ECOLE DES HAUTES ETUDES COMMERCIALES DE PARIS

Ecole Doctorale « Sciences du Management/GODI » - ED 533

Gestion Organisation Décision Information

« ESSAYS IN FINANCIAL MARKET MICROSTRUCTURE »

THESE
présentée et soutenue publiquement le 19 juillet 2013
en vue de l’obtention du
DOCTORAT EN SCIENCES DE GESTION
Par

Jérôme DUGAST
JURY

Présidente de Jury : Madame Carole GRESSE

Professeur des Universités
Université Paris Dauphine - France

Directeur de Recherche : Monsieur Thierry FOUCAULT

Professeur HDR
HEC Paris - France

Rapporteurs : Monsieur Pierre-Olivier WEILL

Professeur Associé
Université de Californie – Los Angeles – Etats-Unis

Madame Sophie MOINAS

Professeur des Universités
Toulouse School of Economics - France

Suffragants : Monsieur Olivier BRANDOUY

Professeur des Universités
IAE de Paris, Université Paris 1 -Panthéon Sorbonne

Monsieur Johan HOMBERT

Professeur Assistant
HEC Paris - France
 Ecole des Hautes Etudes Commerciales

Le Groupe HEC Paris n’entend donner aucune approbation ni improbation aux

opinions émises dans les thèses ; ces opinions doivent être considérées

comme propres à leurs auteurs.

Abstract

This dissertation is made of three distinct chapters. In the first chapter, I show that tradi-
tional liquidity measures, such as market depth, are not always relevant to measure investors’
welfare. I build a limit order market model and show that a high level of liquidity supply
can correspond to poor execution conditions for liquidity providers and to a relatively low
welfare. In the second chapter, I model the speed of price adjustments to news arrival in
limit order markets when investors have limited attention. Because of limited attention, in-
vestors imperfectly monitor news arrival. Consequently prices reflect news with delay. This
delay shrinks when investors’ attention capacity increases. The price adjustment delay also
decreases when the frequency of news arrival increases. The third chapter presents a joint
work with Thierry Foucault. We build a model to explain why high frequency trading can
generate mini-flash crashes (a sudden sharp change in the price of a stock followed by a very
quick reversal). Our theory is based on the idea that there is a trade-off between speed and
precision in the acquisition of information. When high frequency traders implement strate-
gies involving fast reaction to market events, they increase their risk to trade on noise and
thus generate mini flash crashes. Nonetheless they increase market efficiency.

Keywords: liquidity, welfare, limit order market, news, limited attention, imperfect market
monitoring, high frequency trading, mini flash crash, market efficiency.

i
ii
Acknowledgements

First and foremost, I am truly grateful to Thierry Foucault for his invaluable guidance and
support. I would like to thank him for introducing me to such an interesting area of finance
and for teaching me how to think as an economist. My dissertation was greatly inspired by
the innovative research he conducted over the last ten years. It’s been both a pleasure and
an honor to work under his supervision in these past five years.
I would like to thank William Fuchs, Terrence Hendershott, Johan Hombert, Stefano
Lovo, Christine Parlour, Ioanid Rosu and Pierre-Olivier Weill whose advices and comments
greatly stimulated my research. I particularly thank Christine for giving me the opportunity
to spend a fascinating and productive year at the Haas School of Business at UC Berkeley.
I am very glad that Olivier Brandouy, Carole Gresse and Sophie Moinas accepted to take
part in my thesis committee. My dissertation benefited from their work.
Over the last years, I had the chance to meet great researchers and to collaborate with
some of them. I am grateful to Terrence Hendershott and Ryan Riordan for undertaking an
exciting research project with me.
I very much appreciated the friendly and enthusiastic atmosphere in the Finance PhD
office at HEC. I would like to thank my fellow PhD candidates for their support and helpful
comments. I would also like to mention the fruitful interactions I had with PhD candidates
at the Haas School of Business.
Last, but not least, my gratitude goes to my family and friends, for their unconditional
support and love. Thank you all for everything.

iii
iv
Contents

Résumé 1
Chapitre 1 - Les mesures de liquidités sont-elles pertinentes pour mesurer le bien-
être des investisseurs? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
Chapitre 2 - Attention limitée et arrivée de nouvelle . . . . . . . . . . . . . . . . . 15
Chapitre 3 - Trading à haute fréquence, efficience de marché et « mini flash crashes » 20

Introduction 25
Electronic limit order markets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
Algorithmic trading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
Market fragmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
Dissertation overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

1 Are Liquidity Measures Relevant to Measure Investors’ Welfare? 37

1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
1.2 Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
1.2.1 Preferences and asset value . . . . . . . . . . . . . . . . . . . . . . . . 40
1.2.2 Infrequent market monitoring . . . . . . . . . . . . . . . . . . . . . . 41
1.2.3 Limit order market . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
1.2.4 Value function and equilibrium concept . . . . . . . . . . . . . . . . . 43
1.3 Steady state equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
1.3.1 One-tick market . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
1.3.2 Steady state strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
1.3.3 Steady state populations . . . . . . . . . . . . . . . . . . . . . . . . . 49

v
 1.3.4 Micro-level dynamic of the limit order book . . . . . . . . . . . . . . 49
1.3.5 Execution rate and liquidity provision . . . . . . . . . . . . . . . . . . 51
1.3.6 Value functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
1.4 Equilibrium outcomes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
1.4.1 Limit order execution rates . . . . . . . . . . . . . . . . . . . . . . . 54
1.4.2 Market depth . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
1.4.3 Trading intensity/volume . . . . . . . . . . . . . . . . . . . . . . . . . 57
1.4.4 Effects of the market monitoring frequency . . . . . . . . . . . . . . . 58
1.5 Welfare analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
1.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

2 Limited Attention and News Arrival 63

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
2.2 Literature review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
2.3 Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
2.3.1 Asset value dynamic . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
2.3.2 Limited attention . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
2.3.3 Limit order market . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
2.3.4 Value function and equilibrium concept . . . . . . . . . . . . . . . . . 71
2.4 Equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
2.4.1 The symmetric equilibrium . . . . . . . . . . . . . . . . . . . . . . . . 72
2.5 Limit order book in steady state . . . . . . . . . . . . . . . . . . . . . . . . . 74
2.5.1 Value functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
2.5.2 Steady state in the symmetric equilibrium . . . . . . . . . . . . . . . 76
2.6 Limit order book in transition phase . . . . . . . . . . . . . . . . . . . . . . 78
2.6.1 Transition phase strategy . . . . . . . . . . . . . . . . . . . . . . . . . 78
2.6.2 Limit order book dynamics in the transition phase . . . . . . . . . . . 79
2.7 After the transition phase : convergence to a steady state without uncertainty 81
2.8 Empirical implications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
2.8.1 Determinants of the liquidity supply before news arrival . . . . . . . . 83

vi
 2.8.2 Duration between news arrival and price change . . . . . . . . . . . . 85
2.8.3 Order flow decomposition in the price discovery process . . . . . . . . 86
2.9 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87

3 High Frequency Trading, Market Efficiency and Mini Flash Crashes 89

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
3.2 Model setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
3.3 Information, trading strategies, pricing policy and profits . . . . . . . . . . . 97
3.3.1 Period 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
3.3.2 Period 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
3.4 Equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
3.5 Equilibrium price dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
3.6 Market informational efficiency . . . . . . . . . . . . . . . . . . . . . . . . . 110
3.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112

Conclusion 115

A Appendix to chapters 1 and 2 117

A.1 Model Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
A.1.1 Preferences dynamic . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
A.1.2 Limit order market . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
A.1.3 Value function and equilibrium concept . . . . . . . . . . . . . . . . . 118
A.2 Limit order market in steady state w/o fundamental uncertainty . . . . . . . 121
A.2.1 equilibrium conjecture . . . . . . . . . . . . . . . . . . . . . . . . . . 121
A.2.2 Steady state populations . . . . . . . . . . . . . . . . . . . . . . . . . 121
A.2.3 Micro-level dynamic of the limit order book . . . . . . . . . . . . . . 122
A.2.4 Value functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
A.2.5 Equilibrium outcome . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
A.3 Dynamic equilibrium converging to a steady state w/o fundamental uncertainty128
A.3.1 Micro-level dynamic of the limit order book . . . . . . . . . . . . . . 128
A.3.2 Outcome of the dynamic equilibrium . . . . . . . . . . . . . . . . . . 129

vii
 A.4 Limit order book in transition phase . . . . . . . . . . . . . . . . . . . . . . 131
A.4.1 Transition phase strategy . . . . . . . . . . . . . . . . . . . . . . . . . 131
A.4.2 Limit order book dynamics in the transition phase . . . . . . . . . . . 131
A.4.3 Equilibrium in the subgame starting after the fundamental value changed133
A.5 Equilibrium in the perfectly symmetric case . . . . . . . . . . . . . . . . . . 140
A.5.1 equilibrium conjecture . . . . . . . . . . . . . . . . . . . . . . . . . . 140
A.5.2 equilibrium outcome . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
A.5.3 Proof of Proposition A.11 . . . . . . . . . . . . . . . . . . . . . . . . 141
A.5.4 Proof of Proposition A.12 . . . . . . . . . . . . . . . . . . . . . . . . 152
A.6 Comparative statics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
A.7 Empirical Implications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
A.7.1 Liquidity supply in the ”pre-signal” phase . . . . . . . . . . . . . . . 154
A.7.2 Information integration speed . . . . . . . . . . . . . . . . . . . . . . 154
A.7.3 Order flow decomposition of the price adjustment (transition phase) . 154
A.7.4 Risk of being picked-off . . . . . . . . . . . . . . . . . . . . . . . . . . 154
A.7.5 Proof of Corollary A.4 . . . . . . . . . . . . . . . . . . . . . . . . . . 155
A.7.6 Proof of Corollary A.5 . . . . . . . . . . . . . . . . . . . . . . . . . . 155
A.8 Extension to the problem with 2 monitoring rates, λ1 and λ2 , in the symmetric
case, ρ+ = ρ− , s = 1
2
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
A.8.1 steady-state populations . . . . . . . . . . . . . . . . . . . . . . . . . 156
A.8.2 Limit order in steady state without fundamental uncertainty . . . . . 156
A.9 Transition phase . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
A.10 Steady-state with uncertainty in the perfectly symmetric case . . . . . . . . 159

B Appendix to chapter 3 161

B.1 Mini flash crash events . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
B.2 Proofs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161

Bibliography 171

viii
Résumé

L’organisation des échanges sur les marchés financiers a évolué de façon extraordinaire, au
cours des trois dernières décennies, avec l’avènement des nouvelles technologies de l’information
et de la communication. Précédemment, les marchés financiers étaient organisés soit comme
des salles de marchés, où des personnes physiques échangeaient les unes avec les autres,
soit comme des marchés tenus par des intermédiaires, dans lesquels ces derniers se portent
contrepartie à l’échange pour les investisseurs qui les contacteraient par téléphone. Avec
l’évolution des technologies à même de générer, acheminer et exécuter des ordres, les struc-
tures de marché ont progressivement évolué vers l’informatisation des procédures d’échange,
et ce à travers une organisation de marché électronique dirigé par les ordres. Ces nouvelles
technologies ont aussi amélioré la vitesse et la capacité de traitement de l’information des
acteurs de marché. Ceci a donné naissance à un nouveau type de stratégies d’échange, plus
sophistiquées et totalement automatisées : le trading algorithmique. Le trading algorithmique
s’est répandu, depuis lors, et représente désormais plus de 50% du volume des échanges sur
le marché action.

Si l’arrivée de ces technologies ont été nécessaires à l’automatisation des transactions,

la mise en vigueur de régulations, qui promeuvent la compétition entre bourses et autres
plateformes d’échanges, ont catalysé cette évolution. De nouveaux entrants dans l’activité de
cotation et de services boursiers (BATS, Chi-X,. . . ) ont crû en «symbiose» avec le trading
algorithmique. Profitant d’un environnement technologique et d’une structure de coût très
favorable, les traders algorithmiques ont joué le rôle d’offreur de liquidité et, ainsi, ont aidé
ces nouvelles plateformes d’échange à concurrencer les bourses historiques dans l’attrait du
flux d’ordres. Ceci a généré une fragmentation de marché.

1
Les marchés électroniques dirigés par les ordres.

Les marchés électroniques dirigés par les ordres sont des plateformes d’échange boursier
centralisées dans lesquelles l’offre de liquidité peut provenir de chacun des acteurs de marché.
Tout agent peut effectuer une transaction en se portant offreur de liquidité avec des ordres
à cours limité, qui spécifient un prix et une quantité, et sont affichés dans le carnet d’ordre
électronique. Par exemple, dans la Fig.1, les ordres à cours limité à l’achat les plus compétitifs
sont affichés au meilleur prix d’achat, $384.82. Le premier demande 50 parts et le deuxième,
100 parts. Les agents peuvent aussi effectuer une transaction en demandant de la liquidité
avec des ordres au marché, qui sont immédiatement exécutés avec les ordres à cours limité les
plus compétitifs contenus dans le carnet d’ordre. Dans la Fig.1, si un agent envoie un ordre
de marché à la vente de 100 parts, cet ordre sera exécuté au prix de $384.82. Les marchés
électroniques dirigés par les ordres combinent l’aspect centralisé d’une salle de marché et une
large population d’investisseurs potentiels, comme les marchés tenus par des intermédiaires,
et ce grâce aux moyens de communication électroniques.

Figure 1: Vue instantanée d’un carnet d’ordre: ordres à cours limité à la vente du côté bid
(colonne de gauche), ordres à cours limité à l’achat du coté ask (colonne de droite).

Les marchés électroniques dirigés par les ordres ont commencé à se répandre au cours des
années 1980, et d’abord pour les marchés action. Par exemple, la Bourse de Paris ferma sa
salle de marché et devint un marché dirigé par les ordres totalement électronique en 1986.

2
De nos jours, la plupart des marchés action dans le monde suivent cette organisation. Cette
tendance a aussi été suivie par des marchés pour d’autres type de titre financier, comme les
marchés de change ou les marchés de taux d’intérêt.

La conversion massive des marchés financiers à l’organisation autour d’un carnet d’ordre
électronique a motivé tout un domaine de la recherche académique en finance. Les chercheurs
ont initialement voulu comprendre les dynamiques des transactions et de l’offre de liq-
uidité dans ces marchés, ainsi que les stratégies sous-jacentes des acteurs de marché. Les
premières études empiriques ont tout d’abord exposé, de façon factuelle, ces dynamiques,
comme l’illustre la Fig.2 extraite de Biais, Hillion et Spatt (1995).

Figure 2: Prix de transactions et cotations bid, ask pour Elf-Aquitaine, 9 Novembre 1991.
Source: Biais, Hillion and Spatt (1995)

La recherche théorique a étudié comment les motivations usuelles des échanges boursiers,
comme le besoin de liquidité ou bien l’information privée, pouvaient être modélisées dans
le cadre d’un marché dirigé par les ordres et pouvaient générer des prédictions en ligne
avec les observations empiriques. Ces modèles se sont intéressés aux particularités de ces
dynamiques de marché et à leurs implications pour l’efficience informationnelle ou la liquidité
de marché (e.g Glosten (1994), Parlour (1998), Foucault (1999), Foucault, Kadan, Kandel
(2005), Rosu(2010)).

3
Le Trading algorithmique.

L’avancement des technologies de l’information, de même que la conversion massive des

bourses aux marchés électroniques dirigés par les ordres, ont rendu possible et accéléré le
développement du trading algorithmique. Comme le montre la Fig. 3, la participation
d’algorithmes dans les transactions boursières pour les actions, aux Etats-Unis, ont augmenté
de façon constante depuis plus de 5 ans. Aujourd’hui les algorithmes représentent plus de
50% du volume d’échange total.

Figure 3: Part du trading algorithmique dans le volume total des échanges d’actions aux
E-U. Source: Aite Group (2010).

Le trading algorithmique peut être défini comme l’ensemble des stratégies qui s’appuient
sur des algorithmes pour prendre une part, ou l’intégralité, de leurs décisions. Ces stratégies
automatisées conditionnent, généralement, leurs actions sur un ensemble de données de
marché prédéterminées. Les stratégies de trading algorithmiques peuvent être séparées en
deux catégories principales, bien que non exhaustives.

Les algorithmes pour l’execution optimal d’un ordre. Le trading algorithmique

peut aider les investisseurs traditionnels et les intermédiaires, tels les gérants de fonds ou
les courtiers, à optimiser l’exécution de leurs transactions. Ainsi, les courtiers utilisent
couramment des robots qui découpent les ordres de leurs clients et les répartissent dans
le temps et entre différentes plateformes de trading, et ce pour obtenir des coûts de trans-
actions réduits. L’avantage principal de ces stratégies repose sur la capacité des ordina-

4
teurs, premièrement, à surveiller efficacement les fluctuations des conditions de marché et,
deuxièmement, à implémenter, de façon systématique, des procédures d’exécution optimales
basées sur ces conditions de marché.

Le trading à haute fréquence. La seconde, et la plus connue, catégorie de trading al-

gorithmique, est le trading à haute fréquence (HFT à partir d’ici). Les stratégies de HFT
s’appuient sur leur vitesse de réaction et des capacités très importantes de traitement com-
putationnel pour acquèrir de grandes quantités d’information en temps réel et prendre des
décisions à haute fréquence.
Le HFT affecte profondément la façon dont les marchés financiers fonctionnent, et provo-
quent des débats passionnés entre professionnels, académiques et régulateurs de la finance.
Ainsi, dans le New-York Times, Paul Krugman écrit :
«le trading à haute fréquence dégrade, probablement, la fonction du marché financier,
car c’est une sorte de taxe sur les investisseurs qui n’ont pas accès à ces ordinateurs super-
rapides – ce qui signifie que l’argent que Goldman dépense pour ces ordinateurs a un effet
négatif sur la richesse nationale. Comme le grand économiste de Stanford, Kenneth Arrow,
l’écrivit en 1973, la spéculation basée sur de l’information privée impose une «double perte
sociale» : elle consomme des ressources et affaiblit les marchés. » (P. Krugman, «Rewarding
Bad Actors», NY Times, 2 août 2009)
Bien que la recherche académique ait récemment produit des analyses économiques de
l’effet du HFT sur l’efficience informationnelle et la liquidité des marchés financiers, il n’existe
toujours pas de consensus sur son role bénéfique ou non. Une des difficultés est que le HFT
est un «mot valise» qui recouvre des activités très diverses. Certaines firmes (e.g GETCO,
Timberhill, Optiver. . . ) sont des teneurs de marché à haute fréquence et comptent pour une
part importante de l’offre de liquidité à la fois en Europe et en Amérique. D’autres acteurs
(e.g des hedge funds comme Renaissance) utilisent des ordinateurs pour prendre des positions
directionnelles basées sur des «signaux» avant que d’autres investisseurs aient accès à cette
information. Toutes ces activités sont clairement différentes et, de fait, peuvent avoir des
conséquences différentes pour l’efficience et la liquidité des marchés.
Les études empiriques récentes (e.g Hendershott, Jones and Menkveld (2011), Hender-

5
shott and Riordan (2013), Brogaard, Hendershott and Riordan (2012), Chaboud, Chiquoine,
Hjalmarsson and Vega (2009)) ont montré que le HFT avait un effet positif sur les mesures de
qualité de marché. Cependant d’autres études (e.g Hasbourck (2013)) et certains évènements
récents dus au HFT (i.e. le Flash Crash du 6 mai 2010) ont souligné le comportement poten-
tiellement manipulateur et destabilisant de leurs stratégies. Cela laisse ouverte la question
de savoir quel type de HFT a un impact positif pour les marchés financiers.

Le trading algorithmique et les limites cognitives. L’existence du trading algorith-

mique soulève, en elle-même, la question du fondement rationel de l’investissement dans ces
technologies et, implicitement, interroge quelle capacité supplémentaire les ordinateurs ap-
portent aux acteurs de marché. Lorsqu’un investisseur connait un choc de liqudité, il doit
analyser ses positions et son exposition aux risques avant de prendre des décisions d’échange.
Lorsque de l’information nouvelle, apportée par des nouvelles financières, est rendue publique,
les acteurs de marchés doivent interpréter cette information avant de l’utiliser. De ce fait,
ils doivent concentrer leur attention pour accomplir ces tâches spécifiques. Les machines
peuvent obtenir ces informations puis procéder à des transactions beaucoup plus rapidement
que des humains. De plus, elles peuvent surveiller plusieurs sources d’information simul-
tanément et effectuer plusieurs tâches à la fois. Ainsi, le trading algorithmique allège la
contrainte d’attention des investisseurs humains. La recherche théorique peut donc étudier
le trading algorithmique en analysant les effets de l’attention imparfaite pour les marchés
financiers (e.g. Foucault, Roëll et Sandas (2003), Biais, Hombert et Weill (2012), Pagnotta
et Philippon (2012), Foucault, Kadan and Kandel (2013)).
Cependant le trading algorithmique ne se réduit pas à une amélioration des capacités cog-
nitives utilisées pour les stratégies traditionnelles. Premièrement, l’utilisation d’ordinateurs,
en soi, élargit le champ de l’information accessible. Par exemple, la dynamique d’un car-
net d’ordre est difficilement interprétable sans une analyse quantitative informatisée, et la
fréquence élevée de cette dynamique la rend à peine perceptible aux humains. Deuxièmement,
le traitement de l’information par les machines diffère de celle des humains. Les machines
peuvent traiter de l’information « dure » et quantifiable beaucoup plus efficacement alors
qu’elles sont moins à même de gérer des scenarios non-anticipés au moment de leur con-

6
ception et peuvent donc faire des erreurs le cas échéant. Une théorie globale du trading
algorithmique devrait intégrer ces éléments.

La Fragmentation de marché.

Régulation et fragmentation de marché. Les marchés financiers, et spécifiquement les

marchés action, sont maintenant substantiellement fragmentés. Ceci a été induit, princi-
palement, par des actions règlementaires en Europe et aux Etats-Unis. L’Union Européenne
a introduite la Directive sur les Marchés d’Instruments Financier (MiFID) le 1er novembre
2007, ce qui a aboli la règle de concentration dans les pays européens et a promu la concurrence
pour les systèmes et les services boursiers. Les bourses traditionnelles, qui profitaient d’un
pouvoir de marché dans les pays européens (London Stock Exchange en Grande-Bretagne,
Euronext en France, Belgique et Pays-Bas), ont alors du affronter la concurrence des nou-
velles plateformes boursières, telle Chi-X, Turquoise et BATS Europe. Aux Etats-Unis, la
Régulation du Système National de Marché (Reg NMS) a été mise en vigueur en 2007 pour
moderniser et renforcer le système national de marché des actions. Comme avec MiFID,
elle a produit de la concurrence entre plateformes boursières. La Figure 4 illustre comment
le NYSE aux Etats-Unis et le LSE en Europe ont perdu des parts de marché au profit des
nouveaux entrants, respectivement BATS et Chi-X.

Automatisation des échanges et fragmentation de marché. Dans une revue récente

pour le UK Governement Office For Science1 , Carole Gresse écrit :
«Il existe une ancienne croyance commune en théorie économique qui veut que les marchés
de titre sont des monopoles naturels car le coût marginal d’une transaction décroit avec la
quantité d’ordres exécutés dans un marché. Alors que cela a été longtemps dans une certaine
mesure, le progrès technologique a, d’une certaine façon, changé cette réalité. Les coûts fixes
et le temps nécessaire pour mettre en place un nouveau marché ont considérablement diminué
et le trading assisté par ordinateur autorise des stratégies de transaction entre marchés qui
connectent les multiples plateformes d’échange, comme si elles formaient un réseau consolidé
de contreparties avec plusieurs entrées. Ces nouveaux outils amoindrissent l’argument de
1
Market fragmentation in Europe : assessment and prospects for market quality, C. Gresse (2012)

7
Figure 4: Fragmentation de marché en Europe et aux E-U. Le graphique de gauche représente
les parts de marché de la bourse traditionnelle et de la nouvelle entrante pour les actions
listées au NYSE. Le graphique de droite fait la même chose pour des actions européennes.
Source: Menkveld (2012).

l’externalité d’un tel réseau. »

L’automatisation des stratégies d’échanges boursiers et la libéralisation de la concurrence
pour les systèmes et services boursiers ont été concomitants car l’avancement des technologies
de l’information était une condition nécessaire à ces deux évolutions. MiFID, par exemple,
prévoit que les entreprises boursières doivent chercher les meilleures conditions d’exécution
possibles pour les ordres de leurs clients. Dans un environnement de marchés fragmentés,
cela requiert l’utilisation de systèmes informatiques de routage des ordres qui recherchent,
de façon automatique, les meilleurs prix offerts parmi les différentes plateformes boursières.
Ce type de tâche est difficilement réalisable par des humains.
Au-delà de leur rôle clé dans la consolidation des marchés fragmentés, les stratégies
d’échanges automatisés ont probablement eu un effet important pour la croissance des nou-
velles plateformes boursières. Dans un récent document de travail2 , Albert Menkveld ap-
portent des preuves que les nouveaux marchés ont grandi dans une sorte de symbiose avec
certaines entreprises de trading à haute fréquence qui se spécialisaient dans la tenue de
marché à haute fréquence, telle que GETCO. Attirés par une infrastructure adéquate et des
subventions aux ordres à cours limité, ces nouveaux acteurs sont devenus de facto teneurs de
marché, ce qui a aidé les nouvelles plateformes boursières dans la compétition pour attirer le
2
High frequency trading and the new-market maker, A. Menkveld (2012)

8
Figure 5: Le graphique représente la part de marché de l’entrant Chi-X basé sur le nombre
de transactions. Le graphique représente aussi la participation des HFT aux transactions, en
se basant sur leurs échanges à la fois chez Chi-X et Euronext. Source: Menkveld (2012).

flux d’ordres.

Maintenant que le trading algorithmique devient la forme dominante de transaction, les

bourses entrent en concurrence pour attirer leur flux d’ordres en offrant des services attractifs.
Les plateformes boursières ont réduit, de façon drastique, les latences de communication entre
leurs serveurs et ceux de leurs clients. Elles ont massivement investi en bande-passante et
proposent des services de co-location aux traders à haute fréquence. Aujourd’hui, le temps
moyen séparant la soumission d’un ordre de son exécution est inférieur à une seconde (voir
Fig. 6) et de l’ordre de quelques millisecondes pour les traders dont les ordinateurs sont
co-localisés avec le serveur de la plateforme.

L’offre de services dédiés spécifiquement aux traders à haute fréquence est plus large que
celle des plateformes boursières seules. Les médias financiers (Bloomberg, Thomson Reuters,
Dow Jones) offrent des flux de nouvelles financières facilement déchiffrable, et presque en
temps réel, qui visent explicitement les traders à haute fréquence3 .

3
Dow Jones Newswire offre des nouvelles et des données d’évènements avec faible latence pour le trading
électronique. Voir http://www.dowjones.fr/salesandtrading/low-latency-feeds.asp

9
Figure 6: Vitesse moyenne d’exécution d’un ordre petit et immédiatement exécutable, pour
le NYSE. Source: SEC Concept on Equity Market Structure (2010).

Objectif de la thèse

Comprendre les conséquences des changements récents de l’organisation des marchés fi-
nanciers est désormais une question de premier ordre pour la recherche académique. Cette
révolution technologique a notamment souligné la pertinence de l’étude des structures de
marché financier au niveau microéconomique. Dans cette thèse, je traite, avec des modèles
théoriques, trois questions de recherche en lien avec les évolutions récentes de ces marchés.
Ainsi, j’entends contribuer à la littérature académique sur la microstructure des marchés
financiers qui englobe ces questions.

10
Chapitre 1 - Les mesures de liquidités sont-elles perti-
nentes pour mesurer le bien-être des investisseurs?

Le bien-être, au sens économique, des investisseurs est un objectif majeur pour les régulateurs
des marchés financiers. Le bien-être n’étant pas une quantité observable, on pense que la
liquidité de marché peut être un bon concept pour approximer ce bien-être. La liquidité de
marché peut se définir comme la facilité, pour un investisseur, à échanger une quantité donnée
d’actif à un prix qui dévie peu en comparaison d’un prix de référence. De fait, la liquidité de
marché correspond à des coûts de transactions implicites. Dans les marchés centralisés, ces
coûts de transaction implicites sont traditionnellement mesurés avec la fourchette de prix et
la profondeur de marché (le nombre de cotations proches des meilleurs prix proposés à l’achat
et à la vente). Cette définition de la liquidité de marché est biaisée dans le sens du bien-être
des consommateurs de liquidité. La plupart des marchés centralisés (actions, changes,. . . )
sont organisés en marchés dirigés par les ordres dans lesquels tout investisseur peut échanger
en cotant des prix pour offrir de la liquidité. Les mesures de liquidité précédentes prennent
mal en compte le bien-être des offreurs de liquidité. Une grande profondeur de marché peut,
par exemple, être due à un faible taux d’exécution des ordres à cours limité, ce qui, a priori,
n’est pas signe d’un bien-être supérieur pour ceux qui utilisent des ordres à cours limité.
Comment les mesures de liquidité sont déterminées pour les stratégies des investisseurs ?
Comment ces mesures sont-elles reliées au bien-être des investisseurs ?

Afin de traiter ces questions, je construis un modèle dynamique de marché dirigé par les
ordres. La facilité de résolution du modèle me permet d’obtenir des solutions formelles pour
les variables d’équilibre telles que la profondeur de marché, le volume de transaction, le taux
d’exécution des ordres à cours limité et, aussi, le bien-être des investisseurs. Lorsque j’étudie
l’effet de la variation de certains paramètres du modèle, je trouve que (i) la profondeur de
marché co-varie négativement avec le bien-être, (ii) dans la plupart des cas le volume de
transaction co-varie positivement avec le bien-être, à l’exception d’un domaine paramétrique
particulier, et (iii) le taux d’exécution des ordres à cours limité co-varie positivement avec
le bien-être. Ceci montre, premièrement, que le taux d’exécution des ordres à cours limité
et la profondeur de marché peuvent varier dans des directions opposées et, deuxièmement,

11
que les conditions d’exécution des ordres à cours limité peuvent dominées pour le bien-être.
Le corollaire est que des variations ou des chocs sur les mesures de liquidité, telles que la
profondeur de marché ou le volume de transaction, ne correspondent pas forcément à des
changements équivalents pour le bien-être des investisseurs.

La liquidité de marché est habituellement mesurée par les coûts de transaction. Les
coûts de transaction explicites comprennent les commissions des courtiers, les frais de trans-
action,. . . etc. Les coûts de transaction implicites sont mesurés par l’écart entre le prix
d’exécution et un prix de référence qui peut être le prix moyen entre les meilleurs prix offerts
à l’achat et à la vente. La vision traditionnelle de la liquidité de marché trace un lien di-
rect entre les coûts de transaction implicite et l’illiquidité. Dans les marchés centralisés avec
intermédiation, dans lesquels l’exécution des ordres est déléguée à des teneurs de marché,
les coûts de transaction implicites correspondent au surplus que ces teneurs de marché ex-
traient des transactions. L’existence de ces coûts de transaction peut s’expliquer par des
coûts d’inventaire, le risque de sélection adverse ou bien une concurrence imparfaite entre
teneurs de marché. Avec des données de marchés exhaustives, les coûts implicites peuvent
être directement établis à partir des prix observés et des quantités associées cotées par les
teneurs de marchés. Par exemple Chordia, Roll and Subrahmanyam (2000, 2001) étudient
les mouvements agrégés et les co-mouvements de liquidité pour les marchés action du NYSE
qui, à l’époque, étaient tenus par des «spécialistes» (teneurs de marché). Parmi les différents
proxys de liquidité, les auteurs utilisent les fourchettes de prix et les profondeurs de marché.
Dans les marchés considérés, les prix et les quantités observés sont des données de transac-
tion qui étaient annoncés par les spécialistes avant transaction. C’est pourquoi ces proxys
de liquidité correspondaient effectivement aux coûts de transaction implicites auxquels les
investisseurs faisaient face.

Dans les marchés dirigés par les ordres, on peut examiner, de façon similaire, la dy-
namique de l’offre de liquidité avec l’évolution des proxys que sont les fourchettes de prix
et des profondeurs de marché. Jusqu’à récemment certains papiers (Biais, Hillion et Spatt
(1995), Engle, Fleming, Ghysels et Nguyen (2011), Hasbrouck et Saar (2012)) ont étudié
les dynamiques de ces marchés en utilisant des données de carnets d’ordre. Ces données
comprennent typiquement l’évolution du carnet d’ordre, les soumissions, les exécutions et

12
les annulations d’ordres. Dans ce cadre, l’offre de liquidité peut-elle être directement reliée
au bien-être des investisseurs comme dans les anciens marchés action du NYSE avec des
spécialistes ? Dans les marchés dirigés par les ordres, les offreurs de liquidité ne peuvent être
distingués des consommateurs de liquidité comme dans les marchés précédents. Des coûts de
transaction implicites élevés pour les investisseurs qui consomment de la liquidité, avec des
ordres au marché, correspondent à de bonnes conditions d’exécution pour ceux qui offrent
de la liquidité avec des ordres à cours limité. Ce sont des transferts monétaires des consom-
mateurs de liquidité vers les offreurs de liquidité, à l’intérieur d’un ensemble d’investisseurs.
Dans ce type de marché, le bien-être est, intuitivement, élevé lorsque la fréquence, à laquelle
les gains de l’échange sont réalisés entre un offreur et un consommateur de liquidité, est elle
aussi élevée (comme le montrent Colliard et Foucault (2012)). Cette fréquence de transaction
semble être mieux captée par le volume de transaction, par exemple, comme c’est le cas dans
mon modèle. Plus généralement, il n’est pas évident de savoir comment cette fréquence de
transaction devrait être liée aux coûts de transaction implicites pour les ordres au marché,
que mesurent la fourchette de prix et la profondeur de marché.

Dans mon modèle, je considère un cadre en temps continu. L’économie est constituée
d’un continuum d’investisseurs qui peuvent chacun détenir 0 ou 1 unité d’un actif. Leur taux
d’escompte temporel est constant. Chaque investisseur a une valeur privée, haute ou basse,
pour l’actif. La valeur privée d’un agent est aléatoire et idiosyncratique. Sa dynamique est
donnée par une chaine de Markov à deux états en temps continu. Elle passe de haute à basse,
et inversement, avec la même intensité. La différence de valorisation de l’actif entre les agents
génère des motivations pour l’échange et des gains en termes de bien-être lorsque des parts
de l’actif sont transférés d’investisseurs avec une valeur privée basse vers des investisseurs
avec une valeur privée haute. Les transactions ont lieu au sein d’un marché centralisé. Les
investisseurs peuvent échanger soit en offrant de la liquidité avec des ordres à cours limité
soit en consommant de la liquidité avec des ordres au marché.

