MA Louis Mueller

University of St.
Gallen
Master in Banking and Finance
Master’s Thesis
January 2018
Implications of High-Frequency Trading on

the Quality of Electronic Markets
Author Supervisor
Louis Müller Prof. Dr. Karl Frauendorfer
Student Director ior/cf-HSG
Abstract
Technological advancements and regulatory changes facilitated the rise of high-frequency trading in
recent years. This new type of trader takes advantage of low latency technologies and sophisticated
algorithms to obtain an edge over other traders. Even though all high-frequency traders rely on speed,
their strategies are extremely heterogeneous. This thesis utilizes recent literature for the rigorous
investigation of their effects on market quality, namely price discovery and liquidity. At first sight, high-
frequency trading appears to benefit both qualities. However, after closer inspection it has deteriorating
effects on the aggregated level of information. Moreover, the liquidity provided by such traders is not
as abundant as it appears, which poses substantial systemic risks for financial markets.
CONTENTS I
Contents
1 Introduction 1
2 Properties of Algorithmic Trading 3

2.1 Agency Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.2 Proprietary High-Frequency Algorithms . . . . . . . . . . . . . . . . . . . . . . 4
2.2.1 Behavioral Characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.2.2 Heterogeneity of HFT Strategies . . . . . . . . . . . . . . . . . . . . . . 5
2.3 Identification within the Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3.1 Order Flow Proxies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.3.2 Anonymized Account-level Data . . . . . . . . . . . . . . . . . . . . . . 9
2.3.3 Pre-Classified HFT data . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3 Evolution of Financial Market Structure 11

3.1 Implementation of New Regulations . . . . . . . . . . . . . . . . . . . . . . . . 11
3.2 Market Fragmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.3 Modernization of Market Making . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.4 Reduction of Minimum Tick Size . . . . . . . . . . . . . . . . . . . . . . . . . . 15
4 Electronic Trading Platforms and HFT 17

4.1 Fee Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
4.1.1 Maker-Taker Pricing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
4.1.2 Taker-Maker Pricing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
4.2 Optimization of Latency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.2.1 Direct Data Feeds and Co-location . . . . . . . . . . . . . . . . . . . . . 22
4.2.2 Latency Arbitrage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4.3 Interdependence between New Platforms and HFT . . . . . . . . . . . . . . . . 24
5 Speed as Form of Information 25

5.1 Informed Trading and Adverse Selection . . . . . . . . . . . . . . . . . . . . . . 25
5.1.1 Effective Spread Decomposition . . . . . . . . . . . . . . . . . . . . . . . 26
5.1.2 Market Maker Profitability . . . . . . . . . . . . . . . . . . . . . . . . . 27
5.2 High-Frequency Liquidity Consumption . . . . . . . . . . . . . . . . . . . . . . 28
5.2.1 Anticipation of Order Flow . . . . . . . . . . . . . . . . . . . . . . . . . 28
5.2.2 Quote Sniping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
5.3 High-Frequency Liquidity Provision . . . . . . . . . . . . . . . . . . . . . . . . . 32
5.4 Exploitation of Minimum Tick Sizes . . . . . . . . . . . . . . . . . . . . . . . . 34
5.5 The Technological Arms Race . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
6 The Informativeness of Asset Prices 38

6.1 Price Discovery . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
6.1.1 The Vector Autoregressive Framework . . . . . . . . . . . . . . . . . . . 39
6.1.2 Variance Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
CONTENTS II
6.1.3 Information Shares . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

6.1.4 The State-Space Framework . . . . . . . . . . . . . . . . . . . . . . . . . 42
6.2 Efficiency of Prices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
6.2.1 Weak Form of Efficiency . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
6.2.2 Semi-Strong From of Efficiency . . . . . . . . . . . . . . . . . . . . . . . 45
6.3 The Dark Side of Inflated Efficiency . . . . . . . . . . . . . . . . . . . . . . . . 47
6.3.1 Social Costs and Benefits . . . . . . . . . . . . . . . . . . . . . . . . . . 47
6.3.2 Free-Riding on Private Information . . . . . . . . . . . . . . . . . . . . . 47
6.4 Short-term Volatility: Information or Noise? . . . . . . . . . . . . . . . . . . . . 49
7 Liquidity and Transaction Costs 51

7.1 Evolution in Basic Liquidity Measures . . . . . . . . . . . . . . . . . . . . . . . 51
7.1.1 Aggregated Effect of HFT . . . . . . . . . . . . . . . . . . . . . . . . . . 51
7.1.2 Conditional on Trading Strategy . . . . . . . . . . . . . . . . . . . . . . 52
7.2 Phantom Liquidity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
7.2.1 Transaction Costs for Institutional Traders . . . . . . . . . . . . . . . . 53
7.2.2 Selective Liquidity Provision . . . . . . . . . . . . . . . . . . . . . . . . 55
7.3 Systemic Risk and Flash Crashes . . . . . . . . . . . . . . . . . . . . . . . . . . 58
8 Conclusion 60
References 61
A Primer to Limit Order Markets 67
B The State-Space Framework 68
C Inherited Figures 69
D Declaration of Authorship 82
LIST OF FIGURES III
List of Figures
1 Subdivisions of Trading Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . 7
2 Empirical Indicators of AT Activity . . . . . . . . . . . . . . . . . . . . . . . . 8
3 Competition of European Trading Platforms . . . . . . . . . . . . . . . . . . . . 14
4 Average Spreads by Stock Prices . . . . . . . . . . . . . . . . . . . . . . . . . . 15
5 The Underlying Mechanics of Latency Arbitrage . . . . . . . . . . . . . . . . . 23
6 Symbiosis between HFT and ETPs . . . . . . . . . . . . . . . . . . . . . . . . . 24
7 Spread Profits and Positioning Losses of a Market Maker . . . . . . . . . . . . 27
8 Cancel-to-Trade Ratios on Various electronic trading platforms (ETPs) . . . . 32
9 Long-run Floor-Based Trading Activity . . . . . . . . . . . . . . . . . . . . . . 69
10 Value-Weighted Daily Average Effective Spread on the NYSE . . . . . . . . . . 69
11 Dislocations Between Direct Data Feeds and the SIP . . . . . . . . . . . . . . . 70
12 Cumulative HFT vs. non-HFT Net Marketable Buying . . . . . . . . . . . . . . 71
13 Order Book Dynamics of one Security Traded at Euronext Amsterdam . . . . . 71
14 HFT Trading Around Negative and Positive Macro Announcements . . . . . . 72
15 Correlations between S&P500 Futures and ETF at High Frequencies . . . . . . 73
16 Duration of Arbitrage Opportunities between S&P500 Futures and ETF . . . . 73
17 Response of Liquidity Demand and Supply to Information Shocks . . . . . . . . 74
18 Effect of Increasing Tick Size on Trading Activity . . . . . . . . . . . . . . . . . 75
19 Average Variance Ratios by High and Low Update Frequency . . . . . . . . . . 75
20 HFT and nHFT Activity During Extreme Price Movements . . . . . . . . . . . 76
21 HFT Net Flow Throughout the Lifetime of an Institutional Order . . . . . . . 77
22 Abnormal HFT Liquidity Provision to Large Institutional Buys and Sells . . . 78
23 High-Frequency Cross-Market Trading Activity . . . . . . . . . . . . . . . . . . 79
24 Relationship between HFT and Volatility . . . . . . . . . . . . . . . . . . . . . 80
25 Impact of Market Conditions on Market Maker Participation . . . . . . . . . . 81
List of Tables
1 Fee Structures of Selected ETPs Trading in Equities . . . . . . . . . . . . . . . 18
2 Example of an Agency Problem of Broker-Routed Orders . . . . . . . . . . . . 19
3 Percentage of Liquidity Provision by HFT by Size and Price . . . . . . . . . . . 35
LIST OF ACRONYMS IV
List of Acronyms
AT algorithmic trading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
ATS alternative trading system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
bp basis point . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
CAR cumulative abnormal return . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
CBOE Chicago Board Options Exchange . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
CFTC Commodity Futures Trading Commission . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
CME Chicago Mercantile Exchange . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
DMM designated market maker . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
ECN electronic communication network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
ELP endogenous liquidity provider . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
EMH efficient market hypothesis. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .43
EPM extreme price movement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
ETF exchange traded fund . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
ETP electronic trading platform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
EU European Union . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
FOREX foreign exchange. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .45
HFB high-frequency bandit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
HFM high-frequency market maker . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
HFT high-frequency trading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
HHI Herfindahl-Hirschman Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
IRF impulse response function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
IS implementation shortfall . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
IT institutional trader . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
LSE London Stock Exchange . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
MIDAS Market Information Data Analytics System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
MiFID Markets in Financial Instruments Directive . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
MTF multilateral trading facility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
NBBO national best bid or offer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
nHFT non high-frequency trading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
NYSE New York Stock Exchange . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
OPR order orotection rule. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .11
RegNMS Regulation National Market System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
RegATS Regulation Alternative Trading Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
SEC Securities and Exchange Commission . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
SIP Security Information Processor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
TAQ trade and quote database . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
TSX Toronto Stock Exchange . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
VAR vector autoregressive . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
VMA vector moving average . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
VR variance ratio . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
1 INTRODUCTION 1
1 Introduction
”In a high-speed, co-located world, being informed means seeing and acting on
market prices sooner than competitors. Today, it pays to be faster than the average
bear, not smarter. To be uninformed is to be slow. These uninformed traders
face a fundamental uncertainty: they may not be able to observe the market price
at which their trades will be executed. This is like driving through the rear-view
mirror, stock-picking based on yesterday’s prices.”
(Haldane, 2011, p. 4)
Financial markets have undergone substantial changes within the last decade. Not only
have there been major technological advancements and new regulations, but also the emergence
of a new type of market participant: The high-frequency trader. This new trader caused
substantial disruptions and received plenty of (mostly negative) public attention. Owing to
their enormous speed, high-frequency trading (HFT) is often accused of enable an unfair
advantage in comparison with normal traders. It is suspected of being able to predict the
actions of others. This would allow high-frequency traders to take react on incoming orders
by trading ahead of them. Hence, there are numerous allegations of financial markets having
become an uneven playing field, i.e. the collection a quasi speed tax paid by slower traders
(Li, 2017, pp. 15–16).
As a result, academia has taken proactive interest in the topic. Researchers have released
countless theoretical models as well as empirical studies in order to better comprehend the
implications on financial markets. Unfortunately, HFT is an extremely secretive industry and
trading algorithms are heavily guarded. In his recent best-selling book, Lewis (2015, p. 110)
claims that it takes five badge swipes in order to access the trading desks of Citadel LLC, a
major HFT firm. For comparison, only two badge swipes are required to enter the areas of top
secret security clearance at the US Pentagon. This sublime level of secrecy forces researchers
to solely rely on anecdotal evidence. Thus, it can only be speculated about the true nature
of HFT algorithms. Moreover, the choice of datasets is strongly limited and only available for
proprietary academic use. Nevertheless, this has not been discouraging researches to devote
enormous efforts to the HFT phenomenon.
Another complication of this subject is the tremendous complexity of the topic. For this
reason, academic literature is in outright disagreement with respect to some issues. The first
studies conducted on HFT between 2009–2013 were clearly in favor of the new market partic-
ipants, as they primarily focused on the immediate and easily observable effects. Undeniably,
HFT exploits inefficiencies which is directly beneficial. However, from 2013 onward academia
became much more agnostic. Researchers commenced the investigation of alternative channels
through which HFT exerted influence on markets. These studies revealed certain indirect but
nevertheless severe adverse effects, way beyond simple increase in efficiency. This leads to the
two main premises of this thesis. First, HFT does not affect markets in a straightforward
fashion, but rather has countless nuances. Second, HFT literature is in constant evolution by
steadily revealing unseen evidence.
1 INTRODUCTION 2
For these two reasons, this thesis aims to provide a golden thread through the existing
literature on HFT. It strives to provide profound insights into all spheres of the new technology.
Thus, this thesis will only provide the core findings of recent academic studies and present
them in a qualitative manner. In some exceptions, however, the exact empirical methodologies
will be elaborated briefly, more specifically the measures used to approximate HFT activity. To
guarantee the understanding of the concepts elaborated in this thesis, the reader is required to
be familiar with the basic mechanics of limit order markets. If this is not the case, appendix A
will provide a brief introduction and clarification of the most important notions, which serves
as a solid foundation.
To ensure a logical structure throughout this thesis, two main research questions are meant
to be answered at the end. First, it should become clear how HFT works and how it fits into
the contemporary structure of financial markets. Second, this thesis serves as an investigation
of the nuanced effects of HFT on market quality. In accordance with O’Hara (2003, p. 1338),
liquidity and price efficiency are taken as the two main measures of market quality. Even
though both these concepts seem to be distinctive at first, they are ultimately closely related.
Due to the constant emergence of new findings, only recent pieces of literature have been
chosen. Furthermore, the selection of resources is comprised of published articles and working
papers, to reflect the current level of knowledge. To better explain the findings in the literature,
appendix C comprises a collection of important figures by other authors. Since HFT is a topic
of global interest, the studies discussed here use data from trading platforms all over the
world. However, the focus mostly lies on either the US, Canada or Europe. Thus, the main
contribution of this thesis consists is the delivery of a comprehensible as well as comprehensive
primer to the topic of HFT. In contrast to most other existing literature surveys, it does not
only cover a single aspect of HFT but instead its overall effects on financial markets.
As mentioned earlier, the ultimate goal of this thesis is the presentation of a golden thread.
Therefore, it will be contain numerous references and is structured as follows. To begin with,
section 2 will provide an overview about characteristics and strategies of algorithmic trading
and the definition of the term HFT. Next, section 3 reports recent regulatory and technological
developments in financial markets and section 4 shows how HFT seamlessly fits into this new
ecosystem. Thereafter, section 5 explains how HFT uses speed to obtain superior information
over other traders. Subsequently, the focus will shift onto the impact on market quality. More
specifically, section 6 investigates how HFT affects price efficiency and the level of information
in asset prices. Similarly, section 7 analyzes how simple liquidity measures have evolved and
what systemic risks are subsequently provoked by HFT. Finally, section 8 rounds everything
off with a rough recapitulation of all the findings throughout this thesis.
2 PROPERTIES OF ALGORITHMIC TRADING 3
2 Properties of Algorithmic Trading

The term of algorithmic trading (AT) has been abundantly used in recent market microstruc-
ture literature. According to Foucault (2012, p. 6) it can be defined as the usage of fully
automated computer programs designed for data analysis, decision-making and execution. As
a result, the entire trading process is carried out automatically without any human interfer-
ence. Thereby, algorithms are continuously monitoring the market for irregular events (e.g.
sudden arbitrage opportunities) to react by almost immediately with the submission of orders.
Hence, AT is clearly superior to its human counterparts with respect to efficiency. This in-
equality can be attributed to two main advantages of algorithms. First, AT has the ability
to simultaneously assimilate large numbers of signals. The inability to perfectly observe the
markets at all times caused substantial search costs. These monitoring frictions have been
significantly reduced due to technological advances in AT (Hendershott & Riordan, 2013).
With such technology, traders can process vast amounts of information, which reduces the un-
certainty about the current state of the market. Second, improvement of processing efficiency
allowed for the trading process to be extremely accelerated. Relative speeds between traders
have always been important, but suddenly microseconds have become the crucial measure for
trading profitably (SEC, 2010, p. 41).
As a consequence of those two developments, latency has become an important concept,
even though not all ATs are primarily competing for speed. Hasbrouck and Saar (2013, p. 647)
define this concept as the ”time it takes to learn about an event [...], generate a response and
have the exchange act on the response”. In other words, latency is the full reaction time
of a trader to an incoming signal. However, it would be wrong to assume AT to consist of
a homogeneous set of strategies. Hasbrouck and Saar (2013, pp. 648–649) further make a
fundamental distinction between two types of automated traders: Agency- and high-frequency
algorithms. This categorization is widely accepted and has become standard in literature. In
the following, the characteristics of these subsets will be elaborated in more detail.
2.1 Agency Algorithms

The first subset of AT is designed to optimize trade execution. Agency algorithms are normally
used by institutional investors and brokers to hide their true trading intentions and to optimize
the timing of executions. Thereby, the application of such algorithms can create larger profits
by minimizing execution cost. These costs can be induced by multiple factors, for instance
unfavorable quotes and fees, detection by other traders or a lack of market liquidity. As, this
shortfall quickly increases with size of the order, agency algorithms are very popular among
institutions such as pension funds and other large asset managers. (Hasbrouck & Saar, 2013,
p. 652)
One omnipresent strategy among agency algorithms is the splitting of orders, where large
trades are split into so-called child orders (Menkveld, 2016, p. 9). This procedure aims to
hide its true intentions by scattering orders to various exchanges and over a longer period
of time. To other traders, these messages appear to be random noise indicating order flow
from uninformed retail traders. At best, the algorithm gradually enters the desired position
without any other traders noticing. However, agency algorithms are not always being unde-
tected. Other sophisticated market participants might be able to recognize certain patterns
(e.g. autocorrelation), since child orders are not perfectly independent (O’Hara, 2015, p. 264).
In essence, agency algorithms are not very sophisticated and are thus sometimes referred to
as ”slice-and-dice algorithms” (Zhang & Riordan, 2011, p. 4).
Thus, agency algorithms are primarily designed to project a maximum level of randomness
on the surface. This is achieved through a combination of limit- and market orders on multiple
exchanges at once. However, latency does not seem to be crucial. Hasbrouck and Saar (2013,
p. 652) believe that agency algorithms show patterns of clock-time periodicity. Their findings
suggest that most messages persistently arrive during the first 30 milliseconds of each second,
contradicting the uniformity of arrival rate. Even though this is seemingly fast, other traders
operate on significantly lower timescales. Hence, agency algorithms are generally not among
the fastest ATs. The authors further claim that the strategies employed by most agency
algorithms are rather simplistic and do not require minuscule latencies. This renders expensive
investments in faster technologies unnecessary at some point.
2.2 Proprietary High-Frequency Algorithms

The second subset of AT consists of proprietary algorithms and is referred to as HFT. In
contrast to agency algorithm, HFTs pursue proprietary strategies and thus trade with their own
capital. The set of firms operating such algorithms is quite diverse. It consists of proprietary
broker-dealer desks (e.g. Goldman Sachs or Morgan Stanley), hedge funds (e.g. Citadel LLC)
and proprietary trading firms (e.g. GETCO, Optiver or Tradebot) (Goldstein, Kumar, &
Graves, 2014, p. 183). The commonality of all these institutions is extreme technological
sophistication and speed. Another key characteristic of HFT is the lack of intrinsic motivation
to trade. The algorithms are looking for very short-lived opportunities to make profits (Budish,
Cramton, & Shim, 2015, p. 1578) and returns are not generated by long-term prospects or
strategic considerations. Neither fate nor fundamentals of the underlying company are of
any importance to HFTs (Foucault, 2012, p. 21). As a result, their primary objective is the
exploitation of irregularities and inefficiencies in order flow and therefore the generation of
profits from the trading process itself.
2.2.1 Behavioral Characteristics
Akin to agency algorithms, HFTs follow distinct patterns with their strategies. These behav-
iors serve as critical assumptions for the identification of HFT activity in empirical analyses.
However, in the abundance of literature, there have been disputes about the correct definition
of HFT. Finally, the Securities and Exchange Commission (SEC) provided the first quasi
official definition, which is widely accepted among researchers and practitioners. In particular,
the SEC (2010, p. 45) determined five main HFT characteristics.
[1] HFTs trade at extraordinarily high speeds and use sophisticated algorithms for the entire
trading process. Moreover, trade volumes are enormous, whereas profits per share traded
are almost negligible.
[2] The reduction of any form of latency is paramount to HFT. Co-location and raw data
feeds are two common tools to achieve this goal (see section 4.2).
[3] HFTs hold their positions only during very short time-frames before liquidation. This
mitigates inventory risk and reduces exposure to price changes. HFTs might also reverse
their positions multiple times per day by repeatedly crossing the zero-inventory.
[4] HFT algorithms very quickly submit vast amounts of orders which may be canceled again
shortly afterwards.
[5] At the end of each day, HFTs terminate their trading with flat positions. No inventory
or unhedged positions are retained overnight.
Even though the characteristics are clearly defined, they should not be interpreted as a list
of cumulative criteria for HFT classification. It is rather a framework to better understand
the abstract dynamics and to find a common denominator for further research in this field.
Nevertheless, most studies following the publication date of the SEC (2010) have adopted
these assumptions.
Interestingly, the characteristics above allow for a more accurate distinction between HFT
and agency algorithms. Particularly items [3] and [5] are not met by the latter. Agency
algorithms are programmed to execute a predefined trading goal by gradually building the
desired position. This may take several hours to complete. HFTs on the other hand will quickly
reduce their positions after a large trade and, more importantly, close out flat at the end of the
trading day. Similarly, item [2] is rather unlikely to apply to agency algorithms, but it cannot
be ruled out entirely. A few of these might still desire lower latencies to decrease execution cost
even further. Finally, items [1] and [4] are equally applicable to agency algorithms, as discussed
in section 2.1. Alike HFT, agency algorithms are fully automated and produce countless
messages during the trading process. Hence, it may not be sufficient to empirically distinguish
different ATs by only relying on these two criteria, which demonstrates the commonalities
among both types of AT.
Nevertheless, there is more diversity among proprietary strategies, which makes it more
complicated to form a clear opinion (Hasbrouck & Saar, 2013, p. 654). The next subsection
will thus thoroughly elaborate the most common strategies utilized in HFT.
2.2.2 Heterogeneity of HFT Strategies
Even though HFT algorithms have already been quite carefully selected using items [1] to [5]
above, they might be fragmented further into diverse subsets of strategies. In spite of drasti-
cally decreasing latencies and technological advances, none of the strategies are entirely new to
financial markets. Instead, merely their efficiency has drastically improved (Biais & Foucault,
2014, p. 6). Some were already applied by human traders on physical floors before automation
of the trading process. Therefore, the key difference nowadays is latency.
To further investigate the exact implications of HFT on market quality, it is necessary
to refine its strategies more precisely. One way of classification on a higher level can be
achieved by inspection of the preferred order type. For this purpose, Hagströmer and Nordén
(2013, p. 753) proposed the distinction between market-making HFT and opportunistic HFT
strategies, having almost opposing implications for other traders. However, the authors make
this distinction subject to mixed strategies, as HFTs are likely to make extensive use of both
order types at the same time.
Market making strategies. The first subset mainly takes advantage of non-immediately
marketable limit orders (henceforth simply referred to as limit orders). This order type offers
liquidity to others without immediate execution (Biais & Foucault, 2014, p. 5). This is why
market making is often referred to as being passive. It fulfills an essential function for liquidity
provision with positive effect on others. Market makers usually post limit orders on both
sides of the book and earn the spread between bid and ask prices on every round-trip trade
(Jones, 2013, p. 6). This indicates a lack of intrinsic motivation to trade as both buy and sell
orders can be executed profitably. In addition to the spread, market-making can also generate
revenue from liquidity rebates on most trading platforms (Goldstein et al., 2014, p. 185). This
phenomenon will be discussed in section 4.1 with subject to HFT technology.
Opportunistic strategies. Compared to market-making, opportunistic traders show oppo-

site behavior in several ways. They utilize both directly marketable limit orders and market
orders. As there is no qualitative difference, they will henceforth be referred to as market
orders. This type of order takes liquidity from the market through execution against a stand-
ing limit order and are thus often called aggressive. As the trading process is a zero-sum
game, opportunistic traders pay the spread to market makers for each round-trip trade. This
can be seen as compensation for risk of sudden price movements exceeding the half-spread,
which is defined as the difference between bid- respectively ask-price and the midpoint. In
contrast to straightforward market-making, opportunistic strategies are more diverse. The
SEC (2010, pp. 51–57) defines three subgroups: (1) arbitrage-, (2) structural- and (3) direc-
tional strategies. An arbitrage strategy includes constant monitoring of the market for sudden
deviations from parity of related assets (i.e. arbitrage opportunities), such as a price discrep-
ancy of an exchange traded fund (ETF) and its underlying securities. Since these events are
usually rare and disappear quickly, HFTs rely on their speed advantage to exploit the risk-less
profit in time (Biais & Foucault, 2014, p. 7). Moreover, these strategies are computationally
intense. Second, structural strategies take advantage of the market design, particularly the
recent fragmentation of trading venues (see section 3.2). In this setting, speed is an advan-
tageous feature, as fast traders are able see the market ahead of others (Zhang & Riordan,
2011, p. 4), as discussed in section 4.2.2. Finally, directional strategies are speculative and at-
tempt to predict price changes, whereas the time-frame may only be a few milliseconds. Some
HFT algorithms attempt to recognize patterns in the past order flow and take guesses about
incoming orders. With sufficient accuracy, this technique can be used to build the necessary
positions in time. This phenomenon is known as Order anticipation and will be discussed in
section 5.2.1. Other directional algorithms continuously monitor news outlets and trade upon
announcements. Low latencies allow for arrival before other traders, who will adjust prices
into the favorable direction (Jones, 2013, p. 8).
Manipulative strategies. Despite not being severely addressed by the SEC (2010), there
are concerns about a subset of HFT actively manipulating prices. These strategies are usually
a combination of limit- and market orders and can thus not be clearly attributed to either
strategy above. A thorough description of some examples can be found in Biais and Foucault
(2014, 8–9). On the one hand, momentum ignition and spoofing are applied to intentionally
confuse other market participants and cause naı̈ve or even hysteric buy/sell reactions. On
the other hand, quote stuffing consists of the excessive order submission with the objective
to slow down other algorithms or even the trading venue. Unfortunately, there is a lack of
recent empirical studies addressing such manipulations. This might have two reasons: Either
manipulative strategies are a negligible phenomenon and thus require no investigation, or there
has not yet been found a way to quantify such behavior in existing datasets. Exceptions are
Egginton, Van Ness, and Van Ness (2016); Gai, Yao, and Ye (2013) who address quote stuffing.
Nevertheless, manipulative strategies will not be addressed further in this thesis.
AT
Proprietary HFT Agency algorithms
Passive strategies Aggressive strategies
Market-making Manipulation Arbitrage Structural Directional
Momentum ignition Spoofing Quote stuffing News trading Order anticipation
Figure 1: Subdivisions of trading algorithms. AT can be divided into a complicated set of groups. To avoid
confusion, this visualization provides more clarity. It becomes apparent that especially HFT strategies are
very diverse. (Own presentation)
The diversity of algorithmic trading strategies must always be kept in mind when studying
the effects on market quality. Different strategies do not necessarily affect other traders in the
same way. To grasp all subgroups of algorithmic trading and put them into perspective, fig. 1
is supposed to be an assistance for a better understanding.
2.3 Identification within the Data

