physics. In this section, we present the quantitative backgrounds specific to our market microstructure
applications. The notations used in this exposition generally follow those in Durbin and Koopman (2012).45 I
also limit this presentation to linear and Gaussian state space models. Last, this exposition does not cover the
most general case of the linear and Gaussian state space methods, e.g. the error structure is mostly assumed to be
diagonal.

5.4.1 Model representation

Suppose we sequentially observe a set of variables, represented by p × 1 vector y = (y1 , y2 , y3 . . . , yp )0 , which

denotes p−variate response values. The observed p−variate vector at time t, is modeled by the observation
equation, which describes the relationship between the observed vector, the regression effects, and the unobserved
state vector:

yt = Zt αt + Xt βt + t , (13)

where αt is an unobserved m × 1 vector called state vector.

Zt is a p × m matrix that corresponds to the state effect on the observed series. Zt represents effect of
m−dimensional states on the observed p−dimensional response variables.
Xt is a p × k matrix the regression effect on the observed series. For each variable in the p−dimensional vector
yt , there are k predictors.
βt is an k × 1 vector of regression coefficients associated with the predictors in matrix Xt .
t is the error vector called observation disturbances. t is usually assumed to be serially independent, zero-mean,
and Gaussian random vectors.
45
Durbin and Koopman (2012) provides an extensive treatment of the state space approach to time-series analyzes, see chapter 3-6
for a general form of the state space representation. As a more accessible material, Commandeur and Koopman (2007) gives a brief
but intuitive introduction.

99
 The state vector hidden states in vector αt is then modeled as a Markov process by a state transition equation:

αt+1 = Tt αt + ηt+1 . (14)

The state transition equation assumes that a new states vector αt+1 is determined by its previous instance, αt ,
a transition matrix, and an error vector.
Tt is a m−dimensional square matrix called state transition matrix. Tt can depend on the past observations
yt−1 , yt−2 , . . . , y1 .
ηt+1 is a m−dimensional state disturbances. The variance-covariance matrix of the state disturbances is Ω. ηt
is usually assumed to be serially independent, zero-mean, and Gaussian random vectors with covariance matrix
Qt . t and ηt are assumed to be mutually independent. Elements in state effect matrix Zt , regression coefficients
vector βt , state transition matrix Tt , and the variance matrix of the state vector αt can be specified as unknown
parameters of interest.
Kalman filter is applied to Equations (13) and (14) to estimate state vectors α2 , α3 , . . . , αT recursively. Whereas
the initial state vector α1 follows the initial condition equation:

α1 = α + A1 δ + η1 , (15)

where α is a m−dimensional vector of known constants.

δ is a q−dimensional vector of unknown quantities.
A1 is a m by q selection matrix. The initial states are often assumed to be diffuse (the covariances are arbitrarily
large e.g. 107 ).
Although the state space models appear complex in this representation, the presented format incorporates
many salient aspects of the model employed in most empirical uses. In practice, many matrices will be either
omitted or reduced to diagonal or identity matrices. Furthermore, the state space models can be readily adapted
for Box-Jenkins style time-series analysis. Many widely known models, such as Autoregressive Moving Average
(ARMA) models, can be written in state space form.46

46
See Durbin and Koopman (2012) for detailed discussion about ARMA and Autoregressive Intergraded Moving Average (ARIMA)

100
5.4.2 Model estimation

Kalman filter and smoother algorithms are the main tools to estimate the state vectors and the log likelihood
is maximised numerically for any unknown parameters. Kalman filter and smoother are recursive algorithms that
provides forward and backward passes through the data:

1. The Kalman filter is used as the forward pass through the data, from t = {1, 2, . . . , T } using a recur-
sive algorithm that calculates each state vector αt+1 based on the information in observations at time
t = {t, t − 1, t − 2, . . . , 1}.
2. The Kalman smoother is used as the backward pass through the data, from t = {T, T − 1, t − 2, . . . , 1} using
a recursive algorithm that calculates each state αt and disturbance t vectors based on the output from the
Kalman filter.

The Kalman filter is proposed by Kalman (1960), in which the idea of the state space representation is also
formulated. The filtered state alphat+1 and its error variances are obtained using the following array of equations:

υt = yt − Zt αt ,
0
Ft = Zt Pt Zt + Ht ,
0
Kt = Tt Pt Zt Ft−1 ,

Lt = Tt − Kt Zt ,

αt+1 = Tt αt + Kt υt ,
0 0
Pt+1 = Tt Pt Lt + Rt Qt Rt , (16)

where υt is the P × 1 one-step forecast error that is the difference between yt and the expected value of yt
given the information up to t − 1 (E(yt |Yt−1 )).
αt+1 is the m × 1 one-step ahead expected value of the state vector E(αt+1 |Yt ).

models in state space forms.

101
 Pt+1 is the m × m one-step ahead error variance matrix of the state vector V ar(αt+1 |Yt ).
The goal of the kalman filter is to recursively calculate one-step ahead conditional distribution of the state
vector given all past information including past state vector αt and past state vector error variance Pt . The state
vector αt+1 obtained in these equations is called filtered state. Filtered state and its error variances are then used
in a backward recursion:

0 0
rt−1 = Zt Ft−1 υt + Lt rt ,
0 0
Nt−1 = Zt Ft−1 Zt + Lt Nt Pt ,

α̂t = αt + Pt rt−1 ,

Vt = Pt − Pt Nt−1 Pt , (17)

where α̂t is the m × 1 conditional mean of the filtered state vector αt , α̂t is obtained using all available
observations. (i.e. α̂t = E(αt |y1 , y2 , . . . , yT )orE(αt |y))
Vt is the m × m error variance matrix of the filtered state vector, Vt = V ar(αt |y).
The state smoothing algorithms estimates the distribution of the state vectors given the entire series. The
conditional mean state vector α̂t is calculated based on the filtered state vectors via a backward pass. The
disturbance vectors t and ηt in Equation (13) and (14) respectively are then smoothed via disturbance smoothing
algorithm:

0 0
rt−1 = Zt Ft−1 υt + Lt rt ,
0 0
Nt−1 = Zt Ft−1 Zt + Lt Nt Lt ,
0
ˆt = Ht (Ft−1 υt − Kt rt ),
0
V ar(t |y) = Ht − Ht (Ft−1 + Kt Nt Kt )Ht ,
0
η̂t = Qt Rt rt ,
0
V ar(ηt |y) = Qt − Qt Rt Nt Rt Qt , (18)

102
 where ˆt is the smoothed vector of the observation disturbance in Equation (13).
η̂t is the smoothed vector of the state disturbance in Equation (14)
The Kalman smoother for the state vector and disturbances vectors ensures that the smoothed states are
estimated with not only past observations but also current as well as future observations. Unknown parameters
specified in state effect matrix Zt , regression coefficients vector βt , state transition matrix Tt , and the variance
matrix of the state vector αt are estimated numerically via maximising the log likelihood function after the filtering
phase. Once the unknown parameters are fitted into the model, another filtering phase is initiated with the fitted
parameters.

5.5 The empirical analysis of AT and price discovery

This section presents the empirical modeling approach and the estimation results. We first discuss the
motivations and the empirical approach to decompose stocks’ intraday price series into efficient and transient
components. We then measure the permanent and transitory price discovery effects of AT by incorporating AT
order flow. Last, we discuss the empirical result and its implications.