J’étudie une classe d’équilibres stationnaires. Ces équilibres sont tels que l’état agrégé du
marché dirigé par les ordres ne change pas au cours du temps. A l’équilibre, les prix sont
constants au cours du temps et la fourchette de prix est égale au tick, l’écart minimum entre
deux prix d’échange possibles. Tous les ordres à cours limité sont soumis aux meilleurs prix

13
offert à l’achat et à la vente. Les ordres à cours limité à l’achat (resp. à la vente) sont soumis
par des investisseurs avec une haute valeur privée qui ne détiennent pas l’actif (resp. avec
une valeur privée basse et qui détiennent l’actif). Le nombre d’ordre à cours limité de part et
d’autre du carnet d’ordre, c’est-à-dire la profondeur de marché, est tel que les investisseurs
sont indifférents entre l’utilisation d’un ordre à cours limité, pour échanger à un prix attractif
mais avec un délai, ou l’utilisation d’un ordre au marché pour une transaction immédiate mais
avec un coût implicite, la fourchette de prix. Lorsque la taille du tick diminue, l’avantage
comparatif d’un ordre à cours limité diminue. Le délai maximal, pour l’exécution d’un ordre
à cours limité, que les investisseurs sont prêt à accepter, diminue lui aussi. Ceci implique que
la profondeur de marché décroit et que les investisseurs utilisent relativement plus d’ordres
au marché que d’ordres à cours limité.
Le bien-être est relié négativement à la profondeur de marché. Idéalement, tout investis-
seur qui attend avec un ordre à cours limité dans le carnet d’ordre devrait être apparié, et
échanger, avec un investisseur similaire de l’autre côté du carnet d’ordre. Une transaction
entre ces deux investisseurs transférerait un part de l’actif d’un agent à valeur privée basse
vers un agent avec valeur privée haute et, ainsi, améliorerait le bien-être. Le tick permet aux
investisseurs d’utiliser des ordres à cours limité pour extraire plus du surplus de l’échange
que leur contrepartie utilisant un ordre au marché, et sans risquer d’être concurrencer par
des ordres à cours limité plus compétitifs. Ce «pouvoir de marché» relatif qui est donné au
offreur de liquidité est inefficient puisqu’il ralentit le rythme des transactions et la réalisation
des surplus de l’échange associée. C’est pourquoi la taille du tick a un effet négatif sur le
bien-être.
Le niveau de la valeur privée basse d’un investisseur a un effet positif sur le bien-être.
Cet effet est surprenant puisque, toute chose égale par ailleurs, une réduction de l’utilité
qu’un investisseur, avec une valeur privée basse, tire de l’actif devrait affecter négativement
le bien-être global. L’intuition de ce résultat est qu’une diminution de cette valeur privée
basse augmente le coût d’opportunité à attendre dans le carnet d’ordre avec un ordre à cours
limité et de ne pas échanger immédiatement. De ce fait les investisseurs utilisent plus d’ordre
au marché, la profondeur de marché baisse et le bien-être augmente.

14
Chapitre 2 - Attention limitée et arrivée de nouvelle

Les investisseurs ont une capacité d’attention limitée et ne peuvent donc pas surveiller con-
tinument le flux d’information arrivant sur les marchés financiers. En conséquence, ils n’ont
pas la capacité d’obtenir ou d’analyser instantanément les implications des nouvelles fi-
nancières au moment de leur arrivée. Et, de ce fait, le contenu de ces nouvelles ne devient pas
une information commune du marché instantanément non plus. C’est pourquoi, à horizon
court, de l’information publique est en fait de l’information privée pour les investisseurs qui
l’observent les premiers. En raison de l’attention limitée des agents économiques, l’arrivée
d’information publique génère de courtes périodes d’information asymétrique. Comment les
marchés financiers réagissent autour de l’arrivée d’une nouvelle ? Et quel rôle l’attention
limitée joue-t-elle dans ce processus ?

Afin de traiter ces questions, je propose un cadre théorique pour analyser le rôle de
l’attention limitée dans la réaction des marchés aux nouvelles. Je conçois un modèle de
marché dirigé par les ordres en présence d’incertitude sur la valeur de l’actif en raison de
l’arrivée de nouvelles. Ce modèle étend, au marché dirigé par les ordres, le modèle de marché
de gré à gré de Duffie, Garleanu et Pedersen (2005 ,2007). Dans Duffie et al., la principale
imperfection de marché est une friction pour la recherche d’une contrepartie pour une transac-
tion. Dans mon modèle, l’imperfection de marché vient de la capacité d’attention limitée des
investisseurs. Celle-ci est équivalente à une surveillance imparfaite du marché et de l’arrivée
des nouvelles. Les investisseurs ne peuvent pas, en continu, observer l’information publique
et être en contact avec le marché. Ils peuvent y procéder à certains instants, aléatoires, de
surveillance. Ce cadre théorique permet de générer une diffusion graduelle de l’information
parmi les investisseurs à la suite de l’arrivée d’une nouvelle. La surveillance de marché im-
parfaite des investisseurs permet de décrire, de façon jointe, la formation de la liquidité, la
découverte des prix et l’efficience de marché autour de l’arrivée d’une nouvelle.

La réaction des marchés financiers à de l’information publique a motivé toute une ligne de
recherche, à la fois empirique et théorique, en particulier durant les années 1990. Étudier cette
réaction des marchés nous donne entre autres, une meilleure compréhension du processus de
découverte des prix dans ces mêmes marchés. La question de la réaction de marché à de

15
l’information publique a été empiriquement traitée, par exemple, par Eredington et Lee
(1995), Fleming et Remolona (1999) et Green (2004). Ces papiers s’intéressent à la réaction
des marchés pour les titres du Trésor Américain aux annonces macroéconomiques attendues,
c’est-à-dire dont la date et l’heure de publication sont connues. Les deux premiers papiers
montrent que le marché réagit à l’annonce en deux phases successives. Au cours de la première
phase, le prix décale rapidement vers une nouvelle valeur en lien avec les principaux chiffres
contenus dans l’annonce. La deuxième phase de cette réaction se caractérise par une forte
volatilité des prix, suggérant que l’interprétation précise de l’annonce diffère d’un investisseur
à l’autre. Cette phase se termine lorsque ces différentes interprétations convergent. Le papier
de Green montre que ces annonces macroéconomiques exacerbent le problème de sélection
adverse, ce qui suggère que les investisseurs ayant les meilleures capacités de traitement de
l’information peuvent tirer parti de ces évènements.
Mon papier contribue de manière significative à cette littérature en considérant les nou-
velles inattendues. Les arrivées de nouvelles financières sont des évènements quotidiens.
Elles sont publiées par les médias financiers comme Thomson Reuters ou Bloomberg. Elles
délivrent de l’information souvent pertinente à l’évaluation des prix des actifs financiers.
Quasiment toutes ces nouvelles arrivent sur les marchés à des moments non prévus4 . De
plus, la fréquence d’arrivée de ces nouvelles varie beaucoup d’un titre à l’autre5 . La nature
inattendue des évènements peut vraisemblablement empêcher les investisseurs d’être par-
faitement attentif aux nouvelles financières. En introduisant cette attention limitée dans un
modèle de marché dirigé par les ordres, je peux aborder la question de comment les nouvelles
inattendues affectent les décisions de transactions boursières et, par voie de conséquence, la
formation des prix et l’offre de liquidité.
L’évolution actuelle des marchés financiers vient appuyer le choix, fait dans mon modèle,
de considérer l’attention limitée comme une dimension importante pour comprendre les
réactions des marchés, à court terme, à de l’information publique. Ces réactions de court
terme donnent lieu à des questionnements importants depuis que le trading à haute fréquence
4
Dans un échantillon de 40 actions importantes, représentant 70% de la capitalisation boursière du
FTSE100, Gross-Klussmann et Hautsch (2011) trouvent qu’une action est couverte par, en moyenne, 750
nouvelles inattendues sur 1,5 années
5
Dans le même échantillon, la fréquence d’arrivée des nouvelles peut varier de 1 à 10 (de 200 à 2000
nouvelles).

16
s’est développé en utilisant des technologies de surveillance intense des marchés dans le but
d’effectuer des transactions très rapides sur des nouvelles financières. Plus généralement, le
boom du trading algorithmique (qui inclut le HFT) s’explique, en partie, par le besoin des
investisseurs d’améliorer la dimension de surveillance des marchés de leurs stratégies. Cette
tendance montre l’importance, pour les investisseurs, de la capacité d’attention qu’ils allouent
à la surveillance des marchés. De plus, mon modèle considère les marchés électroniques dirigés
par les ordres qui sont utilisés pour la plupart des bourses, cotant des actions et leurs dérivés,
et qui ont rendu possible le développement du trading algorithmique.

Ce papier donne plusieurs implications empiriques pour l’offre de liquidité et la dynamique

des prix autour des arrivées de nouvelle. Lorsque la fréquence d’arrivée de nouvelle augmente,
(i) le niveau de l’offre de liquidité diminue, (ii) les prix s’ajustent plus rapidement à la suite de
l’arrivée d’une nouvelle et (iii) l’importance relative de l’annulation des ordres à cours limité,
dans le processus d’ajustement des prix, diminue. L’intuition pour ces résultats est liée à la
période courte d’asymétrie d’information qui suit une nouvelle et qui est due à l’attention
limitée. Comme la présence d’asymétrie d’information le prévoit habituellement, il existe un
risque de sélection adverse pour les offreurs de liquidité et ce risque varie avec la fréquence
d’arrivée des nouvelles. Les investisseurs peuvent être hésitants à offrir de la liquidité avec
des ordres à cours limité puisque, à la suite de l’arrivée d’une nouvelle, l’attention limitée
retarde leur réaction. Entre temps, leurs ordres à cours limité peuvent être exécutés car
leur niveau de prix n’est plus en lien avec la nouvelle valeur de l’actif et ils offrent donc une
opportunité de profit.

Dans le cadre théorique que je propose, les investisseurs peuvent, à la fois, offrir de la
liquidité avec des ordres à cours limité et consommer de la liquidité avec des ordres au marché.
Avant l’arrivée d’une nouvelle, les investisseurs échangent les uns avec les autres car leur
valeur privée pour l’actif diffère, ce qui génère des gains à l’échange. Pendant cette phase, le
carnet d’ordres est dans un état stationnaire. Le niveau de l’offre de liquidité reste constant
et est déterminé par l’arbitrage suivant. Les ordres au marché permettent une exécution
immédiate alors que les ordres à cours limité offrent un meilleur prix mais supportent un
délai d’exécution et un risque de sélection adverse lorsque la valeur de l’actif change. A
l’équilibre, le niveau de l’offre de liquidité s’ajuste de telle sorte que les investisseurs sont

17
indifférents entre les deux types d’ordre.

Lorsque, à la suite de l’arrivée d’une nouvelle, la valeur de l’actif change, celle-ci est
publiquement accessible mais les investisseurs n’observent pas ce changement immédiatement.
Ils en prennent conscience au bout d’un certain temps qui dépend de l’intensité avec laquelle
ils surveillent le marché. Ceci génère une phase de transition, à la fin de laquelle, les prix
s’ajustent à la nouvelle valeur de l’actif. Ce processus de découverte du prix repose sur
deux dynamiques sous-jacentes. Les investisseurs qui observent la nouvelle valeur de l’actif
assez rapidement peuvent profiter d’opportunités d’arbitrage transitoires en utilisant des
ordres au marché pour exécuter les ordres à cours limité « immobiles » au prix initial. Et
les investisseurs, avec un ordre dans le carnet d’ordre, annulent ces ordres pour éviter la
sélection adverse des ordres au marché précédents. Une fois que les ordres à cours limité, au
prix initial, ont tous été annulés ou exécutés, la phase de transition prend fin et le carnet
d’ordre converge vers un nouvel état stationnaire sans incertitude sur la valeur de l’actif.
Ainsi le modèle offre une description, à haute fréquence, de la dynamique des prix et des
ordres autour des arrivées de nouvelle. Celle-ci devrait être utile aux empiristes6 .

La décision des investisseurs d’utiliser des ordres à cours limité ou au marché pour
échanger, avant l’arrivée d’une nouvelle, dépend du risque de sélection adverse durant la
phase de transition. Toute chose égale par ailleurs, ce risque amplifie la perte anticipée as-
sociée avec la soumission d’un ordre à cours limité, ce qui a un effet négatif sur l’offre de
liquidité. Dans ce cadre, l’effet de la fréquence d’arrivée des nouvelles est intuitif. Des nou-
velles plus fréquentes augmentent la probabilité d’un évènement durant lequel un ordre à
cours limité peut être sujet à la sélection adverse ce qui augmente le risque de sélection ad-
verse. Par conséquent, le niveau de l’offre de liquidité, mesuré par la profondeur de marché
(le nombre d’ordres à cours limité dans le carnet d’ordre) est relié négativement à cette
fréquence. Dans un marché plus fin, la quantité d’ordres à cours limité qui doivent être
annulés ou exécutés, durant la transition, est moindre, ce qui rend l’ajustement des prix plus
rapide.

La capacité d’attention limitée des investisseurs influence ce risque de sélection adverse

6
Engle et al (2009) utilisent des données haute fréquence de carnet d’ordre pour analyser la liquidité et
la volatilité du marché des bons du trésors U.S.

18
et donc l’offre de liquidité avant l’arrivée d’une nouvelle. Une capacité d’attention plus
grande a cependant un effet ambigu. Pour le comprendre, considérons une augmentation
de l’intensité de surveillance des investisseurs7 . D’un côté les investisseurs peuvent annuler
leurs ordres plus rapidement après l’arrivée d’une nouvelle ce qui réduit le risque de sélection
adverse et rend les ordres à cours limité plus profitables. Mais d’un autre côté les investisseurs
peuvent aussi envoyer plus rapidement des ordres au marché pour exécuter des ordres à cours
limité immobiles ce qui aggrave le risque de sélection adverse. Au final les ordres à cours
limité peuvent devenir plus ou moins profitables après une augmentation de l’intensité de
surveillance. Dans le papier, j’identifie des conditions sous lesquelles les ordres à cours limité
deviennent plus profitables. Cependant la magnitude de cet effet sur l’offre de liquidité est
très faible, en particulier si on la compare à l’effet de la fréquence d’arrivée de nouvelle. Ceci
suggère que seul la capacité de surveillance relative, par rapport aux autres participants de
marché, compte réellement pour comprendre la façon dont ce paramètre peut jouer un rôle
quantitatif dans la stratégie des investisseurs autour de l’arrivée d’une nouvelle.

7
Cette augmentation pourrait venir d’une réduction de la latence de marché

19
Chapitre 3 - Trading à haute fréquence, efficience de
marché et « mini flash crashes »

La stabilité des marchés financiers est importante pour attirer les investisseurs. L’instabilité
des prix des titres financiers peut brouiller les anticipations des investisseurs et, au final,
pourrait décourager leur participation aux échanges dans les bourses traditionnelles. Au
cours des dernières années, des anecdotes, provenant de l’industrie financière, rapportent
l’apparition d’un nouveau type d’évènement d’instabilité de marché : le «mini flash crash»
8
(Voir l’Appendix B.1 pour une liste de mini flash crash passés). Un mini flash crash
peut se définir comme un important et brusque changement de prix d’un actif suivi par un
renversement très rapide (voir Figure 1). La fréquence croissante de ces évènements a été
interprétée comme un symptôme de l’ instabilité des marchés et a été attribuée au Trading
à Haute Fréquence (HFT désormais). Entre temps, des papiers récents (e.g, Hendershott,
Jones et Menkveld (2011), Hendershott et Riordan (2013), Brogaard, Hendershott et Riordan
(2012) et Chaboud, Chiquoine, Hjalmarsson et Vega (2009)) suggèrent que le HFT a un effet
positif sur la qualité de marché et son efficience informationnelle. Par quel canal le HFT
pourrait générer des mini flash crashs ? Les marchés financiers peuvent-ils devenir à la fois
plus efficients et moins stables sous l’effet du HFT ?
Afin de traiter ces questions, nous développons une théorie des mini flash crashs. Notre
théorie est basée sur l’idée qu’il existe une tension entre la vitesse et la précision dans
l’acquisition de l’information. Le nouvel environnement de marché permet à des participants
de devenir des HFT et de réagir beaucoup plus rapidement à des nouvelles de différents types
mais au détriment de la précision de ces informations. Nous introduisons cette idée dans un
modèle à deux périodes dans lequel des agents stratégiques peuvent décider d’investir dans
une technologie rapide, qui leur permet d’acquérir, à la période 1, un signal bruité sur la valeur
fondamentale de l’actif et ensuite d’échanger, ou bien ne pas investir et d’attendre la période
2 pour échanger, en acquérant un signal parfait sur la valeur fondamentale. L’avantage des
HFT en termes de vitesse d’acquisition d’information a été étudié par Foucault, Hombert
8
En référence au Flash Crash du 6 Mai 2010. Cf. «The Flash Crash, in Miniature» in the New York Times,
http://www.nytimes.com/2010/11/09/business/09flash.html. Nanex Research rapporte aussi des mini flash
crashs parmi d’autres anomalies de marché, http://www.nanex.net/FlashCrash/OngoingResearch.html

20
et Rosu (2012) mais n’incorpore pas la possibilité d’erreur d’interprétation de l’information
nouvelle. Dans Foucault, Hombert et Rosu (2012), comme dans notre papier, l’avantage de
rapidité est modélisé comme une faculté à échanger une période avant les autres investis-
seurs. Ceci peut se voir comme une forme réduite de la capacité de surveillance de marché
intense des HFT. Celle-ci pourrait-être modélisée dans un cadre où les investisseurs ont des
capacités de surveillance de marché imparfaite, comme dans des papiers récents (e.g. Biais,
Hombert et Weill (2013), Foucault, Kadan et Kandel (2013), Pagnotta et Philippon (2012)).
La littérature sur le HFT considère que le HFT peut aussi bénéficier d’une faculté supérieure
de traitement de l’information qui diffère de l’avantage de rapidité. Pour traiter cette dimen-
sion du problème, des papiers théoriques comme Biais, Foucault et Moinas (2013) modélisent
le HFT comme des agents informés traditionnels (comme dans Glosten (1995)).

Nous trouvons qu’une augmentation de l’activité HFT, due à coût plus faible de la tech-
nologie rapide par exemple, accroit la vraisemblance d’un renversement de prix entre les
périodes 1 et 2. Les renversements de prix se produisent quand les HFT découvrent que le
signal, qu’ils ont acquis à la période 1, était faux et décident donc de corriger leurs trans-
actions à la période 2. Ceci génère des flux de transaction opposés d’un période à l’autre
et, potentiellement, des retournements de prix. L’impact sur les prix des HFT à la période
1 est proportionnel au nombre de HFT. C’est pourquoi la vraisemblance d’un renversement
de prix augmente lorsque le nombre de HFT augmente. En revanche, même si plus de HFT
implique plus de renversements, cela améliore aussi l’efficience informationnelle du marché.
Alors que ces deux implications semblent contradictoires, la présence de HFT permet une
intégration de l’information dans les prix plus rapide lorsque le signal de la période 1 est
informatif, et qui fait plus que compenser le risque d’erreur.

La nouveauté de ce papier est d’introduire un arbitrage, entre la vitesse et la précision pour

le traitement de l’information, et ce pour expliquer pourquoi les HFT pourraient échanger
en se basant sur du bruit et générer des renversements de prix. Il existe des théories de
spéculateurs de court terme qui, ex-ante, se coordonnent rationnellement pour échanger sur
la base de bruit (cf. Froot, Scharfstein et Stein (1992)). Ici, nous pensons à l’échange basé
sur du bruit comme à un risque qui se révèle ex-post. De notre point de vue, il provient
de la compétition pour les échanges lorsque les participants au marché ont la possibilité de

21
Figure 7: «Le 27 septembre 2010, l’action Progress Energy a perdu 90% en quelques secondes
sans raison apparente. Cette chute brutale était la conséquence d’un mini flash crash; une
version réduite du crash de mai...», http://sslinvest.com/news/mini-flash-crash-september-
27th-sends-pgn-shares-down-90.

réagir à des nouvelles, ou d’autres signaux pertinents, en un court laps de temps. A première
vue, l’accélération du traitement de l’information, et de son utilisation, devrait produire une
intégration de l’information dans les prix plus rapide. Cependant, cette accélération accroit
aussi le risque que ces participants basent leurs transactions sur des signaux moins précis.
Bien entendu, pour réduire ce risque, ils pourraient décider de vérifier la précision de la
nouvelle (par l’intervention humaine par exemple). Mais, ce faisant, ils prennent le risque de
perdre une opportunité de profit car ils réagiraient trop tard à un signal informatif. Ainsi,
cette compétition entre participants peut les pousser à réagir trop rapidement, au détriment
de la précision de l’information sur la base de laquelle ils échangent, et peut donc mener à
des renversements de transactions et de prix.

De façon alternative, les renversements de prix, pourraient s’expliquer par la présence

d’investisseurs excessivement confiants qui sur-réagissent à des signaux privés, comme dans
Daniel, Hirshleifer et Subrahmanyam (1998), ce qui génère des corrections de prix suite à
la révélation publique de l’information. Dans ce contexte, les renversements de prix sont
systématiques. Les retours de prix deviennent négativement auto-corrélés et prédictibles.
Le marché y est inefficient contrairement à ce que nous trouvons dans notre modèle, qui ne

22
génère pas d’auto-corrélation. De plus, notre cadre théorique implique des renversements
de prix complets au sens où les prix peuvent revenir à leur niveau original lorsque les HFT
réagissent à du bruit. Dans le papier cité le renversement de prix est une correction partielle
d’un changement précédent excessif mais qui allait dans la bonne direction.

Les renversements de transaction, par des participants qui obtiennent de l’information

plus rapidement que les autres, peuvent se produire lorsqu’un participant bénéficie d’une
fuite anticipée d’information bruitée à propos d’une future annonce publique, comme dans
Brunnermeier (2005). Cela lui permet, premièrement, d’acquérir seul un signal bruité sur
une composante de court terme de la valeur de l’actif, et d’échanger sur cette base, et,
deuxièmement, de savoir de combien son impact sur le prix était due à du bruit, après
que cette composante de court terme est publiquement annoncée. Ainsi, il bénéficie encore
d’un avantage informationnel après l’annonce. Il en profite en renversant partiellement la
part de sa transaction initiale qui était due à la composante bruitée du signal. Cependant
ce renversement de transaction est compensé par des transactions, de sens opposés, par
d’autres participants stratégiques, ce qui rend l’implication pour la dynamique des prix peu
claire, contrairement à notre modèle. Dans le papier de Brunnermeier, le marché est efficient
puisque les prix reflètent toute l’information publique accessible. Cependant l’introduction
d’une fuite d’information a des effets mitigés sur l’efficience, contrairement à notre modèle
ou plus de HFT augmente l’efficience informationnelle.

Les stratégies d’échanges basés sur des signaux informatifs peuvent être aussi diverses
que le spectre de l’information pertinente pour un marché particulier. Les HFT cherchent
des nouvelles financières ou des tendances de marchés informatives qu’ils peuvent traiter et
exploiter le plus vite possible. La source de la précision imparfaite de l’information peut être
endogène ou exogène. Elle peut être endogène car les algorithmes envoient des ordres en se
basant sur l’interprétation d’évènements. Toute chose égale par ailleurs, plus la réaction de
l’algorithme est rapide, moins l’interprétation est précise. Par conséquent les HFT font face à
cet arbitrage lorsqu’ils calibrent leur algorithme. Mais le processus de production de nouvelle
information peut aussi être la source, exogène, de l’imprécision de cette information. S’il y a
une chance, même faible, que certaines nouvelles soient fausses, les HFT doivent alors décider
s’ils prennent le risque de réagir immédiatement à ces nouvelles ou bien s’ils attendent une

23
correction éventuelle. L’anecdote suivante illustre, de façon assez extrême, ce problème de
fausses nouvelles. Le lundi 8 septembre 2008, le prix de l’action United Airlines chuta de
$12 à $3 en, à peu près, quinze minutes. Ensuite le prix rebondit jusque $11 à la fin de la
session de mardi. La cause de ces retournements était un vieil article de presse à propos d’une
procédure de mise en faillite de United Airlines en 2002 et qui par erreur était réapparu en
septembre 2008 dans les titres des pages d’information de Google.

24
Introduction

The organization of trading in financial markets has changed dramatically over the last
three decades, along with the advent of new technologies of information and communication.
Financial market structures were primarily organized as trading floors, where humans are
physically trading with each other, or dealership markets, in which dealers are in charge of
standing as counterparty for investors who would reach them by phone. With the evolution
of technologies for generating, routing and executing orders, market structures progressively
evolved towards the computerization of trading procedure through an electronic limit order
market organization. These new technologies also improved the speed and the information
processing capacity of market participants. It gave birth to a new type of trading strategies,
fully automatized and more sophisticated: algorithmic trading. Algorithmic trading has been
growing since then and now accounts for more than 50% of trading volume in equity markets.
While technologies have been key to enable trading automation, regulatory actions enforcing
competition among exchanges (Reg NMS in the U.S, MiFID in the E.U) have fostered its
evolution. New entrants in the business of exchanges (BATS, Chi-X,...etc) have grown in
«symbiosis» with algorithmic traders. Embedded in an adequate technological and cost
structure environment, they have acted as liquidity providers, and, thus, helped new trading
platforms to compete, with the incumbent exchange, for attracting the order flow. It ended
up generating market fragmentation.
Understanding the consequences of the recent changes of financial markets organization
has become of primary interest for academic research. It has also stressed the relevance
of studying in detail financial market structures. In this dissertation, I aim to investigate
research questions that help addressing the recent evolutions of financial markets and, thus,
to contribute to the financial market microstructure growing literature.

25
 In the following of the introduction, I present some features of the major changes of market
structures: electronic limit order markets, algorithmic trading and market fragmentation.
Then I give an overview of the three chapters of the dissertation.

Electronic limit order markets

Electronic limit order markets are centralized trading venues in which liquidity supply can
be achieved by any market participants. Any agent can trade by supplying liquidity with
limit orders, which specify a price, a quantity of shares and are posted in the electronic limit
order book. For instance, in Fig.1, most competitive buy limit orders are posted at the best
bid price, $384.82, the first one is for 50 shares and the second one for 100 shares. Agents
can also trade by consuming liquidity with market orders, which are immediately executed
against most competitive limit orders displayed in the order book. In Fig.1, if a trader sends
a sell market order for 100 shares, it will be executed at $384.82. Electronic limit order
markets combine the centralized trading location aspect of a trading floor with a large set of
market participants, as in dealership markets, thanks to electronic communication devices.

Figure 8: Instant view of a limit order book: buy limit orders on the bid side (left column),
sell limit orders on the ask side (right column).

26
 Electronic limit order markets started to spread out during the 1980’s, first in equity
markets. For instance the Paris Bourse closed its floor and became a fully electronic limit
order market in 1986. Now most of equity markets around the world are organized as limit
order markets. This trend has now been followed by markets for other types of securities, as
foreign exchange or fixed income markets.

The massive conversion of financial markets to the limit order market organization has
motivated an extensive line of academic research in finance. Researchers have been primarily
interested in understanding the dynamics of trades and liquidity supply in these markets, as
well as the underlying strategies of market participants. First empirical studies have exposed
the dynamics of trades and quotations in limit order markets, as in Biais, Hillion and Spatt
(1995) (see Fig.2).

Figure 9: Transaction prices and bid and ask quotes for Elf-Aquitaine, November 9, 1991.
Source: Biais, Hillion and Spatt (1995)

Theoretical research investigated how usual motives for trading, such as liquidity needs
or private information, could be modelled in a limit order market setup and could generate
market patterns that would match empirical findings. They studied the particularity of these
market dynamics and their implications for informational efficiency or liquidity (e.g. Glosten
(1994), Parlour (1998), Foucault (1999), Foucault, Kadan, Kandel (2005), Rosu (2010)).

27
Algorithmic trading

The advances of information technologies, as well as the massive conversion of exchanges to

electronic limit order markets, have enabled and fostered the rise of automated trading. As
shown in Fig.3, the participation of algorithms to trades in U.S equities transaction have
been constantly growing for more than 5 years. Algorithms now intervene in more than 50%
of the overall trading volume.

Figure 10: Share of algorithmic trading in the total U.S equities trading volume. Source:
Aite Group (2010).

Algorithmic trading can be defined as trading strategies that relies on algorithm to per-
form part, or the entire, of their trading decisions. These automated strategies would usually
condition their actions on a set of predetermined market outputs. Algorithmic trading strate-
gies can be split into two main, though non-exhaustive, categories.

Optimal order execution algorithms. Algorithmic trading can help traditional in-
vestors or intermediaries, as fund managers or brokers, to optimize the execution of their
trading needs. For instance brokers routinely use robots to split the orders of their clients
over time and among multiple trading venues to achieve smaller trading costs. The principal
advantage of these strategies relies on the computer abilities, first, to efficiently monitor fluc-
tuations of market conditions and, second, to systematically implement optimal execution
procedures that load on these market conditions.

28
High Frequency Trading. The second, and most famous, category of algorithmic trading,
is High Frequency Trading (HFT henceforth). HFT strategies relies on intense processing
capacities and reaction speed to acquire a large amount of real time information and take
actions at high frequency.

HFT is profoundly affecting how financial markets work and triggered heated debates
among practitioners, academics and regulators. For instance, in the New-York times, Paul
Krugman writes:

«High-frequency trading probably degrades the stock market’s function, because it’s a kind
of tax on investors who lack access to those superfast computers - which means that the money
Goldman spends on those computers has a negative effect on national wealth. As the great
Stanford economist Kenneth Arrow put it in 1973, speculation based on private information
imposes a «double social loss»: it uses up resources and undermines markets.» (P. Krugman,
«Rewarding Bad Actors», NY Times, August 2, 2009)

Even though academic research has recently produced economic analysis of the effects of
HFT on informational efficiency and liquidity of financial markets, there is still no consensus
on its beneficial, or detrimental, role. One difficulty is that HFT is a catchall phrase for
very diverse activities. Some firms (e.g, GETCO, Timberhill, Optiver etc ...) engage in high
frequency market-making and now account for a large fraction of liquidity supply both in the
U.S. and Europe. Other participants (e.g., hedge funds as Renaissance) use computers to take
directional positions based on «signals» before other investors get access to this information.
Clearly, all these activities are different and as such may have different impact on market
efficiency and market liquidity.

Recent empirical studies (e.g, Hendershott, Jones and Menkveld (2011), Hendershott
and Riordan (2013), Brogaard, Hendershott and Riordan (2012) or Chaboud, Chiquoine,
Hjalmarsson, and Vega (2009)) have shown that HFT had a positive effect on market quality
measures. However other studies (e.g Hasbrouck (2013)) and recent market events ascribed
to HFT (i.e. the Flash Crash of May 6th, 2010) have stressed the potential manipulative and
destabilizing behaviour of their strategies. It leaves open the question of which type of HFT
has a positive impact in financial markets

29
Algorithmic trading and cognitive limitations. The existence of algorithmic trading
itself raises questions about the rationality of investing in these technologies and, implicitly,
asks what additional capacity computers bring to market actors. When a trader is hit by a
liquidity shock, he must analyze his positions and risk exposure prior to take trading decisions.
When new public information, conveyed by financial news, is released, a trader must interpret
this information prior to trade on it. In the previous situations, collecting and processing
information takes time for humans. They must concentrate their attention to accomplish
these specific tasks. Machines can access, process and trade on information much faster than
humans. Moreover they can monitor simultaneously several sources of information and be
multitasking. As a result, algorithmic trading alleviate the attention constraints of human
traders. Consequently, theoretical research can study algorithmic trading by analyzing the
effects of imperfect attention in financial markets (e.g. Foucault, Roell and Sandas (2003),
Biais, Hombert and Weill (2012), Pagnotta and Philippon (2012), Foucault, Kadan and
Kandel (2013)).
However algorithmic trading cannot be reduced to an improvement of cognition abilities
used for traditional trading strategies. First, the use of computer, by itself, extends the
field of available information. For instance, order book imbalances are difficult to interpret
without quantitative computerized analysis, and their high frequency dynamics are barely
perceivable by humans. Second, information processing by computers differs from humans’.
While machines can process hard and quantifiable information much more efficiently, they are
not able to deal with scenarios that were not anticipated at the time of their conception and
can possibly make mistakes. A comprehensive theory of algorithmic trading should include
these features.

Market fragmentation

Regulation and market fragmentation. Financial markets, and specifically equity mar-
kets, are now substantially fragmented. It has mainly been impulsed by regulatory actions
both in Europe and in the U.S. The European Union introduced the Markets in Financial
Instruments Directive (MiFID) on November 1, 2007, which abolished the concentration rule

30
in European countries and promoted competition for trading systems and services. Tradi-
tional exchanges, that used to profit from some market power in European countries (London
Stock Exchange in G.B, Euronext in France, Belgium and the Netherlands), have been facing
competition from new trading platforms, as Chi-X, Turquoise and BATS Europe. In the U.S,
the Regulation National Market System (Reg NMS) was promulgated in 2007 to modernize
and strengthen the national market system for equity securities. As for MiFID, it fostered
competition among trading platforms. Figure 4 illustrates how NYSE in the U.S and LSE
in Europe lost market shares against new entrants as, respectively, BATS and Chi-X.

Figure 11: Market fragmentation in Europe and U.S. The left graph plots incumbent and
entrant market share in NYSE-listed stocks. The right graph does the same for European
listed stocks. Source: Menkveld (2012).

Trading automation and market fragmentation. In a recent review for the UK Gov-
ernment Office for Science9 , Carole Gresse writes:
«There is an old common belief in economic theory that security markets are natural mo-
nopolies because the marginal cost of a trade decreases with the quantity of orders executed in
a market. While this has long been true to a certain extent, technological progress has some-
how changed this reality. The fixed costs and time necessary to launch a new market have
considerably diminished and computer trading now allows cross-market trading strategies that
connect to multiple trading venues as if they were a consolidated network of counterparties
9
Market fragmentation in Europe: assessment and prospects for market quality, C. Gresse (2012)

31
with several entries. Those new tools undermine the network externality argument.»

The automation of trading strategies and the decision to liberalize competition between
trading systems and services happened to be contemporaneous because the advance of infor-
mation technologies was a necessary condition for these two evolutions. For instance MiFID
states that trading firms should aim for the best execution condition possible for their client’s
orders. In a fragmented markets environment, it requires to use smart order routing systems
that automatically look for the best prices available across trading platforms. This kind of
task is barely achievable by human traders.

Figure 12: The graph depicts the market share of the entrant market Chi-X based on the
number of trades. The graph also depicts the high-frequency trader’s participation in trades,
based on its trading in both the entrant (Chi-X) and in the incumbent market (Euronext).
Source: Menkveld (2012).

Beyond their key role in the consolidation of fragmented markets, automated trading
strategies probably had an important impact in the growth of entrant trading platforms.
In a recent working paper10 , Albert Menkveld provides evidences that new entrant markets
grew in a sort of symbiosis with some High Frequency Trading firms that specialized in high
frequency market making, such as GETCO. Attracted by adequate infrastructure and rebates
10
High frequency trading and the new-market maker, A. Menkveld (2012).

32
for limit orders, these new actors became de facto market makers which helped new trading
platforms to compete for the order flow. Fig.5 shows the parallel trends of the market share
growth of Chi-X in Dutch stocks and the growth of High Frequency Trading activity in these
markets.

Now that algorithmic trading is becoming the prevalent form of trading, exchanges com-
pete for attracting their order flows in offering attractive services. Trading platforms have
drastically reduced communication latencies between their servers and those of their clients.
They massively invested in bandwidth and proposed co-location to high frequency traders.
Now the average duration between order submissions and executions are less than a second
(see Fig.6) and of the order of milliseconds for traders whose computers are co-located with
the platform’s server.

Figure 13: NYSE average speed of execution for small, immediately executable orders.
Source: SEC Concept on Equity Market Structure (2010).

The supply of services specifically designed for High Frequency Trading actors have gone
beyond those offered by the only trading platforms. For instance news providers (e.g.
Bloomberg, Thomson Reuters, Dow Jones) propose almost real time and easily readable
financial news that are explicitly aiming HFT clients11 .

11
Dow Jones Newswire offers low-latency news and event-data for electronic trading. see
http://www.dowjones.fr/salesandtrading/low-latency-feeds.asp

33
Dissertation overview

This dissertation tackles, with theoretical models, three important questions in financial mar-
ket microstructure which cover different aspects of the new financial market structures.

The multiplication of trading venues organized as limit order markets raises questions on
how financial markets currently accomplish their assigned role, that primarily is to efficiently
allocate savings across investment opportunities and to allow for an efficient risk sharing
among investors. In order to perform this welfare improving role, financial markets must
be attractive to investors and fund seeking entities. Liquidity and informational efficiency
are the usual attribute of attractive markets. It interrogates how market quality can be
assessed in financial markets. For instance, now that equity markets have migrated toward a
limit order market organization, are traditional liquidity measures still relevant to evaluate
investors’ welfare? I address this question in the first chapter of this dissertation. I show
that usual liquidity measures, such as market depth, may be inversely related to the order
execution quality for limit order users, and thus may not capture well the overall welfare of
investors.