As indicated in the Introduction to this thesis, HFT is an extremely secretive industry. Algo-
rithms are heavily protected by the owners and will not be disclosed under any circumstances.
Moreover, the available data is limited, as not many exchanges are willing to release their tapes
including trader identities for confidentiality reasons. To counter this problem, empirical re-
searchers are forced to find creative solutions. Unfortunately, not all of these methodologies
are equally meaningful. There are qualitative differences in how HFT activity can be identi-
fied. The accuracy of metrics used is heavily depending on the type of data. The following
subsections gives a general overview over three different groups of techniques used for the de-
tection of HFT and AT as well as the respective underlying data (based on Biais & Foucault,
2014, p. 9).
2.3.1 Order Flow Proxies
The first set of empirical studies utilizes publicly available message data. Importantly, these
datasets do not reveal trader identities, but just aggregated order flow. Based on the assump-
tions made about HFT characteristics in section 2.2.1, these techniques identify proxies for
the aggregated AT activity. This approach has several advantages, for instance the abundance
of order-level data. Moreover, these measures are usually straightforward to calculate and
the results easily replicable. However, market-wide proxies can be extremely inaccurate for
HFT detection. As mentioned above, HFT is only a subsample of AT which must be isolated.
However, it is not possible to extract strategies of any given trader using this technique due to
anonymity. It may be possible to detect AT activity on the basis of quick message patterns,
but the exclusion of agency algorithms is unfeasible with absolute confidence. Therefore, such
proxies can only be used to make statements about AT in aggregation, not HFT.
14% 35
Odd-lot volume [%] Cancel-to-trade ratio
12% Trade-to-order volume [%] 30
10% 25
8% 20
6% 15
4% 10
2% 5
0% 0
2012 2013 2014 2015 2016 2017 2018
Figure 2: Empirical indicators of AT activity. The plot shows percentage of odd-lot volume compared to
total volume as well as total traded volume divided by total order volume. The number of cancellations in
relation to the number of trades is shown on the right axis. All of these measures should theoretically be
increasing in HFT activity. The plotted lines show the weekly average over each measure and registered
exchanges. These venues include NYSE Arca, Bats-Y, Bats-Z, Boston Exchange, Edge-A, Edge-X and
NASDAQ. The data are provided by MIDAS. (Own presentation)
The recent evolution of three frequently used measures are reported in fig. 2. The data
was taken from the Market Information Data Analytics System (MIDAS)1 , a comprehensive
database operated by the SEC for HFT regulation purposes. Figure 2 reports the fraction of
odd-lot volume, trading volume in relation to order volume and the number of cancellations
divided by executions. Theoretically, all of these measures should move in accordance with AT
1
Retrieved from https://www.sec.gov/marketstructure/downloads.html (accessed November 14, 2017)
activity. Practically, only odd-lot volume has been increasing in recent years. This might imply
augmented AT activity, as these traders tend to use smaller average order sizes (Litzenberger,
Castura, & Gorelick, 2012, p. 77).
One of the most important contributions to literature using this methodology was made
by Hasbrouck and Saar (2013, p. 658). They attempted to measure strategic runs which
are linked order submissions, cancellations or executions, likely to be caused by a common
trading strategy. Especially important were submissions and subsequent cancellations within
the same millisecond. This also includes split child orders used by agency algorithms. Thus
there is no clear isolation of HFT. Another example was presented by Hendershott, Jones,
and Menkveld (2011, p. 6), who proxy AT activity with the amount of incoming messages per
minute. However, they disregard the fact that there are times of increased human activity, for
instance after news announcements. If algorithms might chose to retreat during these periods,
the can would be strongly biased.
2.3.2 Anonymized Account-level Data
In contrast to the order data presented in section 2.3.1, this category contains a unique trader
identity for each participant. Nevertheless, these identities are just unique numbers and thus
guarantee full anonymity. The dataset provides information about the accounts involved in
each trade. This allows for reverse-engineering of entire trading strategies and even limited
replication of inventories. Using these tools, traders can be assessed individually based on
their behavior with respect to the characteristics introduced in section 2.2.1. Even though
this methodology allows for more exact classification of ATs, it is also just a proxy as the
underlying trader identities remain confidential.
One of the most important examples of this category is the E-mini S&P500 futures market
(E-mini hereafter) dataset, provided by the Commodity Futures Trading Commission (CFTC).
As HFT activity is not just limited to equity markets, index futures are a great opportunity for
either arbitrage or structural strategies. Moreover, the E-mini has one substantial advantage
over equities in terms of data analysis: its trading is fully centralized on the Chicago Mercantile
Exchange (CME) (SEC, 2014, p. 15). The E-mini dataset thus contains all trading activity
in these index futures, allowing for a complete replication of inventories and strategies. In
contrast, equities are simultaneously traded on numerous platforms in a highly fragmented
market (see section 3.2). The data of one equity exchange therefore only covers a fraction of
the entire trading activity in one security. This prevents the accurate replication of strategies.
The first to use the E-mini dataset were Kirilenko, Kyle, Samadi, and Tuzun (2017),
who analyzed HFT activity during the Flash Crash in 2010. They first defined intraday
intermediaries as accounts who build large directional positions and show very high turnover,
however, this is done without regard to the order type. Among those, the 16 most actively
trading accounts are labeled HFTs (p. 976–977). Further examples besides the E-mini dataset
include Hirschey (2017, p. 12), who obtained a set from NASDAQ in the US identifying HFT
together with the underlying trader identification. Breckenfelder (2013, p. 10) and Hagströmer
and Nordén (2013, p. 746) both use millisecond data from the NASDAQ OMX index, consisting
of the 30 largest Swedish stocks. The dataset only includes activity in the Stockholm stock
exchange, but also individual trader identities. Most of these studies are still forced to make
classifying assumptions about different traders based on their activity. Obviously, there are
also other unmentioned studies discussed later in this thesis, using similar datasets from other
sources.
2.3.3 Pre-Classified HFT data
The third and final type of data already includes an indication of HFT activity from scratch.
Therefore, approximation based on certain behaviors becomes unnecessary. One of the few
examples in this category is the NASDAQ HFT dataset, which is exclusively provided to
academics. Nevertheless, it is the most frequently analyzed sample in academic literature
since its first introduction by Brogaard (2010). The dataset is stamped to the millisecond and
includes indicators of HFT activity for all executions on the NASDAQ exchange between 2008
and 2009. Quote data has only become available later on. The dataset gives information on
the type of liquidity supplier and demander using one of the following four codes: NN, NH,
HN and HH (Carrion, 2013, p. 685). As an example, NH implies the execution of an HFT
market order against a standing non high-frequency trading (nHFT) limit order. In this case,
HFT is taking liquidity. The final sample provides the activity of 26 HFT firms with 40 stocks
in each small-cap, mid-cap and large-cap, resulting in a total of 120 equities.
The HFT classification has been performed by the NASDAQ, based partly on information
not included in the final dataset, including soft factors. During the process, five criteria were
mainly decisive to single out HFTs (Brogaard, 2010, p. 8). First, the trader is engaging in
proprietary trading and does not serve customers. Second, sophisticated technologies and
algorithms are used in order to reduce latency. Third, the traders engage in sponsored access
to NASDAQ, allowing them co-location services and large-volume rebates. Fourth, inventories
are constantly oscillating between long and short positions. Finally, limit orders have a very
short life-span implying high order-to-trade ratios. These criteria are almost congruent with
the HFT characteristics introduced by the SEC (2010) in section 2.2.1.
Despite the HFT classification, there are also some caveats to take into consideration with
this dataset (SEC, 2014, p. 13). To begin with, NASDAQ aggregated all information into
the four codes introduced above. During this process, identification of any single trader was
rendered impossible. Thus, statements about isolated HFTs and their strategies are unfeasible.
Moreover, only HFT firms with exclusive engagement in proprietary trading were taken into
account. Large integrated broker-dealer firms were strictly excluded despite potential use of
HFT technology. Therefore, substantial HFT volume might have been overlooked, as large
banks are also suspected to use such technologies. However, the classified HFTs are still
involved in 68.5% of total volume and 74% of all trades on a daily average according to
(Brogaard, 2010, p. 57).
3 EVOLUTION OF FINANCIAL MARKET STRUCTURE 11
3 Evolution of Financial Market Structure

The substantial increase of HFT activity throughout the last decade did not appear out of thin
air. Equity markets around the world were undergoing fundamental changes in their market
structures. This lead to an increased incentive for the application of fast trading technologies.
This section will provide an overview over the relevant developments in regulations and market
structure on a macro- as well as micro level.
3.1 Implementation of New Regulations

In the early 2000s, the SEC was first to realize the necessity of a shift in regulations. As stated
in SEC (2010, p. 11), the consolidation of order flow in one single trading platform enhances a
monopoly structure. This may have detrimental effects on competition and thus the conditions
of all market participants. Beforehand, it was not allowed to trade on venues not officially
classified as exchanges by the SEC. This barrier to entry discouraged the establishment of
new platforms, leaving the entire market to the unchallenged incumbents.
Even though competition among trading platforms is a desirable feature, the regulators
were aware of the fact that it will complicate markets on another level. The lower the barriers
of entry for new platforms, the more fragmented financial market will become, as the entire
volume will be spread over more venues. This leads to an important trade-off: On the one
hand, competition will force established exchanges to become more attractive to traders. On
the other hand, the resulting market fragmentation will entail adverse effects (SEC, 2013,
p. 11). Such disadvantages include increasing complexity of executions.
In the US, it was prohibited for non-exchange entities (so-called alternative trading system
(ATS)) to autonomously match orders without going through registered exchanges. ATS is a
broad term and includes crossing networks, dark pools and broker-dealers. In 2000, Regulation
Alternative Trading Systems (RegATS) represented the first step into the direction of venue
competition (O’Hara, 2015, p. 258). It allowed other trading systems to match incoming buy
and sell orders on their own independent platforms. However, this was apparently not enough
to arouse competition. Most of the order flow was still routed to the large incumbents, despite
potentially worse quotes.
In order to definitely level the playing field, Regulation National Market System (RegNMS)
was approved by the SEC in 2005 (Litzenberger et al., 2012, p. 64). The centerpiece of this
policy change was a revised order orotection rule (OPR), which assigns execution priority to
the best standing limit order nationwide (SEC, 2005, pp. 34–35). This forces the steady obser-
vation of all available quotes in the US in order to determine the current national best bid or
offer (NBBO). This data is continuously collected and published by the Security Information
Processor (SIP). According to RegNMS, each trading platform must either execute an incom-
ing market order at the current NBBO or reroute the order directly to the respective venue.
This policy aims to minimize trade-throughs, i.e. market orders executed against unfavorable
quotes compared to other platforms. The motivation behind the OPR comes from a desire to
enhance liquidity. In that sense, trade-throughs are discouraging liquidity providers to post
limit orders at competitive prices. Without any certainty of execution, liquidity providers have
no incentive to post better quotes (Foucault & Menkveld, 2008, p. 120). As a consequence of
this rule, platforms are competing for the NBBO at all times, to attract volume and profit
from the resulting execution fees. Importantly, the OPR only considers quoted prices and not
the final transaction costs. Thus, the rule is ignoring access fees to trading platforms (see
section 4.1), which can be larger than the quoted spread itself (Angel, Harris, & Spatt, 2011,
p. 43). As a result, a trader might not actually achieve optimal execution, despite trading at
the NBBO.
At roughly the same time, the European Union (EU) implemented Markets in Financial
Instruments Directive (MiFID) for similar purposes, namely the harmonization of financial
markets. Comparable to the US prior to RegATS, the EU had a concentration rule, which
enabled its member states to execute all trading only on established exchanges. MiFID abol-
ished this rule and gave rise to countless new ATSs who started challenging the incumbents for
volume. However, MiFID does not prohibit trade-throughs between platforms like RegNMS
(Kohler & von Wyss, 2012, p. 3). Hence, European markets are likely to be less consolidated
with lower levels of competition, compared to the US. Moreover, smart order-routing systems,
which compare quotes on fragmented exchanges and execute at the best conditions and are
thus a form of AT, are more important in Europe due to the lack of a centralized NBBO
(Zhang & Riordan, 2011, p. 6; Foucault & Menkveld, 2008, p. 120).
3.2 Market Fragmentation

In the aftermath of the regulatory changes discussed above, the entrance rate of new plat-
forms into financial markets drastically increased. Nowadays, there are 212 registered equity
exchanges and 843 ATS legally operating in US equity markets, according to the SEC. This is
a substantial increase compared to the 11 equity exchanges and 50 ATS reported by O’Hara
(2015, p 258).
Shortly after the implementation of RegATS in 2000, one particular type of platform
emerged into financial markets, acting as an electronic crossing-network. Those platforms
are electronically matching third-party buy and sell orders and thus work similarly to con-
ventional exchanges (Angel et al., 2011, p. 28). However, regulations are less restrictive, as
these platforms are only acting as intermediaries for equities listed elsewhere, without having
the authorization to list new stocks themselves. Furthermore, quotes are displayed publicly
as opposed to dark pools. Hence, this new type of platform became the main competitor
to conventional exchanges owing to the changing regulatory environment. Said platforms ex-
ist in both the US and European markets, but under different names. In the US, the SEC
(2010, p. 18) defined them as electronic communication networks (ECNs), whereas they are
referred to as multilateral trading facilitys (MTFs) in Europe (EU, 2004, p. 10). It is not
uncommon for those venues to change their regulatory status after having reached a certain
reputation. Several ECNs have eventually become registered exchanges, for example BATS
(Bats-Y and Bats-Z), DirectEdge (EDGA and EDGX) and NYSE Arca (formerly known as
2
Retrieved from https://www.sec.gov/fast-answers/divisionsmarketregmrexchangesshtml.html (accessed
November 13, 2017)
3
Retrieved from https://www.sec.gov/foia/docs/atslist.htm (accessed November 13, 2017)
ArcaEx). Since both exchanges and ECNs/MTFs are similar in their operations, the aggregate
of those venues will henceforth be referred to as electronic trading platforms (ETPs). This
term therefore comprises all entirely electronic and lit venues.
These recent developments are not invariably beneficial to market participants, but also
came with certain complications. RegNMS made the concept of a centralized market obsolete,
allowing all ATS to trade independently (O’Hara, 2015, p. 258). Obviously, the simultaneous
comparison of quotes among different platforms increases search cost and leads to a trade-off:
Traders must decide between lower search costs or worse execution quality (Foucault, 2012,
p. 16). Meanwhile, advanced technology allows for the minimization of search cost leading to
faster and better trade executions (Menkveld, 2013, p. 717). This is the reason why algorithms
have become a must-have in fragmented markets.
In order to determine the level of concentration in a market, the Herfindahl-Hirschman
Index (HHI) is commonly used as measure of choice. It is defined as
N
X
HHI = s2i
i=1
, where s2i represents the market share of trading platform i and N is the number or total
venues. Since the individual market shares are squared, high fragmentation of volume entails
a lower HHI. In contrast, a high HHI implies the existence of fewer platforms and thus more
consolidation.
This concept can be applied to European equity markets on the basis of data provided
by the Chicago Board Options Exchange (CBOE)4 . It consists of the daily traded dollar-
volume of 37 European trading platforms in 15 countries, including registered exchanges as
well as other ATSs. In a first step, the daily volumes were summed up by week to achieve
more robustness. In a second step, the HHI was calculated based on the aggregated weekly
volumes. Figure 3 plots the resulting HHI in European equity markets starting in 2012. The
graph shows a continuous downward trend, from 0.11 to 0.09 within 6 years. This development
implies growingly fierce competition among platforms as the total trading activity is spread
out more evenly. Hence, the both competition and fragmentation are closely related (SEC,
2013, p. 8). The concept of platform competition will be elaborated further in section 4.
According to (Foucault, 2012, p. 14), AT is the natural consequence of such fragmentation
in security markets. Every market participant has an incentive to automate trading decisions
for different reasons. First, as most platforms only operate electronically, it is easier to partic-
ipate through a computer in a network of computers. Moreover, machines are able to process
incoming data far more efficiently than humans. Second, in order to achieve the optimal exe-
cution, brokers may use algorithmic smart routing technologies. Finally, opportunistic traders
may use HFT technologies to exploit frictions cause by market fragmentation using arbitrage-
or structural strategies (as discussed in section 2.2.2). All of these arguments are based on
one key characteristic of information technology, namely the reduction of overall search cost
in fragmented markets (Haferkorn, 2017, p. 291).
4
Retrieved from https://markets.cboe.com/europe/equities/market share/market/ (accessed on October
28, 2017)
0.12
Herfindahl-Hirschman Index
0.11
0.10
0.09
0.08
2012 2013 2014 2015 2016 2017 2018
Figure
PN 3: Competition of European trading platforms. This plot reports the weekly HHI, which is defined
as i=1 s2i , in European equity markets. A low HHI implies that there are no single exchanges dominating
the market, but rather fierce competition. Daily data was provided by the CBOE for 37 European trading
platforms. (Own presentation)
3.3 Modernization of Market Making

The concept of market making has always been essential to liquidity provision and thus the con-
tribution to better market quality. However, the operations of market makers have significantly
changed together with automation and technological advancements. Among other things, non-
floor trading became much more attractive compared to floor trading due to anonymity, up-
to-date information and speed. This is why on the New York Stock Exchange (NYSE) both
specialists (acting as market makers) and floor brokers continuously decreased their on-floor
exposure and started trading electronically (Hendershott & Moulton, 2011, p 573). Figure 9
graphically shows the decline in floor-trading. As a result, the process has become fully auto-
mated and traditional market makers have been partly replaced by HFTs (SEC, 2010, p. 49;
O’Hara, Saar, & Zhong, 2015, p. 8).
In addition to technology, the regulatory status of market makers also changed. Exchanges
usually offer firms the option to register as designated market makers (DMMs). This special
DMM-status is associated with a number of obligations, such as maintenance of two-sided
markets, stock-specific maximum spreads, moderation of volatility and guaranteed execution
up to a predefined amount of stocks. In other words, DMMs must be present at both sides of
the order book, regardless of market conditions. In return, the special status is compensated
with certain privileges which vary between exchanges, but are unaccessible for ordinary traders.
Such prerogatives include monetary compensation (Euronext Paris), privileged information on
order flow (NYSE) or more favorable trading fees and higher priority at the best quotes and
over odd-lot orders (TSX). (Anand & Venkataraman, 2016, pp.329–330).
An additional type of market maker is the endogenous liquidity provider (ELP). In contrast
DMMs, ELPs do not have special status in the exchange but are ordinary market participants.
Thus, they are neither subject to any special obligations nor privileges. In the context of market
making, ELPs are not required to remain inside the order book and can cancel all outstanding
limit orders in case of unfavorable market conditions. In practice, they mainly take advantage
of rebates to generate profits (see section 4.1). According to O’Hara (2015, p. 261), ELPs are
providing large amounts of liquidity and are essential in the market making process. Moreover,
most HFTs participating as market makers are acting as ELPs without obligations. However,
there are also occasional HFT firms registered as DMMs (O’Hara et al., 2015, p. 9), indicating
that the benefits can indeed outweigh the costs.
3.4 Reduction of Minimum Tick Size

Another change in market structure was practiced on quoting. Since the 1990s, the SEC
has been continuously decreasing the minimum tick size, which is the lowest increment of
acceptable prices. Thereafter, tick size defines the price grid on which stocks can be quoted.
Besides the implications of RegNMS on the market fragmentation, it additionally proposed
the sub-penny rule. Thereafter, stocks above $1 must not be quoted in increments lower than
$0.01 (SEC, 2005, p. 212). Importantly, this rule primarily affects market makers, as it only
restricts quotation of limit orders. In Europe, universal rules about minimum tick sizes have
not yet been implemented. Currently, exchanges are free to chose their tick sizes (FESE, 2011),
which results in unharmonized exchange structures. However, MiFID II will specify universal
rules for all European platforms.
These developments substantially influenced the behavior of order books. For this purpose,
Chordia, Roll, and Subrahmanyam (2008, p. 256) analyzed a long data set of the trade and
quote database (TAQ) with which they reconstructed the aggregated spreads on the NYSE
from 1993 to 2002. They found that decreases in tick size accurately corresponded with
decreasing effective spreads (see fig. 10). This makes sense, as most blue-chip stocks are
already traded with minimal spreads. Thus, the minimum tick size can additionally represent
a lower bound for spreads, rendering further price improvements impossible.
50% Stocks priced below $100

Stocks priced above $100
40%
Share of stocks
30%
20%
10%
0%
1 2 3 4 5 6 7 8 9 10 More
Average spread [USD/100]
Figure 4: Average spreads by stock prices. This figure is a modified version of Avramovic (2012, p. 4),
which was originally based on a proprietary dataset by Credit Suisse. The average spreads were rounded to
result in integers. The calculations were conducted separately for stocks above and below $100. There is
no further information about either data or exact observation period. (Own presentation)
Even clearer evidence for this phenomenon is provided by Avramovic (2012, p. 3–4) for
US markets. She compared the average spreads of all S&P500 stocks by price and the results
are presented in fig. 4. Strikingly, 50% of all stocks priced under $100 are restricted by the
minimum tick size of $0.01 (see section 3.1). This means that in half of those stocks there is no
way for liquidity providers to post competitive quotes in order to jump the limit order queue.
Given that liquidity providers wish to post even narrower spreads than $0.01, this imposes
noteworthy market frictions. In contrast, stocks with higher prices are generally unrestricted,
as their average spread is not below $0.02. In this case, there is still an opportunity to undercut
prices on average.
Overall, fig. 4 shows that the concept of relative tick size has crucial implications on the
restriction of spreads. Relative tick size is defined as the ratio between minimum tick size and
stock price (O’Hara et al., 2015, p. 5). This can be written as
min ∆p
RTSi,t = (1)
pi,t
The relative tick size at time t for stock i is written as RTSi,t , whereas p is the price in
levels and min ∆p the tick size. Thus, low price stocks haver larger relative tick sizes compared
to high price stocks, since they are both restricted at $0.01. Moreover, large cap stocks are
traded more frequently and thus subject to lower spreads. Such stocks are particularly relying
on low relative tick size in order for their spreads to be unrestricted.
In conclusion, this entire section introduced three developments in market structure, which
have substantial influence on HFT. First, market fragmentation increases search costs, as
information is now spread throughout numerous trading platforms. HFT allows for reduction
of those costs owing to its information processing ability. Second, the shift of trading activity
from floors to the electronic domain enables HFTs gather more information. Moreover, it
became possible to participate as an ELP. Third, changes in minimum tick size certainly have
a significant effects on financial markets. Particularly large cap stocks with low prices are
likely to be traded with a stable spread of $0.01.
4 ELECTRONIC TRADING PLATFORMS AND HFT 17
4 Electronic Trading Platforms and HFT

As previously stated in section 3.2, the increasing number of ETPs inevitably leads to stronger
competition among venues. The main source of revenue for ETPs are collected fees from their
members. Those fees are usually due after successful trade execution. However, there have
also been regulatory attempts (e.g. in Canada) to impose fees on all messages, reducing the
excessive quoting by HFTs (Malinova, Park, & Riordan, 2016, p 2). In any case, ETPs are
forced to attract as much order flow as possible to generate profits. This can only be achieved
through appealing conditions for market participants. Two of the most important features
with this respect are low spreads and large depth, whereas the former is most important in the
US due to the OPR. However, both of those criteria enhance cost-efficient execution of large
volumes. In order to provide such favorable conditions, ETPs must necessarily entice reliable
market makers.
According to Menkveld (2013, p. 716–717), the bid-ask spread compensates market makers
for three main components: (1) Order-handling cost/fees, (2) cost of adverse selection and (3)
risk-aversion premium. Thereafter, all ETPs use the same two tools to reduce these costs.
First, lower fees obviously decrease (1). Second, faster systems allow the market maker to
more quickly adjust their limit orders ahead of other traders taking advantage of stale quotes.
Consequently, this reduces (2) by prevention of unfavorable price moves. In the following,
sections 4.1 and 4.2 will separately address those two measures more thoroughly.
4.1 Fee Structure

Until recently, it was common for ETPs to charge both sides of a trade equally after successful
execution. Hence, the fee was two-sided. However, this has significantly changed, as trading
platforms started to attract carefully chosen client bases instead of mass order flow. This
is most easily controlled by fine-tuning the fee structure (O’Hara, 2015, p. 291–292). ETPs
started to differentiate between liquidity makers and takers. The former are passive limit order
traders, who automatically provide liquidity to the order book. The latter is characterized
by the usage of aggressive market orders, thus executing against standing limit orders and
reducing liquidity. The most common pricing scheme nowadays is the maker-taker model,
under which the maker receives a rebate and the taker pays a fee. Overall traders, taker fees
must exceed maker rebates in order for the ETP to remain profitable. This model is thought
to be particularly appealing to market making HFTs, as some of them are only profitable
due to limit order rebates. In the opposite case, the taker-maker model, market orders are
compensated whereas limit orders are subject to a fee. Such platforms are mostly attracting
retail order flow. However, there are still certain venues which do not follow either of both
new models, particularly in Europe. Some platforms still pursue traditional two-sided fees and
others offer subscriptions with unlimited trading volumes5 .
5
Aquis Exchange offers differing monthly subscriptions, each with an average daily message allowance.
Consequently, there are no costs for individual trades left. The detailed fee structure can be viewed on http://
www.aquis.eu/wp-content/uploads/2016/09/Aquis-Fee-Schedule.pdf (accessed on November 17, 2017).
United States [USD per traded share]