5.5.1 The state space decomposition of intraday price discovery

Stock prices do not always reflect the efficient (“true”) value of the company. Many causes are identified in
the literature. For instance, price deviations can arise due to temporary liquidity shocks. Market makers tend to
adjust their prices based on their inventory constraints. Price discreteness and other microstructure features can
also affect stock prices. Agents often under- or overreact to news arrivals and market events. The state space
methods is capable of efficiently incorporate multiple hidden states in a single observed time-series. Many exploit
this feature to disentangle the efficient prices from transitory pricing effects. For example, Hasbrouck (1999)
model the bid and ask quote series via state space procedure. Bid and ask quotes are decomposed into an implicit
efficient price and a quote-exposure cost carried by market makers. Market makers pass on their quote-exposure
cost to the market. And the displayed prices are the sum of the efficient prices and the market making costs.
Menkveld, Koopman, and Lucas (2007) model around-the-clock price discovery process as the unobserved efficient

103
price changes plus transitory price changes.47 In this section, we present a simplified state space decomposition of
stock prices. To minimize the effect of bid–ask bounce in the transacted price series, we measure stock prices using
midquotes. Our model follows the notion that price changes arise from updates to an implicit efficient price series
and a transitory price series. We decompose intraday price series into two components in our observation equation:

pi,t = mi,t + si,t , (19)

where pi,t is the logarithm of the intraday midquote at the time t for stock i. mi,t is the permanent price
component at the time t for stock i and si,t is the transitory price component. Both mi,t and si,t are modeled as
hidden states. The state equation for mi,t is:

mi,t = mi,t−1 + µi,t , (20)

where mi,t is a martingale that captures the efficient price changes. The transitory component of the observed
price series, si,t , is specified as a stationary hidden state:

si,t = φi si,t−1 + νi,t , (21)

where φi si,t−1 is an autoregressive term that characterizes the transitory aspect of si,t . φi is the autoregressive
coefficient. We allow φi to vary between 0 and 1 to capture the mean-reverting feature of the process. µi,t and νi,t
are mutually independent Gaussian process with zero mean and variances of σµ2 and σν2 respectively.
Based on base on the above specifications, we rewrite our empirical model in state space form presented in
Equation (13) and Equation (14). The observation vector yi,t is reduce to a scalar and the state vector αi,t in our
case is a 2 × 1 vector:

yi,t = pi,t ,
0
αi,t = (mi,t , si,t ) . (22)

47
Similar decomposition are featured in Hendershott and Menkveld (2014) and Brogaard, Hendershott, and Riordan (2014).

104
 The state effect Zi,t in Equation (13) is reduced to a 1 × 2 vector; The state transition matrix Ti,t in Equation
(14) is a 2 × 2 matrix; The variance covariance matrix Ωi,t is a diagonal matrix:

Zi,t = (1, 1),

   
2
1 0  σµ 0 
Ti,t =   , Ωi,t =  . (23)
2
0 φi 0 σν

We use Kalman filter and smoothers to estimate the hidden states mi,t and si,t that correspond to the efficient
and transitory component of the intraday midquotes. The autoregressive coefficient in the transitory process (θi ),
and the variances of the error terms in permanent and transitory processes (σµ2 and σν2 , respectively) are treated
as unknown parameters in the likelihood functions. These parameters are estimated numerically via maximum
likelihood estimation.
By way of illustration, Figure (5.2) decomposes a representative intraday stock price series (National Australian
Bank) using the state space approach. The observed logarithm price series is decomposed into a martingale efficient
price series and an autoregressive transient price series. The model leverages the Kalman filter and smoother,
which estimates the unobserved efficient price conditional on all intraday observations. Specifically, the unobserved
efficient price at time j, mi,j , is estimated based on the past prices (pi,t<j ), future prices (pi,t<j ), and the current
price (pi,t=j ).

105
 Figure 5.2: Intraday Price Series and Hidden Efficient Price Series in One Trading Session.
This figure illustrates the intraday dynamics of the observed logarithm price series and the unobserved efficient component
of the price series for the stock National Australia Bank Ltd. (stock symbol NAB) on January 29, 2009.

Figure (5.3) highlights the magnitudes of the two hidden components (mi,t and si,t ) of the observed logprice
series. While the transient component, si,t , contributes to a small fraction of the overall price process, it is
consistently present throughout the trading day.
Overall, Figure (5.2) and (5.3) show that the decomposition allows for a more accurate measure of the permanent
price discovery process and motivates the decomposition of the observed price series and the advantages of the
state space representation.

106
Figure 5.3: Intraday Return of the Efficient Component vs the Transitory Component in One Trading Session.
This figure illustrates the intraday dynamics of the unobserved efficient price series returns and the unobserved transitory
price series for the stock National Australia Bank Ltd. (stock symbol NAB) on January 29, 2009.

5.5.2 The price discovery of AT order flow

To assess the effect of AT and non-AT on the efficient and transitory components of price changes, we extend
our model in Section 5.5.1 by incorporating the net order flow of AT and non-AT.
We model our observation equation as follows:

pi,t = mi,t + si,t , (24)

where pi,t is the log midquote at the time t for stock i. mi,t is the permanent price component at the time t
for stock i and si,t is the transitory price component. Both mi,t and si,t are modeled as hidden states. The state
equation for mi,t is:

107
 mi,t = mi,t−1 + ωi,t , (25)

where mi,t is a martingale. ωi,t characterizes the permanent price increments. The price relevant information
arrival for stock i between time t and time t − 1 is captured by ωi,t . We include the effects of AT and non-AT in
the process of permanent price changes (ωi,t ):

ωi,t = θi iati,t + τi inonati,t + µi,t , (26)

where iati,t and inonati,t are the innovation of AT and non-AT net order flows respectively. The innovation of
AT and non-AT are obtained as the residuals of autoregressive processes following Brogaard, Hendershott, and
Riordan (2014). θi and τi are the main variable of interest, which quantifies the effect of AT and non-AT on the
permanent price process respectively. µi,t is the component of ωi,t unrelated to trading. The transitory component
of the observed price series, si,t , is specified as a stationary hidden state:

si,t = φi si,t−1 + κi ati,t + ϕi nonati,t + νi,t , (27)

where φi si,t−1 is an autoregressive term that characterizes the transitory aspect of si,t . φi is the autoregressive
coefficient. We allow φi to vary between 0 and 1 to capture the mean-reverting feature of the process. ati,t and
nonati,t are the order flow of AT and non-AT at time t, which are calculated as the buy initiated dollar volume
less sell initiated dollar volume at time t for AT and non-AT. The key variables of interest are κi and ϕi , which
measures the effect of AT and non-AT on the transitory price process. µi,t and νi,t are mutually independent
Gaussian process with zero mean and variances of σµ2 and σν2 respectively.
We estimate 9 unknown variables of the state space representation base on the specification in Section 5.5.2.
The impact of AT on the permanent process (θi ), the impact of non-AT on the permanent process (τi ), the impact
of AT on the transitory process (κi ), and the impact of non-AT on transitory process (ϕi ) are treated as elements
in the unobserved state vector (αt ). These hidden states are estimated via Kalman filter and smoothers. The
autoregressive coefficient in the transitory process (θi ), and the variances of the error terms in permanent and
transitory processes (σµ2 and σν2 , respectively) are treated as unknown parameters in the likelihood functions. These

108
parameters are estimated numerically via maximum likelihood estimation.
We then write our empirical model in state space form presented in Equation (13) and Equation (14). The
observation vector yi,t is reduce to a scalar and the state vector αi,t in our case is a 6 × 1 vector:

yi,t = pi,t ,
0
αi,t = (mi,t , si,t , θi , τi , κi , ϕi ) . (28)

The state effect Zi,t is reduced to a 1 × 6 vector; The state transition matrix Ti,t is a 6 × 6 matrix; The variance
covariance matrix Ωi,t has a block structure:

Zi,t = (1, 1, 0, 0, 0, 0),

   
2
1 0 iati,t inonati,t 0 0  σ µ 0 0 0 0 0
   
0 φ 0 0 at nonat   0 σ2 0 0 0 0
 i i,t i,t   ν 
   
0 0 1 0 0 0  0 0 0 0 0 0
   
Ti,t = 
 
 , Ωi,t = 

.
 (29)
0 0 0 1 0 0  0 0 0 0 0 0
   
   
0 0 0 0 1 0  0 0 0 0 0 0
   
   
0 0 0 0 0 1 0 0 0 0 0 0

Overall, the inclusion of AT and non-AT facilitates the interpretation on the relative contribution of the two
trading groups to the price formation processes. We assess and compare both the liquidity demanding trades and
liquidity supplying traders using our state space representation.