The reason why algorithmic trading strategies have been tremendously expanding partly
lies on the fact that algorithmic trading alleviates the limited attention constraint of human
investors. It helps monitoring order book activity more frequently, gathering new public
information more rapidly and subsequently adapting trading strategies more efficiently. Un-
derstanding how limited attention affect trading strategies is thus key to address the effects
of algorithmic trading. In the second chapter of this dissertation, I investigate how investor’s
limited attention affects their trading strategies and the overall market dynamic around news
arrival, in a limit order market. Because of limited attention, investors imperfectly monitor
news arrival. Consequently, in equilibrium, prices reflect news with delay. This delay shrinks
when investors’ attention capacity increases. The price adjustment delay also decreases when
the frequency of news arrival increases. When news arrival frequency is higher, the picking-
off risk increases for limit orders. The limit order book becomes thinner and there are fewer

34
stale limit orders to execute or cancel after news arrival. Thus, it reduces the time it takes
for market prices to reflect news content.

High Frequency Trading strategies use high market monitoring capacities to react to all
kind of market events that are helpful to predict future price changes and thus generate profit
opportunities. By reacting faster to such informative events, HFT can help integrating new
information into prices and thus make markets more efficient. However HFT activity has
come along with a new type of market instability events, mini flash crashes, that can be
defined as a sudden sharp change in the price of a stock followed by a very quick reversal. In
the third chapter of this dissertation, which is based on a joint work with Thierry Foucault,
we investigate the effect of HFT on market efficiency and price stability when reaction speed
to market events comes with a risk of trading on noise. By introducing this trade-off between
reaction speed and information precision, we show that HFT activity can generate mini flash
crashes. However this higher instability of prices comes along with a higher market efficiency.
This finding suggests that HFT helps integrating information into prices more efficiently but
in a less stable way.

35
36
Chapter 1

Are Liquidity Measures Relevant to

Measure Investors’ Welfare?

1.1 Introduction

Investors’ welfare is one main objective of financial market regulators. As welfare is not
observable, market liquidity has been thought as the right concept to approximate welfare.
Market liquidity can be defined as the ease for an investor to trade a given asset quantity
at a price that does not deviate much from a benchmark price. Hence market liquidity
corresponds to some implicit trading costs. In centralized markets these implicit trading
costs are usually measured with bid-ask spreads and market depths (the number of quotes
closed to the best bid and ask prices). This definition of market liquidity and its empirical
measures are biased toward the welfare of liquidity consumers. Most centralized markets
(stock, FX,... etc) are organized as limit order markets in which any investor can trade by
posting quotes and supplying liquidity. Previous liquidity measures do not account well for
the welfare of liquidity suppliers. For instance, a high market depth may arise from a low
limit order execution rate which is, a priori, not welfare improving for limit order users. How
are liquidity measures determined by investors trading strategies? How can these measures
be linked to investors’ welfare?
To address these questions I design a dynamic model of limit order market. The tractabil-
ity of the model allows me to provide closed-form solutions for equilibrium outcomes such

37
as market depth, trading volume, limit order execution rate as well as for welfare. When I
consider variations of several model parameters, I obtain that (i) market depth negatively
co-varies with welfare, (ii) in most cases trading volume positively co-varies with welfare,
except for a specific range of parameters, and (iii) limit order execution rate positively co-
varies with welfare. It shows, first, that limit orders execution rate and market depth may
vary in opposite directions and, second, that execution conditions for limit orders could dom-
inate in the welfare outcome. The corollary is that cross-sectional variations or shocks on
liquidity measures, such as market depth or trading volume, do not necessarily corresponds
to equivalent changes to investors’ welfare.

Market liquidity has been usually measured by trading costs. Explicit trading costs
include brokerage commissions, trading fees,...etc. Implicit trading costs are measured by
the wedge between the execution price and a benchmark that can be the mid-quote (the
best bid and ask prices average). The traditional view on market liquidity makes a direct
link between implicit trading costs and illiquidity. In intermediated centralized markets,
in which trades execution is delegated to dealers or market makers, implicit trading costs
correspond to the surplus that these market makers extract from trades. The existence of this
trading costs can be explained by inventory costs, adverse selection or imperfect competition
among dealers. With comprehensive market data, implicit costs can be directly assessed
from observed prices and corresponding quantities quoted by market makers. For instance,
Chordia, Roll and Subrahmanyam (2000, 2001) study aggregate movement and co-movement
of liquidity for NYSE stocks that, at the time, were run by specialists (market-makers).
Among different liquidity proxies, they use quoted bid-ask spreads and quoted depths. In
the considered markets, quoted prices and offered quantities are transaction data. They are
announced by specialists prior to a trade. Hence liquidity proxies precisely reflect implicit
trading costs that investors were facing.

In limit order markets, one can similarly examines the dynamic of liquidity supply with
the evolution of bid-ask spreads and market depths as proxies. Past and more recent papers
(e.g., Biais, Hillion ans Spatt (1995), Engle, Fleming, Ghysels and Nguyen (2011), Has-
brouck and Saar (2012)) have investigated limit order market dynamics with order book
data. These type of data usually provide the order book evolution, order submissions, ex-

38
ecutions and cancellations. Can liquidity supply be directly linked to investors welfare as
in former NYSE stock markets with specialists? In order driven market, liquidity suppliers
cannot be distinguished from liquidity consumers as in former markets. High implicit trading
costs for investors who consume liquidity, with market orders, correspond to good execution
conditions for investors who supply liquidity with limit orders. These are money transfers,
from liquidity consumers to liquidity suppliers, inside the pool of investors. In this type of
market, welfare is intuitively high when the frequency at which gains from trade are realized,
between a liquidity supplier and a liquidity consumer, is high as well (as shown in Colliard
and Foucault (2012)). This trade frequency would be better captured by trading volume for
instance, which is the case in my model. Generally speaking, it is not clear how this trade
frequency should be linked to implicit trading costs for market orders, measured by market
depth and bid-ask spread.

In my model I consider a continuous-time framework. There is a continuum of competitive

investors who can hold 0 or 1 unit of an asset. They discount time at a constant rate. Each
investor has either a high or low private value for the asset. The asset private value of an agent
is random and idiosyncratic. It is a two-state continuous-time Markov chain. It switches from
high to low or conversely with same intensity. The difference in asset valuations across agents
generates motives for trade and welfare gains when some asset shares are transferred from
investors with a low private value to investors with a high private value. Trading takes place
in a centralized market. Investors can trade by either supplying liquidity with limit orders
or consuming liquidity with market orders.

I study a class of steady-state equilibria. They are such that the aggregate state of the
limit order market does not change over time. At equilibrium, prices are constant over time
and bid-ask spreads are equal to the tick size, the minimal difference between two available
trading prices. All limit orders are submitted at the best bid and ask prices. Buy (resp. sell)
limit orders are submitted by investors with a high private value who do not own the asset
(resp. with a low private value who own the asset). The number of limit orders on each side
of the book, the market depth, is such that investors are indifferent between using a limit
order, to trade at a good price but with a time delay, or a market order to trade immediately
with an implicit cost, the bid-ask spread. When the tick size decreases, the comparative

39
advantage of using limit orders declines. The maximal time delay for limit order execution,
that investors are willing to bear, declines as well. It implies that the market depth decreases
and that investors use relatively more market orders than limit orders.
Welfare is negatively linked to the market depth. Ideally, any investors who is waiting in
the book to have his limit order executed would be matched with a similar investor on the
other side of the book. Trade between two of these investors would transfer the asset from
a low value type to a high value type and thus would increase welfare. The tick-size of the
market allows investors to use limit orders to extract more of the trading surplus than their
counterpart with market orders, without risking to have their limit orders undercut by other
investors. This relative «market power» that is offered to liquidity supplier is inefficient since
it slows down trading and the subsequent trading surplus realization. Hence the tick-size has
a negative impact on welfare.
The level of the private value of a low type investor has a positive impact on welfare.
This effect is surprising since, everything else equal, a reduction of the utility that low type
investors draw from the asset should negatively affect the overall welfare. The intuition for
this result is that a lower asset value for low type increases the opportunity cost for waiting
with a limit order in the book and not trading immediately. Hence investors use more market
orders, market depth declines and welfare increases.
Chapter 1 is organized as follows. Section 1.2 presents the setup and assumptions of the
model. Section 1.3 describes the model equilibrium. Section 1.4 derives model outcomes.
Section 1.5 analyzes welfare implications. Section 1.6 concludes.

1.2 Model

1.2.1 Preferences and asset value

I consider a continuous time framework with an infinite horizon, t ∈ [0, +∞). The economy
is populated with a continuum of investors [0, 1]. They are risk neutral and infinitely lived,
with time preferences determined by a time discount rate r > 0. These investors can trade
an asset with a common value v that is constant over time.

40
Preferences. As in Duffie, Garleanu and Pedersen [2005,2007], an investor is characterized
by an intrinsic type, ”high” or ”low”. A high type investor receives a utility flow v per asset
unit she owns. A low type investor receives a utility flow v − δ per asset unit she owns.
Between time t and time t + dt an investor can switch from one type to another (high to low
or low to high) with probability ρ.dt. Thus, a high type investor has a higher valuation for
the asset than a low type.

Asset holding and supply. As in Duffie et al., investors can own either one or zero unit
of the asset. The asset supply is equal to 12 . So that half of the population owns the asset.
Given the previous assumptions any investor must have a type in the set {ho, hn, lo, ln}
(h: high, l: low, o: owner, n: non-owner). And we can divide the mass of investors in 4
populations: Lho , Lhn , Llo , Lln . They verify the equations

1
Lho + Lhn + Llo + Lln = 1, Lho + Llo = .
2
It is possible to extend the number of possible types by taking into account the limit
order submission status of investors. Indeed, in a limit order book, an owner can either be
out of the market or have an order in the order book. As well for a non-owner. This setting
can generate many subtypes of the previous types. Let’s call T the set of all possible types.
If an investor does not have any limit order submitted in the order book she is out. If she
has a limit order submitted we have to specify at which price it is. For instance a type ln
can be ln − out or ln − B with a buy limit order at price B. Symmetrically a type lo can be
lo − out or lo − A with a sell limit order at price A.

1.2.2 Infrequent market monitoring

In order to impose some structure to the model, I consider that investors can trade at some
random times that follow a Poisson process with a finite frequency. Based on this structure,
I can consider the case where investors can continuously trade in the market by taking the
Poisson process frequency to its infinite limit.

Assumption 1.1. I assume that an investor observes contacts the market at some ran-

41
dom times {ti }i∈N . I call these times ”market monitoring times”. This sequence of market
monitoring times is generated by a Poisson process of intensity λ + ρ.
More specifically between time t and t + dt an investor monitors the market in two types
of situation:

• when she uses the market monitoring technology which occurs with probability λ.dt

• when her private value changes which occurs with probability ρ.dt.

The second assumption states that investors continuously monitor their private value for
the asset and contact the market whenever this private value changes. This assumption
allows to reduce the anticipation problem of the investor who has to take into account the
possibility of future shocks to her private value especially when facing the decision to send
a limit order. Indeed she knows that when a shock occurs she has the possibility to cancel
a previous limit order. Then it prevents her from being executed while it is not optimal
anymore given her new private value.

1.2.3 Limit order market

Trading takes place in a limit order market. Prices at which trades can occur belong to a
countable set of prices, the price grid. The minimum difference between two prices is the
tick size, ∆. Investors can use limit or market orders to trade. Limit orders are orders that
specify a limit price at which the order can be executed. They are stored in the order book
until matched with a market order. The depth of the limit order book at price P , DP , is the
volume of all limit orders submitted at price P . Market orders do not specify a price limit.
They hit the most competitive limit order and get execution immediacy.
For technical reasons I assume that the price grid is bounded. This is reasonable since
trading will not occur at prices higher than a certain threshold since the asset value is bounded
(the corresponding strategies would be strictly dominated by a strategy in which investors
don’t trade). A similar assumption is made in Parlour [1998], Foucault, Kadan and Kandel
[2005] for instance.
Each time an investor monitors the market she can take any number of actions in the
following list under the constraints that she cannot have more than one limit order in the

42
book and that she can hold either 1 or 0 unit of the asset.

• As an owner she can : (i) do nothing and remain an owner; (ii) submit a sell limit order
and remain an owner until her order is executed; (iii) submit a sell market order and
become a non-owner; (iv) cancel a previous sell limit order.

• As a non-owner she can : (i) do nothing and remain a non-owner; (ii) send buy limit
order and remain a non-owner until her order is executed; (iii) send a buy sell market
order and become an owner; (iv) cancel a previous buy limit order.

This defines the action set of an investor as (with some notation abuse)

A = {do nothing, market order, limit orders at the different prices}.

Assumption 1.2. In the limit order book, limit orders are executed following a ”Pro-rata
matching” execution rule1 . In this setup all limit orders submitted at the same price have the
same probability of execution at any point in time, regardless of their submission date.
δ
Assumption 1.3. I assume that r
is big compared to ∆. It ensures that the gains from trade
due to differences in private values, measured by rδ , is bigger that the implicit cost of trading,
the bid-ask spread, which is measured by the tick-size, ∆. More specifically I assume that

δ > (r + 2ρ)∆ (1.1)

1.2.4 Value function and equilibrium concept

An investor chooses a new action at each market monitoring time. The strategy of an agent
is a function σ,

σ :H × Ξ × [0, ∞) → A,

(h, ξ, t) 7→ a.
1
In practice there are some markets where the ”Pro-rata matching” is implemented. However for the
majority of stock markets Time Priority applies. The reality of the Time priority is mitigated by the fact
that there are multiple trading platforms and that agents can use smart order routing technologies for
achieving best trading conditions. The flow of market orders is split among different trading platforms. Then
the time at which a limit order executed is randomized.

43
 The set Ξ gathers all potential state variables. An element of this set ξ ∈ Ξ is defined as
ξ = (θ, v, S) where θ ∈ T is the type of the investor, v is the common value of the asse,t and
S is the aggregate state of the limit order book, that is to say the bid and ask prices and all
the depths at these prices. H is the set of all possible histories of actions and observations
of an investor:

H = {h ∈ (at1 , . . . , atn , ξt1 , . . . , ξtn , t1 , . . . , tn ) ∈ An × Ξn × [0, ∞)n , t1 < . . . < tn , n ∈ N}.

Her strategy, σ, and the strategies of all other investors, Σ, generate her asset holding
process ηt ∈ {0, 1} that is equal to 1 when she holds one unit of the asset, her type process
θt ∈ T and a process of trading prices Pt at which her orders are executed any time she
changes her holding i.e. when ηt switches from 0 to 1 or conversely.

At time t the value function of an investor playing strategy σ is given by

Z ∞
V (ht , ξt , t; σ, Σ) = Et e−r(s−t) dUs ,
t

s.t dUt = ηt (v − δI{θt ∈ lo} )dt − Pt dηt .

The strategy σ is a best response to the other players set of strategies Σ if and only if for all
strategy γ,
∀ht ∀ξt ∀t V (ht , ξt , t; σ, Σ) ≥ V (ht , ξt , t; γ, Σ).

In this paper I focus on Markov perfect equilibria where strategies depend only on
state variables, (θ, v, S).

1.3 Steady state equilibrium

In this chapter, I focus on steady state equilibria in which the aggregate state of the limit
order market is constant over time while trading occurs. More specifically I consider a class
of steady-state equilibria in which the level of liquidity supply (i.e the number of limit orders
in the book) is non zero.

44
Definition 1.1. A limit order market is in a steady state when the displayed depths in the
order book and the different order flows are deterministic and do not change over time.

This steady state is possible in the model because there is a continuum of investors. Each
investor faces idiosyncratic uncertainty on her type. She switches from ”high” to ”low” or
”low” to ”high” with respect to a Poisson process of intensity ρ. By the law of large numbers
applied to the continuum of investors, the share of investors switching from one type to
another is deterministic and equal to ρ.dt at each point in time. For the same reason the
share of investors monitoring the market is deterministic and equal to λ.dt. This generates
a time continuous flow of investors monitoring the market.

Proposition 1.1. For each couple of bid and ask prices (A, B) that verifies the conditions

v δ ρ v ρ
− + ∆ ≤ B < A ≤ − ∆,
r r r r r

A − B = ∆,

there is a unique steady-state equilibrium in which

• all sell limit orders are submitted at price A and all buy limit orders are submitted at
price B.

• The market depths at these two prices are the same and equal to

0 1 ρ∆
DA = DB = αeq = . (1.2)
2 δ − r∆

Proof. see Appendix A.2

This proposition describes trading prices and the aggregate level of liquidity supply in a
specific class of steady-state equilibria. For each pair of bid and ask prices (A, B) satisfying
the stated conditions there is a unique equilibrium in which all liquidity supply is concentrated
at these prices. The collection of all these equilibria generates the class of these equilibria.
In the following of this section, I detail the construction of such an equilibrium and its
underlying trading dynamics.

45
Remark 1.1. The assumption δ − (r + 2ρ)∆ > 0 is necessary to ensure that the interval
h i
v
r
− rδ + ρr ∆, vr − ρr ∆ is non-empty and larger than ∆.

To understand the inequality involving A and B, we can look at the subset of equilibrium
prices " # " #
v δ r+ρ v δ ρ v δ ρ v ρ
− , − ⊂ − + ∆, − ∆
r r r + 2ρ r r r + 2ρ r r r r r
v r+ρ
This inclusion is a consequence of δ − (r + 2ρ)∆ > 0. r
− rδ r+2ρ is the value for a ”low” type
ρ
investor to hold the asset forever and vr − rδ r+2ρ is the value for a ”high” type investor to hold
the asset forever. These are the reserve values for these two types of investor when they hold
the asset. In the case where a low-type owner and a high-type non-owner meet only once
and leave the market afterwards then the trading price has to be in this interval for the two
investors to trade.

In the steady state equilibrium of the limit order market trading also takes place between
low type owners and high type non-owners. However the range of trading prices is wider than
the difference between the two reserve values because investors can trade more than once.

Other equilibria. There are other equilibria than the one described above. Indeed in
order to solve for the equilibrium of this game one must proceed by guess and check. The
first step is to conjecture equilibrium strategies for all agents. The easiest is to assume that
all agents have the same strategy. Given this strategy it is possible to determine the dynamic
of the limit order book. The last step is then to check that it is not profitable to operate a
one-shot deviation from the conjectured strategy for any type, at any point in time of the
game while other agents are playing the conjectured strategy. Solving the problem in that
way is difficult. Defining the set of all equilibria is even harder.

An example of other equilibrium is the empty limit order book equilibrium. In this
equilibrium Investors coordinate on a trading price P where ho’s and ln’s send (marketable)
limit orders. The buy and sell order flows due to lo’s and hn’s are exactly equal which
implies that their limit orders are immediately executed and that the limit order book is
always empty.

46
1.3.1 One-tick market

Proposition 1.2. A limit order market in a steady state at equilibrium is necessarily a one-
tick market. Its bid-ask spread is equal to the tick of the market, A − B = ∆. Moreover
liquidity supply is concentrated at best bid and ask prices:

• all sell limit orders are sent at the price A, generating a depth DA , and there are no
sell limit orders at higher prices than A

• all buy limit orders are sent at the price B, generating a depth DB , and there are no
buy limit orders at lower prices than B

In a steady state at equilibrium limit orders and market orders are sent by investors
following an equilibrium strategy. It generates flows of limit and market orders that are
constant and deterministic over time so that the steady state holds. Let’s consider an in-
vestor for whom it is optimal to send a buy market order at A. If there was a reachable price
A < P < B it would be profitable to send a limit order at P since it would be immediately
hit by the flow of market orders and would get price improvement compare to A. This would
contradict the optimality of the strategy.

This one-tick market result relies on the modelling approach. There is a continuum of
investors and a «zero or one unit» holding constraint. Random idiosyncratic events affect a
deterministic share of investors because of the law of large numbers and finally turn them into
deterministic flows of orders and cancellations. This flows are finite because of the holding
constraint. The key reason for this result is the market order flow that is deterministic and
continuously positive which makes any limit order alone inside the bid-spread immediately
executed. It is also the fact the instantaneous market order flow is infinitesimal and thus is
not big enough to move prices. One might think that in a large market where trades take
place quite continuously and where market orders are small enough to not push prices, for
instance if robots optimize execution by slicing big orders into small ones, then the occurrence
of one-tick bid-ask spreads could be high. Indeed the incentive to send a limit order inside
the best quotes rather than a market order would hold because execution would be almost
immediate.

47
1.3.2 Steady state strategy

When the limit order book is in a steady state trading takes place due to differences in private
values across investors. Investors of types ho and ln do not trade because prices are between
the values of owning the asset for high type and a low type. Given these prices, investors
of type hn and lo are better off after changing their holding status and thus trade. If they
use market orders they directly join the group of «satisfied» agents (ho and ln). If they use
limit orders they become «satisfied» once their order is executed. The consequence of this
strategy is that investors of type lo and hn who have once monitored the market are in the
limit order book. In the steady state they all are in the limit order book.

Proposition 1.3. The equilibrium strategy in the steady-state phase is defined as follows:

• ho: cancel any sell limit order and stay out of the market

• hn: send a buy limit or market order with respect to a mixed strategy. When she
monitors the market she submits a buy market order with probability mA ∈ [0, 1]. It is
executed at the ask price A.

• lo: send a sell limit or market order with respect to a mixed strategy. When she monitors
the market she submits a sell market order with probability mB ∈ [0, 1]. It is executed
at the bid price B.

• ln: cancel any buy limit order and stay out of the market

Proof. see Appendix A.2

In this equilibrium the populations Lho and Lln are not present in the limit order book.
As soon as a ho type switches to a lo type she instantaneously monitors the market: either
she instantaneously switches to a ln type by sending a sell market order or remains a lo type
by sending a sell limit order. Symmetrically as soon as a ln type switches to a hn type she
instantaneously monitors the market: either she instantaneously switches to a ho type by
sending a buy market order or remains a hn type by sending a buy limit order. Consequently
we obtain DA = Llo and DB = Lhn .

48
1.3.3 Steady state populations

In a steady state the levels of aggregate populations do not change over time. Then the flows
of population from high type to low type and from low type to high type must be equal to
each other, ρ(Lho + Lhn ).dt = ρ(Llo + Lln ).dt. Combined with the constraints due to the size
1 of the overall population and the asset supply 1/2 we obtain

1 1
Llo + Lln = Lho + Lhn = , Llo + Lho =
2 2

Proposition 1.4. In a steady state there is one freedom parameter α0 ∈ R such that the
different populations satisfy

1
Lho = Lln = ( − α0 ), Lhn = Llo = α0 .
2

1
It must satisfies the constraints of non-negativity, 2
− α0 ≥ 0 and α0 ≥ 0.

Proof. see Appendix A.2.2

This freedom parameter α0 is determined at equilibrium. It is equal to the liquidity supply

in the limit order book since the depths are equal to DA = Llo = α0 and DB = Lhn = α0 .

1.3.4 Micro-level dynamic of the limit order book

In equilibrium hn and lo investors are indifferent between limit and market orders so that
they submit both market and limit orders. Flows of limit and market orders make the state
of the limit order book sustainable and steady. And these flows must be steady as well.
The flows of buy market orders and buy limit orders are defined by the share mA of
hn investors monitoring the market between t and t + dt who send buy market orders and
the share 1 − mA who send buy limit orders. On the sell side a share mB of lo investors
monitoring the market between t and t + dt send sell market orders and the rest send sell
limit orders.

Ask Side. At time t, on the ask side of the market the depth is constantly equal to DA = Llo
and the order flows going in and out of the ask side of the order book are

49
 • Outflow due limit order executions: execution of buy market orders send by hn’s
monitoring the market, mA (λLhn + ρLln ).dt.

• Outflow due limit order cancellations: investors switching from lo to ho, ρLlo .dt,
lo’s cancelling their sell limit order to send a sell market order, mB λLlo .dt

• Inflow due to limit order submissions: investors switching from ho to lo submitting

a sell limit order, (1 − mB )ρLho .dt

The steady state condition is : ρLlo + mA (λLhn + ρLln ) + mB (λLlo + ρLho ) = ρLho .

limit order cancellations : by lo’s switching to ho’s

ρLlo .dt

by lo’s sending market orders

mλLlo .dt

α0
limit order submissions
by ho’s switching to lo’s :
(1 − m)ρLho .dt
A0

limit order executions by hn’s buy market orders :

m(λLhn + ρLln ).dt

Figure 1.1: Steady-state dynamic of the market depth

Bid Side. At time t, on the ask side of the market the depth is constantly equal to DB = Lhn
and the order flows going in and out of the bid side of the order book are

• Outflow due limit order executions: execution of sell market orders send by lo’s
monitoring the market mB (λLlo + ρLho ).dt.

50
 • Outflow due limit order cancellations: investors switching from hn to ln, ρLhn .dt,
hn’s cancelling their sell limit order to send a sell market order, mA λLhn .dt

• Inflow due to limit order submissions:: investors switching from ln to hn submit-

ting a sell limit order, (1 − mA )ρLln .dt

The steady state condition is : ρLhn + mB (λLlo + ρLho ) + mA (λLhn + ρLln ) = ρLln .

1.3.5 Execution rate and liquidity provision

At any time t in the steady state phase, the flow of market orders hits a share of limit orders
in the order book. Because of the Pro-Rata execution rule, all limit orders on the same side
of the book are equally likely to be executed. Between t and t + dt this probability is equal
to the ratio of the instantaneous flow of market orders over the market depth.
For instance on the ask side, the flow of market orders is equal to mA (λLhn + ρLln ).dt and
the depth is equal to DA = Llo . Hence the instantaneous probability of execution is equal to

mA (λLhn + ρLln )
lA .dt = .dt. (1.3)
Llo

lA is the execution rate for sell limit orders. In the same way we can define the execution
mB (λLlo +ρLho )
rate for buy limit orders, lB = Lhn
.
The execution rates are more natural to handle than the mixed strategy parameters
mA and mB as it clearly appears in the value function subsection. These quantities can
be used equivalently. Indeed, once the execution rates and the state of the limit order
book are defined at equilibrium, the mixed strategies are perfectly defined. For instance
Lhn α0
mB = l
λLlo +ρLho B
= l .
λα0 +ρ( 12 −α0 ) B

Steady state liquidity provision. By incorporating the execution rates, the two steady
state equations can be rewritten as

ρLhn + lB Lhn + lA Llo = ρLln (1.4)

ρLlo + lA Llo + lB Lhn = ρLho (1.5)

51
 These equations are in fact equivalent and define the value of the steady state population
that is to say the value of the depth parameter α0

1 ρ
α0 = (1.6)
2 2ρ + lA + lB

The aggregate properties of the limit order market in this steady state is completely
described by α and the execution rates lA and lB . Indeed they define the steady state
populations, the depths and the aggregate order flows in the limit order book.

1.3.6 Value functions

The equilibrium strategy generates the following system of equations defining the different
value functions for each investor type. Here I only provide the value function for ho and hn
investors as it is very similar for ln and lo.

Type ho. A ho investor stays out of the market until she switches to the lo type. Her
situation is affected when the common value changes. Her value function Vho−out is defined
as follows

Vho−out = v.dt + (1 − r.dt) [(1 − ρ.dt)Vho−out + ρ.dtVlo ]

⇐⇒ (r + ρ)Vho−out = v + ρVlo .

Type hn. A hn investor sends a buy market order with probability mA or limit order with
probability 1 − mA . Sending a buy market order at price A provides her with the value
function Vho−out − A. Indeed she gets execution immediacy by trading at the ask price A and
instantaneously switches to type ho. Sending a buy limit order at price B provides her with
the value function Vhn−B defined as follows

(r + ρ + lB + mA λ)Vhn−B = ρVln−out + mA λ(Vho−out − A) + lB (Vho−out − B).

Once the limit order has been submitted several events can occur: either the investor’s
type changes with intensity ρ and becomes ln or the investor monitors again the market with

52
intensity λ and cancels her limit order to send a market order with probability mA or the
limit order is executed with intensity lB . Each of these events correspond to a change in the
utility function and define the value function of submitting a limit order. Types hn become
indifferent between limit and market orders if and only if Vhn−B = Vho − A. Then the value
function of a type hn is Vhn = Vhn−B = Vho − A.

The equilibrium value for the execution rate for buy limit orders, lB , is defined by the
following condition that makes a hn investor indifferent between using a limit or a market
order:
(r + ρ + lB )(Vho − A) = ρVln + lB (Vho − B). (1.7)

Similarly, the equilibrium execution rate for sell limit order is defined by the following indif-
ference condition for investors with type lo:

(r + ρ + lA )(Vln + B) = v − δ + ρVho + lA (Vln + A). (1.8)

Proposition 1.5. For any couple of equilibrium bid and ask prices (A, B), the equilibrium
limit order execution and value functions are as followed,

v − rA − ρ∆
lB = ,
∆
rB − ρ∆ − (v − δ)
lA = ,
∆
11 1 1
Vho = (v − ρ∆) + (v + ρ(A + B)),
r2 r + 2ρ 2
11 1 1
Vln = (v − ρ∆) − (v + ρ(A + B)),
r2 r + 2ρ 2
Vhn = Vho − A,

Vlo = Vln + B.

Proof. see Appendix A.2

53
1.4 Equilibrium outcomes

1.4.1 Limit order execution rates

The equilibrium limit order execution rates are such that investors with types lo and hn are
indifferent between getting execution immediacy with a market order or having a delayed
execution (for which the loss is measured by the time discount rate r) at a better price with
a limit order. On the ask and on the bid side of the market these equilibrium execution rates
are respectively equal to

rB − (v − δ) v − rA
lA = − ρ, lB = −ρ
∆ ∆
On average a sell (resp. buy) limit order remains in the book during a time 1/lA (resp.
1/lB ) before being executed. This is the maximum average time during which an investor
with type lo (resp. hn) is willing to wait with a limit order in the order book.

Limit order opportunity cost and execution rates. Depending on the equilibrium
prices (A, B), with A − B = ∆, the sell side or the buy side of the market extract more
of the trading surplus. Indeed the higher are A and B, the smaller is lB and the bigger is
lA . An investor with type hn requires a lower execution rate, is willing to wait longer in
the book, because his opportunity cost for not trading immediately, v − rA, declines. This
investor is better-off in an equilibrium with high trading prices. Conversely, an investor with
type lo requires a higher execution rate, is not willing to wait longer in the book, because his
opportunity cost for not trading immediately, rB − (v − δ), increases.

Trading surplus extraction, tick size and execution rates. One dimension of the
incentive to use a limit order is the opportunity to extract more of the trade surplus compared
to the option to use a market order and sell the asset at a lower price (resp. buy at a higher
price). This surplus extraction component is measured by the tick size since it captures the
price difference between market and limit orders. The bigger is ∆, the lower are lA and lB
since investors can extract more of the trading surplus by using limit orders and hence are
willing to wait longer in the book before being executed (cf Fig. 1.2). As we will see in the

54
next subsection, the incentive given to investors to extract trading surplus at the expense of
execution immediacy is welfare deteriorating.

Private value volatility and execution rates. The frequency ρ, at which the preference
for the asset of an investor switches from high to low or low to high, has a negative effect
on the equilibrium execution rates (cf Fig. 1.2). When ρ is high, investors anticipate that,
if they trade immediately after a change in type, they may trade again soon after a switch
back of their type. As a consequence, the incentive for trading is less. Investors suffer less
from waiting with a suboptimal type, lo or hn, with a limit order in the book.
The asset holding cost δ, that low type investors suffer from, has a positive impact on
lA . This effect goes through the opportunity cost channel. A higher δ implies a higher
opportunity cost for lo type investors who use a limit order.

0
Average execution rate. The formula for the average execution rate leq allows to capture
more easily the effects of the model parameters on the execution rates,

0
lA + lB δ − (r + 2ρ)∆ ∂leq ∂l0 ∂l0
0
leq = = and < 0, eq < 0, eq > 0.
2 2∆ ∂∆ ∂ρ ∂δ

2.5 2.5

H ∆L H ΡL H ∆L
2.0 0 2.0
0 l eq 0
l eq l eq
3
1.5 1.5

2
1.0 1.0

1
0.5 0.5

6 7 8 9 10 1 2 3 4 6 7 8 9 10

0
Figure 1.2: Execution rate leq in function of (i) ∆ ∈ [0, δ/(r + 2ρ)] (ρ = 2 δ = 10), (ii)
ρ ∈ [0, (δ − r∆)/2∆] (∆ = 1, δ = 10) and (iii) δ ∈ [(r + 2ρ)∆, 2(r + 2ρ)∆] (∆ = 1, ρ = 2) ,
(r = 1).

0
leq is also the execution rate in the symmetric equilibrium. This is the equilibrium in
which the term of the trade-off, limit order vs. market order, is the same on both side of the
market. The symmetric equilibrium prices are B = 1r (v − 2δ ) − ∆
2
, A = 1r (v − 2δ ) + ∆
2
and
0
makes the execution rates equal lA = lB = leq .

55
1.4.2 Market depth

On both bid and ask sides of the market, the market depth (i.e. the number of limit orders
submitted) is equal, at each point in time, to

0 0 0
0 1 ρ 1 ρ∆ ∂αeq ∂αeq ∂αeq
αeq = = and > 0, > 0, < 0.
0
4 ρ + leq 2 δ − r∆ ∂∆ ∂ρ ∂δ

0.25
0.24 0.24

0.22 0.20 0.22

0.20
0
Α eq H ∆L
0
Α eq H ΡL 0.20
0
Α eq H ∆L
0.15

0.18 0.18

0.10
0.16 0.16

0.14 0.05 0.14

0.12 0.12

6 7 8 9 10 1 2 3 4 6 7 8 9 10

0
Figure 1.3: Market depth αeq in function of (i) ∆ ∈ [0, δ/(r + 2ρ)] (ρ = 2 δ = 10), (ii)
ρ ∈ [0, (δ − r∆)/2∆] (∆ = 1, δ = 10) and (iii) δ ∈ [(r + 2ρ)∆, 2(r + 2ρ)∆] (∆ = 1, ρ = 2) ,
(r = 1).

The tick-size ∆ has a positive effect on the market depth since, everything else equal,
an increase of this parameter increases the trading surplus that one can extract with a limit
order (see Fig. 1.3).

The parameter ρ has a positive effect on the market depth since a higher ρ implies that, at
each time t, there is an increasing fraction of investors whose type have become suboptimal
and thus have trading need, lo and hn, which has a positive effect on the number of limit
order submitted. This effect is mitigated by a higher equilibrium limit order execution rate
but still remains positive (see Fig. 1.3).

δ has a negative effect on the trading intensity. Through the limit order opportunity cost
channel, a higher δ imposes a higher equilibrium execution rate which implies a lower market
depth (see Fig. 1.3).

56
1.4.3 Trading intensity/volume

On the ask and the bid side of the market, the trading intensities, are respectively equal to
0 0
lA αeq and lB αeq . Hence the overall trading intensity is

0 δ − (r + 2ρ)∆ 0 ρ δ − (r + 2ρ)∆
(lA + lB ) × αeq = αeq =
∆ 2 δ − r∆

0.5 0.5 0.5

0.4 0.4 0.4

Trading intensity H ΡL
0.3 0.3 0.3

Trading intensity H ∆ L Trading intensity H ∆ L

0.2 0.2 0.2

0.1 0.1 0.1

6 7 8 9 10 1 2 3 4 6 7 8 9 10

Figure 1.4: Trading intensity in function of (i) ∆ ∈ [0, δ/(r + 2ρ)] (ρ = 2 δ = 10), (ii)
ρ ∈ [0, (δ − r∆)/2∆] (∆ = 1, δ = 10) and (iii) δ ∈ [(r + 2ρ)∆, 2(r + 2ρ)∆] (∆ = 1, ρ = 2) ,
(r = 1).

The overall trading volume can be calculated as the integral of the trading intensity over
the all game period, [0, +∞), discounted at rate r. Thus trading volume is equal to trading
intensity multiplied by a factor 1/r.