Exchange Type Maker fee Taker fee Daily volume [$]
BX Boston Exchange Taker-Maker 0.0020 -0.0017 9’397’346’388
BYX Bats-Y Taker-Maker 0.0018 -0.0008 12’444’071’072
BZX Bats-Z Maker-Taker -0.0025 0.0030 17’425’256’573
EDGA Edge-A Low cost 0.0003 0.0003 4’056’601’809
EDGX Edge-X Maker-Taker -0.0020 0.0030 17’620’645’242
NSDQ NASDAQ Maker-Taker -0.0035 0.0030 47’750’532’107
ARCA NYSE Arca Maker-Taker -0.0031 0.0030 29’563’440’609
Europe [bps of transaction price]

Exchange Type Maker fee Taker fee Daily volume [e]
BXE Bats Europe Low cost 0.000 0.160 2’431’704’431
CXE Chi-X Europe Maker-Taker -0.225 0.240 6’040’405’852
TRQX Turquoise Maker-Taker -0.260 0.280 2’666’915’026
XETR Xetra Traditional 0.360 0.360 5’440’302’140
Table 1: Fee structures of selected ETPs trading in equities. This table presents the conditions of both US
and European ETPs, with reasonably straightforward fee structures. The fees were taken from the individual
websites. Note that negative fees are equivalent to rebates. The listed fees apply to traders in the highest
volume-tier in each ETP. In the US, fees were used for Tape C securities (NASDAQ stocks). The average
daily volume was calculated for each exchange during October 2017 on displayed orders only. Volume data
was provided by the CBOE.
Table 1 provides an overview over a selection of relevant ETPs in terms of fee structure
and volume. The information about fee structures are publicly available on the websites of
each venue. The fees represent expected charges for high-volume liquidity providers, such as
HFTs, and are thus extremely simplified. In practice, there may be other additional factors.
Moreover, rebates are denoted as negative fees, representing a form of credit. Volumes were
calculated using data from the CBOE6 as the average daily volume in October 2017.
The aim of table 1 is to give a rough overview over fee strategies of the largest ETPs.
Most venues are using the maker-taker model, especially on large exchanges in the US. NSDQ
and ARCA go one step further, as their maker rebate exceeds the taker fee for their top-
tier liquidity providers. Moreover, BATS Global Markets is an example for a diversification
strategy with its platforms. In the US it holds BYX, BZX, EDGA and EDGX, all of which
with are clearly distinct in fee structures. In Europe, trading volumes are generally much
lower compared to the US. Moreover, there is no significant European taker-maker market.
Interestingly, all US ETPs use volume-based pricing in dollars per traded share, whereas the
European counterparts are value-based charging basis points (bps) of the transaction price.
Consequently, low priced stocks are relatively more expensive to trade compared to high priced
stocks over the same volume. Furthermore, all of the fees and rebates are a fraction of the
minimum tick size of $0.01, which will be relevant below.
In the following, both maker-taker and taker-maker concepts will be discussed in more
detail and by means of relevant theoretical literature. The main focus therein lies on its effect
on market participants. Note that special attention is given to US markets, since tick sizes are
harmonized already across all venues by the SEC.
6
Retrieved from http://markets.cboe.com/us/equities/market share/ and http://markets.cboe.com/
europe/equities/market share/market/ (accessed on November 16, 2017)
4.1.1 Maker-Taker Pricing
As mentioned earlier, the maker-taker model is predominant in financial markets and intended
to attract liquidity providers. However, there are potential negative effects, especially with
regard to the OPR (see section 3.1). Ultimately, the maker rebate is designed to decrease
quoted spreads and thus competing for the NBBO. Angel, Harris, and Spatt (2011, pp. 39–
43) and Angel, Harris, and Spatt (2015, pp. 21–22) heavily criticize the maker-taker model
for several reasons, including distortion of order routing systems, arising agency problems for
brokers and quoted spreads which do not necessarily represent the true cost of trading. It is
important to understand that they argue in a frictionless and perfectly competitive market.
In this case, liquidity providers receiving maker-rebates are expected to correct the spread
downwards by the same amount in order to remain competitive. Meanwhile, the artificially
decreased bid-ask spread cannot be fully taken advantage of by the aggressive trader, due to
the taker fees. Ultimately, the trading profits will be unaltered as maker-taker pricing is just
a distortion of the bid-ask spread and only the net fee captured by the exchange matters.
This is particularly problematic as the NBBO imposed by RegNMS is based on quoted
spreads instead of transaction costs (SEC, 2005, p. 34). Artificially low spreads are therefore
decisive for all market participants. For instance, dealers who are internalizing their order flow
are forced to execute against market orders at worse prices. Similarly, exchanges are required
to reroute order flow to venues currently holding the NBBO, which is costly. In general,
liquidity providers in traditional two-sided fee structures are systematically disadvantaged.
A B
Fee Structure Traditional Maker-Taker
Make fee 0.05 -0.20
Take fee 0.05 0.30
Net fee 0.10 0.10
Broker routing costs [round-trip]

Limit orders 0.10 -0.40
Market orders 0.10 0.60
Table 2: Agency problem example of broker-routed orders. This is a demonstration of two venues A and
B with alternative fee structures. Importantly, the net fee is identical. Routing costs are measured from
round-trip trades and are defrayed by the brokers themselves. (Own presentation)
Another key issue according to Angel et al. (2011, 2015) is an emerging agency problem
with stock brokers, since fees and rebates are not transferred to their customers but absorbed
themselves. Brokers are usually compensated through fixed commissions. As seen in table 2,
brokers have an incentive to send market orders to venue A and limit orders to venue B. In
practice, this means that brokers route market orders to dealers who internalize them at the
NBBO at low cost. Meanwhile, limit orders are directly sent to maker-taker exchanges with
high exposure to informed order flow. Additionally, the client will not even receive the maker
rebate.
However, these arguments have not remained unchallenged in literature. Foucault, Kadan,
and Kandel (2013, pp. 304–311) presented a theoretical approach by modeling buyer-seller
interactions as liquidity cycles with alternating make- and take-phases. Both active and passive
traders must continuously monitor the market in order to exploit profit opportunities. If one
side is more active, ETPs can inversely adjust maker- and taker fees. This will equilibrate
imbalances in order flow without affecting the platform’s revenues per trade. Hence, an increase
in the maker(taker) fee causes a decrease in maker(taker) activity. However, Foucault et al.
(2013, pp. 329–334) subsequently endogenize transaction prices. Thereafter, they reach the
same conclusion as Angel et al. (2011), but only for a tick size of zero which is equivalent
to a continuous pricing grid. This intuitively makes sense, as a positive tick size is a form
of friction, as it imposes a restriction on optimal prices (see section 3.4). The larger those
frictions are, the less accurately fees can be incorporated into transaction prices. Moreover,
the artificial undercutting of quoted spreads—as intended by design—is only possible if the
minimum tick size is non-restrictive, hence only for spreads of at least $0.02. If all venues
reached the minimum spread, maker rebates grant no advantage anymore (see fig. 4).
Indeed, the argumentation of Angel et al. (2011) is unlikely to hold in practice. In consid-
eration of the fee structures in table 1, it becomes apparent, that US taker fees never exceed
$0.003 per share. In fact, the SEC (2005, p. 28) caps chargeable fees on ETPs at that exact
amount, regardless of maker or taker fee. However, the sub-penny rule (SEC, 2005, p. 29)
does not allow for tick sizes smaller than $0.01 for shares trading above $1. Therefore, tick
size is more than three times larger compared to the maximum chargeable fee. Based on
the argument of Foucault et al. (2013), this implies a major influence of the maker-taker fee
composition, as they can impossibly be accurately reflected in transaction prices.
To address maker-taker pricing empirically, Malinova and Park (2015, pp. 515–516) ana-
lyzed data from Toronto Stock Exchange (TSX) during the revision of its fee structure. After
the update, stocks cross-listed in the US followed volume-based maker-taker pricing: A maker
rebate of $0.00275 and a taker fee of $0.004 per share. The remaining stocks were left with
value-based pricing: No maker rebate and a taker fee of 1.8bps. Malinova and Park (2015,
p. 527) suggest reduced spreads after the introduction of maker rebates. Moreover, total trans-
action costs for aggressive trades remained unchanged. Both results are in line with the theory
of competitive market making. Maker rebates indeed seem to incentivize liquidity providers
to post narrower spreads and constant transaction costs imply reduced spreads to be coun-
terbalanced with higher taker fees. Interestingly, the results also show lower price impact in
maker-taker stocks, indicating reduced information in aggressive orders (see section 5.1). For
this reason, Malinova and Park (2015, pp. 521–522) isolate retail traders, who are presumably
”unsophisticated” and pass taker fees onto their broker. The dataset contains trader identities,
allowing for the approximation of retail accounts by behavioral patterns, i.e. odd-lots trading
and absence of short-selling. In fact, retail traders relatively decrease passive trading. In
contrast, they increase their aggressive volume, indicating that retail traders take advantage
of reduced spreads whereas higher taker fees are internalized by their brokers.
While Malinova and Park (2015) focus on spreads and the activity of retail traders, Bat-
talio, Corwin, and Jennings (2016) examine broker routing behavior and execution quality of
their passive order flow. The authors analyze whether non-marketable limit orders are sent
to favorable venues for the customer. More specifically, three measures are used to proxy
execution quality of limit orders: Likelihood of a fill, speed of a fill and the realized spread
(Battalio et al., 2016, p. 2195). The overall findings indicate a negative relationship between
taker fees and execution quality (pp. 2231–2232). Hence, it is unfavorable to post limit orders
on pronounced maker-taker platforms. Moreover, based on disclosures by the SEC, brokers are
indeed sending limit orders to venues with large maker rebates or even sell it to market makers
(pp. 2206–2208), in order to maximize their profits. Consequently, customers are routed to
venues with worse execution quality for their limit orders without any compensation in the
form of maker rebate in return. This principal-agent problem is perfectly in line with the
theoretical concerns first uttered by Angel et al. (2011, pp. 39–43).
4.1.2 Taker-Maker Pricing
As indicated earlier, taker-maker markets (henceforth inverted markets) are compensating ag-
gressive traders and charging liquidity providers. This approach, might seem unintuitive at
first, since ETPs aim to attract liquidity to their platforms. However, inverted pricing is specif-
ically designed to attract uninformed retail order flow. The model relies on the assumption
that retail traders must go through brokers, who make routing decisions and internalize ac-
cruing fees or rebates. If there are several venues simultaneously offering the NBBO, RegNMS
allows brokers free choice of where to route market orders. This is especially the case for highly
liquid stocks with large relative tick size, as their spreads are usually restricted by tick size (see
section 3.4). As a result, inverted pricing only attracts aggressive order flow of stock brokers,
if their quotes are at the NBBO. This conclusion is following a reversed argumentation as
presented in section 4.1.1.
Despite the maker fees on inverted markets, liquidity provision might still be beneficial
overall owing to certain advantages. First, since brokers usually route market orders of retail
traders, the aggressive order flow is comparatively less informed (O’Hara, 2015, p. 262). Retail
traders are assumed to be less sophisticated than institutions which is manifested in lower price
impact and larger realized spreads (see section 5.1). Therefore, trading against retail traders
is generally more profitable. Second, and even more importantly, limit orders are given an
opportunity to jump the queue and circumvent price-priority on other venues without having
to undercut existing prices by a whole tick (Angel et al., 2015, p. 22). As aggressive order
flow is more frequently routed to platforms with the best taker conditions, limit orders have
higher execution quality at the expense of a maker fee. Posting NBBO-equivalent quotes on
inverted markets is therefore a version of queue jumping other venues (Ye & Yao, 2014, p. 3),
especially when spreads are restricted to one tick. Analogously to maker-taker markets, the
maximum allowed fee on inverted venues is a fraction of the tick size. If pricing was made on
a continuous scale (i.e. no tick size restriction), inverted markets would lose their purpose.
Quoted spreads would always be lower on maker-taker markets due to maker rebates. This
grants absolute execution priority under the OPR. Thus, inverted markets can only exist if
market orders can be freely routed within financial markets.
The empirical results by Battalio et al. (2016), as already presented in section 4.1.1, obvi-
ously apply to inverted fee structures as well. In their example, the inverted EDGA and BX
(see table 1) exceeded all other venues in terms in terms of fill rates and fill speeds (Battalio et
al., 2016, p. 2215). Furthermore, inverted markets were the only ones to offer positive realized
spreads, which is a measure for profitability of passive orders (see section 5.1). This is in line
with the hypothesis about presence of uninformed aggressive order flow on inverted markets.
Overall, the realized spreads on EDGA even exceed the maker fee. According all these results,
liquidity providers might prefer to switch to inverted markets as these offer better execution
quality, especially in stocks where the NBBO is restricted by tick size.
This directly leads to the main conclusion about the difference between both pricing
schemes. In cases of larger spreads, price competition is still possible which gives maker-taker
markets an advantage by artificially decreasing quoted spreads. Thus, maker-taker venues
induce downward pressure on spreads under the OPR. However, once the minimum spread is
reached inverted markets become more attractive to liquidity providers. Since price competi-
tion is not possible anymore, they can jump the queue for a small maker fee.
4.2 Optimization of Latency

The second large powerful tool to attract HFTs to the own market is by offering cutting-
edge technology. Low latency traders are generally dependent on fast machines within ETPs,
because they are otherwise not able to exploit their speed advantage. Such technologies start
with fast matching engines inside the exchanges, able to handle large amounts of orders at once.
However, ETPs have found other ways to reduce latency for individual members. According to
Menkveld (2016, pp. 11–12) and Arnuk and Saluzzi (2009, p. 1), there are two crucial services
offered by ETPs which serve latency reduction purposes: Direct data feeds and co-location.
Even though both services are very distinct, they mutually aim to provide purchasers with a
speed advantage of a few milliseconds compared to others. This delicate difference can have
important implications in today’s fragmented markets through latency arbitrage. All of the
concepts mentioned in this paragraph will be elaborated below.
4.2.1 Direct Data Feeds and Co-location
According to RegNMS, ETPs are required to synchronize the distribution but not the arrival
of their quote data (SEC, 2005, p. 271). Thereafter, venues must immediately send their
quotes to the SIP, but can simultaneously transmit it to whomever they wish (Stone, 2014,
p. 3). It does not matter to regulators where the information arrives first. To boost revenues,
some ETPs sell their raw data feeds to members in exchange for a fee (e.g. NYSE Best Quote
and NASDAQ ITCH). If a subscriber is close by, he will receive the information before the
SIP which determines the NBBO 7 . Thus, some traders have the ability to see future quotes a
few milliseconds ahead of others (Easley, O’Hara, & Yang, 2016, p. 1072). Consequently, they
have more clarity about the current state of the market.
To ensure that the raw data feeds are arriving as quickly as possible, fast traders addi-
tionally purchase co-location services. This allows traders to set up their own systems inside
the ETP’s facilities in order to minimize distance to the matching engine (Jones, 2013, p. 10).
Since the speed of light has become the restrictive upper boundary for transmission speed, the
only variable left to decrease latency with is distance. For instance, the NASDAQ OMX offers
7
See section 3.1
three kinds of co-location services, differing in proximity to the servers and subscription fees
(Brogaard, Hagströmer, Nordén, & Riordan, 2015, p. 3412). It becomes apparent that certain
traders (mostly market makers) are willing to pay large amounts of money to bring latency to
an absolute minimum relative to other traders.
4.2.2 Latency Arbitrage
The combination of both tools discussed in section 4.2.1 above can entail the emergence of
an almost risk free profit opportunity, which was not possible prior to RegNMS: Latency
arbitrage. Arnuk and Saluzzi (2009, pp. 1–2) were the first to pinpoint this phenomenon.
Thereafter, co-located traders with direct data feeds have early access to market information
before the SIP and anybody else. Furthermore, some HFT firms operate proprietary networks
of (preferably straight) fiber-optic cables and microwave towers between different platforms in
order to minimize latency of inter-venue signals (Litzenberger et al., 2012, p. 91).
The mechanics of latency arbitrage are visualized in fig. 5. Let’s assume that an HFT
is continuously monitoring quotes on ETP A. As soon as the quote changes, his co-located
servers are informed through the direct feed. The same signal is simultaneously sent to the
SIP, which is further away and thus receives a delayed signal. In the interim, the HFT decides
whether the quote adjustment can be exploited on other venues. If so, orders are quickly
sent to other platforms, trading into the direction of the incoming change. Once the rerouted
SIP-signal of the new NBBO reaches the other markets, the HFT has already exploited the
profit opportunity.
ETP B
SIP/NBBO
SIP ETP C
via
ETP D
ETP A
ETP B
Connections
Co-located
Direct
ETP C
HFTr
ETP D
t=0 Time
Figure 5: The underlying mechanics of latency arbitrage. Latency arbitrage is only made possible due to the
speed advantage of fast traders. Signals are simultaneously sent out to the SIP and co-located traders with
access to direct data feeds at t = 0. Since HFT communication ways and IT-systems are faster than the
SIP, signals can be processed and transmitted before other venues officially receive the information. (Own
presentation)
This phenomenon is somewhat comparable to a bullet flying at supersonic speeds: It hits

the target out of nowhere and the gunshot is not heard before it is too late. In this context,
the target represents a stale limit order. Since this behavior exploits the fragmented structure
of financial markets, it represents structural HFT strategy discussed in section 2.2.2.
Even though gains from latency arbitrage per trade are almost negligible, this effect can
accumulate over time. For this reason, Ding, Hanna, and Hendershott (2014, pp. 317–321)
constructed an NBBO based on direct data feeds of five8 ETPs (henceforth synthetic NBBO)
and compared it with the official SIP NBBO of Apple stocks. Discrepancies between both
measures are labeled dislocations. The latency between synthetic and SIP NBBO was calcu-
lated as the difference in timestamps of the same recorded event. As shown in fig. 11, the
average latency is approximately 1.5 milliseconds with outliers of up to 3 milliseconds. Fur-
thermore, Ding et al. (2014, p. 322) calculate that quotes are dislocated 0.175% of the time.
Unfortunately, it is almost impossible to exactly determine the gains from latency arbitrage.
However, HFTs could certainly use such strategies owing to their speed advantage.
4.3 Interdependence between New Platforms and HFT

So far, section 4 conveyed the impression that the dependency between ETPs and HFTs is
unilateral, whereas, in fact, this is not the whole truth. HFTs have been shown to profit
from latency reduction and certain fee structures. However, ETPs are just as dependent on
HFT in order to thrive in financial markets (Menkveld, 2014, p. 338). Trading venues can
only be profitable by attracting large volumes which are subject to trading fees. Particularly
the acquisition of market making HFTs is extremely attractive to ETPs, as they provide
substantial liquidity. Moreover, HFT will lead to narrower quoted bid-ask spreads (Foucault,
2012, p. 15), which is crucial under the OPR in the US and smart routing technologies in
Europe. This symbiosis is visually presented in fig. 6.
- Passive volume
- Narrow spreads
- Market share
HFT ETPs
- Favorable fees
- Direct data feeds
- Co-location
Figure 6: Symbiosis between HFT and ETPs. Both HFT and ETPs share a mutual dependency. On the
one hand, HFT relies on the competition of ETPs for market share with attractive conditions. On the other
hand, ETPs cannot thrive in financial markets without the support of liquidity providing HFT. The arrows
above point into the direction of the respective profiteer. (Own presentation)
This shows, that neither HFT nor the current market structure can be examined in isola-
tion. HFT has become an integral part of how financial markets and ETPs work nowadays.
This underlines, once more, the importance of RegNMS and MiFID. Without these regula-
tions, the competition between trading venues would be negligible, making it impossible for
HFT to thrive.
8
Direct feeds from BYX, BZX, EDGA, EDGX and NASDAQ were used (Ding et al., 2014, p. 317). Based
on CBOE data in table 1, their total volume does not exceed 50% of all trading in US markets
5 SPEED AS FORM OF INFORMATION 25
5 Speed as Form of Information

The previous sections were mainly concerned with market structure and its recent evolution.
Henceforth, the focus of this thesis will shift onto HFT itself and how it fits into the new
electronic domain. Most importantly, HFTs have access to advanced knowledge before anyone
else as elaborated in fig. 5. Such an advantage is extremely valuable, because slow traders can
only make decisions based on information that has already become stale (Moallemi & Saǧlam,
2013, p. 1071). This section will therefore technically focus on how HFT takes advantage of
this information and what it means for other market traders.
In this context, the substantial differences between market making and opportunistic strate-
gies as in section 2.2.2 must be taken into account, especially because HFT technology is
beneficial to both categories. It is crucial to distinguish the impact of additional speed onto
aggressive or passive strategies, as they affect markets through different channels. In the
following, both mechanisms are separately discussed in sharp detail, by reference to both the-
oretical and empirical literature. Importantly, this section will only focus on how speed affects
HFTs and their profitability. This will subsequently allow for the evaluation of their impact
on market quality in later sections.
5.1 Informed Trading and Adverse Selection

In the context of trading, the concept of adverse selection is always of utmost importance.
According to Glosten and Milgrom (1985, p. 72), this phenomenon occurs when the aggressive
party of a trade has access to private information, which is not yet public knowledge. Hence,
the trader is informed and has a (more or less accurate) expectation of how the price will
change in the future. In order to exploit this knowledge, he will trade into the direction of his
expectations. Given the informed trader’s presumption is verified, the price will subsequently
move permanently in his favor. Since trading is a zero-sum game, this automatically entails a
loss for the passive trader, as the price change is unfavorable in his perspective. As a result,
market makers prefer to trade with uninformed investors.
Consequently, having private information is desirable to traders, despite the heterogeneity
of information acquisition. According to Easley et al. (2016, p. 1074), information can be
either bought in the form of superior speed (co-location services and direct data feeds), or
it can be acquired through costly fundamental research. Either of those methods will lead
to an edge compared to other traders. In their theoretical model, Goettler, Parlour, and
Rajan (2009, p. 72–73) replicate a dynamic limit order market with informed traders. One
of their key findings shows, that traders without intrinsic motivation to trade (henceforth
speculators) have the highest willingness to pay for information, particularly during phases of
high market volatility. This result make intuitive sense, as speculators are indifferent about
the fundamentals of a given security, as long as it yields short-term profits. Incidentally, this
behavior is congruent with HFT, as discussed in section 2.2.1, which maintains its positions
only over a very brief period of time and even avoids overnight exposure. Therefore, it makes
more sense for HFTs to buy information in the form co-location and direct data feeds instead
of conducting costly research, which requires longer investment horizons for exploitation.
After successful acquisition of private information, it can most efficiently be exploited

against stale limit orders. This implies that market makers are generally adversely selected by
the aggressive order flow. Moreover, market orders have the advantage of immediate execution.
In order to remain profitable, passive traders are compensated by means of the effective spread.
According to Huang and Stoll (1996, p. 324), the effective spread measures the actual transac-
tion price of a trade. In practice, effectively paid spreads can differ from quoted spreads in two
ways. First, some trades might be executed inside the quoted spread by broker-dealers who
internalize order flow. This situation is favorable for the aggressive orders. Second, depth at
the best quoted spreads might be very thin, hence, from the perspective of aggressive traders,
a large market order may be partially executed at worse prices than quoted. In conclusion,
the effective spread delivers a more accurate measure for market maker compensation than the
quoted spread, as it is based on actual execution prices. However, it still does also not include
make-take- or access fees imposed by the venue but only the ”raw” cost of a market order.
5.1.1 Effective Spread Decomposition
Following the prior argumentation, effective bid-ask spreads can be further decomposed. Glosten
(1987, p. 1295) was the first to propose the idea that bid-ask spreads were composed of two
components: Adverse-selection costs and market maker gross-profit. The former compen-
sates passive traders for potential adverse selection of informed aggressive order flow whereas
the latter is what remains as a profit for liquidity provision. Even though the core idea of
Glosten (1987) is still present in recent market microstructure literature, notations have slightly
changed. For this reason, the following definitions are taken from Hendershott et al. (2011,
pp. 10–11), Carrion (2013, pp. 696–697) and Conrad, Wahal, and Xiang (2015, pp. 276–277).
All of these more recent studies intensively examine the impact of HFT on components of the
efficient spread. Nowadays, such measures are always calculated relative to the asset price and
only for a single trade (as opposed to round-trip trades). Hence, the effective half-spread at
time t is defined as follows:
pt − mt
ESt = qt (2)
mt
The transaction price is denoted as pt . If the aggressive order is not executed at one single
price, the volume-weighted average price9 is applied alternatively (Carrion, 2013, p. 689). This
situation can occur if the size of the market order exceeds the depth at the best quote, since the
remaining part will be executed against the next best quotes. Meanwhile, mt is the midpoint
calculated as the average between quoted bid and ask prices (Huang & Stoll, 1996, p. 324).
As discussed before, eq. (2) captures the actual transaction price instead of the quoted price.
Moreover, qt is an indicator variable and can take two values for either buyer-initiated (1)
or seller-initiated (−1) trades. This particularly ensures that ESt remains positive for seller-
initiated trades, since the bid price will always be below the midpoint. To capture adverse
selection, the price impact at time t is defined further as:
9
The volume-weighted average price is a measure to calculate the average execution price of a trade. It is
P N
i=1 (V oli )(pi )
defined as PN where i indexes the ticks at which the order is executed.
i=1 V oli
mt+k − mt
P It = qt (3)
mt
In contrast to the effective spread, the price impact mainly focuses on the evolution of the
midpoint within time-interval k following a given trade. The interval can be chosen arbitrarily.
In theory, this measure should represent the permanent effect of a trade on the midpoint. If P It
is positive, the aggressive trader correctly anticipated the future price change and therefore had
private information not incorporated into prices beforehand. Meanwhile, the passive trader
incurs a loss. As opposed to the price impact, the realized spread is the portion of the spread
that can actually be kept by the passive trader after consideration of the permanent price
impact.
pt − mt+k
RSt = qt = ESt − P It (4)
mt
Apparently, the price impact is equal to eq. (3) subtracted from eq. (2). Intuitively, the
realized spread will thus only be positive if the price impact does not exceed the effective
spread. That is, if the difference between transaction price and midpoint is larger than the
subsequent change in midpoint in absolute terms. Otherwise, the market maker incurs a
negative realized spread, which is equivalent to a loss.
5.1.2 Market Maker Profitability
This idea of realized spreads is further apprehended in Menkveld (2013, pp. 724–725). However,
in contrast to Hendershott et al. (2011) and Carrion (2013), he uses round-trip trades. In other
words, aggressive traders are assumed to use a second market order to exit the accumulated
position after the midpoint has shifted. Hence, only the net trading profit is left. In particular,
Menkveld (2013, p. 724) distinguishes between spread profit and positioning profit. The former
the sum of half-spreads in both trades, whereas the latter stands for the total shift of midpoint
into the favorable direction from the aggressive trader’s perspective. This is equivalent to the
price impact, which represents the ongoing adverse selection.
Ask
Ask
MP MP
Bid Bid
Positive realized spread Negative realized spread
Figure 7: Spread profits and positioning losses of a market maker. The visualized situation is from a market
maker’s perspective, who posts passive limit orders in two periods. The black dots are buy respectively sell
orders of an aggressive trader. The gray area represents the half-spread. The sum of half-spreads on both
trades is the spread profit. Meanwhile, the two-sided arrow represents the positioning loss due to shifting
midquote. (Own presentation, following Menkveld (2013, p. 725) and Aı̈t-Sahalia and Saǧlam (2014, p. 21))
Figure 7 aims to visually explain the concept of realized spreads in eq. (4) from the market
maker’s perspective. If the total spread profits exceed the positioning losses (left panel), then
the market maker profits from the round-trip trade. In contrast, if the adverse selection
component is too large (right panel) the gains the aggressive trader makes a profit on the
round-trip trade. Usually, theoretical models in HFT literature assume market makers to be
risk-neutral and in perfect competition. Under these premises, the spreads will be adjusted
until it is equivalent to the expected price-impact per round-trip trade. This way, the market
maker will incur an expected profit of zero.
5.2 High-Frequency Liquidity Consumption