5.5.3 Estimation results and discussion

The estimation results for AT and non-AT liquidity demanding order flows are presented in Table (5.2). In
general, both AT and non-AT contributes to the permanent price discovery processes. Algorithmic trades provides
more price discovery compared to non-algorithmic trades.

109
 Table 5.2: State Space Model for Intraday Prices and Liquidity Demanding Order Flows
This table presents the estimates for the state space decompositions of large, median, and small stocks using the following
model:

pi,t = mi,t + si,t ,

mi,t = mi,t−1 + ωi,t ,
ωi,t = θi iati,t + τi inonati,t + µi,t ,
si,t = φi si,t−1 + κi ati,t + ϕi nonati,t + νi,t ,

where i denotes stock i and t denotes time t of the day. pi,t is the observed log midquote which is decomposed
into an unobserved efficient price, mi,t and a transient pricing error, si,t . mi,t is modeled as a martingale, and
ωi,t is the efficient price increments. ati,t is the order flow of liquidity demanding AT measured as the buy dol-
lar volume less sell dollar volume at time t for stock i. nonati,t is the order flow of liquidity demanding non-
AT measured analogously. iati,t and inonati,t are the innovations of a VAR of ati,t and nonati,t . The model
is estimated for the intraday stock prices on a per stock-day basis. The p-values are included in parentheses.
Size Groups
Large Medium Small
Panel A: Permanent Component
iat 0.714 2.504 18.550
(0.000) (0.000) (0.000)
inonat 0.598 1.623 15.286
(0.000) (0.005) (0.000)
iat less inonat 0.116 0.833 3.264
(0.000) (0.000) (0.208)
Panel B: Transient Component
at -0.115 0.482 -0.441
(0.000) (0.075) (0.731)
nonat 0.036 1.829 1.592
(0.310) (0.039) (0.453)
at less nonat -0.151 -1.347 -2.033
(0.000) (0.057) (0.425)

110
 Panel A of Table (5.2) reports analyzes of the permanent price component, mi,t . The coefficients for AT and
non-AT order flows (θi , τi , κi , ϕi ) are return per dollar volume scaled by 100 million. The AT order flow coefficient
θi for large stocks equals to 0.714, which implies that one million dollar surprise order flow corresponds to 0.714
basis points increase in the permanent price component for large stocks. Whereas the coefficient for nonAT order
flow of the same amount associates with 0.598 basis points increase. The difference between AT and non-AT
order flows coefficients for large stocks is 0.116 with the p-value of 0.000, which implies that AT order flows are
significantly more informative compared to non-AT order flows for large stocks.
The order flow coefficients for medium and small stocks are progressively larger, consistent with the notion
that prices of smaller, less often traded stocks could be moved with less dollar volume. For medium stocks, the
coefficient of AT order flows is 0.833 higher than that of non-AT. For the group of small stocks, AT order flows
are also more informed compared to non-AT order flows in terms of the coefficient magnitudes. However, the
difference is not statistically significant. Panel A shows that overall, AT contributes more to the efficient price
discovery process, especially in larger stocks.
Panel B of Table (5.2) shows the coefficients of order flows related to the transient price component. The
results indicate that AT mitigates the transient pricing errors while non-AT does not reduce the pricing errors in
large stocks. In smaller stocks, AT contributes less to the transient pricing errors compared to non-AT. Specifically,
the coefficients of AT order flow is -0.115, which suggests that algorithmic traders initiate trades in the opposite
direction of transient pricing errors. On the other hand, the results do not show the mitigating effects of non-AT
for the transient pricing errors. In comparison to non-AT, AT contributes significantly less to the transient pricing
error component.
Table (5.3) presents the estimations for AT and non-AT liquidity supplying order flows. The structure of Table
(5.3) is analogous to that of Table (5.2).
Both AT and non-AT liquidity supplying order flows are negatively related to the permanent price components,
which supports the theoretical prediction that liquidity supplying trades are adversely selected while profiting from
the bid–ask spread. Hendershott, Jones, and Menkveld (2011) shows that AT is more active in larger stocks. In the
large cap group, we find that AT is adversely selected to a larger extent compared to non-AT. In the medium and
small groups, the adverse selection effect is similar between AT and non-AT order flows. Our finding is consistent

111
 Table 5.3: State Space Model for Intraday Prices and Liquidity Supplying Order Flows
This table presents the estimates for the state space decompositions of large, median, and small stocks using the following
model:

pi,t = mi,t + si,t ,

mi,t = mi,t−1 + ωi,t ,
ωi,t = θi iati,t + τi inonati,t + µi,t ,
si,t = φi si,t−1 + κi ati,t + ϕi nonati,t + νi,t ,

where i denotes stock i and t denotes time t of the day. pi,t is the observed log midquote which is decomposed
into an unobserved efficient price, mi,t and a transient pricing error, si,t . mi,t is modeled as a martingale, and
ωi,t is the efficient price increments. ati,t is the order flow of liquidity supplying AT measured as the buy dol-
lar volume less sell dollar volume at time t for stock i. nonati,t is the order flow of liquidity supplying non-
AT measured analogously. iati,t and inonati,t are the innovations of a VAR of ati,t and nonati,t . The model
is estimated for the intraday stock prices on a per stock-day basis. The p-values are included in parentheses.
Size Groups
Large Medium Small
Panel A: Permanent Component
iat -0.623 -2.276 -15.002
(0.000) (0.000) (0.000)
inonat -0.517 -1.876 -18.255
(0.000) (0.000) (0.000)
iat less inonat -0.106 -0.400 3.253
(0.008) (0.178) (0.393)
Panel B: Transient Component
at 0.020 -0.269 -1.694
(0.445) (0.050) (0.440)
nonat 0.034 -0.364 0.038
(0.349) (0.218) (0.990)
at less nonat -0.013 0.095 -1.732
(0.697) (0.767) (0.656)

112
with the notion that the algorithmic traders in our sample are primarily large buy-side institutional institutional
investors who are not specialized electronic market makers (ASX, 2010).

5.5.4 Robustness test: AT on turbulent days

Although algorithmic traders increase market liquidity, it is widely suspected that AT may retreat from the
market in difficult times. To assess this concern, the main analyses are repeated on turbulent days. We have
identified 19 (20) days when the market index increases (decreases) by more than 2% over our sample period.48
Table (5.4) and (5.5) present the estimation results for liquidity demanding order flows and liquidity supplying
order flows on turbulent days, respectively.
Overall, the results on turbulent days are qualitatively and quantitatively similar to those on all trading days.
In particular, Panel A of Table (5.4) reports the association between liquidity demanding order flows and the
permanent price components. Both AT and non-AT order flows contribute to the permanent component of the
prices. One million dollar surprise AT order flows contribute, on average, 0.135 basis points more to the efficient
price component compared to those of non-AT in large stocks. Panel B of Table (5.4) shows the coefficients
of liquidity demanding order flows related to the transient price component. In large stocks, one million dollar
surprise AT order flows contribute -0.115 basis points to the transient price component. Non-AT order flows have
insignificant effects on the transient price component. The difference in AT and non-AT order flow effects are
less pronounced (insignificant) in mediun (small) stocks. Table (5.5) presents the coefficients for AT and non-AT
liquidity supplying order flows. Both AT and non-AT liquidity supplying order flows are negatively related to the
permanent price components, consistent with the notion that liquidity providers are adversely selected.