The tick-size ∆ has a negative effect on the trading intensity since, everything else equal,
an increase of this parameter increases the trading surplus that one can extract with limit
0
order. Thus the equilibrium execution rate declines. The number of limit order αeq increases.
The overall effect on the trading intensity is negative (see Fig. 1.4).

The effect of the parameter ρ is not monotonic. On the one hand a higher ρ implies that,
at each time t, there is an increasing fraction of investors whose type have become suboptimal
and thus have trading need, lo and hn, which has a positive effect. On the other hand, a
higher ρ reduces the intensity of this trading need since investors anticipate that they may
more likely switch back to an optimal type. As a consequence the equilibrium execution rates
decline. The overall effect is positive for low ρ’s and negative for high ρ’s (see Fig. 1.4).

57
 δ has a positive effect on the trading intensity. Through the limit order opportunity cost
channel it imposes a higher equilibrium execution rate. This effect is mitigated by a lower
implied market depth but still remain positive (see Fig. 1.4).

1.4.4 Effects of the market monitoring frequency

Monitoring intensity irrelevance. An interesting feature of this equilibrium is that ag-

0
gregate outcomes, as αeq , do not depend on λ, the monitoring intensity. This is an expected
outcome of the model since trades occur because of differences in private values and because
these private values are monitored continuously. This suggests that market monitoring has
a limited role in a stable market. More specifically market monitoring plays a role when
liquidity supply is, for instance, cyclical as in Foucault et al. [2009]. In my model there is no
cycle since order flows are such that the order book is steady.

Continuous monitoring. The market monitoring rate λ does not impact the aggregate
0
values of the equilibrium, the value functions, the population levels linked to αeq or execution
rates lA and lB . Hence, we can take the model to the limit where investors are continuously
monitoring the market, λ = ∞. Let’s consider the ask side of the book and remind that the
flow of market orders hitting the ask side at t is equal to mA (λLhn + ρLln ).dt = mA (λαeq +
ρ( 12 − αeq )).dt = lA Lhn .dt = lA αeq .dt. This flow is independent of λ. When λ → ∞ we must
have mA → 0 so that this flow remains constant. At the limit, the flow of market orders is
equivalent to mA λαeq .dt which implies that mA λ → lA .
For an investor of type hn, mA λ.dt is the probability that she submits a market order at
time t. Noticing that allows to describe the investors strategy in the limit case. When an
investor switches to type hn, she submits a limit order at price B with probability 1, because
the probability to send a market order is mA that is infinitesimal. At time t her order is
either executed with probability lB .dt, or she decides to cancel it and to send a market order
with respect to a mixed strategy with probability lA .dt, or she cancels it if she switches to
type ln.
For the same reason, when an investor switches to type lo, she submits a limit order at
price A with probability 1. At time t her order is either executed with probability lA .dt, or

58
she decides to cancel it and to send a market order with respect to a mixed strategy with
probability lB .dt, or she cancels it if she switches to type ln.
Taking the limit case leads to an equilibrium in which investors play a Poisson mixed
strategy to choose between limit and market orders. If we were to consider directly the
problem with continuous monitoring we could end up with different types of mixed strategies
where, for instance, investors would submit a market order with positive probability when
their type changes and then play a Poisson mixed strategy. However these strategies should
be such that the flow of market orders and the execution rates are the same as the ones of
the equilibrium with infrequent monitoring, since the terms of the trade-off do not change.

1.5 Welfare analysis

Proposition 1.6. For any steady-state equilibrium with bid and ask prices (A, B), the level
of welfare W is the same and equal to

1 11 v 0 δ
 
0 0 0
W = − αeq × (Vho + Vln ) + αeq × (Vlo + Vhn ) = (v − ρ∆) − αeq ∆= − αeq . (1.9)
2 r2 2r r

The welfare is impacted negatively by ρ and the tick-size ∆, and positively by δ,

∂W ∂W ∂W
< 0, < 0, > 0.
∂∆ ∂ρ ∂δ

50.0
48.88 48.88

48.86 49.5 48.86

48.84 48.84
49.0

W H ∆L W H ∆L
48.82 48.82

W H ΡL
48.5
48.80 48.80

48.78 48.0 48.78

48.76 48.76

6 7 8 9 10 1 2 3 4 6 7 8 9 10

Figure 1.5: Welfare W in function of (i) ∆ ∈ [0, δ/(r + 2ρ)] (ρ = 2 δ = 10), (ii) ρ ∈
[0, (δ − r∆)/2∆] (∆ = 1, δ = 10) and (iii) δ ∈ [(r + 2ρ)∆, 2(r + 2ρ)∆] (∆ = 1, ρ = 2) ,
(v = 100, r = 1).
v1
The maximum level of welfare than can be reached is equal to r2
. It is obtained when
1 1
all the asset supply 2
is owned by high type investors whose population size is 2
as well. In

59
this situation each share of the asset is always offering a utility flow equal to v and then has
a value equal to vr . Reaching this optimum requires that when hn and lo investors come to
the market, after they changed of type, they can trade immediately at one price that is the
same for buyers and sellers.

The role of the tick size. In a steady-state equilibrium this optimum can be reached if
the tick size is nil, ∆ = 0, as we can see in the formula for the welfare (equation 12). The tick
size is the friction that prevents from reaching the optimum. Because there is a difference
in the execution prices for limit and market orders, investors have an incentive to send limit
orders and to wait for execution whereas it would be socially optimal that these orders get
0 δ
immediately executed. The corresponding loss is captured by the term −αeq r
. It shows that
0
the presence of liquidity supply, αeq is suboptimal.
When ∆ = 0, the bid and the ask prices are infinitely close. Then the price improvement
of submitting limit orders is nil and the execution intensities lA and lB must be infinite to
incentivize limit order submission. Because of these infinite execution rates, limit orders
are instantaneously executed and the limit order book is always empty. We can view this
equilibrium as a situation where investors coordinate to trade with each other at a single
price P = A = B and where there is no difference between limit and market orders.

Effects of the private value volatility components. The idiosyncratic preference

switching frequency, ρ, has a negative effect on the welfare (cf Fig. 1.5) through its in-
0
creasing effect on αeq . The asset holding cost δ, for low type investors, has a positive effect
on the welfare.
It is interesting to compare these effects with the benchmark case where there is no
trading. In this case the level of welfare is given by the value function of investors with
type lo and ho, and the initial fractions of these investors type. These value function are as
followed,
v δ r+ρ v δ ρ
Vlo = − , Vho = − .
r r r + 2ρ r r r + 2ρ
Depending on the initial level of populations, the effect of ρ can be positive or negative since
it has a positive effect for lo type (they switch to a high private value faster) and a negative

60
effect for ho type. For instance, in the steady-state case, the initial fraction of each investor
v δ
would be 1/4, the welfare would be equal to 2r
− 4r
and the parameter ρ would have no
effect. In comparison, when there is trading, the welfare become sensitive to this parameter
ρ even in steady-state.
The effect of δ is very clear in the situation without trading since, everything else equal,
increasing this cost induces an actual or an expected utility loss for all investors who own
the asset. Thus it is noticeable that δ has the opposite effect when investors can trade. It
generates an opportunity cost for using limit orders which accelerate trading and make the
asset allocation across investor more optimal.

Model outcomes and proxies for investor’s welfare. To empirically investigate the
source of welfare variations for investors in a given financial market, we need observable
proxies for welfare. Usually liquidity measures are thought as positively related to investor’s
welfare, since higher market liquidity leads to a lower implicit trading cost. My model
provides counter intuitive results in this respect.

0 0
Limit order execution rate (leq ) Market Depth (αeq ) Trading Intensity Welfare

∂
∂∆ − + − −
∂
∂ρ − + +/− −
∂
∂δ + − + +
Figure 1.6: Signs of first order partial derivatives of model’s outcomes with respect to model’s
parameters ∆, ρ and δ.

The previous results for welfare and market depth shows that any variation of parameters
∆, ρ or δ has opposite effects these model outcomes. Market depth, a traditional liquidity
measure, negatively co-varies investors welfare, at least in this model. Trading intensity,
which is also a usual liquidity measures, co-varies much better with the welfare (cf. Fig 1.6)
except for variations of ρ when it has low values. The limit order execution rate is the only
model outcome that always positively co-varies with investor’s welfare.

61
 0 0
Limit order execution rate (leq ) Market Depth (αeq ) Trading Intensity Welfare

∂
∂∆∂ρ 0 + − −
∂
∂∆∂δ − − + +
∂
∂ρ∂δ 0 − + +
Figure 1.7: Signs of second order cross partial derivatives of model’s outcomes with respect
to model’s parameters ∆, ρ and δ.

One could also want to investigate the effect on liquidity and welfare of the change of
one model parameter across different markets that could be sorted with respect to a second
parameter. For instance one could look at the effect of decimalization, a reduction of ∆,
on the cross section of security markets sorted with respect to the holding cost δ (which
would imply using a proxy for such a cost though). To implement these kind of empirical
analysis and draw conclusion on the cross sectional effect of such a shock on welfare, one
would need a welfare proxy for which the second order cross partial derivative, with respect
to the ”shocked” parameter and the parameter used to sort the markets cross section, has
at least the same sign as the corresponding derivative for the welfare. In our setup, trading
intensity would be the best proxy (cf. Fig 1.7).

1.6 Conclusion
Market liquidity measures, as bid-ask spread and market depth, usually focus on implicit
trading costs for liquidity consumers. In limit order markets, these measures do not capture
the execution quality of limit orders, which are used by traders who decide to supply liquidity.
Hence these liquidity measures may not be sufficient to infer investors’ welfare. In this paper,
I show, with a model, that market depth can be negatively related to investors welfare because
high a market depth reflects a low execution rate of limit orders and a relatively low rate
for the gains from trade realization. In this context, the limit order execution rate and the
trading volume better capture investors’ welfare.

62
Chapter 2

Limited Attention and News Arrival

2.1 Introduction

Investors have limited attention capacities and thus cannot monitor continuously the flow
of information in financial markets. As a result they are unable to get or analyze instan-
taneously implications of public financial news when they arrive. And news content cannot
instantaneously turn into common knowledge for the market. Consequently at short horizon
public information is private information for investors who observe it first. Because of limited
attention, public information release generates a short term period of information asymmetry.
How do financial markets react around news arrival? And what role does limited attention
play in this process?
To address these questions I propose a theoretical framework to analyze the role of limited
attention on market reaction to news. I design a model of limit order market in presence of
uncertainty on the asset value due to news arrival. This model extends the OTC markets
framework of Duffie, Garleanu and Pedersen [2005, 2007] to limit order markets. In Duffie et
al., the main market imperfection is the search friction for trading counterparty. In my model
the market imperfection comes from investors’ limited attention capacity. It is equivalent to
an imperfect monitoring of the market and news arrival. Investors cannot continuously
observe public information and contact the market. They do it at some random market
monitoring times. This setup generates a gradual diffusion of new public information among
investors after news arrival. Investors’ imperfect market monitoring allows to jointly describe

63
liquidity formation, price discovery and market efficiency around news arrival.
Financial markets reaction to public information has motivated an extensive line of re-
search both empirical and theoretical especially since the 90’s. Among other goals, studying
the market reaction to public information allows for a better understanding of the price dis-
covery process in financial markets. The question of market reaction to public information
has been addressed empirically by Eredington and Lee [1995], Fleming and Remolona [1999]
and Green [2004] for instance. These papers study the reaction of US Treasury securities
markets to scheduled macroeconomic announcements. The first two papers show that the
market reacts to the announcement in two successive phases. In the first phase, the price
shifts quickly to a new level in line with the main figures of the announcement. The second
phase of this reaction is characterized by a high volatility, suggesting that investors disagree
on the precise interpretation of the announcement. This phase ends when the announcement
interpretations of market participants eventually converge. Green’s paper shows that these
macro announcements increase the level of adverse selection, suggesting that investors with
better processing abilities can take advantage of these events.
With respect to this literature, a significant contribution of this paper is to consider
unscheduled news. Financial news are released everyday by news providers, as Thomson
Reuters or Bloomberg, and deliver relevant information for evaluating asset prices. Virtually
all these news arrive at unscheduled times1 . In addition, the frequency of these news arrival
varies a lot across stocks2 . The unscheduled nature of these events is likely to prevent
investors from perfectly paying attention to financial news. By embedding limited attention
in a limit order book model, I can speak to the question as to how unscheduled news affect
trading decisions and ultimately price formation and liquidity provision.
The current evolution of financial markets supports the choice of limited attention as
an important determinant to address the short term dimension of market reaction to public
information. These short term reactions have become an important issue since some High
Frequency Trading activities have grown by using intensive monitoring technologies to trade
very fast on financial news. More generally the boom of Algorithmic Trading (that includes
1
In a subsample of 40 large stocks, representing 70% of the market cap of the FTSE100, Gross-Klussmann
and Hautsch (2011) find that one stock receives on average 750 unscheduled news over 1.5 year
2
In the same subsample, the news arrival frequency varies ten-fold, from 200 to 2000 news

64
HFT) partly stems from investors’ needs to improve the market monitoring dimension of
their trading strategies. This trend shows how important for investors is the capacity of
attention that they allocate to monitor markets. Moreover my model considers electronic
limit order markets which has been adopted by most equity and derivative exchanges, and
which has enabled the growth of Algorithmic Trading.

The paper has several empirical implications for liquidity supply and price dynamics
around news arrival. When the frequency of news arrival increases, (i) the level of liquidity
supply decreases, (ii) prices adjust faster following news arrival and (iii) the relative impor-
tance of limit order cancellations in the price adjustment process declines. The intuition for
these results stems from the short term period of information asymmetry around news arrival
that is due to limited attention. Consistently with the presence of information asymmetry,
there is a ”picking-off” risk for liquidity suppliers and this risk varies with the frequency of
news arrival. Investors may be reluctant to supply liquidity with limit orders since, following
news arrival, limited attention delays their reaction. In the meantime, their limit orders
can be ”picked-off” because they are not in line with the new asset value and offer a profit
opportunity.

In my framework, investors can both supply liquidity with limit orders and consume
liquidity with market orders. Before news arrival, investors trade with each other because
their private values for the asset are different which generates gains from trade. During this
phase the limit order book is in a steady state. The level of liquidity supply is constant and
determined by the following trade-off. Market orders provide execution immediacy whereas
limit orders provide price improvement but bear execution delay and a picking-off risk when
the asset value changes. At equilibrium the level of liquidity supply adjusts so that investors
are indifferent between market and limit orders.

When, following news arrival, the asset value changes, it is publicly available but investors
do not observe this change immediately. They become aware of it after a while which depends
on their monitoring intensity. This generates a transition phase at the end of which prices
adjust to the new asset value. This price discovery process relies on two underlying dynamics.
Investors who observe the new asset value fast enough can profit from a transitory arbitrage
opportunity by using market orders to ”pick-off” stale limit orders at the initial price. And

65
investors, with limit orders in the order book, cancel these orders to avoid being picked-off by
previous market orders. Once limit orders at the initial price level have all been cancelled or
picked-off, the transition phase stops and the limit order book converges to a new steady-state
without uncertainty on the asset common value. Thus the model provides a high-frequency
description of price and order dynamics around news arrival. This should prove useful for
empiricists3 .
The decision for investors to use limit or market orders to trade before news arrival
depends on the risk to be picked-off during the transition phase. Everything else equal this
risk enhances the expected loss associated with limit order submission and has a negative
impact on the liquidity supply. In this context, the effect of the frequency of news arrival
is intuitive. More frequent news releases increases the likelihood of an event where a limit
orders may be picked-off which turns into a higher picking-off risk. Consequently the level
of liquidity supply measured by the market depth (the number of limit orders in the order
book) is negatively linked to this frequency. In a thinner market, the amount of stale limit
orders that must be cancelled or executed in the transition is less which makes the price
adjustment faster.
Investors’ limited attention capacity influences the level of this risk and thus the liquidity
supply prior to news arrival. More attention however has an ambiguous effect. To see why,
let’s consider an increase of investors monitoring intensity4 . On the one hand investors can
cancel their limit orders faster after news arrival which reduces their risk of being picked-off
and makes limit orders more profitable. However, investors can also send directional market
orders faster to execute against stale limit orders which worsen the risk of being picked-off
for limit orders. Overall limit orders could be more or less profitable after an increase of
monitoring intensity. I identify conditions which make limit orders overall more profitable
after such an increase. However the magnitude of its effect on liquidity supply appears to be
small especially when compared to the effect of the news arrival frequency. This suggests that
only relative monitoring abilities, with respect to other market participant, really matters to
understand how this parameter can quantitatively affect investors trading strategies around
3
Engle et al. (2009) use high-frequency limit order book data to analyze liquidity and volatility in the
U.S. Treasury market.
4
This increase could result from a reduction of latencies.

66
news arrival.
Chapter 2 is organized as follows. Section 2.2 reviews the related literature. Section 2.3
presents the setup and assumptions of the model. Section 2.4 gives the equilibrium and its
general description. Sections 2.5, 2.6 and 2.7 describe the properties of the different phases
involved in the equilibrium dynamic of the limit order market. Section 2.8 discuss empirical
implications of the model. Section 2.9 concludes.

2.2 Literature review

The reaction of financial markets to public news has been and still is an active field of
research. Kim and Verrecchia [1991, 1994] have proposed models to analyze the market
reaction to public information release as earnings announcements. They make predictions
about market liquidity or trading volume around these events. As mentioned above it has
also inspired empirical studies as those of Eredington and Lee [1993, 1995], Fleming and
Remolona [1999] and Green [2004]. More recently Tetlock [2010], using a large data set
on all types of public news, has addressed the effect of public news release on the level of
asymmetric information and how it affects returns around these events. Della Vigna and
Pollet [2009] link market reaction to earnings announcements and investor’s attention to
explain post-earnings announcement drift.
The link between attention capacity and investors’ decisions in financial markets is a fairly
new research topic. Some recent works by Peng and Xiong [2006], Van Nieuwerburgh and
Veldkamp [2009] or Mondria [2010] have developed theories where investors have a limited
capacity of attention and allocate it across assets. The more they allocate attention to an
asset the more precise is their information about its future pay-off. Through this channel
these papers analyze the effect of limited attention on portfolio diversification and asset
prices. My paper contributes to this literature by mapping limited attention capacity to
imperfect market monitoring at the high-frequency level. It allows me to analyze its effect
on trading mechanism.
Market monitoring imperfection has already been stressed as a key determinant of market
dynamics by Darrell Duffie Presidential Address [2010]. Foucault, Kadan and Kandel [2009]

67
address this problem in a limit order book framework. In their paper agents strategically
choose their level of market monitoring but are exogenously considered as limit order or
market order users. Biais and Weill [2009] and Biais, Hombert and Weill [2012] also consider
imperfect monitoring agents. The focus of their papers differs from mine since they consider
limit order market dynamics generated by aggregate liquidity shocks rather than asset value
uncertainty. Moreover I model how asset value uncertainty affect trading strategies before
the change in the asset common value occurs. Whereas, in Biais et al. modelling, the market
dynamic starts with the liquidity shock. In my model the change in the asset common value is
a publicly observable signal but it is not instantaneously observed since market monitoring is
imperfect. Pagnotta and Philippon [2012] study the effect of competition between exchanges
for the market monitoring intensity, or latency, they provide to their customers. They iden-
tify this competition as an incentive for investing in fast trading technologies. Biais et al.
[2009,2012], Pagnotta and Philippon [2012] as well as my paper use and adapt the model of
search friction in OTC markets introduced by Duffie, Garleanu and Pedersen [2005, 2007]
and extended by Lagos and Rocheteau [2009], Lagos, Rocheteau and Weill [2011], Vayanos
and Weill[2008] and Weill [2007,2008].

The effect of information monitoring on market liquidity provision has been studied by
Foucault, Roell and Sandas [2003] in the case of a dealership market. In their model Market
Makers face adverse selection by informed traders and can reduce this risk by monitoring
public information and adjusting their quote. The choice of the monitoring intensity is costly.
In my model monitoring intensity is an exogenous parameter but it affects both liquidity
supply and demand which is more consistent with how limit order markets work. Goettler,
Parlour and Rajan [2009] design a very realistic environment of limit order market that is
not tractable and meant to be solved numerically. In their paper traders do not continuously
monitor the market and decide ex-ante to be privately informed or not about the asset value.

This paper also builds on the dynamic limit order market literature. There are quite a
few papers dealing with this problem when compared to its practical importance. One of
the reasons is that limit order markets are very hard to model. Foucault [1999] and Parlour
[1998] are the first models of limit order markets designed as dynamic games capturing the
inter-temporal aspect of the problem. The tractability of these models is appreciable but is

68
reached at the cost of strong assumptions. Both incorporate private and/or common value
as drivers of trading and price formation processes but do not allow for strategic decision
over the limit order lifetime. Foucault, Kadan and Kandel [2005] focus on the dynamic of the
liquidity supply in a limit order market. In their paper investors trade for liquidity reasons
and solve the market vs. limit order trade-off in function of their preference for immediacy.
Rosu [2009] generalizes Foucault, Kadan and Kandel framework and design a continuous
time model where traders can freely send limit orders at any price and can cancel them.
Rosu [2010] add a common value environment to his previous model. The two papers by
Rosu are build on the fundamental assumption that limit orders are continuously monitored
by their owners. As in Rosu’s models I design a framework that allow for an entire freedom
of choice for investors’ order management at the exception of the zero or one unit holding
constraint (as in Rosu [2009,2010]). Pagnotta [2010] designs a limit order book model with
insider trading where agents optimally choose their trading frequency. At equilibrium they
don’t trade continuously but they continuously observe the market and update their belief
accordingly.

2.3 Model

2.3.1 Asset value dynamic

The model presented in chapter 2 builds on the framework introduced in chapter 1. The
additional modelling block is the introduction of uncertainty for the asset value.

Preferences, Asset holding and supply. See chapter 1, section 1.2.1.

Asset value dynamic. The dynamic of the asset common value vt is the following5 :
5
The modelling of the asset value dynamic is equivalent to other dynamics where the asset pays-off at
some random time in the future and does not provide a continuous flow of utility. For instance the two
following formulation deliver the same result.
• The asset pays off the cash-flow V = vr at a random time that occurs with respect to a Poisson process
of intensity r. And being a low type induces a cost for holding the asset which is equal to δ per unit
of time.
• Or, at a random time that occurs with respect to a Poisson process with intensity r, The asset pays
off a cash-flow vr for high types and v−δ
r for low types.

69
 • at t = 0, the asset common value is equal to v0

• at date τ > t, the common value of the asset changes. This time τ is random and follows
a Poisson distribution of intensity µ, P(µ). At date τ the common value switches to
1
v u = v0 + ω or v d = v0 − ω with equal probabilities 2

• for t > τ the asset value is v u or v d until the end of the game.

• for 0 < t < τ the state of the world is ζ = 0. For τ < t the state of the world is either
ζ = u if vt = v0 + ω or ζ = d if vt = v0 − ω.

Time τ corresponds to the news arrival event in this setup. Consequently µ can be
interpreted as the news arrival frequency since it is the likelihood for such an event to occur
at next period. Parameter ω measures the news surprise that is to say the innovation of the
asset common value that is linked to the news content.

Assumption 2.1. I assume that ω is big compared to δ. It ensures that, if the change in the
common value, of magnitude ω, is not followed by a change in price, the profit opportunity,
measured by ω, is bigger than gains from trade due to differences in private values, measured
by δ. More specifically, I assume that
" #
2r + ρ
ω > 3δ × max 1, . (2.1)
2ρ

2.3.2 Limited attention

Investors have a limited capacity of attention which means that they are not able to process
information instantaneously. As a consequence they can neither track and interpret contin-
uously the flow of public news nor the rest of market activity. Moreover they must allocate
their attention capacity across different tasks which prevents them from being continuously
in contact with the market, able to trade.

Assumption 2.2. I assume that an investor observes the asset value, the market, her private
value and contacts the market at some random times {ti }i∈N . I call these times ”market

70
monitoring times”. This sequence of market monitoring times is generated by a Poisson
process of intensity λ + ρ.
More specifically between time t and t + dt an investor monitors the market in two types
of situation:

• when she uses the market monitoring technology which occurs with probability λ.dt

• when her private value changes which occurs with probability ρ.dt.

The level of investor’s attention allocated to the market monitoring is measured by λ. At

the limit λ = ∞ investors continuously monitor the market. In a sense they are infinitely
attentive to the market and to the flow of news.

2.3.3 Limit order market

See chapter 1, section 1.2.3.

δ
Assumption 2.3. As in chapter 1, I assume that r
is big compared to ∆ but I need this
difference to be higher. More specifically I assume that

δ > (r + 4ρ)∆ (2.2)

2.3.4 Value function and equilibrium concept

See chapter 1, section 1.2.4.

2.4 Equilibrium

In the first part of this section, I provide the equilibrium result of this chapter. I focus on
the symmetric equilibrium. This equilibrium is the core of the chapter. I describe it in more
details in section 2.5, 2.6 and 2.7. I derive empirical implications from this equilibrium in
section 2.8. In the second part of this section, I show that this is not a unique equilibrium.

71
2.4.1 The symmetric equilibrium

Proposition 2.1. For any intensity of news arrival, µ, there exists an equilibrium for which
v− 2δ
bid and ask prices are symmetrical with respect to the average private value of the asset, r
,

1 δ ∆ 1 δ ∆
B = (v − ) − , A = (v − ) + .
r 2 2 r 2 2

There are 3 pairs of these prices at which trades can take place: the one at the beginning of
the game, (A0 , B 0 ), and the one at the ends of the game, (Au , B u ) and (Ad , B d ). At these
prices the equilibrium is unique.
Before news arrival, the depths of the limit order book at price A0 and B 0 are both equal to
the depth parameter α0 (DA0 = DB 0 = α0 ). Limit orders submitted at these prices executes
according to a Poisson distribution of intensity l0 . These two parameters are interdependent:

• l0 is such that investors are indifferent between limit and market orders. It depends,
among other things, on the state of the order book and particularly on α0

• α0 depends on l0 because l0 determines the flow of executed limit orders and consequently
the level of liquidity supply in the order book.

0 0
The equilibrium values (αeq , leq ) are the fixed point solution to this problem of interdependence.

Proof. see Appendix A.5

Equilibrium. The equilibrium strategy and the resulting limit order book dynamic (cf
figure 1) have the following features:

• In the first phase, for 0 ≤ t < τ , when the asset common value is equal to v0 , trading
occurs because of differences in private values across investors: lo’s and hn’s trade
with each other via the limit order book, ho’s and ln’s do not trade. In particular buy
limit orders are submitted by types hn and sell limit orders are submitted by types lo.
During this phase the dynamic of the limit order market is the following:

– the limit order book is in a steady-state phase. Liquidity supply at A0 and B 0

0
is equal to αeq and does not vary over time.

72
 LO : Limit Order
MO : Marker Order Au

Bu
LO cancel
& resubmit

LO cancel
& resubmit
A0
MO
B0

b b
time
News Arrival Transition End

Figure 2.1: Dynamic of the limit order book following ”good” news arrival

– lo type investors submit sell limit or market orders using a mixed strategy. They
choose to submit a market order at price B 0 with probability m or a limit order
at price A0 with probability 1 − m. They do not submit sell limit orders at prices
higher than A0

– hn type investors submit buy limit or market orders using a mixed strategy. They
choose to submit a market order at price A0 with probability m or a limit order
at price B 0 with probability 1 − m. They do not submit buy limit orders at prices
lower than B 0 .

– A sell or buy limit order submitted at price A0 or B 0 executes according to a

0
Poisson distribution of intensity leq such that lo’s and hn’s are indifferent between
limit and market orders.

• Once the asset common value has changed, a transition phase starts for the limit
order book. This transition phase lasts for a finite duration T . During this phase, for

73
 τ < t < τ + T , trading takes place because the new common value of the asset is not
in line with the market prices and thus generates a profit opportunity.

– if the new common value is v0 + ω, investors with type lo cancel their sell limit
order and resubmit them at a higher price Au , investors with types hn and ln send
buy market orders to execute against stale limit orders at price A0 and then stay
out of the order book, investors with type ho stay out of the order book.

– if the new common value is v0 − ω, investors with type hn cancel their buy limit
order and resubmit them at a lower price B d , investors with types ho and lo send
sell market orders to execute against stale limit orders at price B 0 and then stay
out of the order book, investors with type ln stay out of the order book.

• When all limit orders that could potentially be picked-off have been executed or can-
celled the transition phase is over. Trading occurs once again because of differences
in private values across investors. During this last phase ho’s and ln’s do not trade,
lo’s and hn’s trade with each other via the limit order book. The equilibrium strategy
and the limit order book aggregate state converge to a steady-state phase that has
the same features as the first phase except that there is no uncertainty for the future
common value, µ = 0. Bid and ask prices are either Au and B u or Ad and B d depending
on the previous change in the common value. The limit steady-state corresponds to
the equilibrium studied in chapter 1.

Figure 2.1 summarizes the equilibrium dynamic of the limit order book.

2.5 Limit order book in steady state

In this section I explicit the equilibrium strategy in the first phase of the game. This first
stage of the game is a steady state equilibrium similar to the equilibrium studied in chapter
1. The additional feature of this equilibrium is that investors’ strategy takes into account a
possible change of the asset value following news arrival. I first show how the value functions
of investors are affected by the introduction of asset value uncertainty. In a second step I
solve for the equilibrium in a symmetric setup.

74
2.5.1 Value functions

As in chapter 1, I only provide the value function for ho and hn investors as it is very similar
for ln and lo.

Type ho. As in chapter 1, a ho investor stays out of the market until she switches to the lo
type. Her situation is affected when the common value changes. Her value function Vho−out
is defined as follows

1 u 1 d
  
Vho−out = v.dt + (1 − r.dt) (1 − ρ.dt − µ.dt)Vho−out + ρ.dtVlo + µ.dt Vho−out (0) + Vho−out (0)
2 2
µ u d
⇐⇒ (r + ρ + µ)Vho−out = v + ρVlo + [Vho−out (0) + Vho−out (0)].
2

The term µ2 [Vho−out

u d
(0) + Vho−out (0)] corresponds to the change in utility when the asset com-
mon value changes, up or down. These are the values of being a type ho at the beginning,
the time 0, of the transition phase.

Type hn. As in chapter 1, a hn investor sends a buy market order with probability mA
or limit order with probability 1 − mA . Sending a buy market order at price A provides her
with the value function Vho−out − A. Indeed she gets execution immediacy by trading at the
ask price A and instantaneously switches to type ho. The last additional term corresponds,
as above, to the utility change when the asset value changes. Sending a buy limit order at
price B provides her with the value function Vhn−B defined as follows

(r + ρ + lB + mA λ + µ)Vhn−B = ρVln−out + mA λ(Vho−out − A) + lB (Vho−out − B)

µ u d
+ [Vhn−B (0) + Vhn−B (0)].
2

As in chapter 1, the execution rate lB (resp. lA ) is defined by the condition that a hn

investor (resp. lo) is indifferent between using a limit and a market order. It is easy to
check that this value do not depend on the mixed strategy mB . Actually this equilibrium
parameter is well defined by the formula that links the mixed strategy and the execution
rate. Typically lB depends on value functions in the transition phase. These value functions

75
depend on α0 since the level of liquidity provision affects the duration of the transition phase
among other things. Remind that α0 corresponds to the depth of the limit order book and
that the transition phase lasts until this depth has been completely executed or removed.
In the end lB depends on α0 and conversely. Solving for the equilibrium of the game is
equivalent to solve this fixed point problem in the first steady state phase. It also requires
to solve for the game starting with the transition phase.

Compensation for providing liquidity. The equilibrium limit order execution rate can
be seen as a compensation for risk taking. When an investor submits a limit order instead
of a market order, she chooses her execution price but renounces to execution immediacy
and eventually a risk of being picked-off when the asset common value changes. Providing
liquidity requires therefore a compensation for risk taking. This compensation is obtained by
an appropriate execution delay for limit order execution. More precisely the execution rates
lA and lB must incentivize liquidity provision via limit orders. In equilibrium these execution
rates are such that market and limit orders are equally profitable for types hn and lo. This
mechanism appears clearly in agents’ value functions (next subsection).

2.5.2 Steady state in the symmetric equilibrium

In the initial steady state of the symmetric equilibrium, investors coordinate to trade on the
following bid and ask prices:

1 δ ∆ 1 δ ∆
B 0 = (v0 − ) − , A0 = (v0 − ) + .
r 2 2 r 2 2

The symmetry of this equilibrium implies that the term of the trade-off between limit order
and market order is the same on both side of the market. The execution rates that make
investors indifferent between limit and market orders are the same for sell and buy orders:

lA0 = lB 0 = l0 .

The steady state condition for the limit order book implies that the depths of the order

76
book, measured by α0 and the equilibrium execution rate are linked by the following formula:

ρ 1
 
0
α = 0
⇐⇒ l0 = ρ − 1 = g(α0 ). (2.3)
4(ρ + l ) 4α0

The execution rate implied by this formula is infinite for α0 = 0 (g(0) = ∞) and nil for
1
α0 = 4
(g(1/4) = 0).

Proposition 2.2. The execution rate, l0 , that makes investors with types lo and hn indiffer-
ent between limit and market orders is a function of α0 , l0 = f (α0 ). For α0 = 0 and α0 = 41 ,
f is finite and positive. Moreover f is decreasing with respect to α0 ,

∂f
< 0. (2.4)
∂α0

0 0 0
There is a unique αeq ∈ [0, 1/4] such that f (αeq ) = g(αeq ). The intersection point defines
0 0
the equilibrium values αeq and leq (cf Figure 2).

Proof. see Appendix A.5 and A.6

g IΑ 0 M

2000

f IΑ 0 M
1500

1000

0.0002 0.0004 0.0006 0.0008 0.0010

Figure 2.2: The two functions of f and g in function of α0 (λ = 100, µ = 50, r = 1, ρ =

2, ∆ = 1, δ = 10, ω = 50).

77
2.6 Limit order book in transition phase
When there is uncertainty, µ > 0, the first steady state phase lasts a finite time (almost
surely) and is followed by a transition phase. Prior to the asset value change, the world is
in the state ζ = 0. The transition phase starts when the asset common value changes. This
corresponds to the public news arrival event. It occurs at some point in time τ that follows
a Poisson distribution, P(µ). For times t > τ the state of the world is either ζ = u (up) with
v = v 0 + ω or ζ = d (down) with v = v 0 − ω with equal probability. I call T u and T d the
duration of the transition phases in the different states of the world.

2.6.1 Transition phase strategy

Once the common value has switched to a higher level for instance, non-owner turn into
arbitrageurs and have an incentive to buy the asset while it is tradable at a low price, A0 ,
and to resell it at a higher price Au later. At equilibrium investors coordinate on the price
Au at which they will trade the asset in the future.

Proposition 2.3. After the common value has changed, during the transition phase, the
strategy is:

• In the case ζ = u, for τ < t < τ + T u :

- Investor coordinate on a pair of future ask and bid prices (Au , B u )

- lo’s cancel any sell limit order that is not at price Au and submit a limit order at
price Au

- ho’s cancel any sell limit order and stay out of the market

- ln’s send a buy market order and immediately behave with respect to their new
type, lo

- hn’s send a buy market order and immediately behave with respect to their new
type, ho

• In the case ζ = d, for τ < t < τ + T d :

78
 - Investor coordinate on a pair of future ask and bid prices (Ad , B d )

- hn’s cancel any buy limit order that is not at price B d and submit a limit order at
price B d

- ln’s cancel any buy limit order and stay out of the market

- ho’s send a sell market order and immediately behave with respect to their new
type, hn

- lo’s send a sell market order and immediately behave with respect to their new
type, ln

Proof. see Appendix A.4.3

2.6.2 Limit order book dynamics in the transition phase

Before the transition phase begins the limit order book is filled with some limit orders that
supply liquidity. In particular liquidity provisions at best ask and bid prices are defined by
the value of the depths of the limit order book at prices A0 and B 0 . These are equal to DA0
and DB 0 . During the transition phase trading occurs only on one side of the order book. On
this side limit orders offer a profit opportunity. On the other side investors cancel their limit
order and send market orders to hit limit orders offering this opportunity.
For instance when the asset common value makes a positive jump, ζ = u, sell limit orders
submitted at price A0 offer a profit opportunity to buyers. Indeed A0 was an equilibrium
price when the asset common value was equal to v0 but is no longer once this value has
u
moved up to v0 + ω. The liquidity supply on the ask side DA 0 (t) is meant to disappear.

u
It decreases due to two kind of mechanisms. At each time t a mass (λ + ρ)DA 0 (t).dt of

investors monitor the market and cancel their limit order at price A0 . At the same time a
λ+ρ
mass (λ + ρ) × (Lhn (t) + Lln (t)).dt = 2
.dt of investors who do not own the asset monitor
u
the market and send buy market orders that execute at price A0 . The dynamic of DA 0 (t) is

given by the following proposition.