In case of liquidity demanding strategies, speed can be beneficial through multiple channels.
Foucault (2012, p. 19) claims that tradable information for HFT originates from two sources.
First, it can be inherent in the trading process itself, including order flow, (cross-platform)
quotes and imbalances. Therefore, there is no link to underlying fundamentals whatsoever.
Second, it can be derived from news about firm-specific prospects. This comprises earnings
announcements, macroeconomic news or just unofficial gossip (soft information). Both of
those information sources have different implications on other market participants but also
certain commonalities. For instance, in both cases, HFT is somehow able to extract a rent at
the expense of slower market participants. Therefore, aggressive proprietary algorithms are
henceforth referred to as high-frequency bandits (HFBs). The following two subsections will
therefore discuss each of the two information channels separately.
5.2.1 Anticipation of Order Flow
As mentioned above, the trading process itself can serve as a valuable source of information,
especially in fragmented markets. Even though this has nothing to do with firm-fundamentals
as such, it may allow for anticipation of future order flow, which can be traded on in advance.
One example for such anticipation can be found in agency algorithms. As discussed in sec-
tion 2.1, institutional traders often split large parent order into (seemingly) randomized child
orders to disguise true intentions. A proprietary HFB will attempt to decipher these auto-
correlation patterns and trade on it before the entire parent order is executed. In academic
literature, this process is therefore referred to as preying on large orders. On a side note,
successful preying does not necessarily require the usage of low latency technology, but rather
very intelligent pattern recognition software. However, as soon as multiple competing HFBs
are involved, relative speed becomes crucial (Moallemi & Saǧlam, 2013, p. 1071). Furthermore,
HFBs can also employ latency arbitrage for cross-platform order flows.
Following Menkveld (2016, p. 13), it is important to distinguish between preying on in-
formed and uninformed orders, whereas the former provoke a permanent price impact. Un-
informed orders are assumed to be noisy and submitted by liquidity traders without any
underlying systematic. In contrast, informed order flow is a convenient target for HFT algo-
rithms. Principally, the identification of informed trading allows for the prediction of future
permanent price impacts. In other words, it enables HFBs to trade into the direction of future
returns beforehand. Intuitively, large AT-employing financial institutions are the main target
of HFBs. Hence, this can be interpreted as a cat-and-mouse game between agency algorithms
and HFT. In essence, the anticipation of order flow facilitates free-riding on private informa-
tion of others (Baldauf & Mollner, 2015, p. 18). Therefore, anticipatory HFBs mainly extract
their rent from informed traders. This proposition is also confirmed by the theoretical model of
Li (2017, pp. 15–16), where HFBs levy a speed tax on informed but slower traders. The more
accurately HFBs can anticipate informed order flow, the more speed taxes increase. Thus,
informed trading in the presence of anticipatory HFBs becomes unprofitable. Ultimately, an-
ticipation only allows trading on information of others before they have had the chance to do
so. This is why in common speech this phenomenon is often referred to as front-running (Li,
2017, p. 1)
Unfortunately, the exact methods of anticipation have not been revealed due to the extreme
secrecy of HFT algorithms. However, such behavior has been confirmed empirically. The
most momentous study on this issue was delivered by Hirschey (2017). Using an HFT dataset
by NASDAQ10 , he examined patterns of aggressive order flow by both HFTs and nHFTs
throughout 2009. The main goal is the inference of buying- or selling pressure relations between
both groups, which is measured as the net imbalance of marketable orders per second (p. 12).
This measure is derived by subtracting the marketable selling volume from the marketable
buying volume by trader group. The goal was to examine whether HFTs buy(sell) ahead
of nHFTs’ purchases(sales). Notably, Hirschey (2017, pp. 22, 43–46) demonstrates two key
findings. First, he shows that HFTs consistently place their market orders several seconds
ahead of nHFTs, but not the other way around. Figure 12 visualizes this finding through a
time line of net buying pressure around each second. As a test for robustness, the process was
repeated in the absence of relevant public news11 . This controls for the fact that HFT simply
allows for faster reaction to public information, which would logically lead nHFT order flow.
In fact, this alteration had no significant impact on the results, implying that HFT order flow
is leading due to anticipation instead of just faster reaction to common public information.
Second, the net imbalance of marketable orders by HFTs generally predicts the direction of
return over the next 30 seconds. Seemingly, HFTs can identify informed order flow and predict
permanent price changes.
A similar study was conducted by Aquilina and Ysusi (2016), but in European markets.
The authors examined the London Stock Exchange (LSE), BATS Europe and Chi-X Europe
comparatively and slightly adapted the methodology. According to Brogaard, Hendershott,
and Riordan (2014, p. 2272), the NASDAQ dataset is not able to identify all HFTs correctly.
It ignores all trading desks which also offer brokerage services. Therefore, large banks such
as Goldman Sachs or Morgan Stanley are excluded even though they might also employ HFT
technology. In contrast to Hirschey (2017), Aquilina and Ysusi (2016, p. 14) accounted for
this inaccuracy by identifying pure nHFTs as strictly human traders. Nevertheless, the results
are approximately identical with one extension. They show that HFBs react even more pro-
nounced to pure nHFT net buying pressure (pp. 24–25). This indicates that HFBs prefer the
anticipation of human trading.
10
See section 2.3.3
11
News were approximated with 35’000 articles during the observation period retrieved from the Dow Jones
Factiva global news database (https://www.dowjones.com/products/factiva/).
Another approach to identify front-running behavior is elaborated in Brogaard (2010, p. 23–

25), who also utilizes the NASDAQ dataset. However, he explicitly looks at the sequence of
individual trades, whereas Hirschey (2017, p. 13) aggregates the net buying pressure over each
second. According to Brogaard (2010), systematic front-running is occurring if it is more likely
that a non-HFT-initiated trade is preceded by n HFT-initiated trades than vice-versa. Using
this methodology, he found no indication of front-running (p. 63). This might be the case
because his measure seems critically flawed. It assumes that there is no other background
activity in the order book besides the trades in question. In practice, this is not really the
case. NASDAQ reports that its system handles up to 35’000 messages per second12 . Moreover,
the results of Hirschey (2017) indicate that order anticipation can occur several seconds before
the actual trade. Therefore, it is extremely unlikely that the anticipation happens within the
five preceding trades of the nHFT. Assuming that the preceding trades were predominantly
noise, the lack of significant results seems logical.
With regard to the previous results, it can be assumed that HFT is in fact able to anticipate
order flow by other participants. However, it has so far only been suggested that such behavior
exists, but not how it ultimately effects measures of market quality and transaction cost in
particular. This issue will be captured again in the later course of this thesis.
5.2.2 Quote Sniping
In contrast to order flow anticipation, quote sniping is all about speed. More specifically,
the ability to incorporate new public information (henceforth news) as quickly as possible into
prices. The term quote sniping comes from the fact that HFBs will try to execute against limit
orders that have not yet been adjusted, so-called stale quotes. Snipers use public information,
which has neither been received by all market participants nor incorporated into prices. This
is only possible during a very brief time window. Thus, such behavior is very closely related
to the gunshot-analogy in section 4.2.2 about latency arbitrage. Quote sniping can be seen as
an additional adverse selection component for other traders, especially market makers (Biais,
Foucault, & Moinas, 2015, p. 293). Referring to the definitions in section 5.1.1, adverse selec-
tion generally entails larger price impact after execution. Under exogenous effective spreads,
this reduces the realized spread an with it the profitability of market making.
With regard to fig. 1, quote sniping can either be identified as arbitrage- or directional
strategy, depending on the source of the news. In the first case, the HFB is actively monitoring
the market for mispriced securities, which is not an uncommon phenomenon. Budish, Cramton,
and Shim (2015, pp. 1564–1566) compared the midpoints of E-mini S&P 500 futures and an
S&P 500 ETF. As both securities have the same underlying, they should be perfectly coupled.
In practice, however, the correlation starts to break down at high frequencies. This is revealed
in fig. 15, which plots correlations between S&P500 futures and ETFs below 0.1 seconds. There
are two main inferences from this plot. First, correlations strictly monotonically increase over
longer intervals. Nevertheless, it does not exceed the value of 0.45, which means that there
are still large mispricing at this point in time. Second, correlations have strictly increased over
the period 2005–2011. This trend can be reasonably explained by the reduction of arbitrage
12
Retrieved from http://www.nasdaq.com/services/homw.stm (accessed December 4, 2017)
opportunities as a consequence of the rising HFT activity during that period. In the second
case of directional strategies, quick processing of public announcements and their effect on
prices is key. In this context, techniques such as text mining is more important. Figure 14
shows how quickly HFTs react to negative and positive macroeconomic news. Overall, the
underlying logic behind quote sniping is identical for either arbitrage and directional strategies,
since the origin and form of news ultimately does not matter.
The isolated impact of quote sniping on market quality is not obvious. In academic liter-
ature, there are two opposing positions in this regard. On the one hand it is an enhancement
of information incorporation into prices. When HFBs receive a new piece of information and
snipe stale market maker quotes, they are by definition the fastest to incorporate new infor-
mation. Even if the difference is only a fraction of a second, the HFB indeed enhanced price
efficiency by a tiny margin.
On the other hand, quote sniping has significant negative externalities on market makers.
To demonstrate this, Biais et al. (2015) designed a theoretical model about fast traders in
a limit order market. One of their key findings states, that limit order traders will increase
their spreads to compensate for additional adverse selection caused by HFBs, who trade on
advanced information, i.e. at higher speeds (p. 298). As a consequence, liquidity in the market
decreases, making it unattractive for normal traders as well. In essence, HFBs cause negative
externalities on the market which they are not willing to internalize.
A slightly different approach for the same argument is made by Menkveld and Zoican
(2017, p. 1190), who model the effect of decreasing exchange latency on fast trading. This
ultimately results in a race between HFBs and market makers after the revelation of news.
Thereby, HFBs attempt to execute against the stale quotes, whereas market makers try to
cancel them beforehand. An obvious way to compensate potential losses from this race are
larger effective spreads, which increases transaction costs even for ordinary traders. This model
nicely demonstrates how HFBs can extract rents from market makers with inferior speeds.
Furthermore, it also accurately emphasizes crucial characteristics about quote sniping: HFBs
and market makers are dueling in a race to reach the stale limit order and either execute
or cancel it. Menkveld (2016, p. 16) refers to these situations as run games. Therefore,
only relative speed ultimately matters (Menkveld & Zoican, 2017, p. 1190). Consequently, an
HFB must be faster than other traders to remain profitable. As discussed later on, this will
eventually lead to a race-to-the-bottom of HFT latencies.
To isolate quote sniping empirically, Zhang (2017) analyzed the reaction of both HFTs
and nHFTs to hard information shocks. In this context, the term hard refers to quantifiable
shocks in linked securities. Zhang (2017, p. 4) argues that futures generally lead stock prices13 .
Therefore, she examines how shocks in futures affect the passive and aggressive net trading
imbalance of both HFTs and nHFTs. Her key findings are presented in the upper panel of
fig. 17. Apparently, aggressive market participants instantly trade into the direction of the
information shock. However, HFTs are much faster and reach their peak 6 seconds after the
13
The same phenomenon is confirmed by Dobrev and Schaumburg (2017, pp. 19–22) for S&P500 cash and
futures markets. The results are shown in Panel A of fig. 23. They used an unprecedented non-parametric
methodology, which only relies on trading activity and not noisy returns. The study will be elaborated thor-
oughly in section 7.2.1
shock. Thereafter, the net trading imbalance is reversed in order to offset their positions. In
contrast, nHFTs are continuously building positions until 30 seconds after the event (Zhang,
2017, pp. 9–10). These findings are in line with Roşu (2016, p. 6), who assumes HFT to have
extremely high inventory costs. This motivates the immediate offsetting of positions. Further-
more, the mean-reverting inventory also confirms the theoretical predictions of Jovanovic and
Menkveld (2016, pp. 3–4). In their model, HFT is particularly active in times of higher hard
information. However, HFBs are acting as unsolicited middlemen, imposing costs on other
traders. The sniped securities are only held for a brief amount of time, before the position is
offset again.
In conclusion, quote sniping generates rents for HFBs at the expense of other traders
by exploitation of mispriced security prices. Even though the new information can either
be gathered from other fragmented venues or through news announcements, the underlying
mechanism is equivalent, namely the execution against stale limit orders.
5.3 High-Frequency Liquidity Provision

As already suggested in section 5.2.2, quote sniping is primarily accomplished at the expense
of market makers. It is nothing else than adverse selection based on information that has not
been received yet. In the previously introduced racing-analogy by Menkveld and Zoican (2017,
p. 1190), there is obviously a coequal rival to the HFB: The high-frequency market maker
(HFM). Analogously to their opponents, HFMs take advantage of low latency technology,
which enables faster reaction to public news. However, instead of sending marketable orders,
HFMs race to either cancel or adjust their quotes before becoming stale. Essentially, they
aspire to prevent HFBs from sniping their limit orders, thus reducing adverse selection.
NASDAQ
103 NYSE Arca
Cancel-to-trade ratio
Bats-Z
Edge-X
Chi-X
102
101
100
2012 2013 2014 2015 2016 2017 2018
Figure 8: Cancel-to-trade ratios on various ETPs. This plot shows the cancel-to-trade ratios on five
significant ETPs in the US. Chi-X the only venue not registered as an exchange, while showing very unstable
patterns. The remaining exchanges constantly average around a ratio of 20:1 (horizontal line). The data
was provided by the SEC through MIDAS. Note that this graph is plotted on a logarithmic scale with base
10. (Own presentation)
This algorithmization of market making has consequences for the order book. As mentioned
in section 2.2.1, HFTs will post vast amounts of orders, most of which are canceled shortly
afterwards. Similarly, Jones (2013, p. 6) argues that HFMs monitor the market continuously
and adapt their quotes accordingly. This pattern can be observed in fig. 8, which plots cancel-
to-trade ratios of five significant US trading platforms in daily frequency. The data was
extracted from the MIDAS14 . Figure 8 shows, that the four registered exchanges have stable
cancel to trade ratios of around 20:1. This implies that out of 21 limit orders only one is
successfully executed. This pattern indicates a lot of HFM activity in terms of liquidity
provision. Chi-X is the only ETP without registered exchange-status, making it subject to
weaker regulation. This might be a reason for its excessive cancel-to-trade ratio compared to
other venues.
Currently, there is plenty of theoretical literature in circulation about HFMs. Despite
the diversity in methodologies, all of these models fundamentally focus on the profitability of
market making, adverse selection and the effect on effective spreads. Moreover, they roughly
yield identical conclusions. For instance, Hoffmann (2014, pp. 160–162) models a dynamic
limit order market including both fast (i.e. HFMs) and slow market makers. Most notably,
he finds that limit orders are more valuable to HFMs, as they are given the opportunity for
potential adjustments after news. This advantage reduces adverse selection costs caused by
HFBs and increases expected profits ceteris paribus. Hence, HFMs are ultimately able to
post more competitive limit orders with higher execution probability, which obviously reduces
quoted spreads. Slow market makers on the contrary are forced to post quotes with lower
execution probability, as they are unable to react timely to news. This argument is illustrated
in fig. 13. Han, Khapko, and Kyle (2014, p. 5) extend these results with the idea of a winner’s
curse for slow market makers. In their model, they have no ability to react to news, whereas
HFMs do. As a result, limit orders of slow traders are only executed if they are subject
to adverse selection. Moreover, Han et al. (2014, p. 8) conclude that HFT technology is only
valuable in their model if there is plenty of new information appearing in the market. It is even
possible that HFMs also engage in order anticipation strategies as in section 5.2.1. This would
allow them to ajust their quotes before the arrival of marketable order flow. However, there
has not been sufficient empirical proof for this so far. Nevertheless, these findings strongly
imply a crowding-out effect of slow market makers.
Lastly, Aı̈t-Sahalia and Saǧlam (2014, p. 8) include inventory aversion of HFMs in their
theoretical model. Before each quote submission HFMs face a trade-off of between the benefit
of capturing the spread versus the cost of rising inventory. In this case, inventory is costly,
because it entails the risk of changing asset prices. The authors show that due to their speed
advantage, HFMs can decrease the cost of inventory. As a result, HFMs can increase their
liquidity supply (p. 23), which is generally beneficial to market quality.
Conclusively, the previous paragraphs theoretically introduced two ways through which
lower latencies can influence market making behavior: The adverse selection channel and
the inventory management channel. Both of those channels have been empirically addressed
in Brogaard, Hagströmer, Nordén, and Riordan (2015, pp. 3411–3414), using a co-location
14
Retrieved from https://www.sec.gov/marketstructure/downloads.html (accessed November 14, 2017)
upgrade on the NASDAQ OMX. More specifically, the exchange offers three levels of co-
location allowing for differential latencies, the fastest of which includes a 10 Gigabit (10G)
connection. The authors analyzed a dataset with individual trader identifications15 including
indication about the respective co-location status. Among other results, Brogaard et al. (2015)
present two main findings with regard to HFM behavior. First, traders who purchased the 10G
subscription show typical market making behavior such as augmented presence at the NBBO,
larger passive realized spreads and higher order-to-trade ratios (p. 3421). This indicates that
HFMs value speed even more than HFBs. Second, HFMs were able to substantially decrease
the adverse selection component16 following the 10G upgrade. Thus, they are better able to
avoid being picked off after the emergence of news (p. 3432). Furthermore, HFM inventory
crosses the neutral position significantly less frequently after the 10G upgrade (p. 3434). This
generally indicates a buildup of larger positions in either direction compared to beforehand.
This interpretation can be justified with decreasing inventory costs as elaborated above. As a
result, the second finding implies positive impacts of higher speed on both adverse selection and
inventory channels of market making. Furthermore, Hagströmer and Nordén (2013, p. 757)
directly compared the inventory management of HFBs and HFMs on a similar dataset by
NASDAQ OMX. Since the data only comprised orders with anonymized identities, the authors
were required to reconstruct daily net positions based on executed trades. In this context, the
inventory ratio was defined as the absolute value of the accumulated net position at the end of
the day divided by the total daily volume. According to the results, HFBs have significantly
larger inventory ratios than HFMs17 . Apparently, HFBs offset their positions across other
markets, whereas HFMs try to reduce its positions within the same market.
Additional results were provided by Zhang (2017, p. 10), as seen in the bottom panel of
fig. 17. The graph shows how the net position changes for both HFTs and nHFT after the
emergence of news. In contrast to ordinary liquidity providers, HFMs are able to maintain
a neutral position despite enhanced one-sided buying-/selling pressure. Hence, HFMs are
seemingly able to timely withdraw and adjust their stale limit orders, whereas nHFTs incur
substantial adverse selection within the first 6 seconds.
5.4 Exploitation of Minimum Tick Sizes

In a market where a significant amount of stocks is restricted by the minimum tick size 18 ,
there are obviously frictions in quoted spreads. As discussed earlier, restricted spreads are the
consequence of large relative tick size in combination with high liquidity (i.e. for large cap
stocks). Moreover, HFTs being more affected by these restrictions is an intuitive conjecture,
due to exceptionally high turnover. Each trade is coupled with transaction costs in form of
the spread, which cannot be neglected.
15
See section 2.3.2
16
See section 5.1
17
In August 2011(February 2012), when volatility was high(low), the inventory ratio amounted to 6%(13%)
for HFMs and 38%(50%) for HFBs (Hagströmer & Nordén, 2013, p. 756).
18
See section 3.4
One recent study on the influence of tick size restrictions was done by Hagströmer and
Nordén (2013) based on Swedish NASDAQ OMX data. Said venue has three tick size regimes
depending on the stock price in levels. Regimes are adjusted upon each crossing of a thresh-
old19 . In order to make inferences about the impact of tick size, the authors conducted an event
study around threshold crossings (p. 761). The results are presented in fig. 18. Hagströmer
and Nordén (2013, p. 762) show that overall trading activity decreases after an increase in
tick size. This effect is especially pronounced for HFT, indicating that these traders are more
sensitive. However, fig. 18 must be interpreted with caution in two ways. First, the sample
period contained more aggressive than passive HFTs. Second, the plot only shows short-term
changes for the day following the threshold crossing. Therefore, potential overreactions in the
short term are plausible. Despite these caveats, elevated sensitivity of HFT towards tick sizes
seems to be a reasonable inference.
Once again, the effect of tick sizes on HFT is not unambiguous in strategies. One argument
on this issue comes from Weller (2017, pp. 22–23). He pays special attention to relative
tick size which increases in stock price, as defined in eq. (1). If two stocks share the same
fundamental volatility and tick size, the one with the higher price will incur more frequent quote
adjustments. After Weller (2017) low relative tick size is thus preferable for AT. However,
this conclusion is an extreme generalization. In fact, such conditions are only profitable for
aggressive AT strategies, including HFBs. Each quote update embodies a piece of news, which
can be utilized against stale quotes. As a result, lower relative tick size promotes adverse
selection at the expense of limit orders. He justifies this line of reasoning through exceptionally
high correlation between his AT proxies and stock prices in levels.
In contrast to this argumentation Gai et al. (2013, p. 43) use the NASDAQ HFT dataset
20 and examine the ratio of liquidity provided by HFMs grouped by terciles of both market
cap and prices. The results presented in table 3 indicate highest HFM presence in large
cap stocks with low prices, in contradiction to Weller (2017). Importantly, the likelihood
of restricted spreads is driven by two channels. On the one hand, large cap stocks are most
frequently traded, implying narrower spreads compared to smaller stocks, which directly affects
the likelihood of restriction. On the other hand, spreads of cheap stocks are more likely to be
restricted due to higher relative tick size. Both of those effects imply that HFMs in fact prefer
stocks with spreads that are restricted by tick size.
Low price Medium price High price