48
See Table (3.1) and (3.2) for more details on the event days.

113
 Table 5.4: State Space Model for Intraday Prices and Liquidity Demanding Order Flows on Turbulent Days
This table presents the estimates for the state space decompositions of large, median, and small stocks on volatile days.
Volatile days are defined as the days when the absolute values of the market returns exceed 2%.

pi,t = mi,t + si,t ,

mi,t = mi,t−1 + ωi,t ,
ωi,t = θi iati,t + τi inonati,t + µi,t ,
si,t = φi si,t−1 + κi ati,t + ϕi nonati,t + νi,t ,

where i denotes stock i and t denotes time t of the day. pi,t is the observed log midquote which is decomposed
into an unobserved efficient price, mi,t and a transient pricing error, si,t . mi,t is modeled as a martingale, and
ωi,t is the efficient price increments. ati,t is the order flow of liquidity demanding AT measured as the buy dol-
lar volume less sell dollar volume at time t for stock i. nonati,t is the order flow of liquidity demanding non-
AT measured analogously. iati,t and inonati,t are the innovations of a VAR of ati,t and nonati,t . The model
is estimated for the intraday stock prices on a per stock-day basis. The p-values are included in parentheses.
Size Groups
Large Medium Small
Panel A: Permanent Component
iat 0.742 3.015 31.343
(0.000) (0.000) (0.000)
inonat 0.609 1.887 22.586
(0.000) (0.003) (0.000)
iat less inonat 0.135 1.128 8.757
(0.045) (0.036) (0.189)
Panel B: Transient Component
at -0.135 0.500 -5.3019
(0.013) (0.199) (0.265)
nonat 0.037 1.480 1.531
(0.625) (0.018) (0.668)
at less nonat -0.171 -0.980 -6.833
(0.010) (0.075) (0.297)

114
 Table 5.5: State Space Model for Intraday Prices and Liquidity Supplying Order Flows on Turbulent Days
This table presents the estimates for the state space decomposition of large, median, and small stocks on volatile days.
Volatile days are defined as the days when the absolute values of the market returns exceed 2%.

pi,t = mi,t + si,t ,

mi,t = mi,t−1 + ωi,t ,
ωi,t = θi iati,t + τi inonati,t + µi,t ,
si,t = φi si,t−1 + κi ati,t + ϕi nonati,t + νi,t ,

where i denotes stock i and t denotes time t of the day. pi,t is the observed log midquote which is decomposed
into an unobserved efficient price, mi,t and a transient pricing error, si,t . mi,t is modeled as a martingale, and
ωi,t is the efficient price increments. ati,t is the order flow of liquidity demanding AT measured as the buy dol-
lar volume less sell dollar volume at time t for stock i. nonati,t is the order flow of liquidity demanding non-
AT measured analogously. iati,t and inonati,t are the innovations of a VAR of ati,t and nonati,t . The model
is estimated for the intraday stock prices on a per stock-day basis. The p-values are included in parentheses.
Size Groups
Large Medium Small
Panel A: Permanent Component
iat -0.560 -2.541 -20.592
(0.000) (0.000) (0.000)
inonat -0.387 -2.088 -40.975
(0.000) (0.000) (0.008)
iat less inonat -0.173 -0.453 20.383
(0.087) (0.487) (0.202)
Panel B: Transient Component
at -0.059 -0.541 -4.706
(0.442) (0.205) (0.358)
nonat -0.088 -0.572 13.294
(0.418) (0.325) (0.367)
at less nonat 0.029 0.031 -18.000
(0.770) (0.967) (0.270)

115
5.6 Conclusions

In this chapter, we investigate the role of AT in the price formation process. Building on Brogaard, Hendershott,
and Riordan (2014), we apply state space framework and decompose the observed stock prices into unobserved
permanent price series and transient pricing errors. A brief background on the state space models and Kalman
filters and smoothers is provided. These techniques are then applied to analyze 30 large, medium, and small stocks
on the Australian Securities Exchange from October 2008 to October 2009.
We find that AT order flow, calculated as buy dollar volume less sell dollar volume, is positively related to
the permanent price discovery process. AT mitigates transient pricing errors by trading in the opposite direction
of the transient component of the price discovery process. The contributions to the price formation processes
are mainly attributed to algorithmic traders’ market orders. When compared to non-AT, AT contributes more
to the permanent price component and less to the transient pricing errors. Further analyzes indicate that AT’s
contributions remain during turbulent periods, defined as trading days when the stock market’s absolute return
exceeds 2%.
Our study is subject to a few caveats. First, it focuses on a single market and a broad group of algorithmic
traders. Further extensions could be made comparing the longer investment horizon algorithms and the ultra
fast high frequency algorithms. Second, transient pricing errors are identified by exploiting the autoregressive
properties of security prices in the veins of Hasbrouck’s (1991) VAR measures and Brogaard et al.’s (2014) state
space frameworks. Future studies could validate and extend our results by identifying price inefficiencies through
other channels such as news versus rumor comparisons and price manipulation events.

116
 Chapter 6

Conclusion
This thesis presents one survey essay and three empirical essays on computerized trading. This chapter presents
a summary of the main findings and directions for future research.

6.1 Overview and conclusions

The survey essay, presented in Chapter 2, reviews around 100 theoretical, policy, and empirical studies on
the topic of AT and HFT. Besides the common theme of computerized traders being fast, the theoretical studies
employ a wide range of models based on various aspects of CT. Our survey categorizes the theoretical papers in
relation to market maker–taker dynamics, information content of fast traders, recently incurred market structural
changes, and proposed market changes. Overall, CT can leverage its reaction speed to excel as specialized market
makers, liquidity consumers, or both. CT relies on various information such as future order flows, hidden liquidity
state, and machine readable information. Increased market fragmentation, market maker favoring fee structures,
and other market structural shifts facilitate the widespread use of CT. The empirical literature, despite its rapid
growth, is still in the early stage with many papers yet to be published. Therefore, we survey the empirical studies
with an emphasise on data quality, event identification, and establishing causal links. Finally, the empirical effects
of CT are critically summarized based on various market quality dimensions.
The first empirical essay, presented in Chapter 3.4, investigates the characteristics of AT on turbulent day
during the periods from October 2008 to October 2009. Similar to Dennis and Strickland (2002), 19 up swing
days and 20 down swing days when the market moves up or down by more than 2% are identified. We then
analyze the association between individual stock returns and the level of AT intensity in those stocks. We find
that AT intensity is negatively related to individual stocks’ price fluctuation on turbulent days. The effects are
economically significant: on average, 10% (or half standard deviation) more AT buying on days when the market
return is more than 2%, corresponds to 16 basis points less up swing for individual stocks. On days when the

117
market drops by more than than 2%, similar magnitudes are observed between AT sells and individual stock down
swings. The stocks on turbulent days are then matched with those on non-turbulent days based on abnormal
return, market capitalisation, liquidity, and risk. Based on a difference-in-difference analysis, we highlight the
significant association between AT and individual stock returns exists on turbulent days compared to that on
other days. Overall, our findings compliments the empirical findings by Hendershott, Jones, and Menkveld (2011)
and Hasbrouck and Saar (2013) in that AT mitigates price pressures of individual stocks during market swings.
The second empirical essay, presented in Chapter 4.4, examines the determinants of algorithmic execution
strategy and the effects of AT order imbalances. we first apply a logit model to assess the choice of submitting
algorithmic trades versus other trades in relation to the prevailing VWAP movements. VWAP is a common metric
to measure execution performance. Traders aim to buy (sell) at a lower (higher) price compared to VWAP at the
end of the trading day. We construct the intraday VWAP time series as a determinant of AT executions. We
find that AT is more likely to trade when the VWAP–price relation is more favorable. Specifically, algorithmic
traders are more likely to initiate a transaction from the buy (sell) side when VWAP is lower (higher), compared
to other traders. We then extend the analyses by Chordia and Subrahmanyam (2004) by separate the effects of
order imbalances from AT and non-AT. We find that the price impact of AT order imbalance is smaller than those
executed by non-AT.
The last empirical essay, presented in Chapter 5, analyze the role of AT in the price formation process. We
distinguish the impact of AT on the efficient price discovery process and the short-term noises. A state space
frame work is employed to decompose the observed stock prices into unobserved efficient price series and transient
pricing errors. We provide a brief background on the applied models and investigate 30 of the large, medium,
and small Australian stocks during the period between October 2008 and October 2009. We then relate AT and
non-AT order flows, calculated as buy dollar volume less sell dollar volume, to both unobserved processes. We find
that, compared to non-AT, AT contribute significantly more to the permanent price discovery process and less to
the transient pricing errors. Moreover, our result is robust during turbulent days when the absolute value of the
market return exceeds 2%.