Proposition 2.4. When ζ = u during the transition phase the depth at price A0 is

1 1
u
DA 0 (t) = − + [DA0 + ]e−(λ+ρ)(t−τ )
2 2

79
 limit order cancellations : (λ + ρ)DA0 (t).dt

At t = τ the asset value

switches to v 0 + ω.

Initial condtion : DA0 (τ ) = α0 DA0 (t)

A0

limit order executions by hn’s and ln’s buy market orders :

(λ + ρ) × (Lhn (t) + Lln (t)).dt = λ+ρ
2
.dt.

Figure 2.3: Dynamic of the market depth at A0 following ”good” news arrival

which is decreasing and reaches zero at time t = τ + T u , defining the duration T u .

Proof. see Appendix A.4.2

When the asset common value moves down, the process is the same as in the ”up” case
but on the bid side of the order book at price B 0 .

Proposition 2.5. When ζ = d during the transition phase the depth at price B 0 is

1 1
d
DB 0 (t) = − + [DB 0 + ]e−(λ+ρ)(t−τ )
2 2

which is decreasing and reaches zero at time t = τ + T d , defining the duration T d .

Proof. see Appendix A.4.2.4

The dynamic of the order book in the transition phase is given by the initial values of
d u 0
the depths, DB 0 = DB 0 (0) = DA0 (0) = DA0 = α , so that the evolution of the depths in the

80
”up” and the ”down” cases are perfectly similar

1 1
d
∀t DB u
0 (t) = DA0 (t) = D(t) = − + [α0 + ]e−(ρ+λ)(t−τ ) .
2 2

The durations of the transition phases are the same in both states u and d

1
Tu = Td = T = ln(1 + 2α0 ).
ρ+λ

Transition phase in the symmetric equilibrium. The symmetric equilibrium is the

equilibrium where investors coordinate on the following future bid and ask prices in state
ζ = u and ζ = d:

1 δ ∆ 1 δ ∆
B u/d = (v0 ± ω − ) − , Au/d = (v0 ± ω − ) +
r 2 2 r 2 2

The analysis of this transition phase allows to understand the underlying trading mech-
anism for the dynamic of prices in a limit order market. As I develop it in the empirical
implication section we can evaluate the impact of market monitoring or volatility on how fast
prices adjust to new information. This is also possible to determine the role that limit and
market orders play in this price discovery process. In particular we can quantify the effect
of the market monitoring rate on the share of limit order cancellations and market order
executions in the erosion of the initial liquidity supply.

2.7 After the transition phase : convergence to a steady

state without uncertainty
In this section I explicit the strategy and the dynamic of the limit order book that corre-
sponds to the last phase of the game, after the transition phase is over. In the last phase of
the game the limit order book converges to a steady state without uncertainty as in chapter
1. Trading takes place at prices Au and B u if v = v0 + ω or at Ad and B d if v = v0 − ω.

Here I present the general case of this dynamic equilibrium that converges to a steady

81
state without uncertainty. In this equilibrium the terms of the trade-off between limit and
market orders do not change over time and are the same as the ones in the asymptotic steady
state. I use the same notations as in the limit order book in steady state without uncertainty
(αeq , A, B...etc, cf subsection 5.8).

Starting at t = 0 from a one tick market where the depths at prices A and B are DA (0) and
DB (0) constituted respectively by a share of the population Llo (0) and of Lhn (0), investors
follow their corresponding steady state equilibrium strategy described in section 5. The
rates at which hn and lo types send market orders, mA (t) and mB (t), evolve so that the
terms of the trade off are the same as in the steady state equilibrium. More precisely the
intensities at which limit orders are executed are unchanged and equal to lA and lB . In this
framework the dynamic of the different populations is given by the dynamic of the parameter
α,

1
Lho (t) = Lln (t) = ( − α(t))
2
Lhn (t) = Llo (t) = α(t)

and the value functions for each type are the same as in the former steady-state equilibrium.

To fully characterize the level of convergence of the limit order book we look at how its
state is different from the limit steady state. First in the steady state all investors have
positions in line with their optimal strategy. For instance at the limit t = ∞ all types lo have
a limit order in the book at price A. In the dynamic game a lo type investor may have been
out of the market to start with and then has to wait for her first market monitoring time to
submit a limit order. The difference Llo (t) − DA (t) measures the mass of types lo out of the
market. At time t the mass of investors out of the market who would optimally be in the
market are given by the following equations

Llo (t) − DA (t) = (Llo (0) − DA (0))e−(λ+ρ)t

Lhn (t) − DB (t) = (Lhn (0) − DB (0))e−(λ+ρ)t

The second convergence measure is the difference between the population levels at time

82
t and the ones in the steady state equilibrium. This difference is captured by the value
α(t) − αeq . The two dimensions of the convergence level actually reduce to one. Indeed, as
for the steady-state case, describing the evolution of α(t) is enough to describe the dynamic
of the order book since Lho (t), Lln (t), Lhn (t), Llo (t), DA (t) and DB (t) are fully defined when
α(t) is known.

Proposition 2.6. The dynamics of the equilibrium populations are given by the dynamic of
the parameter α,

α(t) = αeq + (α(0) − αeq )e−(2ρ+lA +lB )t

1 − e−[λ−(ρ+lA +lB )]t −(2ρ+lA +lB )t 1 − e−[λ−(ρ+lA +lB )]t −(2ρ+lA +lB )t
+ lA κA e + lB κB e
λ − (ρ + lA + lB ) λ − (ρ + lA + lB )

with κA = Llo (0) − DA (0), κB = Lhn (0) − DB (0)

Proof. see Appendix A.3

2.8 Empirical implications

2.8.1 Determinants of the liquidity supply before news arrival

When investors decide to supply or not liquidity they anticipate that news arrival will trigger
a transition phase where their limit order will bear an adverse selection risk. The model
parameters influence in different ways this risk of being picked-off and thus have an effect on
the size of the liquidity supply.
0
Proposition 2.7. An increase of µ or ω has a negative impact on αeq (cf. Figure 3)

0 0
∂αeq ∂αeq
< 0, <0
∂µ ∂ω

0
Moreover limµ→∞ αeq = 0.
For a value of µ not too low, an increase of the monitoring rate λ has a positive impact
0
on αeq (cf. Figure 3).
0
∂αeq
>0
∂λ

83
Proof. see Appendix A.5.4

0
Prediction 2.1. The liquidity supply before news arrival αeq

• decreases with the frequency of news arrival, µ, or the news surprise, ω.

• decreases with the monitoring rate λ when µ is not too low

0.0010

0.00040695

HΛL
0.0009

0
Α eq

H ΜL
0.0008

0 0.00040690

0.0007
Α eq
0.0006 0.00040685

0.0005

0.00040680
0.0004

0.0003
30 40 50 60 70 200 400 600 800 1000

0 0
Figure 2.4: (i) Evolution of αeq in function µ ∈ [20, 70] for λ = 100 and (ii) evolution of αeq
in function of λ ∈ [0, 1000] for µ = 50 (r = 1, ρ = 2, ∆ = 1, δ = 10, ω = 50).

The first two comparative statics come from the fact that the execution rate that makes
investors indifferent between limit and market order increases with µ and ω. Investors are less
willing to use limit order when the volatility of the asset common value increases, everything
else equal.

∂l0 ∂l0
> 0, >0
∂µ ∂ω
ρ
Mechanically the depth of the order book adjusts because the relation α0 = 4(ρ+l0 )
has to be
satisfied.
0
The effect of λ on αeq also comes from the fact that l0 decreases with respect to λ.
However this dependence of l0 on λ comes from two different channels and has no very
intuitive direction. On the one hand the increase of λ makes investors faster to cancel limit
orders after news arrival. This reduces the picking-off risk and increases the incentive to use
limit orders through a decrease of l0 . On the other when λ investors can send market order
faster after news arrival which increases the picking-off risk and increases l0 . That is why
the effect is overall ambiguous.

84
 We can see that for different values of µ, with the same parametrization as in Figure 3,
the effect be different. Figure 4 shows that it is non monotonic for µ = 0.1 and decreasing
0
for µ = 0.01. The effect is changed for µ fairly small. For instance, for µ = 1, αeq increases
with λ. But overall the effect of λ is small and even negligible compared to the effect of µ.
0.106358

0.0755060

HΛL
0.106358

0
Α eq
HΛL
0.0755059
0.106358
0
Α eq
0.0755058 0.106358

0.0755057 0.106357

0.0755056 0.106357

0.106357
0.0755055

0.106357
0.0755054

200 400 600 800 1000 200 400 600 800 1000

0
Figure 2.5: Evolution of αeq in function of λ ∈ [0, 1000] (i) for µ = 0.1 and (ii) for µ = 0.01
(r = 1, ρ = 2, ∆ = 1, δ = 10, ω = 50).

0
The negligible effect of λ on market depth αeq , and implicitly on investors’ trading strate-
gies, is in line with the intuition that only relative reaction speed matters in a game where
faster trader will earn higher profits following the arrival of new information. In our setup an
increase of λ corresponds to a reaction speed increase for all investors but without a change
in their reaction speed relative to each other. Monitoring abilities are still homogenous in
the investors population.

2.8.2 Duration between news arrival and price change

The duration between news arrival and the change in trading prices in the limit order book
is the duration of the transition phase.

1 0
T = ln(1 + 2αeq )
ρ+λ

Prediction 2.2. For values of λ and µ not too low, the duration of the transition phase T

• decreases with the frequency of news arrival, µ, or the news surprise, ω.

• decreases with the monitoring rate λ when µ is not too low

85
 Prices in the limit order book reflect the new common value of the asset once there is no
arbitrage opportunity left, that is to say that the initial liquidity supply offering this arbitrage
opportunity has disappeared. The populations of potential arbitrageurs is fixed. This is the
group of non-owner if the common value goes up and the group of owner if the common
value goes down. Then the instantaneous flow of directional market orders aiming to profit
from the arbitrage opportunity is proportional to the rate at which this population monitors
the market, λ + ρ, and it does not depend on the parameters that rule the dynamic of the
0
common value. µ and ω only affect the initial liquidity supply αeq . The effect of an increase
of µ or ω is mechanical since it decreases the initial liquidity supply that is consumed and
removed faster in the transition phase. The monitoring intensity λ affects both the liquidity
supply and the flow of directional orders. Since the liquidity supply is a bounded function of
λ, the monitoring rate ends up reducing the duration of the transition phase for λ ”not too
low”.
One should notice that as soon as the asset holding constraint on investors is independent
of µ or ω during the transition phase the flow of directional market orders remains indepen-
dent of these parameters and the result would holds. The ”zero or one unit” assumption is
not key here. However there is a need for a holding constraint otherwise investors could send
infinitely large orders and consume instantaneously the liquidity supply.
The fact that λ and µ or ω are independent is less obvious. As for model of limited
attention allocation, investors could decide of their λ depending on the asset characteristics.
This calls for further extension of the model to endogenize the choice of λ.

2.8.3 Order flow decomposition in the price discovery process

Corollary 2.1. In the transition phase the numbers of limit orders executed and limit orders
cancelled are

0 0
" #
ln(1 + 2αeq ) 0 ln(1 + 2αeq )
LOE = , LOC = αeq −
2 2

Moreover the ratio of limit order cancellations over executed market orders is increasing
with respect to α0 :

86
 ∂ LOC
>0
∂α0 LOE
Prediction 2.3. In the transition phase, the ratio of limit order cancellations over limit
order execution

• decreases with the frequency of news arrival, µ, or the news surprise, ω.

• increases with the monitoring rate λ when µ is not too low

The mechanism behind this result is the following. As mentioned in the previous subsec-
tion the flow of directional market orders during the transition phase is proportional to λ + ρ
0
and does not depend on αeq , µ or ω. On the side of the liquidity supply the instantaneous
probability for an investor to cancel her limit order is also (λ + ρ).dt. The mass of these
0
investors is αeq at the beginning of the transition phase and equal to D(t) afterwards. Then
the flow of limit order cancellations at t during the transition phase is (λ + ρ)D(t).dt which
0 0
depends positively on αeq . If initially αeq is increased, at each point in time the flow of limit
order cancellations is increased whereas the flow of directional market orders is the same.
This explains why the share of limit order cancellation increases. However the transition
phase lasts longer which explains why the number of market orders during the transition
phase increases as well.

2.9 Conclusion
This paper models the effect of limited attention on market reaction to unscheduled news
arrival. Investors’ limited capacity of attention restricts their ability to monitor the market.
This imperfect market monitoring delays price adjustments following news arrival. Because
of their imperfect ability to monitor news, investors take the risk of being picked-off when
they supply liquidity with limit orders. When the frequency of news arrival increases, this
picking-off risk is amplified and consequently (i) the liquidity supply declines, (ii) prices
adjust faster following news arrival and (iii) the share of limit orders cancellations in the
price discovery process decreases.

87
88
Chapter 3

High Frequency Trading, Market

Efficiency and Mini Flash Crashes

3.1 Introduction
Financial markets stability is important to attract investors. Instability of securities prices
can blur investors’ expectations on trading conditions and ultimately may discourage them
to trade in traditional exchanges. In recent years, anecdotes from the trading industry have
reported the rise of a new kind of market destabilizing event: the «mini flash crash»1 (see
Appendix B.1 for a list of past mini flash crashes). A mini flash crash can be defined as
a sudden sharp change in the price of a stock followed by a very quick reversal (see Figure
1). The increasing frequency of these events has been interpreted as a symptom of mar-
ket fragility and ascribed to High Frequency Trading (HFT henceforth). Meanwhile recent
papers (e.g, Hendershott, Jones and Menkveld (2011), Hendershott and Riordan (2013),
Brogaard, Hendershott and Riordan (2012) or Chaboud, Chiquoine, Hjalmarsson, and Vega
(2009)) support that HFT has a positive effect on market quality and market informational
efficiency. Through which channel HFT could generate mini flash crashes? Can financial
markets become jointly more efficient and less stable under HFT actions?
To address these questions we develop a theory of mini flash crashes. Our theory is based
1
In reference to The Flash Crash of May 6th, 2010. See «The Flash Crash, in Miniature» in the New York
Times, http://www.nytimes.com/2010/11/09/business/09flash.html. Nanex Research also reports mini flash
crashes among other market anomalies, http://www.nanex.net/FlashCrash/OngoingResearch.html

89
on the idea that there is a trade-off between speed and precision in the acquisition of in-
formation. The new market environment enables traders to become HFT and react much
faster to news of any type at the cost of the precision of the information used by traders. We
embed this idea in a two periods trading model in which strategic event traders can either
become HFT’s by investing in a fast technology, which allows them to acquire a noisy signal
on the asset fundamental value and trade at period 1, or not invest and only trade at period
2 with a perfect signal on the asset fundamental value. The speed advantage of HFT in
information acquisition has been studied by Foucault, Hombert and Rosu (2012) but it does
not incorporate the possibility for interpretation mistakes of new information. In Foucault,
Hombert and Rosu (2012), as well as in our paper, the speed advantage of HFT is modelled
as an ability to trade a period ahead of other investors. It can be seen as reduced form for a
high market monitoring ability of HFT. It could be modelled in a framework where investors
have imperfect market monitoring capacity, as in some recent papers (e.g. Biais, Hombert
and Weill (2013), Foucault, Kadan and Kandel (2013), Pagnotta and Philippon (2012)). Lit-
erature on HFT considers that HFT may also benefit from a superior information processing
ability that is different from the speed advantage. To address this feature, theoretical pa-
pers, as Biais, Foucault and Moinas (2013), model HFT as traditional informed traders (as
in Glosten (1985)).

We find that an increase in HFT activity, due, for instance, to a lower cost of the fast
technology, increases the likelihood of a price reversal, between period 1 and 2. Price reversals
arise when HFT discovered that the signal, they acquired at period 1, was wrong and then
decide to correct their trade at period 2. It generates opposite trading patterns across the two
periods and, possibly, price swings. The price impact of HFT’s at period 1 is proportional to
the number of HFT’s. Hence the likelihood of a price reversal increases when the number of
HFT’s increases. Even though more HFT’s implies more reversals, it also improves market
informational efficiency. While these two implications seem to be contradictory, the presence
of high frequency traders allows to faster integrate information into prices when period 1’s
signal is informative.

The novelty of this paper is to introduce a trade-off, between speed and precision for
information processing, to explain why HFTs may trade on noise and generate price reversal.

90
Figure 3.1: «On September 27, 2010, Progress Energy, Inc shares plummeted 90% in a matter
of seconds for no apparent reason. The harrowing decline was a consequence of a mini flash
crash; a much smaller version than the crash in May...», http://sslinvest.com/news/mini-
flash-crash-september-27th-sends-pgn-shares-down-90.

There are theories of short-term speculators who ex-ante rationally coordinate to trade on
noise (see Froot, Scharfstein and Stein (1992)). We think of trading on noise as a risk that is
revealed ex-post. In our view, it stems from competition in event trading when traders have
the possibility to react to news, or any other relevant signals, in a very short amount of time.
At first glance, acceleration at which new information is processed, and used, should result
in faster integration of information into prices. However, this acceleration also increases
the risk that event traders base their trading decision on less accurate signals. Of course, to
mitigate this risk, event traders may decide to check news accuracy (for instance using human
intervention). But doing so, they take the risk of losing a profit opportunity by reacting too
late to informative signals. Hence competition among traders may push them to react too
quickly at the expense of the precision of the information on which they trade, and may lead
to subsequent trade and price reversals.

Price reversals may alternatively come from overconfident investors who overreact to
private signals, as in Daniel, Hirshleifer and Subrahmanyam (1998), which generates price
correction following public revelation of information. In this setup, price reversals are sys-
tematic. Price returns become negatively auto-correlated and predictable. The market is

91
inefficient as opposed to what we find with our model, which does not generate negative
auto-correlation. Moreover our setup generates complete price reversals in the sense that
prices can switch back to their original value when HFTs react to noise, while, here, the
price reversal is a partial correction of a previous excessive price change that had the right
direction nonetheless.

Trade reversal, by traders who obtain information faster than others, can arise when a
trader benefit from an anticipated noisy information leakage about future public announce-
ment, as in Brunnermeier (2005). It enables the trader, first, to privately acquire a noisy
signal on a short term component of the asset value, and trade on it; and, second, to know by
how much his price impact was driven by noise, after the short term component is publicly
announced. Hence he benefits from an informational advantage even after the announce-
ment. He profits from it by partially reversing the share of his past trade that happened to
be driven by the noisy component of his signal. However this trade reversal is compensated
by a opposite trades of other strategic traders, which makes the implications for the price
dynamics unclear, contrary to our setup. In Brunnermeier’s paper, the market is efficient
since prices reflect all publicly available information. However the introduction of informa-
tion leakage has mitigated effects on efficiency, contrary to our model in which more HFT
increases market efficiency.

Event trading strategies may be as diverse as the spectrum of relevant information in a

particular market. HFTs look for financial news or informative market patterns that they can
process and trade on as fast as possible. The source of the imperfect information precision
may be either endogenous or exogenous. On the one hand it can be endogenous because
algorithms send orders based on the interpretations of these events. Every thing else equal,
the faster is the algorithm reaction, the less accurate is the interpretation. Consequently
High Frequency Traders face this trade-off when they calibrate their algorithm. In the other
hand the process of new information release may in itself be an exogenous source of noise
for new information. If there is a slight chance that some pieces of news are false, High
Frequency Traders have to decide whether they take the risk to react immediately to the
piece of news or wait for a correction. For instance the following anecdote illustrates, in
a rather extreme way, the false news issue. On Monday Sept. 8, 2008, the stock price of

92
United Airlines dropped to $3 a share from nearly $12 in about fifteen minutes. Then the
price bounced back at $11 at the end of the Tuesday session. The cause of these swings was
an old article about United Airlines’ 2002 bankruptcy-court filing that mistakenly appeared
on September 8, 2008 as a seemingly new headline on Google’s news service.
Chapter 3 is organized as follows. Section 3.2 presents the setup and assumptions of
the model. Section 3.3 gives the equilibrium trading strategies at each period. Section 3.4
gives the equilibrium population sizes and their profits. Section 3.5 describes the equilibrium
price dynamics. Section 3.6 provides results on market informational efficiency. Section 3.7
concludes.

3.2 Model setup

We consider a three periods model (t = 1, 2, 3) of trading in a financial asset. The financial

asset is traded at periods t = 1 and t = 2 and pays-off at t = 3. There are three types of
market participants: a competitive market maker (as in Kyle (1985)), liquidity traders and
event traders. To participate at period t = 1, event traders can pay a cost C, the cost of
being fast. They observe a signal, that is either informative (equal to the asset pay-off) or
noise, and trade. At period t = 2, all event traders participate. They learn the true nature
of the previous signal, as well as its value, and trade. If the signal is informative, they can
trade on this information. If the signal is noise, they learn that the expected value of the
asset is still equal to its ex-ante value and they can take advantage of an eventual mis-pricing.
Figure 2 summarizes the timing of market participants’ decisions. Using this setting we plan
to analyze whether a decrease in the cost of being fast impairs or ameliorates price discovery
and price stability.

Asset Value. The pays-off of the asset is V ∈ {0, 1} at period 3, with equal probabilities.
Thus, prior to the beginning of the game, the asset expected value is E[V ] = 1/2.

Liquidity trading. At each trading period, some liquidity traders send orders for reasons
left exogenous to our model. The order flow of liquidity traders is random and follows a

93
uniform distribution on the interval [−Q, Q]. We note φ the density function associated

1
φ(x) = × I{x∈[−Q,Q]} .
2Q

In the following of the paper, we call ˜l1 and ˜l2 the liquidity traders order flows at periods
t = 1 and t = 2.

Market making. At each trading period, a risk-neutral and competitive market maker
observes the aggregate order flow, that is the sum of event traders and liquidity traders
orders. The aggregate order flows are noted Q̃1 at period t = 1 and Q̃2 at period t = 2. The
market maker executes these order flows at the following trading prices at periods 1 and 2,

P1 = E[V |Q̃1 ] = P r[V = 1|Q̃1 ]

and
P2 = E[V |Q̃2 , Q̃1 ] = P r[V = 1|Q̃2 , Q̃1 ],

as in Glosten and Milgrom (1985) or Kyle (1985).

Event trading. There is a continuum, [0, A], of competitive event traders. Ex-ante an
event traders i ∈ [0, A] chooses either to pay the cost C, so that he can trade at periods
t = 1 and t = 2, or to not pay the cost, and trades only at period t = 2. The mass of «fast»
traders who pay C is α. Traders in the interval [0, α] trade at periods t = 1 and t = 2 while
traders in the interval [α, A] trade only at t = 2. We endogenize α in section 4.4.

Assumption 3.1. In each trading period, trader i sends and order with size Xi that is
bounded, Xi ∈ [−1, 1].

Assumption 3.2. The mass of event traders is smaller than the size of liquidity trading

A<Q

Assumptions 1 and 2 make sure that the impact of event traders is limited. Assumption 1
corresponds to a limit to arbitrage constraint for each event traders. For some reasons (cost

94
 t=0 t=1 t=2 t=3

Event traders - Fast event - All event traders The asset

decide to traders observe learn whether S̃ pays-off its value.
become : the signal S̃, is informative or
- fast, at cost C, then send orders. not.
- or not fast. - Liquidity traders If informative,
send orders. they learn the
- The market asset value.
maker observes Then they send
the aggregate orders.
order flow, then - Liquidity traders
sets a price. send orders.
- The market
maker observes
the aggregate
order flow, then
sets a price.

Figure 3.2: Timing of market participants’ decisions.

of capital, aversion for residual risk,...etc) an event trader cannot trade unbounded quantities.
Assumption 2 states that the mass of event traders is limited as well. The number of trading
desks dedicated to event trading is less than some measures of global trading activity (here
the volume of liquidity trading Q).

Solution concept. In this setup, we look for a Nash equilibrium of the game defined by
strategies of the event traders and the pricing strategies of the market makers: P1 (Q1 ) and
P2 (Q1 , Q2 ).
Given the symmetry of the model, event traders, who trade at the same period, will
obtain the same profit in equilibrium. These profits are equal to π1 at t = 1 and π2 at t = 2.

Determination of α. At equilibrium, the endogenous value of α can be either an interior

solution, 0 < α < A, or a corner solution, α = 0 or α = A. We reach an interior solution
when there exists an α ∈ (0, A) such that each event trader is indifferent between trading
and not trading at t = 1. In this case the equilibrium value of α is defined as the solution of

95
the following equation,
π1 (α) − C = 0.

If each event trader strictly prefer to trade at t = 1, even when all event traders participate
at t = 1, then the equilibrium value is a corner solution, α = A. And the following condition
is satisfied,
π1 (A) − C > 0.

If each event trader strictly prefer to not trade at t = 1, even when no event trader participate
at t = 1, then the equilibrium value is a corner solution, α = 0. And the following condition
is satisfied,
π1 (0) − C < 0.

The determination of α does not involve the profit of trading at t = 2, π2 , because each
investor is infinitesimal. His isolated trades do not move prices. Hence his participation at
t = 1 does not affect his profit at t = 2. That is why his decision to participate at t = 1 only
depends on the profit of trading at t = 1, π1 .

Discussion of the «cost of being fast». We can interpret the «cost of being fast» as
the cost of investing in the technology that allows to react fast to a market event. This
technology is associated with the risk of reacting to noise. This cost can also be interpreted
as an opportunity cost. Event traders must decide on which type of events or on which
market they want to allocate their computing capacity. A third possibility is to interpret
this cost as a technological barrier. When the technology to implement «fast event trading»
strategies was not available, the cost for trading at period 1 was infinite C = ∞. Now that
the technology is available, the cost is finite.

The rise of High Frequency Trading. With this framework we can address some of the
effects High Frequency Trading activities development which stemmed from the improvement
of computing technologies. In our model, it corresponds to a drop of the cost of being fast.
Jointly with the development of algorithmic strategies, the number of high frequency trading
desk has tremendously increased. In our model we can estimate the consequence of this

96
expansion through an increase of the mass of event traders, A, relative to the global trading
activity, Q.

3.3 Information, trading strategies, pricing policy and

profits

In this section we take α as exogenously given. The mass of event traders at t = 2 is set equal
to β. In section 4.4 we endogenize α = αeq and we set β = A since, under our assumptions,
all event traders can participate at period 2 at no cost.

3.3.1 Period 1

Event traders information at t = 1. At t = 1 an event trader i ∈ [0, α] observes a signal

S̃ ∈ {0, 1}, the same for all event traders, defined as:

S̃ = Ũ Ṽ + (1 − Ũ )˜

with Ũ ∈ {0, 1} with respective probabilities P r[U = 0] = 1−δ and P r[U = 1] = δ, ˜ ∈ {0, 1}
with P r[ = 0] = P r[ = 1] = 1/2. The signal distribution conditional on the asset value is
the following,

1+δ
P r[S = 1|V = 1] = P r[S = 0|V = 0] =
2
1−δ
P r[S = 1|V = 0] = P r[S = 0|V = 1] =
2

The unconditional distribution is symmetric P r[S = 1] = P r[S = 0] = 1/2. If we call

P r[S = s|V = 1]P r[V = 1]

µ(s) = P r[V = 1|S = s] =
P r[S = s]

the updated value of the asset for the event trader in period 1 is either

1+δ 1 1−δ 1
µ(1) = > or µ(0) = < .
2 2 2 2

97
The profit of an event trader, who observed the signal and traded a quantity X in period 1,
is
X(µ(s) − E[P1 |S = s])

Event traders strategy. At t = 1 an event trader i ∈ [0, α] sets her strategy based on the
signal value, Xi (s).

Proposition 3.1. At t = 1, the unique equilibrium strategy for an event trader is to buy if
the signal is high and to sell if the signal is low,

X(s = 1) = 1, X(s = 0) = −1.

Proof. The market maker infers the probabilities of the posterior expected asset value to be
either µ(1) or µ(0), from the order flow, and set the price as the expected value of the asset.
Since the market makers information set is smaller the one of event traders, the expected asset
value conditional on the market maker’s information set is necessarily µ(1) ≥ E[V |Q̃1 ] ≥ µ(0).
Moreover, in some state of the world the information set of the market makers is strictly
inferior than the one of event traders. In these cases µ(1) > E[V |Q̃1 ] > µ(0). This implies
that the strategy X(s = 1) = 1 and X(s = 0) = −1 strictly dominates all other strategies.

Depending on the realization of the event traders signal, the aggregate order flow in the
first period is as follows,

Q̃1 = ˜l1 + α if S = 1

Q̃1 = ˜l1 − α if S = 0

Market maker’s pricing policy. At t = 1, the market marker observes the aggregate
order flow Q̃1 and sets a price equal to E[V = 1|Q̃1 ] = P r[V = 1|Q̃1 ].

Proposition 3.2. At t = 1 the competitive market maker pricing policy is

(1 + δ)φ(q − α) + (1 − δ)φ(q + α) 1
P1 (q) = P r[V = 1|Q̃1 = q] = ×
φ(q − α) + φ(q + α) 2

98
or equivalently 
1−δ

if q ∈ [−Q − α, −Q + α)


2





P1 (q) =  1 if q ∈ [−Q + α, Q − α]
 2



 1+δ

 if q ∈ (Q − α, Q + α]
2

This pricing policy reflects two cases of the market maker inference problem. When
the order flow belongs to the interval [−Q + α, Q − α], the market maker cannot infer the
information of event traders and then set the price equal to the ex-ante expected value of
the asset. When the order flow belongs [−Q − α, −Q + α) (resp. (Q − α, Q + α]), the market
maker can infer that the event trader signal is negative (resp. positive) since the order flow
takes «extreme» negative (resp. positive) values.

Event traders profit. With the event traders strategy and the market maker pricing
policy, we can compute the the profit of an event trader.

Proposition 3.3. At t = 1, with a mass α of event traders, the expected profit of one event
trader is
δ Q−α
π1 (α) = × .
2 Q

3.3.2 Period 2

Event traders information. At t = 2, event traders observe the realization of S̃ and Ũ .

In other words they know if the first period signal was informative or not, and know the
value of the asset, if the signal was informative. They are perfectly informed if U = 1. At
t = 2, the profit of an event trader who trades a quantity X is

X(V − E[P2 |V ]) if U = 1
1
 
X − E[P2 |S, U = 0] if U = 0
2

Event trader strategy. At t = 2, an event trader i ∈ [0, β] sets her strategy conditional
on the realizations of the random variables, S̃ and Ũ , and on the realization of the order flow

99
at t = 1, q1 (or equivalently the previous price P1 ). Her strategy is a function Xi (s, u, q1 ).

Proposition 3.4. At t = 2, if U = 1, the unique equilibrium strategy is to buy if the asset

value is high and to sell if the asset value low,

X(1, 1, q1 ) = 1, X(0, 1, q1 ) = −1.

If U = 0, if the t = 1 order flow is q1 ∈ (Q − α, Q + α], which implies that the signal was high
S = 1 and that the asset price went up in the first trading period, then the unique equilibrium
strategy is to sell,
X(1, 0, q1 ) = −1.

If U = 0, if the t = 1 order flow is q1 ∈ [−Q − α, −Q + α), which implies that the signal
was low S = 0 and that the asset price went down in the first trading period, then the unique
equilibrium strategy is to buy,
X(0, 0, q1 ) = 1.

If U = 0, if the t = 1 order flow is q1 ∈ (Q − α, Q + α], which implies that asset price

remained equal to its the ex-ante expected value, 1/2, in the first trading period, then any
strategy X(s, 0, q1 ) ∈ [−1, 1] is an equilibrium strategy. However the only strategy that is
robust to the introduction of a small trading cost is to not trade,

X(s, 0, q1 ) = 0.

Hence there is a unique equilibrium strategy that is robust to the introduction of a small
trading cost. In the following of the paper we consider this strategy.

The equilibrium strategy of event traders at t = 2 is again very intuitive. They trade in
the same direction that first period event traders, if the signal is indeed informative, in order
to take advantage of the remaining profit opportunity. In the other case, they trade in the
opposite direction that first period event traders in order to take advantage of a previous
erroneous price change. Depending on the realizations of the underlying random variables,
the aggregate order flow in the second period is as follows,

100
 • if U = 1 the aggregate order flow is equal to

Q̃2 = ˜l2 + β if V = 1,

Q̃2 = ˜l2 − β if V = 0,

• if U = 0 the aggregate order flow is equal to



β if q1 ∈ [−Q − α, −Q + α)







Q̃2 = ˜l2 + M0 (q1 ) with M0 (q1 ) = 0 if q1 ∈ [−Q + α, Q − α]





−β if q1 ∈ (Q − α, Q + α]



Market maker’s pricing policy. At t = 2, the market marker observes the aggregate
order flow Q̃2 , has already observed Q̃1 , and sets a price equal to E[V = 1|Q̃2 , Q̃1 ] = P r[V =
1|Q̃2 , Q̃1 ].

Proposition 3.5. At t = 2 the competitive market maker pricing policy is

P2 (q2 , q1 ) = P r[V = 1|Q̃2 = q2 , Q̃1 = q1 ] =

δφ(q1 − α)φ(q2 − β) + 1−δ
2
[φ(q1 − α) + φ(q1 + α)]φ(q2 − M0 (q1 ))
δ[φ(q1 − α)φ(q2 − β) + φ(q1 + α)φ(q2 + β)] + (1 − δ)[φ(q1 − α) + φ(q1 + α)]φ(q2 − M0 (q1 ))

101
or equivalently,


0 if q2 ∈ [−Q − β, −Q + β)







if q1 ∈ [−Q − α, −Q + α), P2 (q2 , q1 ) = 1−δ
if q2 ∈ [−Q + β, Q − β]

 2



1 if q2 ∈ (Q − β, Q + β]


2


0 if q2 ∈ [−Q − β, −Q)








1−δ

if q2 ∈ [−Q, −Q + β)


2−δ





if q1 ∈ [−Q + α, Q − α], P2 (q2 , q1 ) =  1 if q2 ∈ [−Q + β, Q − β]
 2



1
if q2 ∈ (Q − β, Q]





 2−δ



1 if q2 ∈ (Q, Q + β]



1

if q2 ∈ [−Q − β, −Q + β)


2





if q1 ∈ (Q − α, Q + α], P2 (q2 , q1 ) = 1+δ
if q2 ∈ [−Q + β, Q − β]

 2



1 if q2 ∈ (Q − β, Q + β]



Remark 3.1. If M0 6= 0 for q1 ∈ [−Q + α, Q − α], the pricing policy is


0 if q2 ∈ [−Q − β, −Q + M0 )








1−δ

if q2 ∈ [−Q + M0 , −Q + β)


 2−δ




P2 (q2 , q1 ) = 1
if q2 ∈ [−Q + β, Q − β]

 2



1
if q2 ∈ (Q − β, Q + M0 ]





 2−δ



1 if q2 ∈ (Q + M0 , Q + β]


Event traders profit. At t = 2, event traders can make profit because they trade on an
informative signal that had not been completely integrated into prices in the first period
either because the market maker has not inferred its value or, if he had, because the signal

102
was noisy which has left some profit opportunity once the informative nature of the signal
is known by event traders. They also make profit by correcting pricing errors due to event
traders at t = 1 when the signal is revealed as uninformative.