Small cap 18.6% 18.9% 18.5%
Medium cap 35.5% 23.4% 22.2%
Large cap 45.4% 37.5% 30.9%
Table 3: Percentage of liquidity provision by HFT by size and price. The table shows the percentages of
all trades where an HFT was on the passive side. Higher percentages are equivalent to elevated liquidity
provision. For this calculation, 120 stocks were included and divided into terciles of both market cap and
price. This results in 13–14 stocks per group. This table was inherited from Gai et al. (2013, p. 43).
19
The thresholds separating all three tick size regimes are located at SEK50 and SEK100. Tick sizes are
increasing with respect to stock prices and defined at SEK0.01, SEK0.05 and SEK0.10 (Hagströmer & Nordén,
2013, p. 760). This approach is meant to somewhat balance relative tick sizes
20
See section 2.3.3
To justify this argument, Gai et al. (2013, pp. 17–18) claim that HFMs prefer speed compe-
tition over price competition with other traders. This consideration is based on the fact that
almost all trading platforms are structured with price-time priority 21 . They argue, that limit
order traders generally aim to maximize execution probability. The easiest way to achieve this
is by undercutting the best price. However, as seen in fig. 4, most stocks are restricted in their
spread which renders price competition impossible. In such cases, the only means to reach the
front of the limit order queue is speed. Thus, HFMs who can quickly adjust their quotes not
only mitigate adverse selection costs, but additionally to increase execution probability at the
new price (Moallemi & Saǧlam, 2013, p. 1071). Inverse markets, as discussed in section 4.1.2,
follow a similar intuition: They provide traders with an opportunity to jump the queue on
other venues in exchange for small fee relative to the tick size. In summary, Gai et al. (2013,
p. 4) see the main problem in the arbitrary divisibility of continuous time as opposed to a
discrete price grid. If both time and price dimensions were either discrete or continuous, speed
would matter much less for liquidity provision. Thus the crowding-out of conventional market
makers would be less of an issue.
The results presented so far are additionally fortified by Hagströmer and Nordén (2013,
p. 764) using the tick size regimes elaborated in section 5.3. In particular, they find that under
increasing tick size aggressive HFTs decrease their volume whereas HFMs become more active.
However, their interpretation differs from the other studies (e.g. Gai et al., 2013), in that the
authors argue with profitability. Under restricted spreads, market making in regimes of high
tick size is more profitable, as spreads are necessarily overvalued. By contrast, aggressive
trading becomes less attractive since higher price impact is required to offset the spread22 .
Even though this section addressed both aggressive and passive strategies in relation to
tick sizes, the effect on HFMs is of higher importance. It has been shown that the race against
HFBs is not the only motivation behind latency reduction. More specifically, restricted spreads
create an opportunity for speed competition against other liquidity providers. This is being
exploited by HFMs as shown empirically.
5.5 The Technological Arms Race

So far, section 5 has highlighted several ways in which low latency can be advantageous in
comparison to other market participants. As a consequence, traders have an incentive to
decrease their latency through large investments. However, it is not quite clear whether such
investments are beneficial to the broad public associated with social costs.
In their model, Biais et al. (2015, p. 304) identify a tendency to over-investment into fast
trading technologies. The reason for this is a certain increase in profitability after investing due
to superior information. These benefits are internalized, whereas adverse selection is imposed
onto slower traders. This reveals a lack of alignment between private and public interests.
Consequently, the model of Biais et al. (2015) can be interpreted as a game of eat-or-to-be-
eaten: Technology investments are only valuable if they give an edge over others. However, the
21
In financial markets, limit orders are usually prioritized by price first. Orders with identical prices are
subsequently ordered with regard to arrival time at the venue. Hence, an undercutting order with the single
best quote will have absolute priority despite delayed arrival. This feature is also explained in appendix A.
22
These assumptions are confirmed in section 7.1.2.
two cases of exclusively fast and exclusively slow traders are equivalent, with exception of the
socially suboptimal investment costs. As a result, only relative speed matters whereas absolute
speed is irrelevant. In practice, there is not only a binary distinction between fast and slow
traders. Rather, HFT can invest arbitrarily into low latency technologies, but ultimately it is
always about being faster than the competition (Moallemi & Saǧlam, 2013, p. 1071; Hoffmann,
2014, p. 164). This point is similarly emphasized by Li (2017, pp. 22–23), who indicates that
speed leads to winner-takes-all situations. This raises the incentives to invest even further.
These developments can be shown empirically based on the occurrence of arbitrage oppor-
tunities. Section 5.2.2 already introduced the concept of mispricings between linked securities
in fig. 15. In addition to this, Budish et al. (2015, pp. 1571–1574) examined the extent of
these frequent arbitrage opportunities. Figure 16 shows the duration of arbitrage profits and
demonstrates that mispricings between S&P500 Futures and ETFs have become increasingly
short-lived. Moreover, the authors link the number of daily arbitrage opportunities to current
market volatility with an R2 of 0.87. Hence, the occurrence of mispricings is mainly driven
by the amount of volatility and thus a consequence of fragmented markets. The only alter-
ation throughout the observation period has been a decrease in duration of such opportunities.
This indicates the continuous investments into faster technologies.Budish et al. (2015, p. 1617)
perceive this to be an unpleasant consequence of continuous markets, where it will always be
possible to undercut competition in terms of latency. This motivates traders to participate in
a never-ending arms race.
In essence, investments into HFT technology only bear social value, if it enables significant
improvements in price efficiency (Jones, 2013, p. 21) and transaction costs for the public.
Otherwise, it is just an internalization of profits, while negative effects are externalized. It
should have been understood so far that speed does matter. Therefore, the impact of HFT on
market quality is of tremendous importance in the current political debate and will be covered
in the following.
6 THE INFORMATIVENESS OF ASSET PRICES 38
6 The Informativeness of Asset Prices

The previous section was mainly concentrated on how speed is an advantage, as it allows for
better information. It has been shown, that HFT technology enables almost instant reaction
to news in the market. This includes quote sniping HFBs exploiting stale quotes and liquidity
providing HFMs. At first sight, it does only seem logical to expect asset prices to reflect
more information in the presence of HFBs and HFMs, as they are adjusted more quickly.
Generally, this would be a very desirable feature for financial markets. However, there are a
few features of HFT which may in fact be hindering this process. More specifically, Foucault
(2012, pp. 21–22) summarizes three important problems:
[i] HFTs hold positions over very short time-periods to reduce their exposure. It is unlikely
that the underlying information is related to long-term fundamental asset values.
[ii] HFT takes information from the trading process itself. As discussed in section 5.2.1,
they anticipate order flow to free-ride on other traders’ information. Ultimately, this
does not increase informativeness of prices.
[iii] Due to the high level of automation, HFT might misinterpret a piece of news and trade
into the wrong direction. This could inject additional noise into markets and thus increase
short-term volatility.
All these problems will be kept in mind throughout this section when assessing the impact
of HFT on asset pricing. Furthermore, two new terms are introduced following Brogaard,
Hendershott, and Riordan (2014, p. 2268). First, price discovery refers to the process of how
new information (news) is incorporated into prices. Second, price efficiency refers to the level
of information which is already reflected in prices. The remainder of this section will put
strong emphasis on empirical findings, whereas theoretical concepts fade into the background.
6.1 Price Discovery

In order to analyze the effects of HFT on price discovery, it is necessary to determine the infor-
mation content of trades. This allows for inferences about which market participants submit
new information. For this purpose, two base methods have emerged in academic literature.
On the one hand, the flow of information can be modeled according to a vector autoregres-
sive (VAR) model with numerous extensions (sections 6.1.1 to 6.1.3). On the other hand,
the state-space model delivers a similar approach, but with numerical optimization techniques
(section 6.1.4). Ultimately, both methods are meant to distinguish transitory component and
permanent component in price changes. The former represents short-term noise which will
eventually revert. However, it induces unnecessary volatility to the price discovery process.
The latter includes actual information and is thus persistent over time. Obviously, the per-
manent component is equivalent to the price impact and of primary interest in terms price
discovery. The following subsections will additionally include an introduction to the method-
ologies used, as it is reasonable to know how information is inferred from simple order-level
data in a statistical sense. Certainly, there are also alternatives to these estimation methods,
such as decompositions of the price impact, but they will not discussed here.
6.1.1 The Vector Autoregressive Framework
The first methodology to be introduced was originally designed by Hasbrouck (1991a, pp. 183–
184). It primarily aims to link order flow together with returns. The VAR approach allows for
bilateral interferences between both variables over a certain lagged time period. This can give
indications about the question of causality. However, the original method was derived long
before the rise of HFT. This is why modern literature adapted the idea of Hasbrouck (1991a)
to facilitate segmentation of different trader types. In spite of certain adjustments, the core
intuition behind the approach has remained unchanged.
The underlying VAR mechanics will be only described in maximum brevity. The following
definitions were taken from Brogaard (2010, pp. 31–32) as a base model. It is self-explanatory,
that this framework can only be applied to transaction level data with indication of trader
type (HFT versus nHFT). The first variable of interest in the VAR system are returns of the
midquote, denoted as rt . Moreover, the order flow must also be included for both HFT (qtH )
and nHFT (qtN ). Importantly, qtH and qtN are only indicator variables taking the values −1 for
a seller-initiated trades, 0 for neutral positions or 1 for a buyer-initiated trades. Importantly,
time t is not measured in clock time, but in event time. The index is increased after each
quote update. Therefore, new information can also be incorporated without actual trades. All
these variables can be combined into the following VAR model:
K
X K
X K
X
H N
rt = αi rt−i + βi qt−i + γi qt−i + ε1,t (5)
i=1 i=1 i=1
K
X K
X K
X
qtH = δi rt−i + H
ρi qt−i + N
ζi qt−i + ε2,t (6)
i=1 i=1 i=1
XK XK XK
qtN = πi rt−i + H
νi qt−i + N
ψi qt−i + ε3,t (7)
i=1 i=1 i=1
The model is estimated over K lags depending on the dataset. The empirical studies referred to
by this thesis use 10 lags and all ε are uncorrelated residuals. Equation (5) models the evolution
of midpoint returns whereas eqs. (6) and (7) represent the order flow of HFT and nHFT. The
systems contains cross-interaction terms to account for causality between the variables. After
estimation, the VAR system can be transformed into vector moving average (VMA) form:
    
rt a(L) b(L) c(L) ε1,t
 H 
qt  = d(L) e(L) f (L) ε2,t  (8)
 
qtN g(L) h(L) i(L) ε3,t
In eq. (8), the coefficients a(L) to i(L) are infinite order lag polynomials of the form a(L) =
P∞ i
i=0 ai L . According to Hasbrouck (1991a, p. 189), the impulse response function (IRF)
resulting from lag polynomials can be interpreted as a measure of private information. This
approach enables the consideration of how order flow affects permanent price changes. More
specifically, the elements in b(L) capture the contribution to price discovery by HFT, whereas
c(L) does the equivalent for nHFTs. The IRF is subsequently defined as the cumulative sum
over the elements of the lag polynomial up to a predetermined lag K, as shown in eq. (9) for
both HFT and nHFT private information.
K
X K
X
IRFH→r = b̂i IRFN →r = ĉi (9)
i=0 i=0
Empirically, there are several studies taking advantage of this methodology. Brogaard
(2010, p. 32–33, 66) was a pioneer in this field, as he was the first to analyze the HFT
dataset by NASDAQ. Therefore, his contribution can be regarded as a pilot study in the HFT
domain. He estimated the IRF with K = 10 lags presented in eq. (9) for 105 stocks from
all market cap terciles. Interestingly, he found significantly higher contribution of HFT over
nHFT in 90 stocks. The remaining 15 did not have significant deviation between both types.
Moreover, Brogaard (2010, pp. 33–34, 67) subtracted lag 1 from lag 10 of the IRFs, where a
positive outcome suggests a persistent impact, since the trade had an increasing effect in the
long-term. Otherwise, the trade implies a short-term overreaction which is quickly adjusted
afterwards. On average, both HFT and nHFT yielded positive results, whereas the former was
significantly larger. An almost identical approach was applied in Deutsche Bundesbank (2016,
pp. 47–49) on German DAX Futures. Likewise, the results are also alike with HFT trades
having significantly larger long-term impact on prices compared to nHFT. Even though both
of these studies are looking at individual assets, it is daring to interpret 10 lags as long-term
effects, even if the VARs were estimated for each security individually. Especially, since the
framework is set in event time instead of clock time.
An extension to this approach can be found in Brogaard, Hendershott, and Riordan (2016,
pp. 11–14), based on message data from Canadian equity markets including masked trader
identities23 . Due lack of classification, the authors used individual behavioral patterns to
proxy HFTs. Additionally, they introduced more variables into the system, including limit
orders and cancellations, which increased the dimensionality of the VAR model. The results
show significantly larger effects on returns by messages originating from HFT (p. 34). This
indicates that HFT does not only enhance price discovery through aggressive trades. Instead,
all types of messages carry on average more information in comparison with nHFT.
6.1.2 Variance Decomposition
So far, this section modeled the impact of trades on future asset returns. However, it is also
possible to decompose the variance of efficient prices into components representing HFT and
nHFT. This approach was originally developed in Hasbrouck (1991b, pp. 573–578). It assumes
the observed midpoint pt to consist of an efficient price mt and a stationary pricing error st ,
which can be written as pt = mt + st . The efficient price is assumed to follow the random walk
mt = mt−1 + wt with identically and independently distributed innovations wt ∼ iid(0, σw 2 ).
Accordingly, the following must hold for the log returns rt :
rt = pt − pt−1 = ∆mt + ∆st

23
See section 2.3.2
If the pricing error st is assumed to be ergodic white-noise without long-term effect on

observed prices, log returns rt can be decomposed into a permanent component ∆mt and a
transitory component ∆st . Following Brogaard (2010, p. 34), the return-affecting lag polyno-
mials a(L)–c(L) estimated in eq. (8) are used for the variance decomposition of efficient price
innovations wt . Under the assumption of orthogonal residuals εt , this can be written as:
N
!2 N
!2 N
!2
X X X
2
σw = ai σε21 + bi σε22 + ci σε23 (10)
i=0 i=0 i=0
According to eq. (10), σw 2 consists of three components: The public component of past returns
(σε21 ) and two private components for each trader type (σε22 ,σε23 ). Brogaard (2010, p. 68) find
19.5% of permanent variance to come from aggressive HFT on average, whereas aggressive
nHFT only contributes 10.5%. This signifies a superior influence of HFB on the price discovery
process. The remaining 70% of variance can be explained by quote updates and other public
information. The same approach was also applied by Benos and Sagade (2016, pp. 70–75) for
the UK equity market during 2011–2012. Interestingly, their results are much different. To
be exact HFT only contribute 6.1% through aggressive trading, whereas the variance share of
nHFT is 36.6%. In this case, HFT only has a minuscule effect on price discovery.
The disparity of these results is hardly explicable, especially since the observation period of
Benos and Sagade (2016) is two years later, since HFT has already become more sophisticated.
One possible explanation for this might be the lack of an OPR in Europe under MiFID.
Therefore, trade-throughs are still possible in the UK (Kohler & von Wyss, 2012, p. 3). This
can potentially impair the ability of HFT to anticipate order flow. Market orders can be
executed straight away without comparison to the current NBBO and possible rerouting.
This reduces the value of latency arbitrage as explained in section 4.2.2. As a result, HFT
appears to have lower contributions to price discovery.
6.1.3 Information Shares
The last VAR extension discussed here was originally introduced by Hasbrouck (1995), namely
information shares. This concept is quite similar to the variance decomposition in section 6.1.2.
However, the underlying presumptions are a different. The information shares approach deter-
mines the individual contributions of n different time series pt = [p1,t , . . . , pn,t ]0 to a common
but unobservable random walk mt = mt−1 + ut . For instance, Brogaard (2010, p. 35) sep-
arately estimates the midpoints of both HFT- and nHFT quotes in the same asset. These
time series are assumed to be co-integrated of order n − 1 with covariance-stationary residuals
(Hasbrouck, 1995, p. 1180). The vector of prices in first differences ∆pt can be modeled as a
vector error correction model. In case of reversibility, this representation can be transformed
into VMA form:
∆pt = Ψ(L)εt (11)
In eq. (11), Ψ(L) represents an n × n matrix containing the lag polynomials. The residuals
εt have no autocorrelation: E [εt εs ] = Ω for t = s and 0 otherwise. Accordingly, Ψ(1) is the
simple summation over all VMA coefficients which is equivalent to the long-term effect on ∆pt .
Due to co-integration, all rows of Ψ(1) are identical and arbitrarily defined as ψ (Hasbrouck,
1995, p. 1181). This leads to the formal definition of the information share:
" #" #
ψj2 Ωjj h
0
i σ12 σ1,2 Ψ1
InfoSharej = and ψΩψ = Ψ1 Ψ2 (12)
ψΩψ 0 σ1,2 σ2 Ψ2
Therein, j ∈ {1, . . . , n} is indexing individual time-series and ψj can be interpreted as the

cumulative long-term effect of a shock in j. Thus, in eq. (12) the numerator stands for the
individual contribution to long-term variance, whereas the denominator represents the total
long-term variance of the underlying unobservable random walk.
Once again, Brogaard (2010, pp. 36, 69) was the first to utilize this methodology on HFT
data. He found HFT to have an average information share of 58.1% in 120 stocks. This is
significantly higher than nHFT. Similarly, Brogaard et al. (2016, p. 42) applied it on Canadian
equity data. Their results show an an even higher contribution of 58–93% on three Canadian
exchanges. Hence, both studies conclude HFT to be essential to price discovery. A somewhat
different approach was chosen by Hasbrouck (2017), who analyzes TAQ data stamped to the
nanosecond as well as direct feed data from various ETPs. A similar study was conducted
by Ding et al. (2014) as presented in section 4.2.2. Even though the data has no indication
of HFT, he estimated the information shares of both SIP and direct feed data over various
time increments ranging from 10 microseconds to 1 second. Importantly, the framework was
fundamentally adapted to allow for clock-time observations. Interestingly, Hasbrouck (2017,
p. 34) found that almost all information comes from direct data feeds in increments below
1 millisecond. Since HFTs usually purchase these services, they are likely to be the main
contributors in such intervals. Moreover, it takes the speed of light several milliseconds to
reach all ETPs in the US. Thus, it makes sense that the centralized SIP cannot participate in
price discovery at even shorter increments.
6.1.4 The State-Space Framework
The second approach to measuring information content in trades is based a state-space mod-
eling. This method is especially useful, because it consists of both an observable measurement
equation and an unobservable transition equation. In fact, the state-space framework is con-
ceptually similar to Hasbrouck (1991b). Akin to section 6.1.2, the unobserved fundamental
value cannot be detected. Additionally, the observed midquote is subject to stationary pricing
errors. However, instead of drawing conclusions from the VAR system, parameters in state-
space models must be optimized using iterative numerical methods, such as the Kalman Filter.
The most influential of these models was provided by Brogaard, Hendershott, and Riordan
(2014, pp. 2278–2280), which analyzed the NASDAQ HFT dataset. Due to the similarities
with previous sections, the approach will not be elaborated here. However, the interested
reader can find all model specifications in appendix B.
The results presented in BHR (2014, pp. 2282–2284) are quite interesting on multiple notes.
To begin with, they estimate the state-space model using aggressive order flow. Most notably,
HF T D trades significantly into the direction permanent price changes, but against transitory
errors. The same is true for nHF T D , but clearly to a lesser extent on both components.
Both of these findings imply the exertion of adverse selection by aggressive traders through
informed trading, as prices are likely to follow their path. However, it is also a sign that the
price discovery process is driven by liquidity demanders. Especially HF T D seems to take
advantage of stale quotes in conformity with HFB behavior. On the contrary, when the model
was estimated over liquidity supplying order flow, the patterns reversed. Both types of passive
traders supply liquidity against the permanent component, but in the direction of transitory
noise. Hence, passive traders are adversely selected. The values were most pronounced for
HF T S implying a larger loss in both components. This finding is inconsistent with HFM
behavior discussed in section 5.3. Due to their speed advantage, HFMs should actually be
able to mitigate adverse selection costs in comparison to nHFT. The authors explain this
discrepancy with inventory management of liquidity providers. Thereby, HF T S are willing to
provide liquidity at worse conditions in order to offset unfavorable positions. For comparison,
(Menkveld, 2013, pp. 733–735) designed a similar state-space model, but focused on intra-day
patterns. He found evidence for significant adverse selection of HFMs.
However, it seems counter-intuitive that BHR (2014, p. 2289) subsequently find HF T S
to be significantly more profitable than nHF T S before and after liquidity rebates. If HF T S
trades are in fact less sophisticated during transitory and permanent shocks, they should
obviously have inferior trading performances. The only plausible explanation seems to lie in
periods without midquote movements. In these cases, HFMs can outcompete nHFTs, allowing
for earnings of the spread on each trade. Akin to section 5.4, this implies that HFMs prefer
stocks with large relative spreads, as quote revisions are less frequent ceteris paribus.
Despite everything, aggregated HF T All is found to have a clearly positive contribution
to price discovery (BHR, 2014, p. 2281). These results are consistent with Brogaard (2010)
discussed in section 6.1.1, who also argues that most price discovery from trade originates in
aggressive HFT order flow, whereas nHFT is of less importance. The results of BHR (2014)
further imply adverse selection cost for passive traders, especially HFT.
6.2 Efficiency of Prices

The next concept to be examined is price efficiency. It has been shown earlier, that HFT
occupies an important role in the enhancement of price discovery. This implies that prices
should be automatically more efficient during HFT presence. This section will give an overview
over empirical evidence on this conjecture. More specifically, it will present studies about the
effects of HFT on the efficient market hypothesis (EMH), originally proposed by Fama (1970,
p. 388). So far, academic literature was mostly concerned with two out of three EMH forms.
First, the weak form assumes asset prices to reflect all historical information. Second, the
semi-strong form is more concerned with the speed of price adjustments after the emergence
of new public information. Both forms will be discussed separately.
Importantly, the results below must be interpreted with caution. Price efficiency is an
overall market condition, where the contribution of individual trades is less important (as
opposed to price discovery). Therefore, most researchers use publicly available message data
(e.g. TAQ) and design indicators to proxy aggregated computerized trading activity24 . For this
24
See section 2.3.1
reason, the proxies can also comprise large amounts AT, as it has certain similarities toHFT in
behavior but entirely different motivations. Exceptions to this are Carrion (2013) and Zhang
(2017) who take advantage of the NASDAQ HFT dataset25 . Fortunately, the results are in
harmony regardless of the dataset used.
6.2.1 Weak Form of Efficiency
For asset prices to be weakly efficient, they must follow a random walk. The underlying
assumption tells that all expectations about future returns are already incorporated into to-
day’s price, hence it is efficient (Fama, 1970, p. 391). In order to test the weak-form EMH
empirically, the most frequently chosen approaches in literature aim at the falsification of au-
tocorrelation in returns, since random walks can neither have momentum nor mean-reversion.
Thus, innovations are carried on forever, but not exacerbated.
Two examples for the detection of autocorrelation are, Carrion (2013, pp. 707–709) and
Zhang (2017, pp. 5, 13) use the measure of price delay. For this purpose, two models are
estimated, where rt are log asset returns and M KTt the contemporary market returns:
Model 1: rt = α + β1 M KTt + ε1,t (13)

L
X
Model 2: rt = α + β1 M KTt + δl M KTt−l + ε2,t (14)
i=1
Model 1 is designed to capture only contemporaneous effects, whereas Model 2 additionally

includes L lags. Both models are estimated independently with focus on the R2 . Since Model
2
2 contains additional independent variables, RModel 2
2 ≥ RModel 1 . The price delay can then be
calculated as follows:
R2
P Dt = 1 − Model
2
1
(15)
RModel 2
According to the definition in eq. (15), P Dt ∈ [0, 1] must hold. Furthermore, higher price
delay implies some kind correlation with past market returns beyond the contemporaneous
effect. Meanwhile, lower price delays are an indication for an underlying random walk.
Both Carrion (2013, p. 709) and Zhang (2017, p. 13) found a significantly negative re-
lationship between price delay and HFT participation rate 26 . This indicates that prices are
more efficient in the presence of HFT and resemble a random walk more closely. Moreover,
both studies re-estimated this relationship using the participation rate for demanding and sup-
plying HFT separately. Apparently, aggressive HFT has a much large impact on price delay.
However, it must be kept in mind, that the causality has not been clarified in either studies.
It is not clear whether HFT makes prices more efficient or is just attracted by such conditions.
According to the argumentation in section 5.2.2, the former option seems to be more likely.
Most notably, HFB activity has the most pronounced negative relation with price delay, even
though their strategies would be unprofitable in already efficient markets. In fact, HFBs strive
on inefficiencies and mispricings.
25
See section 2.3.3
26
The participation rate is defined as the number of HFT trades divided by all trades. Therefore, it gives
an indication of the HFT market share at the moment.
A study based on a pre-classified foreign exchange (FOREX) dataset was delivered by