118
6.2 Suggestions for future research

Finally, we suggest several directions for future research. First, further evidence is required to optimize the
design of financial markets. Given the rapid growth of CT, many market structural changes are being proposed or
introduced without sufficient academic enquiry. For instance, more empirical evidence is needed to determine the
optimal transaction speed. The theoretical literature have offered several competing prediction on whether the
markets should adopt “speed bumps” for trading.49 Empirically, Hendershott, Jones, and Menkveld (2011) find
that the increased reaction speed due to Autoquote upgrade is beneficial to the markets in terms of liquidity and
price discovery. However, Chakrabarty, Jain, Shkilko, and Sokolov (2015) find that the benefits of slower markets
in terms of reduced adverse selection costs outweigh the cost in terms of increased pricing errors.
Second, future research should focus on the heterogeneity of CT since the current research mostly focus on
one or few aggregate groups of AT and HFT. This line of research is further justified by the apparent disconnect
between the poor public perception of CT and the mostly benign effects in the literature. Specifically, several of
the nefarious effects suspected by the public are not found by academic studies.50 The poor publicity of CT may
have been misplaced since the media tends to focus on the negative stories. However, further research is needed
to determine the validity of these suspicions. With the proliferation of account level HFT data, the academia is
equipped to uncover the broad spectrum of HFT and identify potentially hazardous HFT practices.
Third, further research is needed to identify the private information obtained by high frequency traders. Many
studies find that the private information incorporate by HFT is short-lived. O’Hara (2015) notes that short-term
private information and price changes might not only be fundamental value-related but also driven by investors
anticipating each others’ orders. Short-term price discoveries may blur the traditional dichotomy of “informed”
versus “uninformed” traders. Furthermore, to determine the economic contributions of HFT price discoveries,
future studies should establish whether the price changes incorporated by HFT are related to the true value of the
assets.
49
See, e.g., Budish, Cramton, and Shim (2015); Foucault, Kadan, and Kandel (2013); Guo (2015); Bongaerts and Van Achter
(2016).
50
For instance, HFT is widely blamed for “flash crash”, however, Kirilenko, Kyle, Samadi, and Tuzun (2017) find that HFT is not
the cause.

119
 Last, we should re-evaluate the traditional market efficiency measures and propose more market fairness
measures. Specifically, traditional liquidity and price discovery measures are to be re-examined due to the
proliferation of CT. Several studies point out that liquidity measures such as bid–ask spread and market depth
may be influenced by the “quote flickering” practice of HFT.51 Similarly, benefits of price discovery should be
measured in conjunction with its costs (Stiglitz 2014 and Chakrabarty, Jain, Shkilko, and Sokolov 2015, for
example). Moreover, further analyses are needed on market fairness. CT is widely suspected to impose adverse
selection costs on other market participants. However, compared to studies on market efficiency, much fewer
number of papers focus on the negative impact of computerized traders on other investors. For instance, Biais,
Foucault, and Moinas (2015) allow slow traders to become fast traders at a cost. They find that fast traders
impose negative externalities upon slow traders. Therefore, the number of fast traders and the investment into
fast trading technologies in equilibrium is likely to cause a social welfare loss. More empirical evidence is needed
to address the market fairness implications of CT.

51
See, among others, Hasbrouck and Saar (2009), Baruch and Glosten (2013), Hasbrouck and Saar (2013), and Van Ness, Van Ness,
and Watson (2015).

120
References
Aggarwal, N., Thomas, S., 2014. The causal impact of algorithmic trading on market quality. Working Paper.
Indira Gandhi Institute of Development.

Aı̈t-Sahalia, Y., Saglam, M., 2014. High frequency traders: Taking advantage of speed. Working paper. Princeton
University.

Aitken, M., Chen, H., Foley, S., 2017. The impact of fragmentation, exchange fees and liquidity provision on
market quality. Journal of Empirical Finance, forthcoming.

Aitken, M., Cumming, D., Zhan, F., 2015. High frequency trading and end-of-day price dislocation. Journal of
Banking and Finance 59, 330–349.

Aitken, M., Frino, A., 1996. The accuracy of the tick test: Evidence from the Australian Stock Exchange. Journal
of Banking and Finance 20 (10), 1715–1729.

Alampieski, K., Lepone, A., 2011. High frequency trading firms, order book participation and liquidity supply
during periods of heightened adverse selection risk: Evidence from LSE, BATS and Chi-X. Working paper.
Macquarie University.

Alampieski, K., Lepone, A., 2012. High frequency trading in UK equity markets: Evidence surrounding the US
market open. Working paper. Macquarie University.

Anand, A., Chakravarty, S., 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 8 (3), 288–308.

ASIC, 2013. Dark liquidity and high frequency trading. Available at: http://download.asic.gov.au/media/
1338878/info178-published-15-January-2014.pdf(retrieved on March 25, 2014).

ASIC, 2014. Asic market supervision update issue 45. Available at: http://asic.gov.au/about-asic/corporate-
publications/newsletters/asic-market-supervision-update/asic-market-supervision-update-
previous-issues/asic-market-supervision-update-issue-45/(retrieved on June 17, 2015).

121
ASX, 2010. Algorithmic trading and market access arrangements. Available at: http://www.asx.com.au/
documents/media/20100211_review_algorithmic_trading_and_market_access.pdf(retrieved on February
22, 2014).

ASX, 2011a. Getco Australia Pty Ltd - Admission as an ASX market participant. Available at: https:
//www.asxonline.com/intradoc-cgi/groups/participant_services/documents/communications/asx_
030553.pdf(retrieved on June 14, 2014).

ASX, 2011b. Virtu financial Asia Pty Ltd admission as an ASX market (principal trader) participant.
Available at: https://www.asxonline.com/intradoc-cgi/groups/participant_services/documents/
communications/asx_031096.pdf(retrieved on June 14, 2014).

Baldauf, M., Mollner, J., 2015. High-frequency trading and market performance. Working paper. University of
British Columbia.

Banz, R. W., 1981. The relationship between return and market value of common stocks. Journal of Financial
Economics 9 (1), 3–18.

Baron, M., Brogaard, J., Kirilenko, A., 2012. The trading profits of high frequency traders. Working paper. Foster
School of Business.

Baron, M. D., Brogaard, J., Hagströmer, B., Kirilenko, A. A., 2016. Risk and return in high frequency trading.
Working paper, University of Washington.

Baruch, S., Glosten, L. R., 2013. Fleeting orders. Working paper. Columbia Business School.

Bellia, M., Pelizzon, L., Subrahmanyam, M. G., Uno, J., Yuferova, D., 2016. Low-latency trading and price
discovery: Evidence from the Tokyo Stock Exchange in the pre-opening and opening periods. Working Paper.
Goethe University Frankfurt.

Benos, E., Sagade, S., 2016. Price discovery and the cross-section of high-frequency trading. Journal of Financial
Markets 30, 54–77.