Proposition 3.6. With a mass β of event traders in the second period and a mass α of event
traders in the first period, the expected profit of an event trader at t = 2 is

δ Q−α Q−β 1−δ β αQ−β

π2 (α, β) = ×[ ×( + ) + (1 − δ) ]
2 Q Q 2−δQ Q Q

Remark 3.2. Even if M0 6= 0 for q1 ∈ [−Q + α, Q − α], the event traders expected profit at
t = 2 remains unchanged.

3.4 Equilibrium
Equilibrium populations. With given masses, α and β, of event traders at periods t = 1
and t = 2 we obtained the equilibrium strategies of event traders, the market maker pricing
policies and the event traders profits. Here we endogenize these masses of populations.

Proposition 3.7. The masses of event traders at periods t = 1 and t = 2 are uniquely
defined. Trading at t = 2 is costless for event traders, hence β = A. The equilibrium mass
of event traders who pay the cost of being fast, C, to trade at t = 1 is αeq ∈ [0, A], equal to

δ
if C > , αeq = 0,
2

!
δ δ A 2
 
if ≥ C ≥ 1− , αeq = Q 1 − C
2 2 Q δ

!
δ A
if C < 1− , αeq = A
2 Q
 
δ δ A
For C > 2
and C < 2
1− Q
, the equilibrium mass of event traders is a corner solution.
In the first case, the fast technology is too expensive compared to the profit that a trader
alone could extract with, thus no event trader decides to become fast. In the second case,

103
the fast technology is cheap enough so that when all traders invest, they still make a positive
profit.
 
δ δ A
For 2
≥C≥ 2
1− Q
there is an interior solution. the equilibrium mass of event traders
is such that the marginal trader is indifferent between investing and not investing in the fast
technology. The indifference condition is

δ Q − αeq
π1 (αeq ) = × = C.
2 Q

Equilibrium profit of an event trader. At equilibrium, the expected profit of an event

trader, π, is equal to

δ
if C > , π = π2 (0, A)
2

!
δ δ A
if ≥ C ≥ 1− , π = π2 (αeq , A)
2 2 Q

!
δ A
if C < 1− , π = π2 (A, A) + π1 (A) − C
2 Q

Proposition 3.8. The expected profit of an event trader depends on the «cost of being fast»
as follows,

δ ∂π
if C > , =0
2 ∂C

!
δ δ A ∂π
if ≥ C ≥ 1− , >0
2 2 Q ∂C

!
δ A ∂π
if C < 1− , = −1.
2 Q ∂C

104
The expected profit of an event trader is decreasing with respect to the mass of event traders,

∂π
< 0.
∂A

The effect of the mass of event traders, A, on the profit of each trader is very intuitive.
When the mass of event traders increases, the market maker can infer more easily event
traders’ private information which reduces the profit of one of these traders.
When the cost of being fast is high enough so that only a fraction of event traders trade at
t = 1, a drop of this cost generates a decline of the expected profit of event trader. Because of
competition, her expected profit at t = 1 is nil and her expected profit at t = 2 declines. The
increasing number of event traders at t = 1 helps the market maker to infer event traders’
private information which reduces, on average, the level of information asymmetry at t = 2.

Aggregate profit of event traders. At equilibrium the aggregate profit of event traders,
Π, is equal to

δ
if C > , Π = Aπ2 (0, A)
2

!
δ δ A
if ≥ C ≥ 1− , Π = Aπ2 (αeq , A)
2 2 Q

!
δ A
if C < 1− , Π = Aπ2 (A, A) + Aπ1 (A) − AC
2 Q

Proposition 3.9. The aggregate profit of event traders depends on the «cost of being fast»
in the same way that for the profit of one event trader,

∂Π ∂π
=A .
∂C ∂C

The effect of the «cost of being fast» on event traders aggregate profit is mitigated. A
reduction of this cost has negative effect on their profit when it is too expensive so that only
a fraction of event traders can invest in the fast technology. It has a positive effect when the

105
it is cheap enough so that all event traders invest in the technology and when a reduction of
its cost does not drive more competition in event trading.
To draw clearer results, we can study the variation of event traders’ aggregate profit when
the possibility of trading in the first period become possible for free, from C = ∞ to C = 0.
This could be due to a contemporaneous development of information technologies or a change
in the market structure that allows for electronic trading.

Proposition 3.10. The variation of the aggregate profit of event traders when the «cost of
being fast» drops from C = ∞ to C = 0 is equal to

Aπ2 (0, A) − (Aπ2 (A, A) + Aπ1 (A)).

It can be either positive or negative. There is a threshold H(δ) ∈ [0, 1] such that

A
If < H(δ), Aπ2 (0, A) − (Aπ2 (A, A) + Aπ1 (A)) < 0
Q

A
If > H(δ), Aπ2 (0, A) − (Aπ2 (A, A) + Aπ1 (A)) > 0
Q
Given that liquidity traders are not strategic and that the market maker sets competitive
prices, the trading game is a fixed-sum game in which the gross aggregate profit of event
traders corresponds to the implicit trading cost, for liquidity traders, due to information
asymmetry. When the technology to trade fast on market events become available (C = ∞
to C = 0), these implicit trading costs decline when competition among event traders is high
A
(Q > H(δ)), which helps market prices to reflect private information. In this case, HFT is
beneficial to liquidity traders.

3.5 Equilibrium price dynamics

In this section, we study the asset price equilibrium dynamics. We graphically represent,
with probability trees, the different paths that the asset price can take, conditional on S = 1
or S = 0. We decide to represent conditional dynamics for the sake of graphics clarity. To
obtain the unconditional equilibrium tree, one can multiply all branches’ weight by 1/2 in

106
each tree and merge the two trees by adding the weights on the matching branches. In these
price dynamics, we consider an additional period, t = 0, in which the signal is not available
yet and thus the asset is priced at its ex-ante expected value E[V ] = 1/2.

Price dynamics conditional on S = 1 P2 = 1

δβ
δβ 2Q
Q

1+δ 1+δ
P1 = 2
Q−β P2 = 2
Q

α 1
Q
(1−δ)β P2 = 2−δ
Q

β
2Q
1 1 1
P0 = 2
Q−α P1 = 2
Q−β P2 = 2
Q Q

(1−δ)β
2Q

1−δ
P2 = 2−δ

Price dynamics conditional on S = 0

1
P2 = 2−δ

(1−δ)β
2Q
1 1 1
P0 = 2
Q−α P1 = 2
Q−β P2 = 2
Q Q

β
2Q

1−δ
α (1−δ)β P2 = 2−δ
Q Q

1−δ 1−δ
P1 = 2
Q−β P2 = 2
Q

δβ
Q δβ
2Q

P2 = 0

107
Mini flash crash. In our setup, a mini flash crash corresponds to a price path in which,
1 1−δ α 1 1−δ
at t = 1, the price drops, from 2
to 2
with probability 2Q
, or jumps, from 2
to 2
with
α 1 (1−δ)β
probability 2Q
, and then, at t = 2, the price switches back to 2
with probability 2Q
(see
Figure 3.2). The magnitude of a mini flash crash, Mcrash , is equal to difference between prices
at t = 1 and t = 2,

δ
Mcrash = .
2

Proposition 3.11. The probability of a mini flash crash is equal to

α(1 − δ)β αeq (1 − δ)A

Pcrash = 2
= .
2Q 4Q2

The probability of a mini flash crash increases when the mass of event traders increases or
when the «cost of being fast» declines,

∂Pcrash ∂Pcrash
> 0, < 0.
∂A ∂C

The probability of a mini flash crash depends on the signal’s precision δ as follows,
!
δ δ2 δ δ A δ2 ∂Pcrash
If ≥ C ≥ or ≥C≥ 1− ≥ , ≥0
2 2 2 2 Q 2 ∂δ

!
δ δ2 δ A ∂Pcrash
if ≥ ≥C≥ 1− , ≤ 0,
2 2 2 Q ∂δ

!
δ A ∂Pcrash
if 1− ≥ C, ≤ 0.
2 Q ∂δ

The proposition draws a clear link between a more intense HFT activity and a greater
occurrence of mini flash crashes, through the possibility of trading fast with a risk on infor-
mation precision. When the mass of event traders increases or when the «cost of being fast»
decreases, the number of HFTs (event traders who invest in the fast technology and trade at
t = 1) increases. Hence their price impact at t = 1 is higher. The market maker infers more

108
 1+δ
P1 = 2

α (1−δ)β
2Q 2Q

1 1
P0 = 2 P2 = 2

α (1−δ)β
2Q 2Q

1−δ
P1 = 2

Figure 3.3: Price dynamics of mini flash crashes.

easily the signal value and sets more frequently a price different than 1/2, at t = 1. If the
signal was noise then the price eventually reaches back its original value at t = 2.

The signal precision, δ, has an amplifying role on the crash magnitude, Mcrash . When δ is
high, the expected value of the asset, conditional on the signal realization, is closer from one
of the possible realization of the asset value. The difference between the first period price,
1
in the crash path, and the true expected value 2
is thus bigger. And the price reversal is
sharper.

The effect of δ on the frequency of mini flash crashes is mitigated. Intuitively when
the precision of the signal increases, one would expect that the mini flash crash frequency
declines. It is the case when, for instance, the fast technology has a low cost and all event
traders participate at t = 1. In this case it reduces the probability of a price reversal between
t = 1 and t = 2 while the probability of a price change between t = 0 and t = 1 is not
affected by δ. However in some cases, when the «cost of being fast» is such that only a
fraction of event traders invest in the fast technology, a higher δ generates a higher profit for
fast traders. As a consequence the probability that P1 reflects event traders’ signal is higher,
and the probability to be on the path of a crash increases.

109
3.6 Market informational efficiency

For a given trading period, we can measure market efficiency with the unconditional expec-
tation of the square difference between the true value of the asset and its market price at the
considered trading period, E[(Ṽ − Pt )2 ]. As the price is set by a competitive market maker
and always reflects the expected value of the asset conditional on public information, we can
rewrite this efficiency measure as

E[(Ṽ − P1 )2 ] = E[E[(V − P1 (Q̃1 ))2 |Q̃1 ]]

= E[E[(V − E[V |Q̃1 ])2 ]|Q̃1 ]

= E[V[V |Q̃1 ]]

= E[P1 (Q̃1 )(1 − P1 (Q̃1 ))]

and

E[(Ṽ − P2 )2 ] = E[V[V |Q̃1 , Q̃2 ]]

= E[P2 (Q̃2 , Q̃1 )(1 − P2 (Q̃2 , Q̃1 ))]

This measure corresponds to the expectation of the asset pay-off’s conditional variance,
at a given trading period. The smaller are these variances, the more efficient is the market.

Proposition 3.12. Given the masses of strategic traders at periods 1 and 2, the variances
at period 1 and 2 are " #
1 α
V1 = E[(Ṽ − P1 ) ] = × 1 − δ 2
2
,
4 Q
" # " #
1α Q−β β 1Q−α 1−δ β Q−β
V2 = (1 − δ) (1 + δ) + + 2 + ,
4Q Q Q 4 Q 2−δQ Q
that can be rewritten as

δ 1 − δ2 1 1−δ
V1 = π1 (α) + and V2 = π2 (α, β) + .
2 4 2 4

Corollary 3.1. At equilibrium, period 1 and 2 variances are equal to

110
 δ
• if C > 2
1 1 1−δ
V1 = and V2 = π2 (0, A) +
4 2 4
consequently
∂V1 ∂V2
= 0, <0
∂A ∂A
.
 
δ δ A
• if 2
≥C≥ 2
1− Q

δ 1 − δ2 1 1−δ
V1 = C + and V2 = π2 (αeq , A) +
2 4 2 4

consequently
∂V1 ∂V1 ∂V2 ∂V2
= 0, > 0, < 0, >0
∂A ∂C ∂A ∂C
 
δ A
• if C < 2
1− Q

δ 1 − δ2 1 1−δ
V1 = π1 (A) + and V2 = π2 (A, A) +
2 4 2 4

consequently
∂V1 ∂V1 ∂V2 ∂V2
< 0, = 0, < 0, =0
∂A ∂C ∂A ∂C

First, let’s notice that the risk, for fast traders, to trade on uninformative on signals, does
not affect the efficient nature of the price process. The competitive market maker sets the
asset price equal to the conditional expected value of its pay-off. The asset price process is
a martingale.
However informational efficiency and the existence mini flash crashes seem to contradict
each other. It would be the case if the effect of HFT was to add mini flash crashes in the
range of possible price dynamics, leaving the conditional distribution of other price dynamics
unchanged. Then mini flash crashes would be equivalent to systematic price reversals. Said
differently, HFT would generate price swings that could be anticipated. Negative auto-
correlation would arise which would contradict the martingale property of an efficient price
dynamic. The effect of the competitive pricing, rather than a mechanical pricing policy, is

111
that mini flash crashes also affect the levels and likelihoods of other price dynamics.
In our model, prices remain efficient. Moreover informational efficiency improves when
High Frequency Trading activity increases, whether it is triggered by a reduction of the «cost
of being fast» or by an increase of the mass of event traders, because it increases their price
impact at t = 1 and the probability that their private information is revealed. The effect is
the same at t = 2 when the mass of event traders, A, increases. It implies that mini flash
crashes are compensated by «momentum» to keep and enhance efficiency.
The effect of HFT is, ultimately, to incorporate into prices the signal they observe at
αeq
t = 1, with probability Q
. Following such a price change, at t = 2, the price will «crash»
(1−δ)A
back to the initial expected value of the asset, 1/2, with probability Q
, it will stay at
Q−A
its t = 1 level,with probability Q
, or it will keep on converging towards the true asset
δA
value, with probability Q
. The former possible dynamic, a price «momentum», can neither
be anticipated. The frequencies of crashes and momentum increase under the action of HFT.
The joint increases of these two types of price dynamic compensate each other so that they
cannot be anticipated, and therefore it keeps market efficiency.
Our model implication is in line with empirical findings on the effect of High Frequency
Trading on market quality. Moreover we find that informational efficiency improvement goes
along with a higher frequency of mini flash crashes and «momentum», which suggests that
HFT allows for a faster integration of new information into prices, but in a less stable way.
Our theory of mini flash crashes can reconcile the empirical findings on the beneficial effect
of HFT on market efficiency with the concerns raised by the increasing frequency of price
instability bursts, such as mini flash crashes.

3.7 Conclusion

High Frequency Trading can enhance market efficiency by processing and incorporating faster
new information into prices, as shown by several empirical works. However HFT has also
triggered the emergence of mini flash crashes, punctual events of intense price instability. We
show with a model that, when the high information processing speed of HFT is associated
with a risk of trading on noise due to erroneous information interpretation, HFT gener-

112
ates price reversals unrelated to underlying change of the asset value. We also show that
while HFT increases the frequency of mini flash crashes, it simultaneously improves market
efficiency.

113
114
Conclusion

I briefly conclude this dissertation by emphasizing some directions in which I would like to
take my research further.

There is still a lot to be done to understand the effects of algorithmic trading in financial
markets. This broad question can be tackled by studying the implications of heterogene-
ity in attention capacities among investors, since algorithmic trading can be seen as a way
to alleviate limited attention constraints. The model, I introduced in the first and second
chapters of this dissertation, offers a way to theoretically pursue this investigation. In my
framework, attention capacity heterogeneity could be modelled by assuming that a fraction
of investors, algorithmic traders, can monitor the market more intensively than the others
and thus can react faster following news arrival. This idea is already at the heart of a joint
on going project, with Profs. Terrence Hendershott and Ryan Riordan, in which we plan to
investigate the informational advantage of investors using Algorithmic Trading (AT) tech-
nologies over the rest of the market. We model this advantage by assuming that AT’s have
a higher information processing speed. Our goal is to derive empirical implications related
to investors’ trading decisions and to test these implications using a database provided by
Deutsche Boërse which contains all AT orders in DAX stocks over 13 days.

Economic agents have cognitive limitations and, in particular, they have a limited capacity
of attention. In the case of financial markets, investors must be attentive to all sort of
information or market events, so that they can take trading decisions. This fact raises the
fundamental question of how investors allocate their attention capacity across markets and
information sources. Provided that investors make this choice rationally, solving for models of

115
optimal attention allocation can help deriving testable implications on, for instance, the joint
dynamics of asset prices, or the difference of trading strategies across markets. The theoretical
framework, I introduced in first and second chapters, is a natural modelling platform to study
how investors allocate their attention capacity across markets and the consequences on asset
prices joint dynamics. Theoretical research on this question has been recently growing. These
models usually consider limited attention as a constraint on the precision of signals on asset
pay-offs. Investors allocate their attention by choosing the variances of the signals they
acquire. The way I model limited attention is different and corresponds to the frequency at
which investors monitor markets. To investigate the endogenous level of investors’ attention,
I would extend my model by allowing investors to decide on how they allocate a finite
monitoring capacity across markets with different characteristics.

116
Appendix A

Appendix to chapters 1 and 2

In this appendix I provide all the proofs of my job market paper ”Limited Attention and News Arrival in
Limit Order Markets”. I recall the model assumptions that are slightly more general than the ones of the
paper.

A.1 Model Setup

A.1.1 Preferences dynamic
The economy is constituted by a continuum of investors [0, L]. They are all risk neutral and infinitely lived,
with time preferences determined by a constant discount rate r > 0 which is also equal to the risk-free interest
rate.
The asset supply is equal to S = sL where 0 < s < 1 that is initially distributed among investors who
can hold either 1 unit or 0 unit of this asset.
As in Duffie, Garleanu and Pedersen, an investor is characterized by whether she owns the asset and by
an intrinsic type that is ”high” or ”low”. A high type owner enjoys a utility flow of v by owning this asset
whereas a low type owner receives a utility flow of v − δ. Then she can consume in the present or save this
utility flow for future consumption.
Between time t and time t + dt a high type investor can ”suffer” from an idiosyncratic shock and switches
to the low type with a probability ρ− .dt. And reciprocally a low type investor switches to the high type with
a probability ρ+ .dt.
Given the previous assumptions any investor must have a type in the set {ho, hn, lo, ln} (h: high, l: low,
o: owner, n: non-owner). And we can divide the mass of investors in 4 populations: Lho , Lhn , , Llo , Lln .
They verify the equations
Lho + Lhn + Llo + Lln = L
Lho + Llo = sL, Lhn + Lln = (1 − s)L
However the number of possible types can be greater than the four aforementioned by taking into account
their limit order submission status. Indeed in a limit order book an owner can either be out of the market or
have an order in the order book at any price reachable. As well for a non-owner. This setting can generate
many subtypes of the previous types. Let’s call T the set of all possible types. However to precise the status
of an investor in the limit order book we will precise if she is out or has sent a limit order. For instance a
type ln can be ln − out or ln − B with a limit order at price B. Symmetrically a type lo can be lo − out or
lo − A with a limit order at price A.

A.1.2 Limit order market

Trading takes place through a limit order market. Such a market is characterized by a price grid at which
limit orders can be sent. We assume that this price grid is bounded which is reasonable since, if the asset

117
utility flow is bounded, trading cannot happen at some high prices (the corresponding strategies would be
strictly dominated by a strategy where investors don’t trade). The investor choices are the following
• owners can : do nothing and remain an owner, send sell limit order in the order book and remain an
owner until his/her order is executed, send a sell market order and become a non-owner or cancel a
previous sell limit order.
• non-owners can : do nothing and remain a non-owner, send buy limit order in the order book and
remain a non-owner until his/her order is executed, send a buy sell market order and become an owner
or cancel a previous buy limit order.
This defines the action set of an investor,

A = {do nothing, market order or marketable limit orders, limit orders at all other prices}

Assumption A.1. In the limit order book, limit order are executed following a ”Pro-rata matching” execution
rule. Given that all limit orders are of size 1 it means that at one price (A or B), at any time, all limit
orders have the same probability of execution.

Assumption A.2. At time t an investor can make a trading choice any time he contacts the market
• if she monitors the market, which occurs with probability λ.dt
• if she ”suffers” from an idiosyncratic shock (as described above), which occurs with probability ρ+/− .dt

Proposition A.1. A limit order market is in a steady state when the displayed depth in the order book and
the different order flows do not change over time.
A steady state equilibrium, for this limit order market definition, is necessarily a one-tick market in the
sense that
all sell limit orders are sent at the price A generating a depth DA , and we don’t observe any other sell limit
order at higher prices than A
all buy limit orders are sent at the price B generating a depth DB , and we don’t observe any other buy limit
order at lower prices than B
A > B and there is no accessible exchange price between A and B. A − B = ∆ is equal to the tick of the
market.

Proof. In the steady state, a sell limit order that is send at a higher price than A will never be executed
because the liquidity supply at the price A keep the same positive value and is never consumed. Symmetrically
for buy limit orders. Then there is no incentive to send limit orders further from the best quotes A and B.
If there was a reachable price A < P < B it would be profitable to send limit orders at P rather than
market orders.

A.1.3 Value function and equilibrium concept

An investor is choosing her actions at each random time when she is contacting the market. The strategy σ
of an agent is a function

σ :H × Ξ × [0, ∞) → A
(h, ξ, t) 7→ a

Where any element Ξ is the set of all potential state variables. If ξ ∈ Ξ then ξ = (θ, v, S) where θ ∈ T is
a type, v is the fundamental value of the asset and S is the aggregate state of the limit order book, that is to
say the bid and ask prices and all the depths at these prices. H is the set of all possible histories of actions
and observations of an investor:

H = {h ∈ (at1 , . . . , atn , ξt1 , . . . , ξtn , t1 , . . . , tn ) ∈ An × Ξn × [0, ∞)n , t1 < . . . < tn , n ∈ N}

118
 Her strategy, σ, and the strategies of all other investors, Σ, are generating an asset holding process
ηtσ,Σ ∈ {0, 1}, a type process θtσ,Σ ∈ T and a trade price Ptσ,Σ any time ηtσ,Σ switches from 0 to 1 or conversely.

At time t the value function of an investor playing strategy σ is given by

Z ∞
V (ht , ξt , t; σ, Σ) = Et e−r(s−t) dUs
t
s.t dUt = ηtσ,Σ (v − δI{θσ,Σ ∈ lo} )dt − Ptσ,Σ dηtσ,Σ
t

The strategy σ is a best response to the other players set of strategies Σ if and only if for all strategy γ,

∀ht ∀ξt ∀t V (ht , ξt , t; σ, Σ) ≥ V (ht , ξt , t; γ, Σ).

Lemma A.1. If the following conditions are fulfilled

• for any contacting time this is not optimal to make more than one trade, given Σ.
• the sequence of contacting time with the market {T̃n }n∈N is such that ∀t limn→∞ P[T̃n < t] = 0
• the process Zt that counts the number of contacting time is such that after for any strategy σ, any
point in time t and any type θt there is a M > 0 such that
Z ∞
e−r(s−t) dZs < M
t

then a strategy σ is a best response to Σ if and only if the one-shot deviation principle holds. The one-shot
deviation principle is verified when the value function generated by σ and Σ, for all type and time, cannot be
improved by deviating from σ only at one contacting time with the market and then playing σ at any future
contacting time.

Remark A.1. The first condition of the lemma implies that, at equilibrium, this is not possible to generate
an infinite profit. This is actually a natural feature of a limit order book where there are limit orders waiting
at the best quotes which prevent anyone to buy and sell with limit orders an infinite numbers of time in an
instant. The only way to trade more than once in a limit order market is to send buy and sell market orders
in a row which clearly not profitable. When contacting times are ”Poissonian” the second and the first point
are verified.

Proof of Lemma A.1

Let’s assume σ is not improvable by a one-shot deviation. A profitable deviation γ is
such that we can find an history ht , a state ξt and a time t that verify

V (ht , ξt , t; γ, Σ) > V (ht , ξt , t; σ, Σ)

Let’s call
dKtσ = ηsσ,Σ (v − δI{θsσ,Σ ∈ lo} )ds − Psσ,Σ dηsσ,Σ

Let’s assume that we can find a profitable deviation γ that is different from σ over a finite number of
contacting time. The set of these kind of deviations is the non-empty and we can consider a deviation with
a minimal number deviation N > 0 by definition. It means that after N contacting time σ is played. Now
let’s consider γ ∗ the strategy of playing γ over the N − 1 contacting time and then to play σ. Necessarily we
have
V (ht , ξt , t; γ ∗ , Σ) ≤ V (ht , ξt , t; σ, Σ)
Let’s call {Tnγ } the sequence of contacting time generated by γ. By definition γ ∗ would generate the same
distribution for the N − 1 contacting time. And until TNγ −1 it will also generates the same consumption path.

119
then we can write
t+Tiγ
"N −1 Z #
X γ
−r(t+TN
∗
V (ht , ξt , t; γ , Σ) = Et e −r(s−t)
dKsγ +e −1
)
V (hγt+T γ , ξt+T
γ
γ ,t + TNγ −1 ; σ, Σ)
γ N −1 N −1
i=1 t+Ti−1

t+Tiγ
"N −1 Z #
X γ
−r(t+TN
(one-shot unimprovability) ≥ Et e −r(s−t)
dKsγ +e −1
)
V (hγt+T γ , ξt+T
γ
γ ,t + TNγ −1 ; γ, Σ)
γ N −1 N −1
i=1 t+Ti−1

≥ V (ht , ξt , t; γ, Σ)

There is a contradiction.

Now let’s consider any profitable deviation γ.

First let’s show that for any strategy and any sequence of contacting time such that limn→∞ P[Tn < t] = 0
we have h i
lim Et e−r(t+Tn ) V (hγt+Tn , ξt+T
γ
n
, t + Tn ; γ, Σ) =0
n→∞

The assumption of the lemma gives us that V (ht , ξt , t; γ, Σ) is bounded whatever the strategy and that the
bound does not depend on the strategy. Let’s call f (T ) = e−r(t+T ) V (hγt+T , ξt+T
γ
, t + T ; γ, Σ). We can find A
such that f < A because f is bounded. Let  > 0, let T > 0 such that P[Tn < T ] <  and such that f (T 0 ) < 
for all T 0 > T . Then

Et [f (Tn )] = Et [f (Tn )|T < Tn ]P[Tn < T ] + Et [f (Tn )|T> n]P[Tn > T ] < A + 

and we have the convergence result

let β = V (ht , ξt , t; γ, Σ) − V (ht , ξt , t; σ, Σ) > 0. Because of the previous convergence result we can find
N0 such that for N > N0
γ
X t+Tiσ
"N Z # "N Z #
X t+Ti β
−r(s−t) γ −r(s−t) σ
Et e dKs − Et e dKs >
γ
i=1 t+Ti−1
σ
i=1 t+Ti−1
2

we can then consider γN the strategy of playing γ until TN and σ afterwards. Then for N big enough,
β
|Et [e−r(t+TN ) V (hγt+T
N γN
, ξt+T , t + TN ; γN , Σ)] − Et [e−r(t+TN ) V (hσt+TN , ξt+T
σ
, t + TN ; σ, Σ)]| <
N N N
4
and then V (ht , ξt , t; γN , Σ) > V (ht , ξt , t; σ, Σ) This is a contradiction with the first result of the proof.

120
A.2 Limit order market in steady state w/o fundamen-
tal uncertainty
The first type of equilibria I am implementing is a steady state equilibrium when the fundamental value of
the asset does not change. In these equilibria the four types {ho, hn, lo, ln} are sufficient to describe the
micro dynamic of the order book. This imply that investor are pooled by type in the order book.

A.2.1 equilibrium conjecture

Conjecture A.1. I am looking for the set of parameters such that the equilibrium strategies are
• ho: cancel any sell limit order and keep the asset’s unit
• hn: send a buy limit or market order (indifferently)
• lo: send a sell limit or market order (indifferently)
• ln: cancel any buy limit order and stay without the asset
This conjectured equilibrium is a Markov Perfect Equilibrium with state variable ξt = (θt , vt , St ).
In this potential equilibrium the populations Lho and Lln are not present in the limit order book. As soon
as a ho type switches to a lo type she instantaneously contacts the market: either she instantaneously switches
to a ln type by sending a sell market order or stay a lo type by sending a sell limit order. Symmetrically as
soon as a ln type switches to a hn type she instantaneously contacts the market: either she instantaneously
switches to a ho type by sending a buy market order or stay a hn type by sending a buy limit order. This is
straightforward that in a steady state equilibrium, DA = Llo and DB = Lhn .

A.2.2 Steady state populations

In a steady state the aggregate populations stay at the same level. Then the flows of population from high
type to low type and from low type to high type must be equal to each other, ρ− (Lho + Lhn ) = ρ+ (Llo + Lln ).
Then we must have
ρ+ ρ−
Lho + Lhn = + L, L lo + Lln = L
ρ + ρ− ρ+ + ρ−
Proposition A.2. In a steady state equilibrium there is an α ∈ R such that the different types of population
verify
Lho = (s − α)L
ρ+
 
Lhn = −s+α L
ρ+ + ρ−
Llo = αL
ρ−
 
Lln = −α L
ρ+ + ρ−
Conditions on the possible values taken by α are:
ρ+ ρ−
s − α ≥ 0, −
− s + α ≥ 0, α ≥ 0, + −α≥0
ρ+ +ρ ρ + ρ−
Assumption A.3. The asset supply is less than the high type population in steady state and more than the
− +
low type population, ρ+ρ+ρ− ≤ s ≤ ρ+ρ+ρ− , and the high valuation population is bigger than the low valuation
population, ρ+ > ρ− .
ρ−
Given this assumption the condition on α is now 0 ≤ α ≤ ρ+ +ρ−

Remark A.2. A simplification of the problem parametrization will be the symmetric case where ρ+ = ρ−
and s = 12

121
Proof of proposition A.2
The steady state is defined by the system
    ρ+ 
1 1 0 0 Lho ρ+ +ρ− L
  ρ− L 
 Lhn  = 
0 0 1 1 
 ρ+ +ρ− 
 
1 0 1 0  Llo   sL 
0 1 0 1 Lln (1 − s)L
+ −
First it is to check that Lho = sL, Lhn = ( ρ+ρ+ρ− − s)L, Llo = 0, Lln = ρ+ρ+ρ− L is a particular solution of
this system. The general space of solutions of this system is equal to
    
sL 
1 1 0 0 sL  
−1
+ +
( ρ ρ
 ρ+ +ρ− − s)L 0 0 1 1  ( − s)L  1 
 
+ −
ρ +ρ
 
1 0 1 0 = 
 + ker   + V ect 
  
0 0  1 

   
ρ− 0 1 0 1 ρ− −1
ρ+ +ρ− L ρ+ +ρ− L

A.2.3 Micro-level dynamic of the limit order book

In the equilibrium conjecture we need that hn and lo types are indifferent between limit and market order so
that we observe in the same time flows of liquidity demand and supply that make the state of the limit order
book sustainable and steady. Given that we look for a steady state equilibrium, the flows must be steady as
well. That is why a share 0 < mA < 1 of the population of hn type contacting the market at t sends a buy
market order and the rest of them send a buy limit order. For the same reason a share 0 < mB < 1 of the
population of lo type contacting the market at t sends a sell market order and the rest of them send a sell
limit order.

Ask Side. At time t, on the ask side of the market the depth is constantly equal to DA = Llo and the
order flows sustaining this steady state are
• Outflow: people switching from lo to ho, ρ+ Llo .dt, lo type cancelling their sell limit order to send a
sell market order, mB λLlo .dt, execution of buy market orders send by hn type contacting the market
mA (λLhn + ρ+ Lln ).dt.
• Inflow: people switching from ho to lo type sending a sell limit order, (1 − mB )ρ− Lho .dt
The steady state condition is then:

ρ+ Llo + mA (λLhn + ρ+ Lln ) + mB (λLlo + ρ− Lho ) = ρ− Lho

Bid Side. At time t, on the ask side of the market the depth is constantly equal to DB = Lhn and the
order flows sustaining this steady state are
• Outflow: people switching from hn to ln, ρ− Lhn .dt, hn type cancelling their sell limit order to send
a sell market order, mA λLhn .dt, execution of sell market orders send by lo type contacting the market
mB (λLlo + ρ− Lho ).dt.
• Inflow: people switching from ln to hn type sending a sell limit order, (1 − mA )ρ+ Lln .dt
The steady state condition is then:

ρ+ Lhn + mB (λLlo + ρ− Lho ) + mA (λLhn + ρ+ Lln ) = ρ+ Lln

A.2.4 Value functions

The conjectured equilibrium generates the following system of equations defining the different value functions
for each investor types.

122
type ho. An investor of type ho keeps its asset until he/she switches to the lo type.
Vho = v.dt + (1 − r.dt)[(1 − ρ− .dt)Vho + ρ− .dtVlo ] ⇐⇒ (r + ρ− )Vho = ρ− Vlo + v

type ln. An investor of type ln does not sen any order until he/she switches to the hn type.
Vln = (1 − r.dt)[(1 − ρ+ .dt)Vln + ρ+ .dtVhn ] ⇐⇒ (r + ρ+ )Vln = ρ+ Vhn

type hn. An investor of type hn send a buy market or limit order indifferently. To be consistent with
the fact that a share mA < 1 of the hn type on the market are sending a market order, we can consider that
they are playing a mixed strategy between market and limit order with probability mA for the market order.

As we assume the steady state, being a hn type during a certain amount of time means being on the bid
side of the limit order book during this time. At time t, the outflow of the bid side due to market order is
mB (λLlo + ρ− Lho ).dt, then the probability for a limit order to be executed at t is lB .dt with

mB (λLlo + ρ− Lho )
lB =
Lhn

Indeed because the priority rule in the limit order book is ”Pro-Rata”, the instantaneous probability of
execution is equal to the ratio of the market order flow over the depth of the limit order book.
Given that these investor are indifferent between sending a limit or a market order the two associated
value function must be equal:
Vhn = Vho − A

and

Vhn = (1 − r.dt)[(1 − ρ− .dt − λ.dt − lB .dt)Vhn + ρ− .dtVln + λ.dt(mA (Vho − A) + (1 − mA )Vhn ) + lB .dt(Vho − B)]
⇐⇒(r + ρ− + lB + mA λ)Vhn = ρ− Vln + mA λ(Vho − A) + lB (Vho − B)

leading to, given the indifference result

(r + ρ− + lB )(Vho − A) = ρ− Vln + lB (Vho − B)

type lo. An investor of type lo send a sell market or limit order indifferently. To be consistent with the
fact that a share mB < 1 of the lo type on the market are sending a market order, we can consider that they
are playing a mixed strategy between market and limit order with probability mB for the market order.

For the same reason as for type hn, at time t, the outflow of the ask side due to market order is
mA (λLhn + ρ+ Lln ).dt, then the probability for a limit order to be executed at t is lA .dt with

mA (λLhn + ρ+ Lln )
lA =
Llo

Because these investor are indifferent we have

Vlo = Vln + B

and
(r + ρ+ + lA + mB λ)Vlo = v − δ + ρ+ Vho + mB λ(Vln + B) + lA (Vln + A)

leading to
(r + ρ+ + lA )(Vln + B) = v − δ + ρVho + lA (Vln + A)

123
Proposition A.3. Solving the system implied by these equations gives

v − rA − ρ− (A − B)
lB =
A−B
rB − ρ+ (A − B) − (v − δ)
lA =
A−B
1 ρ +
1 ρ−
Vho = −
(v − ρ− (A − B)) + (v + ρ+ A + ρ− B))
+
rρ +ρ r + ρ + ρ ρ + ρ−
+ − +

1 ρ+ 1 ρ+
Vln = −
(v − ρ− (A − B)) − (v + ρ+ A + ρ− B))
+
rρ +ρ r + ρ + ρ ρ + ρ−
+ − +

Vhn = Vho − A
Vlo = Vln + B

Corollary A.1. The following condition is sufficient for lA and lB to be take positive values

v δ r + ρ− v δ ρ−
− ≤ B < A ≤ −
r r r + ρ+ + ρ− r r r + ρ+ + ρ−

This implies that δ − (r + ρ+ + ρ− )∆ > 0.