Chaboud, Chiquoine, Hjalmarsson, and Vega (2014). However, their quote and transaction
data solely distinguish between humans and machines. Thus, their study is directed to the
entirety of AT, instead of HFT in isolation, and return autocorrelation. The methodology is
mainly based on structural VAR systems enabling Granger-Causality tests with contempora-
neous effects. Their evidence suggests a strongly negative causal effect of AT participation rate
onto autocorrelation, but not vice-versa. The same holds for isolated aggressive AT, which is
in line with the results in the previous paragraph. Interestingly, passive AT and autocorrela-
tion have mutual causality (Chaboud et al., 2014, pp. 2071–2072). In the realm of HFT, this
phenomenon can be easily explained. On the one hand, efficient prices are attracting HFM
as there is less risk of adverse selection. On the other hand, HFM themselves significantly
contribute to price efficiency by quickly updating their quotes. Even though only AT was
considered, these results can also provide valid insights into HFT as a significant component.
The remaining studies with respect to the weak EMH are based on widely available TAQ
data without trader identification27 . Therefore, HFT activity must be approximated, which is
usually done under consideration of messaging rates. Excessive usage of messages is commonly
associated with elevated HFT activity. Importantly, this approach is very vague and only
proxies passive HFMs, as HFBs tend towards market orders. For instance, Boehmer, Fong,
and Wu (2015) use public message data of 42 European exchanges. They define the negative
total volume divided by the number of messages as a proxy. The resulting measure increases
in messages. Their findings indicate a negative relationship between HFT and autocorrelation
over 10 and 30 minutes. However, this timespan is extremely long and even 10 minutes is
perceived as an eternity by actual HFTs. The effect of HFT on such horizons are unclear.
Finally, random walks can additionally be tested using variance ratios (VRs)28 . Under
the assumption of zero autocorrelation, VRq = 1 must hold (Lo & MacKinlay, 1988, p. 47),
since variances are additive over multiple periods. Conrad et al. (2015, pp. 277–280) test this
presumption using Japanese stock market data and find improved VRs (closer to 1) during
times of high messaging intensity. The authors thus consider HFT to positively affect price
efficiency. The results are visually presented in fig. 19. Overall, this study is qualitatively al-
most identical to Boehmer et al. (2015). In conclusion, all evidence presented above is strongly
in favor of the positive relation between HFT and the EMH in its weak form. Particularly,
aggressive HFBs remove the predictability of asset prices based on historical results.
6.2.2 Semi-Strong From of Efficiency
The second form of the EMH addressed here is more challenging to investigate empirically.
Since the semi-strong form claims instant incorporation of all public information into asset
prices (Fama, 1970, p. 404), it requires data from multiple venues or linked securities with
27
See section 2.3.1
28
Variance ratio is defined as follows, where yt is the one-period log return and q ∈ N is a multiplier
determining the number of lags of the VR:
Pq−1
Var s=0 yt−s
VRq =
q × Var [yt ]
exact timestamps. This allows a verdict about the level of market integration. The underlying
assumption for the analysis of this EMH form is simultaneous activity of HFTs on multiple
trading platforms. It has already been demonstrated in section 5.5, that the process of infor-
mation incorporation between linked securities has been accelerating. This was demonstrated
by Budish et al. (2015) at the example of S&P500 Futures and ETFs in figs. 15 and 16. How-
ever, they did not connect this phenomenon to HFT. By contrast, this section will address
the impact of HFT (respectively AT) directly.
The first example to be presented is Haferkorn (2017, pp. 287–289) who compares quotes
of French blue chip stocks across Euronext Paris (venue 1) and BATS Chi-X Europe (venue
2). The dataset is stamped to the millisecond across venues, hence there might be certain
disparities in time-keeping. However, exchanges should be sufficiently synchronized to permit
precise insights at millisecond level. The author further used order-to-trade ratios as proxy
for HFT activity. More importantly, the level of price efficiency was measured as midpoint
dispersion across venues, whereas midpoints are again calculated as the arithmetic average
between bid and ask prices. This approach enables the detection of price disparities and must
be minimized for efficiency:
|M P1,t − M P2,t |
M Ptdisp = 2 ∗
M P1,t + M P2,t
In fact, the fixed-effects panel regression across stocks showed a significantly negative relation
between order-to-trade ratio and midpoint dispersion. For robustness and to account for
issues of endogeneity, the regression was repeated but with the introduction of the German
HFT Act29 as exogenous event instead. The results continue to show an increase in midpoint
dispersion after the event, which implies deteriorating efficiency after the restriction of HFT
(Haferkorn, 2017, p. 290). Overall, HFT seems to actively contributed to the transmission of
information between venues.
In addition to the tests on autocorrelation discussed in section 6.2.1, Chaboud et al. (2014,
pp. 2063–2065) examined triangular arbitrage opportunities in FOREX markets. More specif-
ically, they analyzed USD($), EUR(e) and JPY(¥) and the respective exchange rates. Ar-
bitrage opportunities were registered, if a trader can earn a strictly positive profit by only
trading these three currencies. Such instances are an indicator for mispricings and thus in-
efficient markets. As in section 6.2.1, the authors analyze the connection between AT and
arbitrage opportunities by means of Granger-Causality tests with contemporaneous effects.
The results from these tests are not unambiguous and can most accurately be interpreted
as mutual causality, especially between aggressive AT and triangular arbitrage opportunities.
This can once again be explained with HFB behavior. On the one hand, HFBs are lured by
inefficiencies. On the other hand, their quote sniping behavior will make markets more efficient
by exploitation of arbitrage opportunities.
29
The German HFT Act was introduced in 2013 and is designed to limit HFT activity through additional
legal requirements and the introduction of a mandatory message fees (Haferkorn, 2017, p. 286)
6.3 The Dark Side of Inflated Efficiency

So far, HFT seems to play a very crucial role in market efficiency, making it a seemingly
indispensable trader type. However, price efficiency is only half of the story. There have been
numerous concerns in literature despite the improvements in efficiency. Particularly, they
address items [i] and [ii] presented at the beginning of section 6. Both imply a lack of meaning
in HFT-incorporated information in terms of fundamental value. On the one hand, HFT
position are too short-dated to carry any interest in the long-term prospects of the underlying
company. On the other hand, HFT does not participate in costly information acquisition
process. Eventually, this raises the question of where HFTs take their information from.
6.3.1 Social Costs and Benefits
Price efficiency is undoubtedly a desirable trait in financial markets and directly beneficial to
all market participants. However, at one point the marginal benefit of faster price discovery
declines (Stiglitz, 2014, p. 6). For instance, it does not matter to long-term investors whether
prices are adjusted several milliseconds earlier or later, because they have intrinsic value to
allocate their capital anyway. Moreover, HFT does not necessarily bring new information to
the table, but rather trades into the direction of upcoming price changes. Hence, the additional
social value seems rather limited, especially for long-term fundamental investors.
In contrast, such slight enhancements are indirectly associated with tremendous costs for
all market participants. For instance, there is a trade-off between price efficiency and adverse
selection costs (Jones, 2013, p. 20; Menkveld, 2014, p. 341). As seen previously, it is mostly
HFBs who contribute to price discovery, whereas their speed advantage poses information
asymmetries. Especially limit order traders will have to carry the adverse selection costs,
making liquidity provision less attractive in presence of frequent news. Furthermore, over-
investments into fast technologies are another cost of speed, leading to a wasteful arms race
among HFTs as in section 5.5. In the end, the costs of these socially non-beneficial expenses
are borne by the general public.
All of these reasons encourage careful balancing of the trade-off between benefits and costs.
Surely, HFT can have positive impact on market quality, but it must not be forgotten that it
comes at a certain price. Especially at very low latencies, marginal costs become very high
whereas marginal benefits are diminished from a social perspective.
6.3.2 Free-Riding on Private Information
So far, the costs and benefits have been quite straightforward. Yet, the increase in price
efficiency by HFT might ultimately lead to less informative prices, even though this seems un-
intuitive at first. The underlying reasoning can be traced back to Grossman and Stiglitz (1980,
p. 404). They argue, that a market participant is only willing to acquire costly information
if he is being compensated for it. However, as seen in section 5.2.1, HFT is seemingly able
to anticipate informed order flow. This enables partial extraction of information-based rent
at the expense of originally informed traders (Stiglitz, 2014, p. 7). The reduced profitability
of private information will eventually decrease the incentive for market research. Finally, the
total amount of fundamental information in prices will diminish despite enhanced efficiency.
Therefore, price efficiency and price informativeness are ultimately distinct concepts (Weller,
2017, p. 1).
This idea has been inherited by the theoretical model of Baldauf and Mollner (2015) and
adapted into the realm of fast trading. In their model, analysts must decide whether to
acquire costly information (i.e. their individual research intensity), which can be traded on
subsequently, whereas incentive to do so strongly depends on the expected payoff (p. 10).
The model assumes that increasing speeds enhance the ability to anticipate informed order
flow through latency arbitrage30 (p. 14). Order-anticipation deters the analysts rent twofold.
First, HFBs are able to free-ride on the private information. Second, HFMs can adjust their
quotes ahead of the analyst’s informed market orders thus reducing the price impact. This is
equivalent to the prevention of adverse selection. Hence, Baldauf and Mollner (2015, pp. 15–
16) argue, that HFT decreases the incentives to acquire fundamental information through both
mechanisms. In conclusion, the effect of HFT must be contemplated with caution. On the
one hand, it allows faster incorporation of already existing information. On the other hand,
it deters overall research intensity in financial markets (p. 18).
Unfortunately, empirical evidence on the acquisition of fundamental information is scarce,
as it is hardly quantifiable. One exception to this, however, was produced by Weller (2017)
using a rather long-term approach with AT. For his analysis, he defined the price-jump ratio
in security i as a measure for information incorporation around public announcement t at date
T (p. 10–11), which is defined as:
CARTit−1,T +b
PJRit = (16)
CARTit−a,T +b
Equation (16) is essentially a ratio between two cumulative abnormal returns (CARs). Ab-
normal returns were calculated in relation to expected returns by means of the Fama-French
three-factor model. More specifically, a > 0(b ≥ 0) captures the observation period be-
fore(after) the announcement at T . According to the definition in eq. (16), the numerator
represents the total post-announcement CAR, whereas the CAR over the entire period is in
the denominator. Consequently, a higher price-jump ratio indicates that most information is
incorporated after the news release, thus the price ”jumps”. Conversely, low price-jump ratios
imply significant amounts of private information even before the official announcement. In
this case, prices are adjusted more smoothly. As proxy for AT, Weller (2017, p. 16) different
indicators31 presented in fig. 2 earlier with microsecond quote data across venues provided by
MIDAS. Therefore, it might have certain distortions due to timing issues. However, this can
still be seen as a relatively accurate measure, as the AT proxies are very rough and do not
rely on the exact ordering of messages in any way.
Interestingly, Weller (2017, pp. 18–19) found significant evidence for a positive relation
between price-jump ratio and AT. Hence, more information is incorporated at the announce-
ment date in the presence of algorithms. However, the direction of causality has not yet been
30
See section 4.2.2
31
The measures in question are the ratio of odd-lot trading, trade-to-order ratio, cancel-to-trade ratio and
average trade size. These proxies should encompass both aggressive and passive AT strategies.
resolved. For this purpose, all AT proxies are instrumentalized in a two-stage regression us-
ing various exogenous control variables (p. 24). The results are robust, which leads to the
interpretation of AT reducing the level of overall price informativeness.
Clearly, these findings are not entirely representative of HFT, as it is only a subset thereof.
However, HFT and agency algorithms follow similar mechanisms, as both types appreciate
latency advantages which allows for quick exploitation of profitable trading opportunities at
the cost of other traders. Therefore, it is very likely that HFT substantially contributes to the
deterioration of price informativeness.
6.4 Short-term Volatility: Information or Noise?

Finally, this section will address item [iii] presented at the beginning of section 6: The concept
of excessive short-term volatility through misinterpretation of signals. Importantly, it focuses
on the combined effect of both passive and aggressive HFT. It is a consensus among academic
literature, that volatility and price efficiency are positively related, as argued by Stiglitz (2014,
p. 8). This makes intuitive sense, since the incorporation of new information will necessarily
move prices. The higher the frequency and magnitude of these adjustments, the higher the
stock volatility. This statement has been proven in a simple fashion by Boehmer et al. (2015,
pp. 25, 39). They calculated the Spearman Rank correlation between return autocorrelations
and various measures of volatility, which resulted in significantly negative values. Hence, price
efficiency and volatility are complements.
Consequently, HFT is obviously somehow contributing to increasingly volatile prices. For
this reason, there have been many studies linking both fast trading and volatility by simple
regression32 , but these methods is plagued by endogeneity issues. It is not clear whether the
level fast trading has an influence on volatility or vice-versa. Fortunately, in few instances,
the causality between both factors has been taken into account. For example, Brogaard (2010,
pp. 42, 74) used the 2008 short-sale ban as an exogenous event. He argues that the restriction
is especially harmful for HFT strategies due to frequent crossings of the flat inventory. He finds
a weak increase in volatility after the event and infers HFT to have a mitigating impact on
volatility. However, this approach is quite dangerous, since the short-sale ban was implemented
due to fragile markets in the first place. It may very well be that volatility was primarily driven
by other factors instead of HFT. Similarly, Hagströmer and Nordén (2013) use an increase in
minimum tick-size as exogenous event for increased HFM activity. They find that larger tick-
sizes significantly decrease volatility and thus infer a positive effect of passive HFT strategies.
However, it is foreseeable that larger tick-sizes entail less frequent quote updates as discussed
in section 5.4, which implies lower volatility. Hence, both of the approaches above are flawed
in some way. One event study that seems more reasonable was done by Gai et al. (2013,
pp. 5, 25, 45) on the basis of two technological shocks, which decreased minimum latencies
from microseconds to nanoseconds on NASDAQ. Obviously, such a difference only matters to
extremely fast traders, i.e. HFTs, whereas it has practically no impact on all other market
participants. Interestingly, the authors found a slight but significant increase in volatility
32
An example for this are Hasbrouck and Saar (2013, pp. 665–666), who find a negative relation between
high-minus-low ranges over 10-minute intervals and AT activity.
after the upgrades. Therefore, a negative impact of HFT on volatility is concluded. Likewise
Boehmer et al. (2015, p. 41) found the introduction of co-location to significantly increase
volatility measures at several venues.
For the further investigation of this problem, it is therefore sensible to assess volatility more
distinctively. According to Carrion (2013, p. 705), Boehmer et al. (2015, p. 25) and Brogaard et
al. (2017, p. 12), there must be clearly distinguished between fundamental volatility and excess
volatility. Whereas the former contains new information and has a permanent price impact,
the latter represents transitory noise which is quickly reversed. Obviously, the minimization
of transitory noise is desirable for market quality.
The influence of HFT on permanent price impact and transitory noise has already been
touched on in section 6.1.4 using the state-space methodology by BHR (2014). They found
evidence that overall HFT trades into the same(opposite) direction as price innovation(noise).
This implies a desirable contribution to the reduction of noise. In confirmation of this pre-
sumption, similar results were presented by Brogaard et al. (2017). For their analysis, they
labeled the top 99.99 quantile of per-second returns as extreme price movements (EPMs) and
analyzed the trading activity of HFT around these events. Figure 20 shows that aggressive
nHFT is most active with aggregated HFT trading against the jump. This is true regardless
of whether the EPM is transitory or permanent. Even more interestingly, by means of probit
regressions Brogaard et al. (2017, pp. 23, 37) find EPMs to be significantly less likely following
high HFT activity, particularly when jumps are transitory. In contrast, permanent EPMs re-
main unaffected. This is strong evidence for HFTs detaining unnecessary noise while allowing
for permanent price innovations. In contrast to this, Boehmer et al. (2015, pp. 24, 38) came to
different results. They regressed volatility on a lagged AT proxy and found a positive relation,
even after controlling for lagged autocorrelation. This rather implies an increasing effect of
AT on noise.
In conclusion, HFT plays a significant role in the process of price discovery. It trades
against transitory noise but into the direction of new information. Therefore, it trades against
undesirable volatility. This leads to markets being efficient by a variety of measures. However,
HFT will prevent analysts from the costly acquisition of new information, due to anticipation
of their order flow. This effect will deter absolute price informativeness in financial markets,
even though the available information is incorporated more efficiently. In terms of HFT impact
on volatility, academic literature is quite ambiguous without definite verdict.
7 LIQUIDITY AND TRANSACTION COSTS 51
7 Liquidity and Transaction Costs

Besides price discovery, the second main purpose of financial markets is the provision of liquid-
ity (O’Hara, 2003, p. 1338), since it allows for efficient asset allocation by reduction of trans-
action costs. However, liquidity is easy to recognize but hard to define. There are countless
measures for different aspects of liquidity but none of them is universally accepted (Krishna-
murti, 2009, p. 25). Furthermore, there are potentially large qualitative differences between
those measures. Therefore, this section will provide a nuanced perspective over different facets
of liquidity and the influence exerted by algorithmic market participants.
Notably , not all types of HFT have the same effect on liquidity (Breckenfelder, 2013,
pp. 8–9). On one hand, competition between passive market makers is generally viewed as
beneficial to long-term investors. On the other hand, competition between liquidity-demanding
strategies has detrimental consequences for the broad public. The implications of aggressive
HFBs have already been explicitly discussed in section 6. Therefore this section will focus
particularly on liquidity provision by HFMs and its effects on others.
7.1 Evolution in Basic Liquidity Measures

As mentioned before, there is a variety of ways to measure liquidity. Nevertheless, many
academic studies focus their attention on easily accessible measures, such as quoted/effective
spreads as well as depth. Both of these proxies can be obtained by simple reconstruction of the
limit order book. This can be achieved with any TAQ dataset and is thus a popular approach.
However, it may not be appropriate from the perspective of large institutions, who want to
trade under the radar. This section will be held rather short in comparison with the amount
of available literature for two reasons. First, the results are almost homogenous across studies
and second, they have already been well predicted through theoretical literature.
7.1.1 Aggregated Effect of HFT
Angel et al. (2015, p. 4) have shown that a massive decrease in spreads coincided with the rise of
both AT and HFT. Even though it is harder to find evidence for a causal link, this phenomenon
can be explained by Glosten and Milgrom (1985, p. 97) and Glosten (1987, p. 1295). They
claim the level of adverse selection of market makers to be reflected in spreads. This logic
was already introduced in section 5.1, where the price impact is a component of the effective
spread. Therefore, a reduction in adverse selection would lead to a decrease in trading costs,
hence more liquidity. As discussed in section 5.3, fast technology allows HFMs more agility in
terms of quote updates which entails a reduction of quoted spreads and effective spreads. In
this regard, Conrad et al. (2015, p. 282) found stocks with high quote update frequency to have
lower spreads. Even though this indicates lower adverse selection costs for market makers, the
results have to be regarded with caution. Namely, highly traded stocks are obviously updated
more frequently, as there are more messages incoming. At the same time, their spreads tend
to be narrower. Thus, the coherence does not necessarily have to be caused by HFMs.
In addition to the two spread measures, this section will also draw attention to quoted
depth around the midpoint, which is defined as the limit order volume at or close to the
NBBO. Brogaard (2010, p. 72) found nHFT to post significantly larger executable volume
than HFT. He does this by separately removing all HFT(nHFT) limit orders from the book,
which results in lower(higher) price impact following a market order of given size. However,
despite the lower absolute depth, the share of HFT depth increases with proximity to the
NBBO (Deutsche Bundesbank, 2016, p. 54). This is consistent with section 5.3, where nHFT
are forced to post wider spreads to protect themselves against adverse selection. effect is also
visualized in fig. 13. Conclusively, all quoted spread, effective spread and quoted depth are
very closely related.
Causality between fast trading and liquidity was also confirmed in practice by the majority
of studies. For instance, Boehmer et al. (2015, p. 36) found a negative relationship between
their AT proxy33 and effective/quoted spreads. To claim causality, they used the introduction
of co-location services at various trading venues as exogenous events. In fact, these upgrades
and AT seem to have an overall reducing effect on spreads. A similar event was chosen by
Brogaard et al. (2015, pp. 3424–3426), namely a co-location upgrade on NASDAQ OMX. In
succession of the upgrade, liquidity measures such as depth within 0.5% of the NBBO, quoted-
and effective spreads all improved. In contrast, depth at the NBBO slightly decreased. Lastly,
Hendershott et al. (2011, pp. 20–21) examined the impact of Autoquote, a new system on
the NYSE which enabled the automation of quote updates. In accordance with the previous
studies, they found a significant decrease(increase) in spreads(depth).
All of the studies in the previous paragraph have good reason to chose such latency up-
grades. It directly influences low latency traders in a positive way34 , whereas slower partic-
ipants are not directly affected. Moreover, these events are implemented by the ETPs and
can thus be considered exogenous from the trader’s perspective. It can be said with high
confidence, that AT has a positive causal effect on spreads and depth in aggregation.
7.1.2 Conditional on Trading Strategy
So far, section 7.1.1 only discussed the aggregated effect of fast trading onto liquidity measures.
However, it has been stressed frequently over the course of this thesis that aggressive HFBs and
passive HFMs have fundamentally different strategies. Therefore, their presence in limit order
books will differ. Much of the analysis in section 5.4 is based on the fact that aggressive(passive)
traders prefer low(high) spreads, because it makes trading less expensive(more profitable).
This logic has also be proven to be correct empirically. Zhang and Riordan (2011, p. 10),
Carrion (2013, pp. 700–701) and Hendershott and Riordan (2013, p. 1019) followed very similar
methodologies for this reason. All of them found aggressive HFT orders to be significantly
less common whilst spreads are high. In the meantime, the opposite has been shown for HFT
limit orders. It can thus be interpreted, that HFT tend to supply liquidity while it is scarce
and consume when it is abundant. Hence, they play a stabilizing role in the supervision of
liquidity.
33
See section 6.2.1
34
Brogaard, Hendershott, Hunt, and Ysusi (2014, p. 360) find evidence that the market share of HFT
significantly increases around four latency-reducing technology upgrades on the LSE.
7.2 Phantom Liquidity