122
Berkowitz, S. A., Logue, D. E., Noser, E. A., 1988. The total cost of transactions on the NYSE. Journal of Finance
43 (1), 97–112.

Bernales, A., 2014. Algorithmic and high frequency trading in dynamic limit order markets. Working paper.
Universidad de Chile.

Bershova, N., Rakhlin, D., 2013. High-frequency trading and long-term investors: A view from the buy-side.
Journal of Investment Strategies 2 (2), 25–69.

Biais, B., Foucault, T., Moinas, S., 2015. Equilibrium fast trading. Journal of Financial Economics 116 (2), 292–313.

Biais, B., Woolley, P., 2011. High frequency trading. Working paper. Toulouse School of Economics, IDEI.

Blocher, J., Cooper, R. A., Seddon, J., Van Vliet, B., 2016. Phantom liquidity and high-frequency quoting. Journal
of Trading 11 (3), 6–15.

Boehmer, E., Fong, K., Wu, J., 2013. Algorithmic trading and changes in firms’ equity capital. Working paper.
EDHEC Business School.

Boehmer, E., Fong, K., Wu, J., 2015. International evidence on algorithmic trading. Working paper. EDHEC
Business School.

Boehmer, E., Shankar, R., 2014. Low latency trading and comovement of order flow, prices, and market conditions.
Working paper. EDHEC Business School.

Bongaerts, D., Van Achter, M., 2016. High-frequency trading and market stability. Working paper. Erasmus
University.

Breckenfelder, J. H., 2013. Competition between high-frequency traders, and market quality. Working paper.
European Central Bank.

Brogaard, J., Carrion, A., Moyaert, T., Riordan, R., Shkilko, A., Sokolov, K., 2016. High-frequency trading and
extreme price movements. Working paper. University of Washington.

123
Brogaard, J., Garriott, C., Pomeranets, A., 2016. High-frequency trading competition. Working paper. University
of Washington.

Brogaard, J., Hagströmer, B., Nordén, L., Riordan, R., 2015. Trading fast and slow: Colocation and liquidity.
Review of Financial Studies 28 (12), 3407–3443.

Brogaard, J., Hendershott, T., Hunt, S., Ysusi, C., 2014. High-frequency trading and the execution costs of
institutional investors. The Financial Review 49 (2), 345–369.

Brogaard, J., Hendershott, T., Riordan, R., 2014. High frequency trading and price discovery. Review of Financial
Studies 27, 2267–2306.

Brolley, M., Malinova, K., 2012. Maker-taker fees and informed trading in a low-latency limit order market.
Working paper. University of Toronto.

BSE, 2014. Bse may update, 2014. Available at: http://www.sebi.gov.in/cms/sebi_data/attachdocs/

1470393485587.pdf (retrieved on December 15, 2016).

Budish, E., Cramton, P., Shim, J., 2015. The high-frequency trading arms race: Frequent batch auctions as a
market design response. The Quarterly Journal of Economics 130 (4), 1547–1621.

Carrion, A., 2013. Very fast money: High-frequency trading on the NASDAQ. Journal of Financial Markets 16 (4),
680–711.

Cartea, A., Payne, R., Penalva, J., Tapia, M., 2016. Ultra-fast activity and market quality. Working paper.
University of Oxford.

Cartea, Á., Penalva, J., 2012. Where is the value in high frequency trading? Quarterly Journal of Finance 2 (3),
1–46.

Cespa, G., Vives, X., 2016. High frequency trading and fragility. Working Paper. Cass Business School.

124
CFTC/SEC, 2010. Preliminary findings regarding the market events of may 6, 2010. Available at: http://www.
cftc.gov/ucm/groups/public/@otherif/documents/ifdocs/opa-jointreport-sec-051810.pdf (retrieved
on February 2, 2013).

Chaboud, A. P., Chiquoine, B., Hjalmarsson, E., Vega, C., 2014. Rise of the machines: Algorithmic trading in the
foreign exchange market. Journal of Finance 69 (5), 2045–2084.

Chakrabarty, B., Jain, P. K., Shkilko, A., Sokolov, K., 2015. Speed of market access and market quality: Evidence
from the SEC naked access ban. Working paper. Wilfrid Laurier University.

Chakrabarty, B., Moulton, P. C., Shkilko, A., 2012. Short sales, long sales, and the Lee–Ready trade classification
algorithm revisited. Journal of Financial Markets 15 (4), 467–491.

Chakraborty, S., 2016. High-speed trading hits limit in india. Financial Review. Available at:
http://www.bloomberg.com/news/articles/2015-05-04/toronto-exchange-confronts-controversial-
trading-rebate-system(retrieved on June 10, 2015).

Chan, K., Fong, W.-M., 2000. Trade size, order imbalance, and the volatility–volume relation. Journal of Financial
Economics 57 (2), 247–273.

Chordia, T., Green, C., Kottimukkalur, B., 2016. Low latency trading on macroeconomic announcements. Working
paper. Emory University.

Chordia, T., Roll, R., Subrahmanyam, A., 2002. Order imbalance, liquidity, and market returns. Journal of
Financial Economics 65 (1), 111–130.

Chordia, T., Subrahmanyam, A., 2004. Order imbalance and individual stock returns: Theory and evidence.
Journal of Financial Economics 72 (3), 485–518.

Clark-Joseph, A. D., 2014. Exploratory trading. Working paper. University of Illinois.

Cochrane, J. H., 1994. Permanent and transitory components of GNP and stock prices. The Quarterly Journal of
Economics, 241–265.

125
Colliard, J.-E., Foucault, T., 2012. Trading fees and efficiency in limit order markets. Review of Financial Studies.

Colliard, J.-E., Hoffmann, P., 2017. Financial transaction taxes, market composition, and liquidity. Journal of
Finance, forthcoming.

Commandeur, J. J., Koopman, S. J., 2007. An introduction to state space time series analysis. Oxford University
Press.

Cvitanic, J., Kirilenko, A. A., 2010. High frequency traders and asset prices. Working paper. MIT Sloan School.

Dennis, P. J., Strickland, D., 2002. Who blinks in volatile markets, individuals or institutions? Journal of Finance
57 (5), 1923–1949.

Dierkens, N., 1991. Information asymmetry and equity issues. Journal of Financial and Quantitative Analysis
26 (02), 181–199.

Domowitz, I., Yegerman, H., 2005. The cost of algorithmic trading: A first look at comparative performance, in
Brian R. Bruce ed. Algorithmic Trading: Precision, Control, Execution (Institutional Investor London).

Duong, H. N., Kalev, P. S., Sun, Y., 2016. Algorithmic trading in non-announcing rival firms. Working paper.
UniSA Business School.

Durbin, J., Koopman, S. J., 2012. Time series analysis by state space methods, 2nd Edition. Vol. 38. Oxford
University Press.

Easley, D., Lopez de Prado, M., O’Hara, M., 2012. The volume clock: Insights into the high frequency paradigm.
Journal of Portfolio Management 39, 680–711.

Egginton, J. F., Van Ness, B. F., Van Ness, R. A., 2016. Quote stuffing. Financial Management.

Ellis, K., Michaely, R., O’Hara, M., 2000. The accuracy of trade classification rules: Evidence from NASDAQ.
Journal of Financial and Quantitative Analysis 35 (04), 529–551.

Engle, R. F., Ferstenberg, R., Russell, J., 2012. Measuring and modeling execution cost and risk. The Journal of
Portfolio Management 38 (2), 14–28.

126
ESMA, 2014. High-frequency trading activity in EU equity markets. Available at: http://www.esma.europa.eu/
system/files/esma20141_-_hft_activity_in_eu_equity_markets.pdf (retrieved on May 22, 2015).

Fama, E. F., MacBeth, J. D., 1973. Risk, return, and equilibrium: Empirical tests. Journal of Political Economy,
607–636.