Proof of proposition A.3

First, by replacing Vhn by Vho − A and Vlo by Vln + B this is easy to obtain that

(r + ρ− )Vho − ρ− Vln = v + ρ− B
(r + ρ+ )Vln − ρ+ Vho = −ρ+ A

and then to get the expression of Vho and Vln .

Replacing Vln by Vlo − B and Vho by Vhn + A in the equation of indifference between market and limit orders
we obtain

(r + ρ− + lB )(Vho − A) = ρ− (Vlo − B) + lB (Vho − B)

(r + ρ+ + lA )(Vln + B) = v − δ + ρ+ (Vhn + A) + lA (Vln + A)

which gives

v + ρ− B − (r + ρ− )A = lB (A − B)
− ρ+ A + (r + ρ+ )B − (1 − δ) = lA (A − B)

Proof of Corollary A.1

lA and lB must be positive numbers. Then the following conditions must hold:

v + ρ− B v − δ + ρ+ A
A≤ −
, B≥
r+ρ r + ρ+

Sufficient conditions for these conditions to hold is that

ρ− ρ+ ρ−
 
v v−δ v δ
A≤ −
+ −
+ A ⇔A≤ −
r+ρ r+ρ r+ρ + r+ρ + r r r + ρ+ + ρ−

and then
ρ+ ρ− v δ r + ρ−
 
v−δ v
B≥ + −
+ −
B ⇔B≥ −
r+ρ+ r+ρ + r+ρ r+ρ r r r + ρ+ + ρ−

124
A.2.5 Equilibrium outcome
The two steady state equations can be rewritten as

ρ− Lhn + lB Lhn + lA Llo = ρ+ Lln

ρ+ Llo + lA Llo + lB Lhn = ρ− Lho

Proposition A.4. The conjecture equilibrium is indeed an equilibrium for a set of parameters value (close
enough to the symmetric case). The equilibrium populations are characterized by the value
+
ρ− s − lB ( ρ+ρ+ρ− − s)
αeq =
ρ+ + ρ− + lA + lB

The aggregate properties of the limit order market in this steady state equilibrium is completely described
by the equilibrium value αeq since it gives the equilibrium populations, the depths and the aggregate order
flow in the limit order book.

Proof of Proposition A.4

Given the equilibrium strategy σ (and Σ generated by σ) the random times at which an investor i is going to
contact the market are generated by the Poisson process corresponding to her valuation of the asset (intensity
ρ), the process of his market monitoring(intensity λ) and the processes of his limit order executions if she
sends a limit order at A or B (intensity lA and lB ). In this framework this is obvious that assumptions of
lemma 1 can apply.

In this purpose I consider ”one-shot” deviations where an investor can deviate from the assumed strategy
when she contacts the market but not considering to deviate from the assumed strategy in her future decision.
By backward induction, checking if this kind deviation is not profitable is also checking that a finite number
of deviation in a row is not profitable as well.

type hn. A type hn has only one way to deviate which is to stay out of the market and not sending a
buy market or limit order. Actually we could consider a change of the mixed strategy parameter between
limit and market orders but given that the investor is infinitesimal this deviation does not change the value
of these two actions. Then he/she is indifferent between all mixed strategies which makes this deviation not
profitable.
The value of not trading when contacting the market is

Vhn−out = (1 − r.dt)[(1 − ρ− .dt − λ.dt)Vhn−out + ρ− .dtVln + λ.dtVhn

⇐⇒(r + ρ− + λ)Vhn−out = ρ− Vln + λVhn ≤ (ρ− + λ)Vhn

if Vln ≤ Vhn which is true if and only if Vln ≥ 0 because (r + ρ+ )Vln = ρ+ Vhn . In this case this deviation is
not profitable. It is easy to verify that Vln ≥ 0 iff

1 + ρ− B
A≤
r + ρ−

which is verified as soon as lB > 0.

type lo. As in the previous case the only deviation we have to consider is the case where a type lo in
contact with the market decides to keep the asset. In this case the value function is

Vlo−out = (v − δ).dt + (1 − r.dt)[(1 − ρ+ .dt − λ.dt)Vlo−out + ρ+ .dtVho + λ.dtVlo

⇐⇒(r + ρ+ + λ)Vlo−out = v − δ + ρ+ Vho + λVlo

125
As we know that (r + ρ+ )(Vlo − B) = ρ+ (Vho − A) we get

(r + ρ+ + λ)Vlo−out = v − δ + ρ+ (A − B) − rB + (r + ρ+ + λ)Vlo

Then the deviation is not profitable iff

1 − δ + ρ+ A
v − δ + ρ+ (A − B) − rB ≤ 0 ⇐⇒ B ≥
r + ρ+

which is verified as soon as lA > 0.

type ho. A type ho can deviate in two different ways:

• by sending a sell market order. The value function associated is then V = Vhn +B = Vho −A+B < Vho .
This is not profitable.
• by sending a sell limit order. In this case the value function is given by

Vho−A = v.dt + (1 − r.dt)[(1 − ρ− .dt − λ.dt − lA .dt)Vho−A + ρ− .dtVlo + λ.dtVho + lA .dt(Vhn−out + A)]
⇐⇒(r + ρ− + λ + lA )Vho−A = v + ρ− Vlo + (λ + lA )Vho + lA (Vhn−out − Vhn ) < (r + ρ− + λ + lA )Vho

This deviation is not profitable.

type ln. A type ln can deviate in two different ways:

• by sending a buy market order. The value function associated is then V = Vlo −A = Vln +B −A < Vln .
This is not profitable.
• by sending a buy limit order. In this case the value function is given by

Vln−B = (1 − r.dt)[(1 − ρ+ .dt − λ.dt − lB .dt)Vln−B + ρ+ .dtVhn + λ.dtVln + lB .dt(Vlo−out − B)]

⇐⇒(r + ρ+ + λ + lB )Vln−B = ρ+ Vhn + (λ + lB )Vln + lB (Vlo−out − Vlo ) < (r + ρ+ + λ + lB )Vln

This deviation is not profitable.

The two steady state equations can be rewritten as

ρ− Lhn + lB Lhn + lA Llo = ρ+ Lln

ρ+ Llo + lA Llo + lB Lhn = ρ− Lho

replacing by the possible value of these masses at equilibrium we obtain the equations in function of αeq

ρ+ ρ+ ρ−
ρ− ( − s + αeq ) + lB ( − s + αeq ) + lA αeq = ρ( − αeq )
ρ+ + ρ− ρ+ + ρ− ρ+ + ρ−

ρ+
ραeq + lA αeq + lB ( − s + αeq ) = ρ(s − αeq )
ρ+ + ρ−
that both give the same result for αeq
+
ρ− s − lB ( ρ+ρ+ρ− − s)
αeq =
ρ+ + ρ− + lA + lB
ρ+ ρ+ ρ−
For s = ρ+ +ρ− , αeq = ρ+ +ρ− ρ+ +ρ− +lA +lB and in this case we verify that

ρ−
0 ≤ αeq ≤
ρ+ + ρ−

126
 +
For s close enough to ρ+ρ+ρ− these inequalities are (strictly) verified. Equilibrium value of steady populations
are given by the equations

ρ+
 
Lho = (s − αeq )L, Lhn = − s + αeq L,
ρ+ + ρ−
ρ−
 
Llo = αeq L, Lln = − αeq L
ρ+ + ρ−

To be sure that the equilibrium exists we need to check that

0 ≤ mA ≤ 1, 0 ≤ mB ≤ 1

For instance
Lhn
mB = lB
λLlo + ρ− Lho
Because lB , Lhn , Llo and Lho do not depend on λ, we can always a high enough value of λ so that mB < 1.
As well for
Llo
mA = lA
λLhn + ρ+ Lln

We can also noticed that market order flows are independent of the monitoring parameter λ. Indeed the
sell market order is equal to mB (λLlo +ρ− Lho ) = Lhn lB and the buy market order flow is mA (λLhn +ρLln ) =
Llo lA .

Remark A.3. In the symmetric case

1 ρ
αeq =
2 2ρ + lA + lB
αeq αeq lA
mA = 1 lA < 1 lA = <1
λαeq + ρ( 2 − αeq ) ρ( 2 − αeq ) ρ + lA + lB
and for the same reason
lB
mB < <1
ρ + lA + lB
Whatever is the value of λ the equilibrium exits

127
A.3 Dynamic equilibrium converging to a steady state
w/o fundamental uncertainty
To complete the former steady state class of equilibria this is possible to design a dynamic equilibrium, con-
verging toward one of these equilibria, in which the terms of the trade-off between limit and market orders
are unchanged. This will be useful for the next sections in order to define a broader class of equilibria where
the limit order market switch from one steady state to another.

Starting at t = 0 from a one tick market where the depth at prices A and B are DA (0) and DB (0)
constituted respectively by a share of the population Llo (0) and of Lhn (0), agents follow their corresponding
steady state equilibrium strategy described earlier. The rate at which hn and lo types are sending market
orders, mA (t) and mB (t), are evolving so that the terms of the trade off are the same as in the steady state
equilibrium. More precisely the intensity rate at which limit orders are executed are unchanged and equal
to lA and lB . In this framework the dynamic of the different population is given by the dynamic of the
parameter α,

Lho (t) = (s − α(t))L

ρ+
 
Lhn (t) = − s + α(t) L
ρ+ + ρ−
Llo (t) = α(t)L
ρ−
 
Lln (t) = − α(t) L
ρ+ + ρ−

and the value function for each type are the same as in the former steady-state equilibrium.

A.3.1 Micro-level dynamic of the limit order book

Ask Side. At time t, on the ask side of the market the depth is equal to DA (t) and the order flows
sustaining this steady state are
• Outflow: people switching from lo to ho, ρ+ DA (t).dt, lo type cancelling their sell limit order to send
a sell market order, mB (t)λDA (t).dt, execution of buy market orders send by hn type contacting the
market mA (t)(λLhn (t) + ρ+ Lln (t)).dt = lA DA (t).dt.
• Inflow: people switching from ho to lo type sending a sell limit order, (1 − mB (t))ρ− Lho (t).dt, lo type
outside of the order book sending a sell limit order (1 − mB (t))λ(Llo (t) − DA (t)).dt

Bid Side. At time t, on the ask side of the market the depth is constantly equal to DB (t) and the order
flows sustaining this steady state are
• Outflow: people switching from hn to ln, ρ− DB (t).dt, hn type cancelling their sell limit order to
send a sell market order, mA λDB (t).dt, execution of sell market orders send by lo type contacting the
market mB (t)(λLlo (t) + ρ− Lho (t)).dt = lB DB (t).dt.
• Inflow: people switching from ln to hn type sending a sell limit order, (1 − mA (t))ρ+ Lln (t).dt, hn
type outside of the order book sending a sell limit order (1 − mA (t))λ(Lhn (t) − DB (t)).dt.
Then the dynamics of the limit order books depths are given by the first order differential equations,

dDA
= ρ− Lho (t) − ρ+ DA (t) − lA DA (t) − lB DB (t) + λ(Llo (t) − DA (t))
dt
dDB
= ρ+ Lln (t) − ρ− DB (t) − lB DB (t) − lA DA (t) + λ(Lhn (t) − DB (t))
dt

128
A.3.2 Outcome of the dynamic equilibrium
Proposition A.5. The dynamics of the equilibrium populations are given by the dynamic of the parameter
α,
ρ+
  
dα − + −
= ρ s − lB + −
− s − [ρ− + ρ+ + lA + lB ]α(t) + lA κA e−(λ+ρ )t + lB κB e−(λ+ρ )t
dt ρ +ρ
with solution
−
+ρ+ +lA +lB )t
α(t) = αeq + (α(0) − αeq )e−(ρ
−
1 − e−[λ−(ρ +lA +lB )]t −(ρ− +ρ+ +lA +lB )t
+ lA κA e
λ − (ρ− + lA + lB )
+
1 − e−[λ−(ρ +lA +lB )]t −(ρ− +ρ+ +lA +lB )t
+ lB κB e
λ − (ρ− + lA + lB )
Llo (0)−DA (0) Lhn (0)−DB (0)
with κA = L , κB = L
ρ− ρ− ρ−
Corollary A.2. If 0 ≤ α(0) ≤ ρ− +ρ+ and 0 ≤ αeq ≤ ρ− +ρ+ then ∀t ≥ 0, 0 ≤ α(t) ≤ ρ− +ρ+

Proposition A.6. For a set of parameter values this limit order book dynamic is an equilibrium dynamic
that converges toward a steady state equilibrium at the same ask and bid prices, A and B. In this equilibrium
the value functions for the different agents types are constant and equal to the value function corresponding
to the limit steady state equilibrium.

Proof of Proposition A.5

In addition the lo type agents who are not in the order book at t = 0 enter the market as soon as soon as
they contact the market. The dynamic of this population is then
+
Llo (t) − DA (t) = (Llo (0) − DA (0))e−(λ+ρ )t

For the same reason the dynamic of hn type not in the limit order book is given by
−
Lhn (t) − DB (t) = (Lhn (0) − DB (0))e−(λ+ρ )t

As in the steady state case

DB (t) DA (t)
mB (t) = −
lB , mA (t) = lA
λLlo (t) + ρ Lho (t) λLhn (t) + ρ+ Lln (t)
are well defined (∈ [0, 1]) for all t > 0 for some value of λ high enough.

To obtain the differential equation that drives the dynamic of α, we use the differential equation for
+
DA (t) and the equalities DA (t) = Llo (t) − (Llo (0) − DA (0))e−(ρ +λ)t . We obtain
dLlo (t) + +
+ (ρ+ + λ)(Llo (0) − DA (0))e−(ρ +λ)t = ρ− Lho (t) − (ρ+ + lA )(Llo (t) − (Llo (0) − DA (0))e−(ρ +λ)t )
dt
− +
− lB (Lhn (t) − (Lhn (0) − DB (0))e−(ρ +λ)t ) + λ(Llo (0) − DA (0))e−(ρ +λ)t
which gives
dLlo (t) + −
= ρ− Lho (t)−(ρ+ +lA )Llo (t)−lB Lhn (t)−lA (Llo (0)−DA (0))e−(ρ +λ)t −lB (Lhn (0)−DB (0))e−(ρ +λ)t
dt
 + 
and then use that Lho (t) = (s − α(t))L, Lhn (t) = ρ+ρ+ρ− − s + α(t) L, Llo (t) = α(t)L. to get the final
differential equation
ρ+
  
dα − + −
= ρ − lB + −
− s − [ρ− + ρ+ + lA + lB ]α(t) + lA κA e−(λ+ρ )t + lB κB e−(λ+ρ )t
dt ρ +ρ

129
 − +
To obtain the general solution to this ODE, we look for the functional form α(t) = c(t)e−(ρ +ρ +lA +lB )t
.
Then
ρ+
  
− + dc + −
e−(ρ +ρ +lA +lB )t = ρ − − lB + −
− s + lA κA e−(λ+ρ )t + lB κB e−(λ+ρ )t
dt ρ +ρ
and then
h  + i
ρ− − lB ρ+ρ+ρ− − s −
−
e−[λ−(ρ +lA +lB )]t − 1
+ρ+ +lA +lB )t
c(t) = c(0) + × (e(ρ − 1) − lA κA
ρ− + ρ+ + lA + lB λ − (ρ− + lA + lB )
+
e−[λ−(ρ +lA +lB )]t − 1
− lB κB
λ − (ρ+ + lA + lB )

Proof of Corollary A.2

First we can see that α(t) > 0 obviously. Since α is converging it can have extrema. Given the ODE that
defines α these extrema must verify

ρ+
  
+ −
[ρ− + ρ+ + lA + lB ]α(t) = ρ− s − lB + −
− s + lA κA e−(λ+ρ )t + lB κB e−(λ+ρ )t
ρ +ρ

This gives

ρ+ ρ+
  
−
− − −(λ+ρ+ )t
+
[ρ + ρ + lA + lB ]α(t) ≤ ρ s − lB + −
− s + l A α(0)e + l B (α(0) + + −
− s)e−(λ+ρ )t
ρ +ρ ρ +ρ
we can rewrite
ρ− s ρ+
 
−
[ρ− + ρ+ + lA + lB ]α(t) ≤ (ρ− + ρ+ ) − lB − s (1 − e−(λ+ρ )t )
ρ− + ρ+ +
ρ +ρ −
+ −
+ (lA + lB )α(0) − α(0)lA (1 − e−(λ+ρ )t
) − α(0)lB (1 − e−(λ+ρ )t
)

And then
ρ− s
[ρ− + ρ+ + lA + lB ]α(t) ≤ (ρ− + ρ+ ) + (lA + lB )α(0)
ρ− + ρ+

130
A.4 Limit order book in transition phase
Before we turn to the description of the limit order book dynamic in the transition phase. Preceding the
beginning of the transition phase the world is in the state ζ = ∅. The transition phase starts when the
asset fundamental value changes. It changes at some point in time τ . It is stochastic and follows a Poisson
distribution, P(µ). After τ the state of the world is either ζ = u (up) and v = v 0 + ω or ζ = d and
v = v 0 − ω(down) with equal probability. In this section I start time back to 0 when the fundamental value
changes. Here time t corresponds to time t + τ of the overall game. I call T u and T d the duration of the
transition phases in the different states of the world.

A.4.1 Transition phase strategy

To define properly the strategy in the transition phase it is necessary to decide what will be the steady state
phase in the last phase. Different level of prices can define an equilibrium. If all investors know that a
particular steady state equilibrium will be played during the last phase they coordinate on this equilibrium.
Conjecture A.2. After the fundamental value has changed if ζ = u, for t > T u , investors coordinate on the
steady state equilibrium over the bid-ask prices (Au , B u ) and if ζ = d, for t > T d , investors coordinate on
the steady state equilibrium over the bid-ask prices (Ad , B d ).
During the transition phase, the strategy is:
• In the case ζ = u, for t < T u :
- lo cancel any sell limit order that is not at price Au and submit a limit order at price Au
- ho cancel any sell limit order and stay out of the market
- ln send a buy market order and immediately behave as their new type, lo
- hn send a buy market order and immediately behave as their new type, ho
• In the case ζ = d, for t < T d :
- hn cancel any buy limit order that is not at price B d and submit a limit order at price B d
- ln cancel any buy limit order and stay out of the market
- ho send a sell market order and immediately behave as their new type, hn
- lo send a sell market order and immediately behave as their new type, ln
The rational here is that when the fundamental value changes to become higher for instance, non-owner
become arbitrageurs and have an incentive to buy the asset while it is tradable at a low price, A0 , and to
resell it at a high price Au later. This is what we are conjecturing.

A.4.2 Limit order book dynamics in the transition phase

Before the transition phase begins the limit order book is filled with some liquidity. In particular liquidity
provision at best ask and bid prices is defined by the value of the depth of the limit order book at prices A0
∅ ∅
and B 0 . These are equal DA 0 and DB 0 . During the transition phase the limit orders at prices A
0
and B 0
are being cancelled or executed which generates a dry-out of liquidity at these prices. Here we are giving
the dynamics of the liquidity supply which offers a free option opportunity after the fundamental value has
changed.

In the case S = u
Proposition A.7. When S = u during the transition phase the depth at price A0 is
+ − +
∅ −(λ+ρ
u
DA 0 (t) = −(1 − s)L + [DA0 + (1 − s)L]e
)t
+ L∅hn (e−(λ+ρ )t
− e−(λ+ρ )t
)
which is decreasing and has a unique zero, defining the time T u .

131
Proof of Proposition A.7
Following the possible equilibria conjectured, for t < τ there are potentially sell limit orders at prices A0 and
∅ ∅ ∅ ∅
Au and buy limit orders at prices B 0 and B d . DA 0 , DAu , DB 0 and DB d are the associated depths. Given

the strategy for t < τ , we must have

∅ ∅ ∅ ∅ ∅ ∅
DA 0 + DAu = Llo , DB 0 + DB d = Lhn

The limit order book dynamics is given by :

u
• hn and ln type agents cancel their limit orders and send a buy market order while DA 0 > 0

u
∂DB
= −(λ + ρ− )DB
0 u
0 (t)
∂t
u
∂DB d
= −(λ + ρ− )DB
u
d (t)
∂t
This implies that
∂Luhn
= −(λ + ρ− )Luhn (t)
∂t
because when ln types switch to hn they immediately send a market order and become ho.
• lo type with limit orders at A0 cancel their limit orders or are executed by market orders send by hn
and ln u
∂DA 0 u − u u
= −(λ + ρ+ )DA +
0 (t) − (λ + ρ )Lhn (t) − (λ + ρ )Lln (t)
∂t
or u
∂DA 0 u − u u
= −(λ + ρ+ )DA +
0 (t) − (λ + ρ )Lhn (t) − (λ + ρ )[(1 − s)L − Lhn (t)]
∂t
• lo types formerly ln or in the limit order book at A0 send sell limit orders at Au and ho types cancel
their limit orders
u
∂DA u u − u u − u + u
= λDA 0 (t) + ρ Lhn (t) + λLln (t) + ρ Lho (t) − ρ DAu (t)
∂t
The transition phases ends at T u such that DA
u u
0 (T ) = 0.
u
The dynamic of DA 0 (t) is given by the

differential equation :
u
∂DA −
− ρ− )L∅hn e−(λ+ρ )t
0 u
= −(λ + ρ+ )DA +
0 (t) − (λ + ρ )(1 − s)L + (ρ
+
∂t
+
u −(λ+ρ )t
Let’s find a solution of the ODE of the type DA 0 (t) + (1 − s)L = c(t)e . This gives
+ −
⇔ċe−(λ+ρ )t
= (ρ+ − ρ− )L∅hn e−(λ+ρ )t

+
−ρ− )t
⇔ċ = (ρ+ − ρ− )L∅hn e(ρ
+
−ρ− )t
⇔c(t) = c0 + L∅hn e(ρ

Then + −
−(λ+ρ
u
DA 0 (t) + (1 − s)L = c0 e
)t
+ L∅hn e−(λ+ρ )t

∅ ∅
and given the initial condition c0 = DA 0 + (1 − s)L − Lhn .

In the case S = d
Proposition A.8. When S = d during the transition phase the depth at price B 0 is
− + −
∅ −(λ+ρ
d
DB 0 (t) = −sL + [DB 0 + sL]e
)t
+ L∅lo (e−(λ+ρ )t
− e−(λ+ρ )t
)

which is decreasing and has a unique zero, defining the time T d .

132
Proof of Proposition A.8
The limit order book dynamics is given by :
d
• lo and lo type agents cancel their limit orders and send a sell market order while DB 0 > 0

d
∂DA 0 d
= −(λ + ρ+ )DA 0 (t)
∂t
d
∂DA u d
= −(λ + ρ+ )DA u (t)
∂t
This implies that
∂Ldlo
= −(λ + ρ+ )Ldlo (t)
∂t
because when ho types switch to lo they immediately send a market order and become ln.
• hn type with limit orders at B 0 cancel their limit orders or are executed by market orders send by lo
and ho
d
∂DB
= −(λ + ρ− )DB −
0 d + d d
0 (t) − (λ + ρ )Llo (t) − (λ + ρ )Lho (t)
∂t
or
d
∂DB
= −(λ + ρ− )DB −
0 d + d d
0 (t) − (λ + ρ )Llo (t) − (λ + ρ )[sL − Llo (t)]
∂t
• hn types formerly ho or in the limit order book at B 0 send buy limit orders at B d and ln types cancel
their limit orders
d
∂DB d d + d d + d − d
= λDB 0 (t) + ρ Llo (t) + λLho (t) + ρ Lln (t) − ρ DB d (t)
∂t

The transition phases ends at T d such that DB

d d
0 (T ) = 0.
d
The dynamic of DB 0 (t) is given by the

differential equation :
d
∂DB +
= −(λ + ρ− )DB − −
− ρ+ )L∅lo e−(λ+ρ )t
0 d
0 (t) − (λ + ρ )sL + (ρ
∂t
The ODE solving is as for proposition D.1.

A.4.3 Equilibrium in the subgame starting after the fundamental

value changed
Once the common value v has changed (after τ ) we already know that after the transition phase (t > T u/d )
we are playing an equilibrium strategy solved in section III. For instance if S = u at t = T u we begin to
play the dynamic equilibrium with the different population at time T u and the depth DA u u
u (T ) at A
u
and
u
0 at B . To obtain the equilibrium in the subgame we need to show that during the transition phase the
conjecture strategy is indeed optimal.

The conjecture strategy implies that there is no possibility to send any sequence of limit order (executed)
at the same point in time which show that this is not profitable to have more than on trade at one point
in time. Given that contacting times are Poissonian we clearly are in the condition of the lemma 2.1. The
two following propositions confirm that the strategies we proposed as equilibrium strategies candidates are
indeed generating an equilibrium.

Proposition A.9. When S = u the conjecture strategy in the subgame starting at τ is an equilibrium strategy.

Proposition A.10. When S = d, the conjecture strategy in the subgame starting at τ is an equilibrium
strategy.

133
Proof of Proposition A.9
During the transition phase the observed types under the conjecture strategy are lo − Au (lo with a limit
order at price Au ), lo − A0 , ho − out, ln − out, hn − B 0 and hn − B d . But hn − B 0 and hn − B d can be
gathered under the label hn − out because their limit orders are not executed under the conjecture strategy
so they would get the same outcome if they were out of the order book. In this framework we can define the
system of ODE’s defining the value function of these types:

• lo − Au stay in the limit order book until they switch of type

u
u u ∂Vlo−Au
u
(r + ρ+ )Vlo−Au (t) = v −δ+ + ρ+ Vho−out (t)
∂t
u u u u
with Vlo−Au (T ) = V̄lo−Au which is the value for a lo of having a limit order at price A once the last
dynamic equilibrium is played (the optimal strategy in the last phase).
• ho − out stay out until they switch of type
u
∂Vho−out
(r + ρ− )Vho−out
u
(t) = v u + + ρ− Vlo−A
u
u (t),
u
Vho−out (T u ) = V̄ho−out
u
∂t

• ln − out send a buy market order and immediately behave as their new type: they send a sell limit
order and become lo − Au
u
u ∂Vln−out u u u
(r + ρ+ + λ)Vln−out (t) = + ρ+ (Vho−out (t) − A0 ) + λ(Vlo−A 0
u (t) − A ), Vln−out (T u ) = V̄ln−out
u
∂t

• hn − out send a buy market order and immediately behave as their new type
u
∂Vhn−out
(r + ρ− + λ)Vln−out
u
(t) = + ρ− (Vlo−A
u 0 u 0
u (t) − A ) + λ(Vho−out (t) − A )
∂t
with (r + ρ− + λ)Vhn−out
u
(T u ) = (r + ρ− + λ)V̄hn−out
u
= ρ− V̄ln−out
u u
+ λV̄ln−B u because as soon as this
u u
type contact the market being hn − B is optimal after T .
• lo − A0 cancel their limit order or are executed
u
u u
∂Vlo−A0
u u u
(r +ρ+ +λ+kA0 (t))Vlo−A0 (t) = v −δ + +ρ+ Vho−out (t)+λVlo−A 0
u (t)+kA0 (t)(Vln−out (t)+A )
∂t
u u u 0
with Vlo−Au (T ) = V̄ln−out + A and the intensity rate for the execution of the limit order

(λ + ρ− )Luhn (t) + (λ + ρ+ )Luln (t) + D˙A

u
0
kA0 (t) = u (t) = −(ρ + λ) − u (t)
DA 0 DA 0

Let’s call
r + ρ− −ρ−
 u
−v u
    
0 0 Vho−out
 −ρ+ r + ρ+  u u u
0 0 ~ =  Vlo−A ~ =  −(v − δ) 
 
M =  −ρ+
, V u
 and K 
−λ r + ρ+ + λ 0   Vln−out   (ρ + λ)A0 
+

−λ −ρ− 0 r + ρ− + λ u
Vhn−out (ρ− + λ)Au

Then the previous system of differential equations can be written as

∂ ~ ~ +K ~
V =M ×V
∂t
u
plus the non-homogeneous differential equation that rules Vlo−A u in which the source term is a combination

of the other value functions.

134
 The matrix M ’s eigenvectors and associated eigenvalues are
1
               
1 0 0
 ρ+
~ 4 = − ρ+
1   0 0 ρ− 
~    ~ +  ~ − + −
      
E
 1 = 1 , r , E2 = 1 , r + ρ + λ , E3 = 0 , r + ρ + λ ,
E
− −  , r + ρ + ρ 
    

ρ
1 0 1 1
and the solution of the previous vectorial ODE is
~ = M −1 × K
V ~ + C1 E
~ 1 ert + C2 E ~ 3 e(r+ρ− +λ)t + C4 E
~ 2 e(r+ρ+ +λ)t + C3 E ~ 4 e(r+ρ+ +ρ− )t

with the constants Ci ’s determined by the conditions at time T u .

Now let’s check the possible deviations.

type ln. During the transition phase the only way a type ln can deviate from the conjecture strategy is
u
to stay out until the next contacting time, and then the get the value Vln−out (t), rather than sending a buy
u u u
0
market order and get the value Vlo−Au (t) − A . If we call X(t) = Vlo−Au (t) − A0 − Vln−out (t) we obtain the
ODE
∂X
(r + ρ+ + λ)X(t) = + v u − δ − rA0
∂t
the solution of this equation is of the type
+ v u − δ − rA0
X(t) = C × e(r+ρ +λ)t
+
r + ρ+ + λ
Given that for t > T u we have V̄lo−A u u
u = V̄ln−out + B
u
then X(T u ) = B u − A0 > 0. We also have
v − δ − rA > 0. Then if C > 0 X(t) is positive for all t, if C < 0 X(t) is decreasing and positive in T u and
u 0

then positive over [0, T u ]. The deviation is not profitable.

u u
type hn. We are in the same case than the type ln. let’s call X(t) = Vho−out (t) − A0 − Vhn−out (t) and
the corresponding ODE is
∂X
(r + ρ− + λ)X(t) = + v u − rA0
∂t
with v u − rA0 > 0 and X(T u ) = V̄ho−out
u u
− A0 − V̄hn−out u
> V̄ho−out u
− A0 − V̄hn−B u 0
u = A − A > 0. We get

the same result.

type ho.
u
• Instead of staying out a type ho could send a sell market order at price B 0 and get Vhn−out (t) + B 0 .
This clearly not profitable given what have been said for type hn
• Another deviation could be to send a limit order at A0 . The corresponding value function would be
define by
∂V
(r + ρ− + λ + kA0 (t))V (t) = v u + + ρ− Vlo−A
u u u 0
u (t) + λVho−out (t) + kA0 (t)(Vhn−out (t) + A )
∂t
u
∂V ∂Vho−out
= + (r + ρ− + λ + kA0 (t))Vho−out
u
(t) − u
− kA0 (t)(Vho−out u
(t) − Vhn−out (t) − A0
∂t ∂t
u
Calling X(t) = Vho−out − V we obtain the ODE
∂X
(r + ρ− + λ + kA0 (t))X(t) = u
+ kA0 (t)(Vho−out u
(t) − Vhn−out (t) − A0 )
∂t
The solution of this ODE is of the type
Rt Z t Rs
(r+ρ− +λ+kA0 (s))ds u u − (r+ρ− +λ+kA0 (l))dl
X(t) = e 0 [C − kA0 (s)(Vho−out (s) − Vhn−out (s) − A0 )e 0 ds]
0

135
 Rt
(r+ρ− +λ+k (s))ds
Then X(t) × e 0 A0
is decreasing and X(T u ) = V̄ho−out
u u
− A0 − V̄hn−out > Au − A0 > 0
u
. X(t) is positive over [0, T ]. This deviation is not profitable.
• A type ho could send a limit order at any price A0 < A < Au The corresponding value function would
be define by
∂V
(r + ρ− + λ)V (t) = v u + + ρ− Vlo−A
u u
u (t) + λVho−out (t)
∂t
u
∂V ∂Vho−out
= + (r + ρ− + λ)Vho−out
u
(t) −
∂t ∂t
that would give
−
u
Vho−out (t) − V (t) = C × e(r+ρ +λ)t

and this order would be executed at T u , then Vho−out

u
(T u ) − V (T u ) = V̄ho−out
u u
− A − V̄hn−out >
u
A − A > 0. The deviation is not profitable.
• Lastly a type ho could send a limit order at price Au . The corresponding value function would be
define by
∂V
(r + ρ− + λ)V (t) = v u + + ρ− Vlo−A
u u
u (t) + λVho−out (t)
∂t
u
∂V ∂Vho−out
= + (r + ρ− + λ)Vho−out
u
(t) −
∂t ∂t
that would give again
−
u
Vho−out (t) − V (t) = C × e(r+ρ +λ)t

And at T u , Vho−out
u
(T u ) − V (T u ) = V̄ho−out
u u
− V̄ho−Au > 0.

type lo.
• Instead of staying at Au a type lo could send a sell market order at price B 0 and get Vln−out
u
(t) + B 0 .
This clearly not profitable given what have been said for type ln
u
• Another deviation could be to send a limit order at A0 . The corresponding value function is Vlo−A0 (t)
u u
Calling X(t) = Vlo−Au (t) − Vlo−A0 (t) we obtain the ODE

∂X u u
(r + ρ+ + λ + kA0 (t))X(t) = + kA0 (t)(Vlo−A 0
u (t) − Vln−out (t) − A )
∂t
The solution of this ODE is as before
Rt Z t Rs
(r+ρ+ +λ+kA0 (s))ds u u
+
0 − 0 (r+ρ +λ+kA0 (l))dl
X(t) = e 0 [C − kA0 (s)(Vlo−Au (t) − Vln−out (t) − A )e ds]
0
Rt −
(r+ρ +λ+k (s))ds
Then X(t) × e 0 A0
is decreasing and X(T u ) = V̄lo−A
u 0 u
u − A − V̄ln−out = B
u
− A0 > 0 .
u
X(t) is positive over [0, T ]. This deviation is not profitable.
• A type lo could send a limit order at any price A0 < A < Au The corresponding value function would
be define by
∂V
(r + ρ+ + λ)V (t) = v u − δ + u
+ ρ+ Vho−out u
(t) + λVlo−Au (t)
∂t
u
∂V u ∂Vlo−A u
= + (r + ρ+ + λ)Vlo−A u (t) −
∂t ∂t
that would give
+
u (r+ρ +λ)t
Vlo−Au (t) − V (t) = C × e

and this order would be executed at T u , then Vlo−A

u u u u u u
u (T )−V (T ) = V̄lo−Au −A− V̄ln−out = B −A > 0.

The deviation is not profitable.

136
 • Lastly a type lo could stay out. The corresponding value function would be define by
∂V
(r + ρ+ + λ)V (t) = v u − δ + u
+ ρ+ Vho−out u
(t) + λVlo−Au (t)
∂t
u
∂V u ∂Vlo−Au
= + (r + ρ+ + λ)Vlo−A u (t) −
∂t ∂t
that would give again
+
u (r+ρ +λ)t
Vlo−Au (t) − V (t) = C × e

And at T u , Vlo−A
u u u u u
u (T ) − V (T ) = V̄lo−Au − V̄lo−out > 0.