The results in section 7.1 seem quite in favor of automated trading. However, it must be kept
in mind that all of these studies relied on simple order book variables. Therefore, they are
just valid for a short amount of time and not robust to major disruptions within the market.
For instance, HFTs certainly have the ability to withdraw all the liquidity on request owing
to low latency. In turbulent markets the displayed liquidity may not actually be as available
as it seems. For this reason, the term phantom liquidity is often utilized in combination
with HFMs. This concept will be elaborated throughout the remainder of section 7. The
following two subsections will provide insights on how HFMs react to different circumstances.
Section 7.2.1 examines the effect of significant events, whereas section 7.2.2 concentrates on
general market conditions.
7.2.1 Transaction Costs for Institutional Traders
So far, the only measures of transaction cost discussed were depth, quoted- and effective spread.
These measures may be accurate for small and uninformed liquidity traders, who are overall
randomly engaging in stock markets with respect to future price changes. In contrast, this
might not be the case for institutional traders (ITs). As mentioned in section 2.1, institutions
tend to split their trades into numerous child orders, which are being executed throughout the
trading day. Active HFMs might be able to detect such patterns and adjust their quotes to the
disadvantage of ITs. On the one hand, HFMs are raising transaction costs for such informed
trades. On the other hand, they embody an acceleration of the price discovery process. This
trade-off has already been discussed in section 6.3. It must be kept in mind that this section
focuses only on isolated events, such as large trades.
To analyze this empirically, academic literature attempted to measure transaction costs of
large trades scattered throughout the trading day. One measure of choice is the implementation
shortfall (IS), which is also known as execution shortfall in literature. Even though most
studies use slightly different definitions, the core idea is always identical. The IS measures
the difference between initial price and ultimate execution prices of all child orders from the
perspective of the IT and can be defined after Brogaard et al. (2014, p. 350) as follows:
N
X Pjtn − Pj,t−
ISjt = ωjtn × Djtn × (17)
Pj,t−
n=1
Equation (17) represents the IS at day t for trader j and child order n. The respective volume
is denoted as ω and D is an indicator variable, which is 1 for buys and −1 for sales. Pjtn stands
for the execution price of child order n and Pj,t− represents the price just before the arrival of
the first child order. Note that eq. (17) calculates the IS dollar-volume but one could just as
well define it in terms of basis points. Either way, the IS ought to increase if the transaction
price rises(declines) during a large buy(sell) order.
The first study to be discussed here was done by Brogaard et al. (2014, pp. 360–264),
who used a 2SLS approach35 to analyze the coherence between latency, HFT and the market-
adjusted IS. To estimate institutional trades, the Abel/Noser dataset was used, which monitors
trading costs of 204 ITs. They used four latency-reducing technology upgrades on the LSE as
exogenous events and found significant evidence for increasing HFT activity after the upgrades,
but no effect on overall IS. The authors thus conclude HFT to thrive under low latency without
having negative effects on other market participants. However, it must be noted that every
institutional order is considered and not just the very large ones. Possibly, HFT only change
their behavior when they detect numerous child orders, which in this case is not enough to
obtain statistical significance in terms of IS. A comparable setup was used by Tong (2015)
on the Abel/Noser and NASDAQ HFT datasets. In a simple panel regression using a variety
of control variables, she found a significantly positive relationship between HFT intensity
and IS of institutional investors (p. 38). This is even true under consideration of individual
institutional trading skills in the form of firm-fixed effects (p. 42, 46). To address causality,
Tong (2015, p. 40) made use of a VAR model between HFT and IS. The results imply HFT to
be clearly Granger-causing increasing IS. Apparently, both studies above yield opposite results
despite using almost identical datasets. This may be explained by the fact that Brogaard et
al. (2014, p. 350) use the market-adjusted IS, by deducting the same-day market return which
is not shown in eq. (17). Since markets are likely to move into the direction of informed
trading, the raw IS is likely to be compensated when adjusted with market returns. Hence,
their measure might referred to as ”excess” IS. In the context addressed here, Tong (2015)
seems to be more reliable.
The next two studies explicitly aim for the aggregation of small child orders to simulate
the impact of large informed trades. First, van Kervel and Menkveld (2016) examine whether
HFMs trade against- or with the wind in presence of institutional orders. For this purpose,
the authors aggregated all child order executions per institution per stock-day to institutional
meta-orders (p. 12). They find HFMs to initially provide liquidity to institutions, i.e. trade
against the meta-order. However, after over 6(2) hours, they lean with the buy(sell) order by
trading into the same direction (p. 17). These results are visualized in fig. 21. These findings
imply HFM not being able to anticipate order flow. Instead, it recognizes the one-sided order
flow after several hours and then decides to trade into the same direction to hedge permanent
losses (p. 23). Furthermore, van Kervel and Menkveld (2016, p. 28) find with-wind trading
to be significantly increasing IS for institutions, whereas against-wind trading is beneficial.
Even though these results are intuitive, they explain a shift in profitability over the course of
a meta-order. According to fig. 21, institutional trades experience much more IS towards the
end. On another note, Korajczyk and Murphy (2015, p. 11) only examine large institutional
trades above $1 million. Similar to van Kervel and Menkveld (2016), they find HFMs to
initially lean against large orders and subsequently reverse direction. In contrast, aggressive
35
The 2SLS approach consists of a two-stage regression to find causal inferences of an instrumental variable.
In the first stage, the instrument is regressed on a set of control variables. In the second stage, the variable of
interest is regressed on the estimated instrument. For the study in question, the instrument was HFT activity
whereas the IS represents the variable of interest.
HFBs recognize the pattern even earlier, especially for sales, and trade with it (Korajczyk &
Murphy, 2015, p. 31). Using a 2SLS approach, as introduced earlier, they find a significantly
negative causal influence of HFT liquidity provision on IS (p. 46)
The last study discussed here was presented by Dobrev and Schaumburg (2017), using an
innovative non-parametric methodology based on simple TAQ data. They base their analysis
on millisecond data and match trading activity across venues given a certain offset36 (pp. 5–
7). This approach is quite advantageous, at it enables the examination of spillover effects
without requirement of HFT classification. To test for phantom liquidity, the authors matched
the execution of trades on the CME with the cancellation of quotes on BrokerTec, a large
dealer network. Strikingly, Dobrev and Schaumburg (2017, pp. 15–17) found that liquidity
is withdrawn following trades in futures markets. This is particularly true for large trades
exceeding 50 contracts. The interim period is exactly 5 milliseconds, which coincides with the
minimum latency between both venues as restricted by the speed of light. The results can be
seen in Panels B and C in fig. 23. As a results, it seems reasonable to interpret this as clear
evidence for phantom liquidity. However, the results must be consumed with greatest caution.
These findings have two major weaknesses which were omitted by Dobrev and Schaumburg
(2017). First, it only reveals that there has been at least one cancellation during the interval of
interest, but neglects the withdrawn volume. Possibly, the cancellations might have no impact
on the NBBO whatsoever. Second, the approach also does not consider resubmission of new
limit orders. It might be possible that HFMs expect a large order to arrive. Therefore, they
post limit orders within the best quotes to jump the queue and earn the spread. Overall, the
results presented in this study only provide clear evidence for almost minimal cross-market
latency of HFMs.
In summary, HFM only provides liquidity to large institutional orders up to a certain
point. Thereafter, it will trade with the wind to hedge its losses, which results in much higher
transaction costs for ITs in the form of IS. Therefore, institutions in practice care much more
about the IS and cumulative price impact instead of the half-spread in terms of transaction
cost (van Kervel & Menkveld, 2016, p. 1). This behavior cannot not be blamed on HFM only.
Conventional market makers will also eventually adjust their quotes to compensate the risk of
building inventory (Grossman & Miller, 1988, p. 618). However, the effect is still significantly
stronger during high HFM presence to the disadvantage of institutional traders.
7.2.2 Selective Liquidity Provision
Unfortunately, the results in section 7.2.1 do not provide the entire story. More specifically,
HFTs have no obligation whatsoever in terms of liquidity provision. In contrast to DMMs,
HFMs can withdraw all of their limit orders at will37 , e.g. when conditions are unfavorable.
In isolation, it may be perfectly reasonable to allow the freedom of posting quotes. However,
as discussed in section 5.3, HFM is crowding-out slower market makers due to their speed
36
The matching process is based on the occurrence of certain events during the same millisecond across
venues. For instance, if limit orders are canceled simultaneously, it is counted as a match. Obviously, the
distance between venues is associated with a minimum latency, given by the speed of light. Therefore, the
intervals are examined with offsets from −100 to +100 milliseconds.
37
See section 3.3
advantage. Particularly in periods of market distress, nHFT DMMs carry the entire cost in
the form of a winner’s curse (Han et al., 2014, p. 8). Meanwhile, HFMs only make markets
more liquid when market making conditions are easy. Empirical analyses to these concerns are
presented below. In contrast to section 7.2.1, this section only covers the reaction to general
market conditions.
To begin with, Korajczyk and Murphy (2015), who were already mentioned briefly in
section 7.2.1, expanded their research design through the inclusion of DMMs. Moreover, the
authors defined stressful trades as trades which make up a particularly large fraction of total
daily volume during a given stock-day (p. 13). It must be kept in mind that these trades
are split into countless child orders and disseminated throughout the trading day. This is
certainly a complication for inventory management. It is harder to offset inventory when a
large institution is exerting one-sided pressure. As it shown in fig. 22, HFT is recording much
less activity during stressful institutional trades. Moreover, they are trading more with the
wind in these situations (pp. 32–33). For example, unlike when the trade is non-stressful,
HFT will buy much more than it sells during stressful institutional buys. The reverse is
true for institutional sells. Furthermore, Korajczyk and Murphy (2015, pp. 40–41) show
that stressful trades coincide with lower(higher) liquidity provision of HFTs(DMMs). This
effect is reinforced for high volume stocks. This finding indicates that DMMs must provide
liquidity under unfavorable trading conditions as HFMs reduce their contribution. To ensure
the robustness of stressful trades, the authors additionally defined stressful days as days with
unusually large open-to-close returns (p. 18), which coincide with higher volatility. Indeed,
HFMs still provide less liquidity throughout stressful days (Korajczyk & Murphy, 2015, p. 44).
A very similar approach can be found in Boehmer et al. (2015, p. 27). However, they define
market making to be hard when there are two consecutive days with extreme one-sided buying-
or selling pressure. Nevertheless, this definition is quite similar to the stressful trades explained
above. Subsequently, they found AT to provide less liquidity and increase effective spreads
compared to normal days (Boehmer et al., 2015, p. 40). According to both of these studies,
HFMs seem to withdraw from the market under tougher conditions. Thus these results are in
line with the apprehensions at the beginning of this section.
In their study, Korajczyk and Murphy (2015) only defined stress if trading activity in a
given stock-day is concentrated in one institution. However, there are also other indicators
which may make market making inconvenient. One often discussed factor in academia is
intraday volatility. Unfortunately, academic literature is in constant disagreement about the
exact interaction. Even theoretical models are leading to contradicting conclusions. On the
one hand, there are models such as Foucault (1999, p. 101). He claims that volatile markets
would increase spreads which, in turn, makes limit orders more profitable. Thus, HFMs should
be more active in periods of high volatility. On the other hand, Aı̈t-Sahalia and Saǧlam (2014,
pp. 23–24) see volatility as an additional threat of being picked off by HFBs, who exploit
arbitrage opportunities. More specifically, there are more frequent changes in quotes which
entails the risk of limit orders becoming stale. A similar argument claims that inventory
management is much more difficult in volatile markets. HFMs would thus take advantage of
market orders to offset positions immediately. As a consequence, HFMs would decrease their
activity in volatile markets.
Similar to the theoretical concepts, empirical literature is also divided on this issue. For
instance, BHR (2014, p. 2294–2295) recalibrate the state-space model from section 6.1.4. They
attempt to find different patterns between days with high permanent volatility (i.e. much
information) and normal days. Indeed, HFM appears to be providing even more liquidity to
permanent price changes than usual. Thus, there is even larger willingness by HFMs to provide
liquidity against the wind. Furthermore, Brogaard et al. (2017, p. 31) yielded comparable
results. As elaborated in section 6.4, the authors use EPMs to examine the behavior of HFT.
They find evidence for HFMs providing by far the most liquidity during such extreme events.
However, this is only true if the EPM is idiosyncratic. As soon as multiple EPMs occur at the
same time or in short succession, HFMs will not provide liquidity but reverse trading into the
direction of the price change. Thus, HFMs are unwilling to absorb systemic shocks.
In contrast to this, several studies allege HFM to withdraw liquidity in times of high
volatility. Brogaard (2010, p. 51) provides evidence for decreasing liquidity provision by HFT
during periods of elevated volatility. Instead, HFT is rather demanding liquidity. The results
are visualized in fig. 24. In addition to this, Menkveld (2013, pp. 734–736) examines HFM
activity by time of the day. He finds HFM to increase their activity after the intense phase
of price discovery in the morning has ended. This indicates that HFMs tend to avoid volatile
times during the day.
All the contradicting studies can be quite frustrating, but it accurately reflects the dis-
agreement among academics on this topic. However, there is one final study by Anand
and Venkataraman (2016) which exactly addresses behavioral differences between ELPs and
DMMs. Importantly, the majority of HFMs are not obliged to provide liquidity and are thus
labeled ELPs as in section 3.3. The authors have access to a TSX dataset with distinct user
identifiers as well as DMM indicators. They use propensity scores to identify normal traders
closely imitating the behavior of DMMs. Such traders voluntarily provide liquidity without
any obligation to do so and are subsequently defined as ELPs (pp. 331–332). Throughout
their study, the authors distinguish between situations in which both DMMs and ELPs are
active as opposed to only DMMs. Importantly, Anand and Venkataraman (2016, pp. 338, 342)
claim two main findings. First, ELPs are generally much more active during periods of high
volatility. However, spreads tend to be lower during phases of ELP activity, which is possibly
due to elevated competition among market markets. Figure 25 illustrates these points graph-
ically. Second, ELPs are only present when market making profits are higher, thus liquidity
provision is easier. DMMs realize higher profits in the presence of ELPs despite the additional
competition. Most strikingly, DMMs cross the zero position much more often and are able to
maintain lower inventory during high-profit phases. This implies a lack of directional pressure
within the order book. This fact becomes also visible in Panel B of fig. 25, where ELPs are
only active at lower levels of order imbalance. Apparently, inventory management is more
important to ELPs than volatility for the decision of whether to enter a market.
In conclusion, the results of both Korajczyk and Murphy (2015) and Anand and Venkatara-
man (2016) seem to be the most reasonable due to well-considered research design. Both imply
that HFMs in fact prefer transitory noise. However, they actively avoid one-sided pressure
due to aggravated inventory management. Hence, HFMs take advantage of their option to
withdraw liquidity provision at will, creating phantom liquidity. Under certain circumstances,
this may eventually lead to severe consequences for market stability.
7.3 Systemic Risk and Flash Crashes

The previous section discussed phantom liquidity, more precisely the permission of HFMs to
withdraw limit orders at will. In combination with the crowding-out effect on traditional mar-
ket makers, HFT might introduce new fragility into financial markets (Foucault, 2012, p. 24).
This effect is especially severe if HFT strategies are correlated. In a theoretical model, Jarrow
and Protter (2012, p. 4) assume HFTs to react in unison with respect to the same signals. As
mentioned earlier, an HFT receives much of his information from the order flow itself, which
can be seen by all other HFTs as well. Therefore, all HFTs can collectively be viewed as one
single enormous trader. Consequently, their actions are unlikely to be clearly distinguishable.
In case of HFMs adhering to such correlated strategies, a coordinated withdrawal of liquidity
can lead to severe market failure.
One prominent example of such a catastrophe is the Flash Crash on May 6, 2010. According
to CFTC-SEC (2010, p. 2), the crash was caused by a large algorithmic sell order of around $4.1
billion in S&P500 Futures. The order provoked a draining of buy-side limit orders which lead
to cross-security contagion and a price drop of 55%(20%) in S&P500 Futures(ETFs) within
less than 30 minutes. Interestingly, this entire breakdown was merely initiated by one-sided
selling pressure. Ever since, regulators have been genuinely concerned about the prophylaxis
of similar events in the future. Therefore, this section aims to provide an overview over the
empirical evidence for correlation in HFT strategies as well as the potential fragility associated
with it.
To begin with, Chaboud et al. (2014, pp. 2055–2058) examined trading patterns of ATs in
three currency pairs. They attempted to find out whether the independent matching assump-
tion holds, i.e. whether ATs and human traders are matched randomly. Interestingly, they
found both groups to be trading less with participants from the same group. In other words,
ATs are more likely to engage with a human trader instead of a fellow AT. These results hold
for both liquidity demanding and supplying trades. Hence, AT strategies seem to be positively
correlated as they pursue the same objectives. This notion is reassured by Hendershott and
Riordan (2013, p. 1013) who analyze the conditional probabilities following a predefined mes-
sage. Therefore, they examine a proprietary dataset from Xetra message-per-message. For
instance, they compare the observed likeliness of two consecutive AT trades with the expected
probability derived from the independent matching assumption. The results clearly show that
AT as well as human trading occur in clusters and are not randomly ordered. This is another
indication of correlated strategies.
A somewhat different approach was chosen by Anand and Venkataraman (2016, p. 335).
As already discussed in section 7.2.2, they analyzed the trading behavior of ELPs. More specif-
ically, they calculated the current level liquidity provision relative to the individual average
per ELP. This allows the authors to interpret periods of below-average liquditiy provision as
ELPs withdrawing from the market. In the next step, they regressed the individual liquidity
contribution onto the contemporaneous aggregated ELP activity in the market. Apparently,
75% of ELPs have a significantly positive coefficient with their peers. This implies that they
tend to withdraw from the market simultaneously as they respond to the same signals. Hence,
such strategies are of systemic relevance.
Apparently, academic literature is consistent in terms of correlated behavior. Despite
the heterogeneity of HFT strategies, they all seem to be quite homogeneous within. This
is particularly true for liquidity providing HFMs. Therefore, HFT can be seen as a potential
source of fragility. To understand this more clearly, Kirilenko, Kyle, Samadi, and Tuzun (2017)
present an in-depth analysis of intermediaries during the Flash Crash. For their research
design, the authors used the E-mini dataset38 , which includes all messages on the S&P500
Futures with individual trader identification. Unlike equities, these securities do not have
DMMs, which forces the authors to approximate intermediaries as accounts with consistently
low inventories. The 16 most active of which are automatically defined as HFM, whereas the
remainder is reported as market makers (pp. 976–977). Using this classification, Kirilenko
et al. (2017, p. 984) initially found both market makers and HFMs trading against the price
decrease and thereby absorbing the early shock. Thereafter, HFMs started to offset their
position by trading into the direction of the price drop. Thus, they withdrew liquidity from
the market in order not to keep their inventories low. Even under extreme conditions, HFMs
were unwilling to build large positions.
These findings are also in line with the official explanation delivered by CFTC-SEC (2010).
Therefore, it is often concluded that HFMs did not initiate the Flash Crash, as they initially
provided liquidity during the intense selling-pressure. However, they did exacerbate the price
drop by eventually trading with the wind. Similar behavior has already been documented
in other markets by Korajczyk and Murphy (2015) and van Kervel and Menkveld (2016) as
explained in section 7.2. In conclusion, phantom liquidity seems to be a serious threat for
stability, since markets appear to be much more liquid than they actually are.
38
See section 2.3.2
8 CONCLUSION 60
8 Conclusion
HFT is certainly a complex topic with countless nuances. It is a recent phenomenon which
has emerged within the last decade. HFT did no appear out of thin air but was the logical
consequence of technological innovations and new regulations that shaped electronic market
structure. Especially RegNMS and MiFID paved the way for this unprecedented trader. Over-
all, HFT can process large amounts of information almost instantly and has the ability to trade
huge volumes. Latencies of HFTs are currently on the level of microseconds which might de-
crease even further in the future. Nevertheless, there is high heterogeneity in HFT behavior,
but it is mainly distinguished between aggressive and passive strategies.
Regardless of HFT strategy, speed poses an essential advantage. On the one hand, aggres-
sive HFB will use this feature to eliminate arbitrage opportunities across markets or securities.
This can be done by exploitation of stale limit orders. Moreover, it has also been found that
HFBs have the ability to anticipate order flow of others, which allows for the adaption of
trading pattern beforehand. On the other hand, passive HFMs can adjust their quotes more
quickly than conventional market makers. Hence, other market participants are put at a
disadvantage in both cases. This is especially true for nHFT liquidity providers.
Obviously, such behavior has important implications on overall market quality, which has
been the core topic of this thesis. In terms of price efficiency, the presence of HFT poses
a significant enhancement. This is particularly true for aggressive HFB strategies, since the
exploitation of arbitrage opportunities has a large impact on the price discovery process.
Unfortunately, this also entails increasing adverse selection for other market participants and
makes them less likely to post competitive quotes. Even though HFT enhances price discovery
through a direct channel, it is negatively influences the total amount of informativeness through
a second indirect channel. Order anticipation strategies impose a free-rider problem onto
informed traders. As the incentives for them to acquire costly information decrease, the
absolute amount of informativeness will too.
The increasing adverse selection also impacts market liquidity, whereas conventional market
makers are victim to a crowding-out effect: They are not fast enough to adjust stale limit
orders. In contrast, HFMs are able to post much more competitive quotes, as they are subject
to less adverse selection. This development lead to an overall decrease in transaction costs.
However, the fact that HFMs have no obligation to supply liquidity entails potential market
fragility. It has been shown that HFMs only supply liquidity when conditions are pleasant, i.e.
when there is no one-sided pressure and inventory management is easy. Moreover, algorithmic
strategies have been shown to be correlated and thus systemically relevant. Taken together,
these findings imply that HFT has the potential to exacerbate a market collapse under stress,
such as during the Flash Crash.
For this reason, regulators need to find the right tools to keep the advantages of HFT while
averting potential harm. Several tools have been discussed in this thesis to which HFT reacts
very sensitively. Such measures include the adaption of fee structures on trading venues or
minimum tick sizes. Unfortunately, HFT affects financial markets in many different ways and
there is a lot more research to be made before the implementation of new regulations.
REFERENCES 61
References
Aı̈t-Sahalia, Y., & Saǧlam, M. (2014). High frequency market making: Taking advantage
of speed (NBER Working Paper No. w19531). National Bureau of Economic Research.
Retrieved from https://ssrn.com/abstract=2342011
Anand, A., & Venkataraman, K. (2016). Market conditions, fragility, and the eco-
nomics of market making. Journal of Financial Economics, 121 (2), 327–349.
doi:10.1016/j.jfineco.2016.03.006
Angel, J. J., Harris, L. E., & Spatt, C. S. (2011). Equity trading in the 21st century. Quarterly
Journal of Finance, 1 (1), 1–53. doi:10.1142/S2010139211000067
Angel, J. J., Harris, L. E., & Spatt, C. S. (2015). Equity trading in the 21st century: An
update. Quarterly Journal of Finance, 05 (01), 1–39. doi:10.1142/S2010139215500020
Aquilina, M., & Ysusi, C. (2016). Are high-frequency traders anticipating the order flow?
cross-venue evidence from the uk market. Market Microstructure and Liquidity, 2 (3&4),
1–35. doi:10.1142/S2382626617500058
Arnuk, S., & Saluzzi, J. (2009). Latency arbitrage: The real power behind predatory high
frequency trading (White Paper). Themis Trading LLC. Retrieved from https://pdfs
.semanticscholar.org/0f6d/8a58f1ec09fc107e7df18f786b55605c5af5.pdf
Avramovic, A. (2012). Manufacturing volume: The stock split solution (Market Commentary).
Credit Suisse. Retrieved from https://edge.credit-suisse.com
Baldauf, M., & Mollner, J. (2015). High-frequency trading and market performance (Working
Paper). Retrieved from https://ssrn.com/abstract=2674767
Battalio, R., Corwin, S. A., & Jennings, R. (2016). Can brokers have it all? On the relation
between make-take fees and limit order execution quality. The Journal of Finance, 71 (5),
2193–2238. doi:10.1111/jofi.12422
Benos, E., & Sagade, S. (2016). Price discovery and the cross-section of high-
frequency trading. Journal of Financial Markets, 30 (Supplement C), 54–77.
doi:10.1016/j.finmar.2016.03.004
Biais, B., & Foucault, T. (2014). HFT and market quality. Bankers, Markets & Investors,
128 (1), 5–19.
Biais, B., Foucault, T., & Moinas, S. (2015). Equilibrium fast trading. Journal of Financial
Economics, 116 (2), 292 - 313. doi:10.1016/j.jfineco.2015.03.004
Boehmer, E., Fong, K. Y., & Wu, J. J. (2015). International evidence on algorithmic trading
(AFA 2013 San Diego Meetings Paper). doi:10.2139/ssrn.2022034
Breckenfelder, J. H. (2013). Competition between high-frequency traders, and market quality
(Working Paper). European Central Bank. NYU Stern Microstructure Meeting 2013.
doi:10.2139/ssrn.2264858
Brogaard, J. (2010). High frequency trading and its impact on market quality (Working
Paper). Northwestern University Kellogg School of Management. Retrieved from http://
ssrn.com/abstract=1641387
Brogaard, J., Carrion, A., Moyaert, T., Riordan, R., Shkilko, A., & Sokolov, K.
(2017). High-frequency trading and extreme price movements (Working Paper).
doi:10.2139/ssrn.2531122
REFERENCES 62
Brogaard, J., Hagströmer, B., Nordén, L., & Riordan, R. (2015). Trading fast and
slow: Colocation and liquidity. The Review of Financial Studies, 28 (12), 3407-3443.
doi:10.1093/rfs/hhv045
Brogaard, J., Hendershott, T., Hunt, S., & Ysusi, C. (2014). High-frequency trading
and the execution costs of institutional investors. Financial Review , 49 (2), 345–369.
doi:10.1111/fire.12039
Brogaard, J., Hendershott, T., & Riordan, R. (2014). High-frequency trading and price
discovery. The Review of Financial Studies, 27 (8), 2267–2306. doi:10.1093/rfs/hhu032
Brogaard, J., Hendershott, T., & Riordan, R. (2016). Price discovery without trading: Evidence
from limit orders (Working Paper). doi:10.2139/ssrn.2655927
Budish, E., Cramton, P., & Shim, J. (2015). The high-frequency trading arms race: Fre-
quent batch auctions as a market design response. The Quarterly Journal of Economics,
130 (4), 1547–1621. doi:10.1093/qje/qjv027
Carrion, A. (2013). Very fast money: High-frequency trading on the NASDAQ. Journal of
Financial Markets, 16 (4), 680 - 711. doi:10.1016/j.finmar.2013.06.005
Chaboud, A. P., Chiquoine, B., Hjalmarsson, E., & Vega, C. (2014). Rise of the machines:
Algorithmic trading in the foreign exchange market. The Journal of Finance, 69 (5),
2045–2084. doi:10.1111/jofi.12186
Chordia, T., Roll, R., & Subrahmanyam, A. (2008). Liquidity and market efficiency. Journal
of Financial Economics, 87 (2), 249 - 268. doi:10.1016/j.jfineco.2007.03.005
Commodity Futures Trading Commission–Securities and Exchange Commission. (2010). Find-
ings regarding the market events of may 6, 2010 (Report). The Joint Advisory Commit-
tee on Emerging Regulatury Issues. Retrieved from https://www.sec.gov/news/studies/
2010/marketevents-report.pdf
Conrad, J., Wahal, S., & Xiang, J. (2015). High-frequency quoting, trading,
and the efficiency of prices. Journal of Financial Economics, 116 (2), 271–291.
doi:10.1016/j.jfineco.2015.02.008
Deutsche Bundesbank. (2016). Bedeutung und wirkung des hochfrequenzhandels
am deutschen kapitalmarkt (Monatsberichtsaufsatz Oktober 2016). Retrieved
from https://www.bundesbank.de/Redaktion/DE/Downloads/Veroeffentlichungen/
Monatsberichtsaufsaetze/2016/2016 10 hochfrequenzhandel.pdf
Ding, S., Hanna, J., & Hendershott, T. (2014). How slow is the nbbo? a comparison with
direct exchange feeds. Financial Review , 49 (2), 313–332. doi:10.1111/fire.12037
Dobrev, D., & Schaumburg, E. (2017). High-frequency cross-market trading: Model free
measurement and applications (Working Paper). Retrieved from https://editorialexpress
.com/cgi-bin/conference/download.cgi?db name=SNDE2017&paper id=226
Easley, D., O’Hara, M., & Yang, L. (2016). Differential access to price information in fi-
nancial markets. Journal of Financial and Quantitative Analysis, 51 (4), 1071–1110.
doi:10.1017/S0022109016000491
Egginton, J. F., Van Ness, B. F., & Van Ness, R. A. (2016). Quote stuffing. Financial
Management, 45 (3), 583–608. doi:10.1111/fima.12126
REFERENCES 63
European Union. (2004). Directive on markets in financial instruments (Directive