Fong, K. Y., Liu, W.-M., 2010. Limit order revisions. Journal of Banking & Finance 34 (8), 1873–1885.

Foster, F. D., VISWANATHAN, S., 1993. Variations in trading volume, return volatility, and trading costs;
evidence on recent price formation models. Journal of Finance 48 (1), 187–211.

Foucault, T., 1999. Order flow composition and trading costs in a dynamic limit order market. Journal of Financial
Markets 2 (2), 99–134.

Foucault, T., Hombert, J., Roşu, I., 2016. News trading and speed. The Journal of Finance 71 (1), 335–382.

Foucault, T., Kadan, O., Kandel, E., 2013. Liquidity cycles and make/take fees in electronic markets. Journal of
Finance 68 (1), 299–341.

Foucault, T., Kozhan, R., Tham, W. W., 2017. Toxic arbitrage. Review of Financial Studies, forthcoming.

Frino, A., Mollica, V., Monaco, E., Palumbo, R., 2017. The effect of algorithmic trading on market liquidity:
Evidence from earnings annoucements on Borsa Italiana. Pacific-Basin Finance Journal, forthcoming.

Frino, A., Prodromou, T., Wang, G. H., Westerholm, P. J., Zheng, H., 2017. An empirical analysis of algorithmic
trading around earnings announcements. Pacific-Basin Finance Journal, forthcoming.

Gai, J., Yao, C., Ye, M., 2014. The externalities of high frequency trading. AFA 2014 Philadelphia Meetings
Papers.

Gao, C., Mizrach, B., 2011. High frequency trading in the equity markets during large-scale asset purchases.
Working paper. Rutgers University.

Gao, C., Mizrach, B., 2016. Market quality breakdowns in equities. Journal of Financial Markets 28, 1–23.

127
Garman, M. B., 1976. Market microstructure. Journal of Financial Economics 3 (3), 257–275.

Garvey, R., Wu, F., 2010. Speed, distance, and electronic trading: New evidence on why location matters. Journal
of Financial Markets 13 (4), 367–396.

Gerig, A., 2015. High-frequency trading synchronizes prices in financial markets. Working paper. U.S. Securities
and Exchange Commission.

Glosten, L. R., 1994. Is the electronic open limit order book inevitable? Journal of Finance 49 (4), 1127–1161.

Glosten, L. R., Milgrom, P. R., 1985. Bid, ask and transaction prices in a specialist market with heterogeneously
informed traders. Journal of Financial Economics 14 (1), 71–100.

Goldstein, M., Kwan, A., Philip, R., 2016. High-frequency trading strategies. Working paper. The University of
Sydney.

Goldstein, M. A., Kumar, P., Graves, F. C., 2014. Computerized and high-frequency trading. The Financial Review
49 (2), 177–202.

Grossman, S. J., Stiglitz, J. E., 1980. On the impossibility of informationally efficient markets. The American
Economic Review 70, 393–408.

Guo, M., 2015. Smooth liquidity trading, super impatient speculator, and price spikes. Working paper. Shanghai
Advanced Institute of Finance.

Hagströmer, B., Nordén, L., 2013. The diversity of high-frequency traders. Journal of Financial Markets 16 (4),
741–770.

Hagströmer, B., Nordén, L., Zhang, D., 2014. How aggressive are high-frequency traders? The Financial Review
49 (2), 395–419.

Han, J., Khapko, M., Kyle, A. S., 2014. Liquidity with high-frequency market making. Working paper. Swedish
House of Finance.

128
Harris, F. H. d. B., 2015. The effects of algorithmic trading on security market quality. The Journal of Trading
10 (2), 41–53.

Hasbrouck, J., 1991. Measuring the information content of stock trades. The Journal of Finance 46 (1), 179–207.

Hasbrouck, J., 1999. The dynamics of discrete bid and ask quotes. Journal of Finance 54 (6), 2109–2142.

Hasbrouck, J., 2002. Stalking the efficient price in market microstructure specifications: An overview. Journal of
Financial Markets 5 (3), 329–339.

Hasbrouck, J., 2015. High frequency quoting: Short-term volatility in bids and offers. Working paper. NYC Stern
School of Business.

Hasbrouck, J., Saar, G., 2009. Technology and liquidity provision: The blurring of traditional definitions. Journal
of financial Markets 12 (2), 143–172.

Hasbrouck, J., Saar, G., 2013. Low-latency trading. Journal of Financial Markets 16 (4), 646–679.

Hasbrouck, J., Seppi, D. J., 2001. Common factors in prices, order flows, and liquidity. Journal of Financial
Economics 59 (3), 383–411.

Hendershott, T., Jones, C. M., Menkveld, A. J., 2011. Does algorithmic trading improve liquidity? Journal of
Finance 66 (1), 1–33.

Hendershott, T., Menkveld, A. J., 2014. Price pressures. Journal of Financial Economics 114 (3), 405–423.

Hendershott, T., Riordan, R., 2013. Algorithmic trading and the market for liquidity. Journal of Financial and
Quantitative Analysis 48 (04), 1001–1024.

Hirschey, N., 2016. Do high-frequency traders anticipate buying and selling pressure? Working paper. London
Business School.

Hoffmann, P., 2014. A dynamic limit order market with fast and slow traders. Journal of Financial Economics
113 (1), 156–169.

129
Holden, C. W., Subrahmanyam, A., 1992. Long-lived private information and imperfect competition. Journal of
Finance, 247–270.

Ito, T., Yamada, M., 2015. High-frequency, algorithmic spillovers between NASDAQ and Forex. Working paper.
NBER.

Jarnecic, E., Snape, M., 2014. The provision of liquidity by high-frequency participants. The Financial Review
49 (2), 371–394.

Jarrow, R. A., Protter, P., 2012. A dysfunctional role of high frequency trading in electronic markets. International
Journal of Theoretical and Applied Finance 15 (03).

Jarrow, R. A., Protter, P., 2015. Liquidity suppliers and high frequency trading. SIAM Journal on Financial
Mathematics 6 (1).

Jiang, G., Lo, I., Valente, G., 2014. High-frequency trading around macroeconomic news announcements: Evidence
from the US treasury market. Working paper. Bank of Canada.

Johnson, H., Van Ness, B. F., Van Ness, R. A., 2016. Are all odd-lots the same? Odd-lot transactions by order
submission and trader type. Working paper. The University of Mississippi.

Jones, C. M., 2013. What do we know about high-frequency trading? Working paper. Columbia University.

Jovanovic, B., Menkveld, A. J., 2016a. Dispersion and skewness of bid prices. Working paper. VU University
Amsterdam.

Jovanovic, B., Menkveld, A. J., 2016b. Middlemen in limit-order markets. Working paper. New York University.

Kalman, R. E., 1960. A new approach to linear filtering and prediction problems. Journal of Fluids Engineering
82 (1), 35–45.

Kang, J., Shin, J., 2012. The role of high frequency traders in electronic limit order markets. Working paper.
KAIST College of Business.

130
Karpoff, J. M., 1987. The relation between price changes and trading volume: A survey. Journal of Financial and
Quantitative Analysis 22 (01), 109–126.

Kauffman, R. J., Hu, Y., Ma, D., 2015. Will high-frequency trading practices transform the financial markets in
the Asia Pacific region? Financial Innovation 1 (1), 1–27.

Kirilenko, A., Kyle, A. S., Samadi, M., Tuzun, T., 2017. The flash crash: The impact of high frequency trading on
an electronic market. Journal of Finance, forthcoming.

Kirilenko, A. A., Lo, A. W., 2013. Moore’s Law versus Murphy’s Law: Algorithmic trading and its discontents.
Journal of Economic Perspectives, 51–72.

Korajczyk, R. A., Murphy, D., 2016. High frequency market making to large institutional trades. Working paper.
Kellogg School of Management.

Kyle, A. S., 1985. Continuous auctions and insider trading. Econometrica 14, 1315–1335.

Lam, E., 2015. Toronto exchange confronts controversial trading rebate system. Bloomberg Business.
Available at: http://www.bloomberg.com/news/articles/2015-05-04/toronto-exchange-confronts-
controversial-trading-rebate-system(retrieved on June 10, 2015).

Laughlin, G., Aguirre, A., Grundfest, J., 2014. Information transmission between financial markets in Chicago and
New York. The Financial Review 49 (2), 283–312.

Leach, J. C., Madhavan, A. N., 1992. Intertemporal price discovery by market makers: Active versus passive
learning. Journal of Financial Intermediation 2 (2), 207–235.

Leach, J. C., Madhavan, A. N., 1993. Price experimentation and security market structure. Review of Financial
Studies 6 (2), 375–404.

Lee, C., Ready, M. J., 1991. Inferring trade direction from intraday data. Journal of Finance 46 (2), 733–746.

Lee, S. S., Mykland, P. A., 2012. Jumps in equilibrium prices and market microstructure noise. Journal of
Econometrics 168 (2), 396–406.

131
Lepone, A., Sacco, A., 2013. The impact of message traffic regulatory restrictions on market quality: Evidence
from Chi-X Canada. Working paper. University of Sydney.

Lewis, M., 2014. Flash Boys: a Wall Street Revolt. WW Norton & Company.

Li, S., 2015. Long lived flow of private information, high frequency competition, and market efficiency. Working
paper. U.S. Securities and Exchange Commission.

Liu, H., Kaniel, R., 2006. So what orders do informed traders use? Journal of Business 79 (4).

Longstaff, F. A., 2010. The subprime credit crisis and contagion in financial markets. Journal of Financial Economics
97 (3), 436–450.

Madhavan, A., Richardson, M., Roomans, M., 1997. Why do security prices change? A transaction-level analysis
of NYSE stocks. Review of Financial Studies 10 (4), 1035–1064.

Madhavan, A. N., 2002. VWAP strategies. Institutional Investor Inc, 32–39.

Malinova, K., Park, A., Riordan, R., 2013. Do retail traders suffer from high frequency traders? Working Paper.
University of Toronto.

Martinez, V. H., Roşu, I., 2013. High frequency traders, news and volatility. Working paper. HEC Paris.

Menkveld, A. J., 2013. High frequency trading and the new market makers. Journal of Financial Markets 16 (4),
712–740.

Menkveld, A. J., 2016. The economics of high-frequency trading: Taking stock. Annual Review of Financial
Economics 8, 1–24.

Menkveld, A. J., Koopman, S. J., Lucas, A., 2007. Modeling around-the-clock price discovery for cross-listed stocks
using state space methods. Journal of Business & Economic Statistics 25 (2).

Menkveld, A. J., Zoican, M. A., 2017. Need for speed? Exchange latency and market quality. Review of Financial
Studies, forthcoming.

132
Michaels, D., 2016. IEX gains SEC approval for new stock exchange. Bloomberg Business. Available at:
http://www.bloomberg.com/news/articles/2015-05-04/toronto-exchange-confronts-controversial-
trading-rebate-system(retrieved on August 21, 2016).

Moosa, I., Ramiah, V., 2015. The profitability of high-frequency trading: Is it for real. The Handbook of High
Frequency Trading, 25–45.

O’Hara, M., 2003. Presidential address: Liquidity and price discovery. Journal of Finance 58 (4), 1335–1354.

O’Hara, M., 2015. High frequency market microstructure. Journal of Financial Economics 116 (2), 257–270.

O’Hara, M., Saar, G., Zhong, Z., 2015. Relative tick size and the trading environment. Working paper. Cornell
University.

O’Hara, M., Yao, C., Ye, M., 2014. What’s not there: Odd lots and market data. The Journal of Finance 69 (5),
2199–2236.

Putniņš, T. J., Barbara, J., 2016. Heterogeneity in the effects of algorithmic and high-frequency traders on
institutional transaction costs. Working Paper. University of Technology Sydney.

Riordan, R., Storkenmaier, A., 2012. Latency, liquidity and price discovery. Journal of Financial Markets 15 (4),
416–437.

Roşu, I., 2016a. Fast and slow informed trading. Working paper. HEC Paris.

Roşu, I., 2016b. Liquidity and information in order driven markets. Working paper. HEC Paris.

Scholtus, M., van Dijk, D., Frijns, B., 2014. Speed, algorithmic trading, and market quality around macroeconomic
news announcements. Journal of Banking and Finance 38, 89–105.

SEBI, 2016. Discussion paper on ‘strengthening of the regulatory framework for algorithmic trading & co-location’.
Available at: http://www.sebi.gov.in/cms/sebi_data/attachdocs/1470393485587.pdf (retrieved on De-
cember 15, 2016).

133
SEC, 2010. Concept release on equity market structure. Available at: http://www.sec.gov/rules/concept/
2010/34-61358.pdf (retrieved on February 22, 2014).

SEC, 2013. In the matter of Knight Capital Americas LLC respondent. Available at: https://www.sec.gov/
litigation/admin/2013/34-70694.pdf (retrieved on March 22, 2015).

SEC, 2014. Equity market structure literature review part II: High frequency trading. Available at: https://www.
sec.gov/marketstructure/research/hft_lit_review_march_2014.pdf (retrieved on March 27, 2015).

Shkilko, A., Van Ness, B., Van Ness, R., 2012. Short selling and intraday price pressures. Financial Management
41 (2), 345–370.

Skjeltorp, J. A., Sojli, E., Tham, W. W., 2015. Trading on Algos. Working paper. Central Bank of Norway.

Smidt, S., 1971. Which road to an efficient stock market: Free competition or regulated monopoly? Financial
Analysts Journal, 18–69.

Stiglitz, J. E., 2014. Tapping the brakes: Are less active markets safer and better for the economy? Working paper.
Columbia University.

Tetlock, P. C., 2007. Giving content to investor sentiment: The role of media in the stock market. Journal of
Finance 62 (3), 1139–1168.

The Committe on Financial Services, 2014. Oversight of the SEC’s agenda, operations, and FY 2015 budget request.
Available at: http://financialservices.house.gov/uploadedfiles/113-75.pdf (retrieved on March 20,
2015).

Thompson, S. B., 2011. Simple formulas for standard errors that cluster by both firm and time. Journal of Financial
Economics 99 (1), 1–10.

Tong, L., 2015. A blessing or a curse? The impact of high frequency trading on institutional investors. Working
paper. Fordham University.

Uno, J., Shibata, M., 2011. Speed of trade and liquidity. Working paper. Waseda University.

134
Van Kervel, V., Menkveld, A., 2016. High-frequency trading around large institutional orders. Working paper. VU
University Amsterdam.

Van Ness, B. F., Van Ness, R. A., Watson, E. D., 2015. Canceling liquidity. Journal of Financial Research 38 (1),
3–33.

Weller, B. M., 2016. Efficient prices at any cost: does algorithmic trading deter information acquisition? Working
Paper. Duke University.

Yang, L., Zhu, H., 2016. Back-running: Seeking and hiding fundamental information in order flows. Working Paper.
Rotman School of Management.

Ye, M., Yao, C., Gai, J., 2013. The externalities of high frequency trading. Available at SSRN 2066839.

Zhang, S., Riordan, R., 2011. Technology and market quality: the case of high frequency trading. Working paper.
ECIS 2011 Proceedings.

Zhang, S. S., 2013. Need for speed: An empirical analysis of hard and soft information in a high frequency world.
Working paper. Manchester Business School.

Zhou, H., Kalev, P. S., Lian, G., 2016. Algorithmic trading in turbulent markets. Working paper. UniSA Business
School.

135