2004/39/EC). Retrieved from http://eur-lex.europa.eu/LexUriServ/LexUriServ.do
?uri=OJ:L:2004:145:0001:0044:EN:PDF
Fama, E. F. (1970). Efficient capital markets: A review of theory and empirical work. The Jour-
nal of Finance, 25 (2), 383–417. Retrieved from http://www.jstor.org/stable/2325486
Federation of European Securities Exchanges. (2011). Updated FESE tick size tables as
of september 2011 (Official Document). Retrieved from http://www.fese.eu/images/
documents/UPDATED FESE TICK SIZE TABLES AS OF OCT 2012.pdf
Foucault, T. (1999). Order flow composition and trading costs in a dynamic limit order market.
Journal of Financial Markets, 2 (2), 99–134. doi:10.1016/S1386-4181(98)00012-3
Foucault, T. (2012). Algorithmic trading: Issues and preliminary evidence. In F. Abergel, J.-
P. Bouchaud, T. Foucault, C.-A. Lehalle, & M. Rosenbaum (Eds.), Market microstructure
(pp. 1–40). John Wiley & Sons Ltd. doi:10.1002/9781118673553.ch1
Foucault, T., Kadan, O., & Kandel, E. (2013). Liquidity cycles and make/take fees in
electronic markets. The Journal of Finance, 68 (1), 299–341. doi:10.1111/j.1540-
6261.2012.01801.x
Foucault, T., & Menkveld, A. J. (2008). Competition for order flow and smart or-
der routing systems. The Journal of Finance, 63 (1), 119–158. doi:10.1111/j.1540-
6261.2008.01312.x
Gai, J., Yao, C., & Ye, M. (2013). The externalities of high frequency trading (Working Paper).
doi:10.2139/ssrn.2066839
Glosten, L. R. (1987). Components of the bid-ask spread and the statistical properties of
transaction prices. The Journal of Finance, 42 (5), 1293–1307. doi:10.2307/2328528
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.
doi:10.1016/0304-405X(85)90044-3
Goettler, R. L., Parlour, C. A., & Rajan, U. (2009). Informed traders and limit order markets.
Journal of Financial Economics, 93 (1), 67–87. doi:10.1016/j.jfineco.2008.08.002
Goldstein, M. A., Kumar, P., & Graves, F. C. (2014). Computerized and high-frequency
trading. Financial Review , 49 (2), 177–202. doi:10.1111/fire.12031
Grossman, S. J., & Miller, M. H. (1988). Liquidity and market structure. The Journal of
Finance, 43 (3), 617–633. doi:10.2307/2328186
Grossman, S. J., & Stiglitz, J. E. (1980). On the impossibility of informationally efficient
markets. The American Economic Review , 70 (3), 393-408. Retrieved from http://
www.jstor.org/stable/1805228
Haferkorn, M. (2017). High-frequency trading and its role in fragmented markets. Journal of
Information Technology, 32 (3), 283–296. doi:10.1057/s41265-016-0030-6
Hagströmer, B., & Nordén, L. (2013). The diversity of high-frequency traders. Journal of
Financial Markets, 16 (4), 741–770. doi:10.1016/j.finmar.2013.05.009
Haldane, A. G. (2011). The race to zero (Speech). Beijing. International Economic Association
Sixteenth World Congress. Retrieved from http://www.bis.org/review/r110720a.pdf
REFERENCES 64
Han, J., Khapko, M., & Kyle, A. S. (2014). Liquidity with high-frequency market making
(Research Paper No. 14-06). Swedish House of Finance. doi:10.2139/ssrn.2416396
Hasbrouck, J. (1991a). Measuring the information content of stock trades. The Journal of
Finance, 46 (1), 179–207. doi:10.2307/2328693
Hasbrouck, J. (1991b). The summary informativeness of stock trades: An econometric analysis.
The Review of Financial Studies, 4 (3), 571–595. doi:10.1093/rfs/4.3.571
Hasbrouck, J. (1995). One security, many markets: Determining the contributions to
price discovery. The Journal of Finance, 50 (4), 1175–1199. doi:10.1111/j.1540-
6261.1995.tb04054.x
Hasbrouck, J. (2017). Price discovery in high resolution (Working Paper). Stern School of
Business. Retrieved from https://www.rsm.nl/fileadmin/home/Department of Finance
VG5 /LQ2017/Hasbrouck Joel.pdf
Hasbrouck, J., & Saar, G. (2013). Low-latency trading. Journal of Financial Markets, 16 (4),
646–679. doi:10.1016/j.finmar.2013.05.003
Hendershott, T., Jones, C. M., & Menkveld, A. J. (2011). Does algorithmic trading improve
liquidity? The Journal of Finance, 66 (1), 1–33. doi:10.1111/j.1540-6261.2010.01624.x
Hendershott, T., & Moulton, P. C. (2011). Automation, speed, and stock mar-
ket quality: The nyse’s hybrid. Journal of Financial Markets, 14 (4), 568–604.
doi:10.1016/j.finmar.2011.02.003
Hendershott, T., & Riordan, R. (2013). Algorithmic trading and the market for
liquidity. Journal of Financial and Quantitative Analysis, 48 (4), 1001—1024.
doi:10.1017/S0022109013000471
Hirschey, N. (2017). Do high-frequency traders anticipate buying and selling pressure? (Work-
ing Paper). London Business School. doi:10.2139/ssrn.2238516
Hoffmann, P. (2014). A dynamic limit order market with fast and slow traders. Journal of
Financial Economics, 113 (1), 156–169. doi:10.1016/j.jfineco.2014.04.002
Huang, R. D., & Stoll, H. R. (1996). Dealer versus auction markets: A paired comparison of
execution costs on NASDAQ and the NYSE. Journal of Financial Economics, 41 (3),
313–357. doi:10.1016/0304-405X(95)00867-E
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),
1–15. doi:10.1142/S0219024912500227
Jones, C. M. (2013). What do we know about high-frequency trading? (Research Paper No.
13-11). Columbia Business School. doi:10.2139/ssrn.2236201
Jovanovic, B., & Menkveld, A. J. (2016). Middlemen in limit order markets (Working Paper).
doi:10.2139/ssrn.1624329
Kirilenko, A., Kyle, A. S., Samadi, M., & Tuzun, T. (2017). The flash crash: High-
frequency trading in an electronic market. The Journal of Finance, 72 (3), 967–998.
doi:10.1111/jofi.12498
REFERENCES 65
Kohler, A., & von Wyss, R. (2012). Fragmentation in european equity markets and market
quality - Evidence from the analysis of trade-throughs (Working Paper No. 2012/10).
Swiss Institute of Banking and Finance (s/bf HSG). Retrieved from http://ux-tauri
.unisg.ch/RePEc/usg/sfwpfi/WPF-1210.pdf
Korajczyk, R. A., & Murphy, D. (2015). High frequency market making to large institutional
trades (Working Paper). doi:10.2139/ssrn.2567016
Krishnamurti, C. (2009). Introduction to market microstructure. In C. Krishnamurti &
R. Vishwanath (Eds.), Investment management: A modern guide to security analysis
and stock selection (pp. 13–29). Berlin, Heidelberg: Springer. doi:10.1007/978-3-540-
88802-4 2
Lewis, M. (2015). Flash boys: Cracking the money code. United Kingdom: Penguin Books
Ltd.
Li, W. (2017). High frequency trading with speed hierarchies (Working Paper). Johns Hopkins
University. doi:10.2139/ssrn.2365121
Litzenberger, R., Castura, J., & Gorelick, R. (2012). The impacts of automation and high
frequency trading on market quality. Annual Review of Financial Economics, 4 (1),
59–98. doi:10.1146/annurev-financial-110311-101744
Lo, A. W., & MacKinlay, A. C. (1988). Stock market prices do not follow random walks:
Evidence from a simple specification test. The Review of Financial Studies, 1 (1), 41–66.
Retrieved from http://www.jstor.org/stable/2962126
Malinova, K., & Park, A. (2015). Subsidizing liquidity: The impact of make/take fees on
market quality. The Journal of Finance, 70 (2), 509–536. doi:10.1111/jofi.12230
Malinova, K., Park, A., & Riordan, R. (2016). Taxing high frequency market making: Who
pays the bill? (Working Paper). doi:10.2139/ssrn.2183806
Menkveld, A. J. (2013). High frequency trading and the new market makers. Journal of
Financial Markets, 16 (4), 712–740. doi:10.1016/j.finmar.2013.06.006
Menkveld, A. J. (2014). High-frequency traders and market structure. Financial Review ,
49 (2), 333–344. doi:10.1111/fire.12038
Menkveld, A. J. (2016). The economics of high-frequency trading: Taking stock. Annual
Review of Financial Economics, 8 , 1–24. doi:10.1146/annurev-financial-121415-033010
Menkveld, A. J., & Zoican, M. A. (2017). Need for speed? Exchange latency and liquidity.
The Review of Financial Studies, 30 (4), 1188-1228. doi:10.1093/rfs/hhx006
Moallemi, C. C., & Saǧlam, M. (2013). OR Forum—The cost of latency in high-frequency
trading. Operations Research, 61 (5), 1070-1086. doi:10.1287/opre.2013.1165
O’Hara, M. (2003). Presidential address: Liquidity and price discovery. The Journal of
Finance, 58 (4), 1335–1354. doi:10.1111/1540-6261.00569
O’Hara, M. (2015). High frequency market microstructure. Journal of Financial Economics,
116 (2), 257–270. doi:10.1016/j.jfineco.2015.01.003
O’Hara, M., Saar, G., & Zhong, Z. (2015). Relative tick size and the trading environment
(Working Paper). doi:10.2139/ssrn.2463360
Roşu, I. (2016). Fast and slow informed trading (Working Paper). HEC Paris - Finance
Department. doi:10.2139/ssrn.1859265
REFERENCES 66
Securities and Exchange Commission. (2005). Regulation NMS (Release No. 34-51808; File
No. S7-10-04). Retrieved from https://www.sec.gov/rules/final/34-51808.pdf
Securities and Exchange Commission. (2010). Concept release on equity market structure
(Release No. 34-61358; File No. S7-02-10). Retrieved from https://www.sec.gov/rules/
concept/2010/34-61358.pdf
Securities and Exchange Commission. (2013). Equity market structure literature review part I:
Market fragmentation (Report). Retrieved from https://www.sec.gov/marketstructure/
research/fragmentation-lit-review-100713.pdf
Securities and Exchange Commission. (2014). Equity market structure literature review part II:
High frequency trading (Report). Retrieved from https://www.sec.gov/marketstructure/
research/hft lit review march 2014.pdf
Stiglitz, J. E. (2014). Tapping the brakes: Are less active markets safer and better for the
economy? (Tech. Rep.). Federal Reserve Bank of Atlanta. 2014 Financial Markets Con-
ference. Retrieved from https://www8.gsb.columbia.edu/faculty/jstiglitz/sites/jstiglitz/
files/2014 Tapping Brakes pub.pdf
Stone, G. (2014). SIP and feeds (White Paper). Bloomberg Tradebook. Retrieved from
https://www.bloomberg.com/professional/blog/sip-vs-direct-feeds-latency-rules/
Tong, L. (2015). A blessing or a curse? The impact of high frequency trading on institutional
investors (Working Paper). European Finance Association. doi:10.2139/ssrn.2330053
van Kervel, V., & Menkveld, A. J. (2016). High-frequency trading around large institutional
orders (Working Paper). doi:10.2139/ssrn.2619686
Weller, B. (2017). Does algorithmic trading reduce information acquisition? (Working Paper).
Duke University. doi:10.2139/ssrn.2662254
Ye, M., & Yao, C. (2014). Tick size constraints, market structure, and liquidity (Working
Paper). doi:10.2139/ssrn.2359000
Zhang, S. (2017). Need for speed: Hard information processing in a high-frequency world.
Journal of Futures Markets, 1–19. doi:10.1002/fut.21861
Zhang, S., & Riordan, R. (2011). Technology and market quality: The case of
high frequency trading (ECIS 2011 Proceedings). European Conference of Infor-
mation Systems. Retrieved from http://aisel.aisnet.org/cgi/viewcontent.cgi?article=
1094&context=ecis2011
A PRIMER TO LIMIT ORDER MARKETS 67
A Primer to Limit Order Markets

Limit order markets are the most prevalent design for ETPs. This section will give a brief
introduction into the trading rules in such markets. The following explanations were developed
with reference to Krishnamurti (2009, pp. 14–15).
Prices. In limit order markets, there is no single transaction price. Instead, it is distinguished
between the bid and aks prices. The bid(ask) refers to the maximum(minimum) for which the
market is willing to buy(sell) an asset. Thereby, bid and ask will always be separated by a
strictly positive bid-ask spread. The arithmetic average of bid and ask prices is referred to as
midpoint.
Orders. Fortunately, limit order markets can be explained by means of only two order types.
First, limit orders sits in the book awaiting execution. Such order consist of a quantity and a
price at which it is executed. Second, market orders only consist of a quantity, which will be
executed against the best limit orders. Specifically, a market order to buy(sell) will be executed
against the lowest(highest) ask(bid) among all standing limit orders. Since limit orders are
basically an offer for other traders, they are referred to as passive or liquidity supplying. In
contrast, market orders reduce the amount of limit orders and are thus called aggressive or
liquidity demanding.
Depth. The concept of depth is generally desirable for liquidity. It stands for the volume,
which is available at the best quotes or within a certain range of the best quotes. Markets are
said to be thick (thin) when there is large(small) limit order volume available for execution.
Priority. Most limit order markets follow price-time priority governing the execution prob-
ability of limit orders. Thereafter, passive orders are primarily sorted by price and secondarily
by time of arrival. Therefore, a large market order will first execute against all limit orders at
the best price. If two limit were posted at the best quotes, priority is given to the one with
prior arrival.
Queue Jumping. In order to improve the execution probability of limit orders, traders
can employ queue jumping strategies to get ahead of everyone else. This can be achieved
by posting a limit order inside the best quotes and thus undercutting competition. Due to
price-time priority, this order will be filled by the first incoming market order regardless of
belated arrival.
B THE STATE-SPACE FRAMEWORK 68
B The State-Space Framework

The explanations provided here are a direct continuation of section 6.1.4. The state-space
approach in context of HFT was originally provided by BHR (2014, pp. 2278–2280). For this
reason, the brief definitions below are held closely to this study. To begin with the state-space
model can be written as follows:
pi,t = mi,t + si,t (18)

mi,t = mi,t−1 + wi,t (19)
D D
wi,t = κD
i,H ∗ HF
\ T i,t + κD
i,N ∗ nHF
\T i,t + µi,t (20)
D D D D
si,t = φsi,t−1 + ψi,H ∗ HF Ti,t + ψi,N ∗ nHF Ti,t + vi,t (21)
The first two equations are identical to the variance decomposition approach following Has-
brouck (1991b, p. 574) as discussed earlier. Importantly, eq. (18) is the measurement equation
of observable logarithmic midpoints, whereas eq. (19) is the transition equation of hidden
fundamental values. The true value is modeled according to a random walk with innovations
wi,t and does not fulfill ergodicity. Hence, each wi,t has a permanent effect. As before, mid-
points consist of both a permanent component mi,t based on fundamentals and a transitory
component si,t which represents undesirable noise. Equation (20) models the permanent inno-
D
vations in the true value and is based on exceptional demand of both trader groups. HF \ T
D
\T measure surprise innovation 39 in aggressive order flow. Permanent innovations
and nHF
unrelated to trading are captured by µi,t . Lastly, eq. (21) represents the composition of the
transitory and stationary noise. In contrast to Hasbrouck (1991b), this measure is not white
noise, due to the inclusion of an autoregressive element si,t−1 . Moreover, HF T D and nHF T D
is measured in absolute order flow unlike eq. (20). With regard to the complete model, indi-
vidual contribution to price discovery can be inferred from parameters κD D
i,H and κi,N , where
a positive(negative) value indicates informed trading(adverse selection costs). Meanwhile, the
D and ψ D in reversed form.
influence on noise is analogously represented in ψi,H i,N
As defined in eqs. (20) and (21) the independent variables include all trading activity from
perspective of aggressive liquidity demand. Alternatively, the authors also estimated the same
state-space model using the passive liquidity supply of both groups (HF T S and nHF T S ).
Since each trade has both one aggressive and passive participant, both variants consider the
entire set of trades.
39
Surprise innovation of order flow estimated using a VAR(p) system containing only the order flow HFT
and nHFT as variables. The surprise innovation per group is defined as the residual. This approach aims to
account for autocorrelation in order flow. Therefore, only deviations from the expected order flow are thought
to influence price discovery
C INHERITED FIGURES 69
C Inherited Figures
Figure 9: Long-run floor-based trading activity. This plot shows the ratio of trading activity of on-floor
traders across all NYSE stocks from January 1999–May 2006. A declining ratio implies an increase in off-
floor trading. It must be kept in mind that the observation period ends before the introduction of HFT to
financial markets. This figure was taken from Hendershott and Moulton (2011, p 574).
Figure 10: Value-Weighted daily average effective spread on the NYSE. This plot shows how effective
spreads have assimilated over time. Moreover, the plot contains indications of decreases in minimum tick
size in 1997 and 2001. Obviously, lower tick sizes lead to cheaper transaction costs from trading. However, it
must be kept in mind that the observation period ends before the introduction of HFT to financial markets.
This figure was taken from Chordia, Roll, and Subrahmanyam (2008, p. 256).
Figure 11: Dislocations between direct data feeds and the SIP. The plot represents a histogram of latency
differences between the NBBO constructed from direct data feeds and the official SIP NBBO. The horizontal
axis shows time difference in milliseconds, which is defined as the difference between timestamps of the same
recorded event. A larger offset means higher latency advantage of direct feeds over the official SIP quotes.
The three panels report latency advantages from direct feeds provided by BATS, NASDAQ and Direct Edge.
This figure was taken from Ding, Hanna, and Hendershott (2014, p. 320).
Figure 12: Cumulative HFT vs. non-HFT net marketable buying. This plot shows cumulative marketable
buying and selling for stocks sorted into portfolios by HFTs net marketable buying at the one-second horizon.
This procedure is repeated for every single second, whereas the current second is denoted as t = 0. The
plot only shows the relative evolution of the highest and lowest net marketable buying decile. This figure
was taken from Hirschey (2017, p.36).
Figure 13: Order book dynamics for one security traded at Euronext Amsterdam. For both plots, the
horizontal axis is time and prices are on the vertical axis. Limit buy(sell) orders are in green(red) and current
spreads are marked in black. The left panel shows all quoting activity, whereas the right panel only captures
limit orders by HFTs. This figure was taken from Menkveld (2016, p.12).
Panel A: Negative Macro News
Panel B: Positive Macro News
Figure 14: HFT trading and portfolio returns around negative and positive macro announcements. The
net order flow is shown on the left axis, whereas value-weighted sample portfolio returns are plotted on the
right axis. Time is measured in seconds and the news announcement takes place at t = 0. Events are
negative(positive) when they underperform(overperform) compared to analysts forecasts. Additionally, the
plot shows 95% confidence bands as dotted lines. Panel A shows the reaction to negative news, whereas
Panel B shows positive news. These plots was taken from Brogaard, Hendershott, and Riordan (2014,
pp. 2296–2297).
Figure 15: Correlations between S&P500 futures and ETF at high frequencies. The correlation between
both assets is presented over different high-frequency time horizons. The correlations on the vertical axis is
calculated based on the midpoint returns of both securities. The horizontal axis shows the time interval in
milliseconds. Each line represents the results of a specific year between 2005–2011. This figure was taken
from Budish, Cramton, and Shim (2015, p. 1566).
Figure 16: Duration of arbitrage opportunities between S&P500 futures and ETF. This figure aims to show
the decrease in duration of arbitrage opportunities over time. It plots the duration of arbitrage opportunities
on the vertical axis against the proportion of mispricings exceeding it. Each line represents the results of a
specific year between 2005–2011. This figure was taken from Budish, Cramton, and Shim (2015, p. 1570).
Figure 17: Response of liquidity demand and supply to information shocks. Liquidity demanders and
suppliers react differently to new hard information. This figure plots the trading responses of both HFTs
and nHFTs. The top panel shows the aggregate effect on aggressive trading, whereas the bottom Panel
focuses on passive trading. The vertical axes show the cumulative net trading and the horizontal axes the
lags in seconds. As the plot is limited to t = 0, inferences about potential anticipation patterns cannot be
made. This plot was taken from Zhang (2017, p. 10).
Figure 18: Effect of increasing tick size on trading activity. Changes in tick size regimes obviously influence
market participants and their behavior. The plot shows how an increase in tick size affects trading within
ceteris paribus. It shows the median change on the day after the event in selected trading indicators and
is based on 89 threshold crossings. The results are based on data NASDAQ OMX Sweden in 2010–2012,
which also contains indicators of HFT. Importantly, before 2011, there was only negligible HFM activity in
Swedish markets. This plot was taken from Hagströmer and Nordén (2013, p. 762).
Figure 19: Average variance ratios by high and low update frequency. The stocks are grouped into quintiles
by both size and the number of trades, resulting in 25 groups. The stocks are further divided by low and
high quoting activity to approximate HFT presence. This plot was taken from Conrad, Wahal, and Xiang
(2015, p. 278).
Figure 20: HFT and nHFT activity during extreme price movements in a second-by-second view. Extreme
price movements are defined as per-second returns above the 99.99 quantile. The left axis shows order
imbalance by total HFT, demanding HFT and demanding nHFT. The right axis shows the cumulative
return for t ∈ [−10, 10]. Panel A shows the entire sample, whereas Panel B(C) shows permanent(transitory)
price changes. This plot was taken from Brogaard et al. (2017, p. 46).
Figure 21: HFT net flow throughout the lifetime of an institutional order. The solid line represents the
HFT cumulative net order flow sampled at 30-minute intervals. Notably, the patterns are somehow reversed
in both plots because the left(right) panel shows buy(sell) orders, therefore, positive(negative) net order flow
implies trading with the large order. The dotted line plots the progress of the large institutional order. This
plot was taken from van Kervel and Menkveld (2016, p. 17).
Panel A: Large Institutional Buys
Panel B: Large Institutional Sells
Figure 22: Abnormal HFT liquidity provision to large institutional buys and sells. This graph reports
mean abnormal HFT buying and selling activity to stressful and non-stressful large institutional trades. The
horizontal axis shows the time decile, whereas the bars on the bottom represent the density of the execution
percentage of institutional trades. Large institutional buys(sells) are shown in Panel A(B). This plot was
taken from Korajczyk and Murphy (2015, pp. 32–33).
Panel A: S&P500 Cash (NYSE) vs. Futures (CME) Markets
Panel B: CME Trades vs. Depth Reductions on BrokerTec
Panel C: Large CME Trades vs. Depth Reductions on BrokerTec
Figure 23: High-frequency cross-market trading activity. The horizontal axis shows the time offset in
milliseconds. A positive offset means that the cash market follows the future market (CME) and vice-versa
for negative offsets. There are vertical lines are drawn at ±5 and 0 milliseconds. The reason for this being
the speed of light, which takes around 5 milliseconds to travel one-way between Chicago and New York. The
vertical axis shows the absolute cross market activity. Panel A represents trading activity between the NYSE
and CME. Panels B and C show how trades on the CME lead to a withdrawal of liquidity on BrokerTec.
In particular, Panel C only includes large trades with a size of at least 50 contracts. This plots were taken
from Dobrev and Schaumburg (2017, pp. 13, 16–17).
Figure 24: Relationship between HFT and volatility. The horizontal axis shows 100 volatility quantiles,
whereas the vertical axis reports the deviation of the mean trading activity. The top panel represents the
aggregated HFT behavior, whereas the two lower panels report HFT supply and demand. This plot was
taken from Brogaard (2010, p. 51).
Figure 25: Impact of market conditions on market maker participation. Panel A shows log volume, Panel
B shows the percentage bid-ask spreads, Panel C shows buy-sell imbalance, and Panel D shows intraday
volatility. The horizontal axis plots market cap quintiles. The plot differentiates between situations where
only DMMs supply liquidity (black) and when both DMMs and ELPs are involved (light gray). This plot
was taken from Anand and Venkataraman (2016, p. 337).
D DECLARATION OF AUTHORSHIP 82
D Declaration of Authorship
I hereby declare
• that I have written this thesis without any help from others and without the use of
documents and aids other than those stated above;
• that I have mentioned all the sources used and that I have cited them correctly according
to established academic citation rules;
• that I have acquired any immaterial rights to materials I may have used such as images
or graphs, or that I have produced such materials myself;
• that the topic or parts of it are not already the object of any work or examination of
another course unless this has been explicitly agreed on with the faculty member in
advance and is referred to in the thesis;
• that I will not pass on copies of this work to third parties or publish them without the
Universitys written consent if a direct connection can be established with the University
of St.Gallen or its faculty members;
• that I am aware that my work can be electronically checked for plagiarism and that I
hereby grant the University of St.Gallen copyright in accordance with the Examination
Regulations in so far as this is required for administrative action;
• that I am aware that the University will prosecute any infringement of this declaration
of authorship and, in particular, the employment of a ghostwriter, and that any such
infringement may result in disciplinary and criminal consequences which may result in
my expulsion from the University or my being stripped of my degree.
Date and Signature
By submitting this academic term paper, I confirm through my conclusive action that I
am submitting the Declaration of Authorship, that I have read and understood it, and that it
is true.

MA Louis Mueller

Uploaded by

Copyright:

Available Formats

MA Louis Mueller

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

MA Louis Mueller

Uploaded by

Copyright:

Available Formats

University of St.

Implications of High-Frequency Trading on

2 Properties of Algorithmic Trading 3

3 Evolution of Financial Market Structure 11

4 Electronic Trading Platforms and HFT 17

5 Speed as Form of Information 25

6 The Informativeness of Asset Prices 38

6.1.3 Information Shares . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

7 Liquidity and Transaction Costs 51

A Primer to Limit Order Markets 67

B The State-Space Framework 68

2 Properties of Algorithmic Trading

2.1 Agency Algorithms

2.2 Proprietary High-Frequency Algorithms

2.2.1 Behavioral Characteristics

2.2.2 Heterogeneity of HFT Strategies

Opportunistic strategies. Compared to market-making, opportunistic traders show oppo-

Proprietary HFT Agency algorithms

Passive strategies Aggressive strategies

Market-making Manipulation Arbitrage Structural Directional

Momentum ignition Spoofing Quote stuffing News trading Order anticipation

2.3 Identification within the Data

2.3.1 Order Flow Proxies

2.3.2 Anonymized Account-level Data

2.3.3 Pre-Classified HFT data

3 Evolution of Financial Market Structure

3.1 Implementation of New Regulations

3.2 Market Fragmentation

3.3 Modernization of Market Making

3.4 Reduction of Minimum Tick Size

50% Stocks priced below $100

4 Electronic Trading Platforms and HFT

4.1 Fee Structure

United States [USD per traded share]

Europe [bps of transaction price]

4.1.1 Maker-Taker Pricing

Broker routing costs [round-trip]

4.1.2 Taker-Maker Pricing

4.2 Optimization of Latency

4.2.1 Direct Data Feeds and Co-location

4.2.2 Latency Arbitrage

This phenomenon is somewhat comparable to a bullet flying at supersonic speeds: It hits

4.3 Interdependence between New Platforms and HFT

5 Speed as Form of Information

5.1 Informed Trading and Adverse Selection

After successful acquisition of private information, it can most efficiently be exploited

5.1.1 Effective Spread Decomposition

5.1.2 Market Maker Profitability

5.2 High-Frequency Liquidity Consumption

5.2.1 Anticipation of Order Flow

Another approach to identify front-running behavior is elaborated in Brogaard (2010, p. 23–

5.2.2 Quote Sniping

5.3 High-Frequency Liquidity Provision

2012 2013 2014 2015 2016 2017 2018

5.4 Exploitation of Minimum Tick Sizes

Low price Medium price High price

5.5 The Technological Arms Race

6 The Informativeness of Asset Prices

6.1 Price Discovery

6.1.1 The Vector Autoregressive Framework