Proof of Proposition A.10

During the transition phase the observed types under the conjecture strategy are hn − B d , hn − B 0 , ln − out,
ho − out, lo − A0 and lo − Au . But lo − A0 and lo − Au can be gathered under the label lo − out because
their limit orders are not executed under the conjecture strategy so they would get the same outcome if they
were out of the order book. In this framework we can define the system of ODE’s defining the value function
of these types:

• hn − B d stay in the limit order book until they switch of type

d
∂Vhn−B d
(r + ρ− )Vhn−B
d
d (t) = + ρ− Vln−out
d
(t)
∂t
d d d d
with Vhn−B d (T ) = V̄hn−B d which is the value for a hn of having a limit order at price B once the
last dynamic equilibrium is played (the optimal strategy in the last phase).
• ln − out stay out until they switch of type
d
d ∂Vln−out d d
(r + ρ+ )Vln−out (t) = + ρ+ Vhn−B d (t), Vln−out (T d ) = V̄ln−out
d
∂t

• ho − out send a sell market order and immediately behave as their new type: they send a buy limit
order and become hn − B d
d
∂Vho−out
(r+ρ− +λ)Vho−out
d
(t) = v d + +ρ− (Vln−out
d d
(t)+B 0 )+λ(Vhn−B 0 d d d
d (t)+B ), Vho−out (T ) = V̄ho−out
∂t

• lo − out send a buy market order and immediately behave as their new type
d
d ∂Vlo−out
(r + ρ+ + λ)Vlo−out (t) = v d − δ + d
+ ρ+ (Vhn−B 0 d 0
d (t) + B ) + λ(Vln−out (t) + B )
∂t
d
with (r + ρ+ + λ)Vlo−out (T d ) = (r + ρ+ + λ)V̄lo−out
d
= v d − δ + ρ+ V̄ho−out
d d
+ λV̄lo−Ad because as soon
d d
as this type contact the market being lo − A is optimal after T .
• hn − B 0 cancel their limit order or are executed
d
∂Vhn−B 0
(r + ρ− + λ + kB 0 (t))Vhn−B
d
0 (t) = + ρ− Vln−out
d d
(t) + λVhn−B d 0
d (t) + kB 0 (t)(Vho−out (t) − B )
∂t
d d u 0
with Vhn−B 0 (T ) = V̄ho−out − B and the intensity rate for the execution of the limit order

(λ + ρ− )Ldho (t) + (λ + ρ+ )Lulo (t) − D˙B

d
0
kB 0 (t) = d (t)
= −(ρ + λ) − d (t)
DB 0 DB 0

Possible deviations:

137
type ho. During the transition phase the only way a type ho can deviate from the conjecture strategy is
d
to stay out until the next contacting time, and then the get the value Vho−out (t), rather than sending a sell
d 0 d d
market order and get the value Vhn−B d (t) + B . If we call X(t) = V hn−B d (t) + B 0 − Vho−out (t) we obtain
the ODE
∂X
(r + ρ− + λ)X(t) = + rB 0 − v d
∂t
the solution of this equation is of the type
+ v u − δ − rA0
X(t) = C × e(r+ρ +λ)t
+
r + ρ+ + λ
Given that for t > T d we have V̄hn−B
d d d d 0 d
d = V̄ho−out −A then X(T ) = B −A > 0. We also have rB −v
0 d
> 0.
d
Then if C > 0 X(t) is positive for all t, if C < 0 X(t) is decreasing and positive in T and then positive over
[0, T d ]. The deviation is not profitable.

d d
type lo. We are in the same case than the type ho. let’s call X(t) = Vln−out (t) + B 0 − Vlo−out (t) and the
corresponding ODE is
∂X
(r + ρ+ + λ)X(t) = + rB 0 − (v d − δ)
∂t
with rB 0 − (v d − δ) > 0 and X(T d ) = V̄ln−out
d d
+ B 0 − V̄lo−out d
> V̄ln−out u
+ B 0 − V̄lo−A 0
d = B − B
d
> 0. We
get the same result.

type ln.
d
• Instead of staying out a type ln could send a sell market order at price A0 and get Vlo−out (t) − A0 .
This clearly not profitable given what have been shown for type lo
• Another deviation could be to send a limit order at B 0 . The corresponding value function would be
define by
∂V d d d
(r + ρ+ + λ + kB 0 (t))V (t) = + ρ+ Vhn−B 0
d (t) + λVln−out (t) + kB 0 (t)(Vlo−out (t) − B )
∂t
d
∂V d ∂Vln−out d d
= + (r + ρ+ + λ + kB 0 (t))Vln−out (t) − − kB 0 (t)(Vln−out (t) − Vlo−out (t) + B 0 )
∂t ∂t
d
Calling X(t) = Vln−out − V we obtain the ODE
∂X d d
(r + ρ+ + λ + kB 0 (t))X(t) = + kB 0 (t)(Vln−out (t) − Vlo−out (t) + B 0 )
∂t
The solution of this ODE is of the type
Rt Z t Rs
(r+ρ+ +λ+kB 0 (s))ds d d − (r+ρ+ +λ+kB 0 (l))dl
X(t) = e 0 [C − kB 0 (s)(Vln−out (t) − Vlo−out (t) + B 0 )e 0 ds]
0
Rt +
(r+ρ +λ+k (s))ds
Then X(t) × e 0 B0
is decreasing and X(T d ) = V̄ln−out
d d
+ B 0 − V̄lo−out > B0 − Bd > 0
. X(t) is positive over [0, T d ]. This deviation is not profitable.
• A type ln could send a limit order at any price B 0 > B > B d The corresponding value function would
be define by
∂V d d
(r + ρ+ + λ)V (t) = + ρ+ Vhn−B d (t) + λVln−out (t)
∂t
d
∂V d ∂Vln−out
= + (r + ρ+ + λ)Vln−out (t) −
∂t ∂t
that gives
+
d
Vln−out (t) − V (t) = C × e(r+ρ +λ)t

and this order would be executed at T d , then Vln−out

d
(T d )−V (T d ) = V̄ln−out
d d
+B−V̄lo−out > B−B d > 0.
The deviation is not profitable.

138
 • Lastly a type ln could send a limit order at price B d . The corresponding value function would be
define by
∂V d d
(r + ρ+ + λ)V (t) = + ρ+ Vhn−B d (t) + λVln−out (t)
∂t
d
∂V d ∂Vln−out
= + (r + ρ+ + λ)Vln−out (t) −
∂t ∂t
that would give again
+
d
Vln−out (t) − V (t) = C × e(r+ρ +λ)t

And at T d , Vln−out
d
(T d ) − V (T d ) = V̄ln−out
d d
− V̄ln−Ad > 0.

type hn.
• Instead of staying at B d a type hn could send a buy market order at price A0 and get Vho−out
d
(t) − A0 .
This clearly not profitable given what have been said for type ho
d
• Another deviation could be to send a limit order at B 0 . The corresponding value function is Vhn−B 0 (t)
d d
Calling X(t) = Vhn−B d (t) − Vhn−B 0 (t) we obtain the ODE

∂X
(r + ρ− + λ + kB 0 (t))X(t) = d
+ kB 0 (t)(Vhn−B d 0
d (t) − Vho−out (t) + B )
∂t
The solution of this ODE is as before
Rt Z t Rs
(r+ρ− +λ+kB 0 (s))ds d d
−
0 − 0 (r+ρ +λ+kB 0 (l))dl
X(t) = e 0 [C − kB 0 (s)(Vhn−B d (t) − Vho−out (t) + B )e ds]
0
Rt
(r+ρ− +λ+k (s))ds
Then X(t) × e 0 B0
is decreasing and X(T d ) = V̄hn−B
d 0 d 0 d
d + B − V̄ho−out = B − A > 0

. X(t) is positive over [0, T d ]. This deviation is not profitable.

• A type hn could send a limit order at any price B 0 > B > B d The corresponding value function would
be define by
∂V
(r + ρ− + λ)V (t) = δ + + ρ− Vln−out
d d
(t) + λVhn−B d (t)
∂t
d
∂V ∂Vhn−B d
= + (r + ρ− + λ)Vhn−B
d
d (t) −
∂t ∂t
that would give
−
d (r+ρ +λ)t
Vhn−B d (t) − V (t) = C × e

and this order would be executed at T d , then Vhn−B

d d d d d d
d (T )−V (T ) = V̄hn−B d +B − V̄ho−out = B −A >

0. The deviation is not profitable.

• Lastly a type hn could stay out. The corresponding value function would be define by
∂V
(r + ρ− + λ)V (t) = + + ρ− Vln−out
d d
(t) + λVhn−B d (t)
∂t
d
∂V ∂Vhn−B d
= + (r + ρ− + λ)Vhn−B
d
d (t) −
∂t ∂t
that would give again
−
d (r+ρ +λ)t
Vhn−B d (t) − V (t) = C × e

And at T d , Vhn−B
d d d d d
d (T ) − V (T ) = V̄hn−B d − V̄hn−out > 0.

139
A.5 Equilibrium in the perfectly symmetric case
In order to obtain an equilibrium strategy for the entire model I am focusing on the perfectly symmetric case.
The conditions for this equilibrium to hold in a more general case would probably be that the parameter’s
values make the model close enough to the perfectly symmetric case. This would generates model outcomes
similar to the ones in the perfectly symmetric case.

The perfectly symmetric setup is defined as follows

• for investor types: ρ+ = ρ− = ρ, s = 1
2
1
• for the asset value and dynamic: v u = v 0 + ω, v d = v 0 − ω, p = 2

• for the targeted prices: B ∗ = 1r (v ∗ − 2δ ) − ∆

2, A∗ = 1r (v ∗ − 2δ ) + ∆
2

and in term of magnitude we assume that

ω δ
>> >> ∆
r r
which means that gains from arbitraging are higher than gain from trading for liquidity reason and the last
one being higher that the implicit cost of trading, the bid-ask spread.

∅ d u
In this case the dynamics of the order book in the transition phase is given by DB 0 = DB 0 (0) = DA0 (0) =
∅ ∅
DA 0 = α L,
1 ∅ 1
∀t DBd u
0 (t) = DA0 (t) = D(t) = − L + [α + ]Le−(ρ+λ)t
2 2
and the transition phases last the same time in the states u or d
1
Tu = Td = T = ln(1 + 2α∅ ).
ρ+λ

A.5.1 equilibrium conjecture

I consider equilibria where the limit order market dynamic is in a steady-state for t < τ and converges to
another steady state for t > τ . The idea is to conjecture (and to solve) the class of equilibria by backward
induction:
• For S = ∅ (before τ ): with a model setup sufficiently symmetric, whatever the value of µ, this is a
steady state equilibrium over a pair of prices (A0 , B 0 ) where ho and ln stay out of the market, lo and
hn indifferently send limit or market orders.
• For T + τ > t > τ this is the transition phase
• For t > T + τ if S = u we are playing the equilibrium dynamic converging to a steady state over
the bid-ask prices (Au , B u ) (as described in the previous section), and if S = d we are playing the
equilibrium dynamic converging to a steady state over the bid-ask prices (Ad , B d ).

A.5.2 equilibrium outcome

Proposition A.11. In the perfectly symmetric case, whatever the value of µ, there is a unique steady state
equilibrium defined by the pair of prices (A0 , B 0 ) where lo and hn indifferently send limit or market orders
and where ho and ln stay out of the market. The rate at which limit orders are executed are the same for
∅ ∅ ∅
sell and buy orders lA 0 = lB 0 = l . And the equilibrium populations are characterized by the value

∅ ρ
αeq =
4(ρ + l∅ )
∅
Moreover limµ→∞ αeq = 0.

140
Proposition A.12. For a value of µ high enough, an increase of the monitoring rate λ has a positive impact
∅
on αeq .
∅
∂αeq
>0
∂λ
∅
An increase of the fundamental volatility, µ or ω has a negative impact on αeq

∅ ∅
∂αeq ∂αeq
< 0, <0
∂µ ∂ω

A.5.3 Proof of Proposition A.11

Value functions in the game stage prior to the utility flow change
In the state S = ∅ when the limit order book is in the steady state, the value functions corresponding to the
types in the conjecture equilibria are defined by
∅ ∅
for the type ho, Vho = Vho−out with

(r + ρ− + µ)Vho−out
∅
= v 0 + ρ− Vlo∅ + µpVho−out
u d
(0) + µ(1 − p)Vho−out (0)

for the type lo, value function of sending limit orders at Au and A0 are
∅ ∅ + ∅ ∅ ∅ u d
(r + ρ+ + lA 0 0
0 + µ)Vlo−A0 = v − δ + ρ Vho + lA0 (Vln + A ) + µpVlo−A0 (0) + µ(1 − p)Vlo−A0 (0)

∅ + ∅ u d
(r + ρ+ + µ)Vlo−A 0
u = v − δ + ρ Vho + µpVlo−Au (0) + µ(1 − p)Vlo−Au (0)

If lo types only send limit orders at Au , Vlo∅ = Vlo−A

∅
u . If lo types only send indifferently market orders
0 ∅ ∅ ∅ 0
and limit orders at A , Vlo = Vlo−A0 = Vln + B . And if they are indifferent between the three actions
Vlo∅ = Vlo−A
∅ ∅ ∅
u = Vlo−A0 = Vln + B .
0

∅ ∅
for the type ln, Vln = Vln−out with

∅ ∅ u d
(r + ρ+ + µ)Vln−out = ρ+ Vhn + µpVln−out (0) + µ(1 − p)Vln−out (0)

for the type hn, value function of sending limit orders at B d and B 0 are

(r + ρ− + lB
∅ ∅ − ∅ ∅ ∅ 0 u d
0 + µ)Vhn−B 0 = ρ Vln + lB 0 (Vho − B ) + µpVhn−B 0 (0) + µ(1 − p)Vhn−B 0 (0)

(r + ρ− + µ)Vhn−B
∅ − ∅ u d
d = ρ Vln + µpVhn−B d (0) + µ(1 − p)Vhn−B d (0)

∅ ∅
If hn types only send limit orders at B d , Vhn = Vhn−B d . If hn types only send indifferently market
∅ ∅ ∅
orders and limit orders at B , Vhn = Vhn−B 0 = Vho − A0 . And if they are indifferent between the
0

three actions Vlo∅ = Vhn−B

∅ ∅ ∅ 0
d = Vhn−B 0 = Vho − A .

In this case the strategy conjectured at equilibrium is such that lo and hn are indifferent
between sending market orders and limit orders at A0 and B 0 and do not send limit orders at
Au or B d

To have indifference for lo and hn types the rates of limit orders execution must be equal to

∅ 1 0
lB 0 = [v − rA0 − ρ− ∆ + µp(Vho−out
u u
(0) − A0 − Vhn−B d 0 d
0 (0)) + µ(1 − p)(Vho−out (0) − A − Vhn−B 0 (0))]
∆
∅ 1 u u d d
lA 0 = [rB 0 − ρ+ ∆ − (v 0 − δ) + µp(Vln−out (0) + B 0 − Vlo−A 0
0 (0)) + µ(1 − p)(Vln−out (0) + B − Vlo−A0 (0))]
∆

141
 To show the existence of an equilibrium we need to show that the conjecture strategy is optimal and
to solve for the equilibrium (steady-state) populations in the state S = ∅ which is equivalent to prove the
existence of an acceptable α∅ such that:
+
∅ ρ
∅
ρ− s − lB 0 ( ρ+ +ρ− − s)
α = ∅ ∅
ρ + + ρ − + lA 0 + lB 0

We must also verify that the rate at which market orders are send m∅A0 and m∅B 0 are indeed between 0 and
1.

general case
type lo.
• A type lo can deviate by sending a limit order at price Au > A > A0 and gets the value
∅
(r + ρ+ + λ + µ)V = v 0 − δ + ρ+ Vho + λVlo∅ + µpVlo−A
u d
(0) + µ(1 − p)Vlo−out (0)
u u
because Vlo−A (0) < Vlo−Au (0) this deviation is less profitable than the one shot deviation of sending
u
a limit order at A .
• a type lo can also deviate by staying out. For the same reason this is less profitable than the one shot
deviation of sending a limit order at Au
• when lo sends a limit order at Au for a one shot deviation the induced value function is defined by
∅
(r + ρ+ + λ + µ)V = v 0 − δ + ρ+ Vho + λVlo∅ + µpVlo−A
u d
u (0) + µ(1 − p)Vlo−out (0)

and we have

(r + ρ+ + λ + µ)(Vlo∅ − V ) = lA
∅ u u
0 ∆ + µp(Vlo−A0 (0) − Vlo−Au (0)) = (r + ρ
+
+ µ)(Vlo∅ − Vlo−A
∅
u)

u u ∅
Vlo−A0 − (0)Vlo−Au (0) < 0 then as soon as the strategy conjectured for the type lo is optimal then lA0 ∆ +
u u ∅
µp(Vlo−A0 (0) − Vlo−Au (0)) > 0 and then lA0 > 0 which a necessary condition for the equilibrium. Now we
have to get the necessary condition on µ to make it hold and then study the function

(r + ρ+ + µ)(Vlo∅ − Vlo−A
∅ 0 + 0
u ) = rB − ρ ∆ − (v − δ)

u u d d
+ µp(Vln−out (0) + B 0 − Vlo−A 0
u (0)) + µ(1 − p)(Vln−out (0) + B − Vlo−A0 (0))

type hn.
• A type hn can deviate by sending a limit order at price B d < B < B 0 and gets the value

(r + ρ− + λ + µ)V = ρ− Vln
∅ ∅
+ λVhn u
+ µpVhn−out d
(0) + µ(1 − p)Vhn−B (0)
d d
because Vhn−B (0) < Vhn−B d (0) this deviation is less profitable than the one shot deviation of sending
d
a limit order at B .
• a type hn can also deviate by staying out. For the same reason this is less profitable than the one shot
deviation of sending a limit order at B d
• when lo sends a limit order at B d for a one shot deviation the induced value function is defined by

(r + ρ− + λ + µ)V = ρ− Vln
∅ ∅
+ λVhn d
+ µpVhn−out d
(0) + µ(1 − p)Vhn−B d (0)

and we have

(r + ρ− + λ + µ)(Vhn
∅ ∅
− V ) = lB d d
0 ∆ + µ(1 − p)(Vhn−B 0 (0) − Vhn−B d (0)) = (r + ρ
− ∅
+ µ)(Vhn ∅
− Vhn−B d)

142
 d d
Vhn−B 0 − (0)Vhn−B d (0) < 0 then as soon as the strategy conjectured for the type hn is optimal then
∅ d d ∅
lB 0 ∆ + µ(1 − p)(Vhn−B 0 (0) − Vhn−B d (0)) > 0 and then lB 0 > 0. Now we have to get the necessary condition

on µ to make it hold and then study the function

(r + ρ− + µ)(Vhn
∅ ∅
− Vhn−B 0 0 −
d ) = v − rA − ρ ∆

u u d d
+ µp(Vho−out (0) − A0 − Vhn−B 0
0 (0)) + µ(1 − p)(Vho−out (0) − A − Vhn−B d (0))

type ho.
• A type ho can deviate by sending a market order at B 0 . This is clearly not profitable because this is
optimal for a type hn to send a market order at A0 .
• A type ho can deviate by sending a limit order at a price A > A0 and gets the value defined by

(r + ρ− + λ + µ)V = v 0 + ρ− Vlo∅ + λVho

∅ u
+ µpVho−A d
(0) + µ(1 − p)Vho−out (0)

u u ∅
because Vho−A (0) < Vho−out (0) (by optimality in the transition phase) we obtain V < Vho .
• the last deviation possible for a type ho is to send a limit order at A0 and gets the value defined by

(r + ρ− + λ + lA
∅ 0 − ∅ ∅ ∅ ∅ 0 u d
0 + µ)V = v + ρ Vlo + λVho + lA0 (Vhn−out + A ) + µpVho−A0 (0) + µ(1 − p)Vho−out (0)

u u ∅ 0 ∅ 0 ∅ ∅
because Vho−A0 (0) < Vho−out (0) and Vhn−out + A < Vhn + A = Vho we obtain V < Vho .

type ln.
• A type ln can deviate by sending a market order at A0 . This is clearly not profitable because this is
optimal for a type lo to send a market order at B 0 .
• A type ln can deviate by sending a limit order at a price B < B 0 and gets the value defined by
∅ ∅ u d
(r + ρ+ + λ + µ)V = ρ+ Vhn + λVln + µpVln−out (0) + µ(1 − p)Vln−B (0)

d d ∅
because Vln−B (0) < Vln−out (0) we obtain V < Vln .
• the last deviation possible for a type ln is to send a limit order at B 0 and gets the value defined by
∅ + ∅ ∅ ∅ ∅ u d
(r + ρ+ + λ + lB 0
0 + µ)V = ρ Vhn + λVln + lB 0 (Vlo−out − B ) + µpVln−out (0) + µ(1 − p)Vln−B 0 (0)

∅
d
because Vln−B d
0 (0) < Vln−out (0) and Vlo−out − B
0
< Vlo∅ − B 0 = Vln
∅ ∅
we obtain V < Vln .
Now let’s show first the 4 following formulas:
Z Tu u −rs
u 0 u u 0 DA 0 (s)e
Vln−out (t) +A − Vlo−A0 (t) = −[v − δ − rA ] u (t)e−rt ds
t DA 0

d d rB 0 − (v d − δ) (λ + ρ+ )(B 0 − B d ) − ρ+ ∆ −(r+ρ+ +λ)T d +

Vln−out (t) + B 0 − Vlo−A0 (t) =
+
+ +
e × e(r+ρ +λ)t
r+ρ +λ r+ρ +λ
Z Td d −rs
d d d DB 0 (s)e
Vho−out (t) − B 0 − Vhn−B 0
0 (t) = −[rB − v ]
d (t)e−rt
ds
t DB 0

u u v u − rA0 (λ + ρ− )(Au − A0 ) − ρ− ∆ −(r+ρ− +λ)T u −

Vho−out (t) − A0 − Vhn−B 0 (t) =
−
+ −
e × e(r+ρ +λ)t
r+ρ +λ r+ρ +λ
u u
Let’s call X1 (t) = Vln−out (t) + A0 − Vlo−A0 (t)

u
ḊA 0 ∂X1 u 0
(r − u (t) )X1 (t) = ∂t − [v − δ − rA ]
DA 0

143
 D u (0)
The solution of the homogeneous ODE is t 7→ c DAu0 (t) ert . Then we look for a solution of the type X1 (t) =
A0
D u (0)
c(t) DAu0 (t) ert and we obtain
A0

Z t u
DA 0 (s) −rs
c(t) = c0 + [v u − δ − rA0 ] u e ds
0 DA0 (0)

and because X1 (T u ) = 0 we must have

Z Tu u
u 0 DA 0 (s) −rs
c0 = −[v − δ − rA ] u e ds
0 DA0 (0)

and we obtain
Z Tu u −rs
u 0 DA 0 (s)e
X1 (t) = −[v − δ − rA ] u ds
t DA0 (t)e−rt

We also know that

d
ḊB 0 u u ∂
(r − )(Vho−out (t) − B 0 − Vhn−B 0 (t)) = (V u u
(t) − B 0 − Vhn−B 0 d
0 (t)) − [rA − v ]
d
DB 0 (t) ∂t ho−out

which leads to the fourth equation of the lemma.

d d
Let’s call X2 (t) = Vln−out (t) − Vlo−A0 (t)

∂X2
(r + ρ+ + λ)X2 (t) = − (λ + ρ+ )B 0 − (v d − δ)
∂t

and (r + ρ+ + λ)X 2 (T d ) = ρ+ ((V̄hn−B

d d d d d + d d d
d − V̄ho ) + λ(V̄ln−out − V̄lo−Ad ) − (v − δ) = −ρ A − λB − (v − δ).

This can be solved easily.

And finally

∂
(r + ρ− + λ)(Vho−out
u d
(t) − Vhn−B 0 (t)) = (V u d
(t) − Vhn−B − 0
0 (t)) + (λ + ρ )A + v
u
∂t ho−out

with (r + ρ− + λ)(V̄ho−out
u d
− V̄hn−B 0) = v
u
+ ρ− B u + λAu

Symmetric parametrization
ρ− = ρ+ = ρ, s = 1/2
In this case the conditions on m∅A0 and m∅B 0 are verified.
We can now calculate
Z T
11 1 1
D(t)e−rt = L(e−rT − 1) + [α∅ + ]L(1 − e−(r+ρ+λ)T )
0 r2 r+ρ+λ 2
11 −rT 1 ∅ 1 1 1 −rT
= L(e − 1) + [α + ]L − Le
r2 r+ρ+λ 2 r+ρ+λ2
1 1 ρ+λ 1 ρ+λ 1
= α∅ L − L+ L r
r+ρ+λ 2 r(r + ρ + λ) 2 r(r + ρ + λ) (1 + 2α∅ ) ρ+λ

144
 Then we get

∅ ∅ δ
lB 0 + lA0 + 2ρ = −r
∆   
1 u 0 1 1 ρ+λ 1 ρ+λ 1
+ µp −∆ − (v − δ − rA ) − + r
∆ r + ρ + λ 2α∅ r(r + ρ + λ) 2α∅ r(r + ρ + λ) (1 + 2α∅ ) ρ+λ
" #
1 rB 0 − (v d − δ) (λ + ρ)(B 0 − B d ) − ρ∆ 1
+ µ(1 − p) + r+ρ+λ
∆ r+ρ+λ r+ρ+λ (1 + 2α∅ ) ρ+λ
" #
1 v u − rA0 (λ + ρ)(Ad − A0 ) − ρ∆ 1
+ µp + r+ρ+λ
∆ r+ρ+λ r+ρ+λ (1 + 2α∅ ) ρ+λ
  
1 0 d 1 1 ρ+λ 1 ρ+λ 1
+ µ(1 − p) −∆ − (rB − v ) − + r
∆ r + ρ + λ 2α∅ r(r + ρ + λ) 2α∅ r(r + ρ + λ) (1 + 2α∅ ) ρ+λ

Perfectly symmetric parametrization

In the perfectly symmetric case we have
u u d d
Vln−out (t) + A0 − Vlo−A 0
0 (t) = Vho−out (t) − B − Vhn−B 0 (t)
Z T
δ ∆ D(s)e−rs
= −[ω − + r ] ds
2 2 t D(t)e−rt
d d u u
Vln−out (t) + B 0 − Vlo−A 0
0 (t) = Vho−out (t) − A − Vhn−B 0 (t)

ω + 2δ − r ∆ (λ + ρ) ωr − ρ∆ −(r+ρ+λ)T
= 2
+ e × e(r+ρ+λ)t
r+ρ+λ r+ρ+λ

∅ ∅
which implies that lA = lB = l∅ and that

δ
2l∅ + 2ρ = −r
∆   
1 δ ∆ 1 1 ρ+λ 1 ρ+λ 1
+ µ −∆ − (ω − + r ) − + r
∆ 2 2 r + ρ + λ 2α∅ r(r + ρ + λ) 2α∅ r(r + ρ + λ) (1 + 2α∅ ) ρ+λ
" #
1 ω + 2δ − r ∆
2 (λ + ρ) ωr − ρ∆ 1
+ µ +
∆ r+ρ+λ r + ρ + λ (1 + 2α∅ ) r+ρ+λ
ρ+λ

To prove the existence of the equilibrium we have to show that there is an α∅ solution to the equation
ρ
G(α∅ ) = α∅ × (2l∅ + 2ρ) =
2
such that
∅ ∅
(r + ρ + µ)(Vhn − Vhn−B d) > 0

(r + ρ + µ)(Vlo∅ − Vlo−A
∅
u) > 0

and l∅ > 0. We can notice that

∅ ∅ ∅ µ d d ∅
(r + ρ + µ)(Vhn − Vhn−B d ) = lB ∆ − (V d (0) − Vhn−B 0 (0)) < lB ∆
2 hn−B
∅ ∅ ∅ ∅ ∅ ∅
so as soon as (r + ρ + µ)(Vhn − Vhn−B d ) > 0, lB > 0. As well for (r + ρ + µ)(Vlo − Vlo−Au ) and lA .

145
Analysis of function G. Let’s prove that G(α) − ρ2 has a unique zero on [0, 1/4].

µ ρ + λ ω − 2δ + r ∆
  
∅ δ µ δ − r∆
G(α ) = 2
+ α∅ × −r+ −∆ +
∆ r 2(r + ρ + λ) ∆ ∆ r+ρ+λ
µ ρ + λ ω − 2δ + r ∆
2 1 µ (λ + ρ) ωr − ρ∆ α∅
− r + r
∆ r 2(r + ρ + λ) (1 + 2α∅ ) ρ+λ ∆ r + ρ + λ (1 + 2α∅ ) ρ+λ +1

The second derivative of G is the second derivative of

µ ρ + λ ω − 2δ + r ∆ 1 µ (λ + ρ) ωr − ρ∆ α∅
f (α∅ ) = − 2
r + r
∆ r 2(r + ρ + λ) (1 + 2α∅ ) ρ+λ ∆ r + ρ + λ (1 + 2α∅ ) ρ+λ +1

that can be rewritten

µ ρ + λ ω − 2δ + r ∆ 1 µ (λ + ρ) ωr − ρ∆ 21 + α∅ − 12
f (α∅ ) = − 2
r + r
∆ r 2(r + ρ + λ) (1 + 2α∅ ) ρ+λ ∆ r + ρ + λ (1 + 2α∅ ) ρ+λ +1
ρ+λ
µ r δ−r∆
2 − ρ∆ 1 µ (λ + ρ) ωr − ρ∆ 1
= r − r
∆ 2(r + ρ + λ) (1 + 2α∅ ) ρ+λ ∆ 2(r + ρ + λ) (1 + 2α∅ ) ρ+λ +1

The derivatives of f are

   
∂f µ 1 r ρ + λ δ − r∆ 1
= −2 − ρ∆ r
∂α∅ ∆ 2(r + ρ + λ) ρ+λ r 2 (1 + 2α∅ ) ρ+λ +1
    
µ 1 r λ+ρ 1
+ 2 +1 ω − ρ∆ r
∆ 2(r + ρ + λ) ρ+λ r (1 + 2α∅ ) ρ+λ +2

∂2f
    
µ 1 r r ρ + λ δ − r∆ 1
= 4 +1 − ρ∆ r
∂(α∅ )2 ∆ 2(r + ρ + λ) ρ + λ ρ + λ r 2 (1 + 2α∅ ) ρ+λ +2
     
µ 1 r r λ+ρ 1
− 4 +1 +2 ω − ρ∆ r
∆ 2(r + ρ + λ) ρ+λ ρ+λ r (1 + 2α∅ ) ρ+λ +3

∂2f
The sign of ∂(α∅ )2
is the sign of
    
r ρ + λ δ − r∆ ∅ r λ+ρ
s(α) = − ρ∆ × (1 + 2α ) − +2 ω − ρ∆
ρ+λ r 2 ρ+λ r

δ−r∆
Since ω > 2 we have
    
λ+ρ r ∅ r
s(α) < ω − ρ∆ × × (1 + 2α ) − +2
r ρ+λ ρ+λ

And because at equilibrium 0 ≤ α∅ ≤ 1/4 < 1

   
λ+ρ r
s(α) < ω − ρ∆ × 2 −2 <0
r ρ+λ
2
∂ f ∂G
On [0, 1/4], ∂(α∅ )2 < 0 then on [0, 1/4] ∂α∅ is either always positive, always negative, or positive and then

negative. G can at most cross the ρ/2 horizontal line on [0, 1/4] twice. And if G(1/4) > ρ2 it means that it
is crossed only once.

146
 ρ 1
G(1/4) − = [δ − (r + 2ρ)∆]
2 4∆" #
µ ρ + λ ω − 2δ + r ∆ 2 ∆ δ − r∆
+ − +
∆ r 2(r + ρ + λ) 4 4(r + ρ + λ)
" #
ρ+λ δ−r∆
µ r 2 − ρ∆ 1 (λ + ρ) ωr − ρ∆ 1
+ r − r
∆ 2(r + ρ + λ) (3/2) ρ+λ 2(r + ρ + λ) (3/2) ρ+λ +1
1
First δ − (r + 2ρ)∆ > 0. Moreover 1 > r > 23 . Then
(3/2) ρ+λ

" #
ρ µ ρ + λ ω − 2δ + r ∆
2 (r + ρ + λ)∆ δ − r∆
G(1/4) − > − +
2 (r + ρ + λ)∆ r 2 4 4
" #
ρ+λ δ−r∆
µ r 2 − ρ∆ (λ + ρ) ωr − ρ∆
+ −
(r + ρ + λ)∆ 3 3

then

     
(r + ρ + λ)∆ ρ 1ρ+λ δ 1ρ+λ 1 1ρ+λ 1
(G(1/4) − ) > ω × + × − + + r∆ × − − >0
µ 2 6 r 2 6 r 2 6 r 2

Because G(0) = 0 there is unique α ∈ [0, 1/4] such that G(α∅ ) = ρ2 . Moreover ∂G ∅
∂α (αeq ) > 0.

Other equilibrium conditions. First, because we are in the perfectly symmetric case, we have
∅ ∅ ∅ ∅
(r + ρ + µ)(Vhn − Vhn−B d ) = (r + ρ + µ)(Vlo − Vlo−Au )

Indeed
∅ ∅ 0 0
(r + ρ + µ)(Vhn − Vhn−B d ) = v − rA − ρ∆

1 u u 1 d d
+ µ (Vho−out (0) − A0 − Vhn−B 0
0 (0)) + µ (Vho−out (0) − A − Vhn−B d (0))
2 2

(r + ρ + µ)(Vlo∅ − Vlo−A
∅ 0 0
u ) = rB − ρ∆ − (v − δ)

1 u u 1 d d
+ µ (Vln−out (0) + B 0 − Vlo−A 0
u (0)) + µ (Vln−out (0) + B − Vlo−A0 (0))
2 2
with v 0 − rA0 − ρ∆ = rB 0 − ρ∆ − (v 0 − δ) = δ−(r+2ρ)∆
2
d
> 0 and Vho−out d
(t) − A0 − Vhn−B u
d (t) = Vln−out (t) +
0 u
B − Vlo−Au (t) because these two expressions are solutions of the same ODE
dX
(r + ρ + λ)X = +c
dt
with c = v d − rA0 − (ρ + λ)∆ = rB 0 − (v u − δ) − (ρ + λ)∆ = −ω + 2δ − ( 2r + ρ + λ)∆ and the conditions at
t=T
d d u u ω
V̄ho−out − A0 − V̄hn−B 0
d = V̄ln−out + B − V̄lo−Au = −
r
then we have
d d u u
Vho−out (t) − A0 − Vhn−B 0
d (t) = Vln−out (t) + B − Vlo−Au (t)
" ! #
ω − 2δ + ( 2r + ρ + λ)∆ ω ω− δ
+ ( 2r + ρ + λ)∆
=− + − 2
e−(r+ρ+λ)(T −t)
r+ρ+λ r r+ρ+λ

147
and finally

X(α∅ ) ∅ ∅ ∅ ∅
= (r + ρ + µ)(Vhn − Vhn−B d ) = (r + ρ + µ)(Vlo − Vlo−Au )
2
δ − (r + 2ρ)∆
=
" 2 #
µ ω + 2δ − r ∆ 2 (λ + ρ) ωr − ρ∆ 1
+ +
2 r+ρ+λ r + ρ + λ (1 + 2α∅ ) r+ρ+λ
ρ+λ
" ! #
µ ω − 2δ + ( 2r + ρ + λ)∆ ω ω − 2δ + ( 2r + ρ + λ)∆ 1
− + − r+ρ+λ
2 r+ρ+λ r r+ρ+λ (1 + 2α∅ ) ρ+λ
That can be rewritten
δ − (r + ρ + λ)∆
X(α∅ ) = δ − (r + 2ρ)∆ + µ
r+ρ+λ
(r + 2λ)∆ − δ 1
+µ
2(r + ρ + λ) (1 + 2α∅ ) r+ρ+λ
ρ+λ

Given that (r + ρ + µ)X(α∅ ) < 2l∅ then if we know that for all t X(t) > 0 then lim∞ G(α∅ ) = ∞ and since
G(0) = 0 there is an equilibrium.

case where X(α∅ ) > 0 ∀α∅

• if δ > (r + 2λ)∆ > (r + ρ + λ)∆ then X(α∅ ) is increasing and
1
X(0) = (δ − (r + 2ρ)∆)(1 + µ )>0
2(r + ρ + λ)

• if (r + 2λ)∆ > δ > (r + ρ + λ)∆ then X(α∅ ) is decreasing and obviously X(α∅ ) > 0∀α∅
• if (r + 2λ)∆ > (r + ρ + λ)∆ > δ then X(α∅ ) is decreasing and

(r + ρ + λ)∆ − δ
X(∞) = δ − (r + 2ρ)∆ − µ
r+ρ+λ
δ−(r+2ρ)∆
which is positive iff µ < δ
∆− r+ρ+λ
.

case where X(α∅ ) can be negative

δ−(r+2ρ)∆
If (r + 2λ)∆ > (r + ρ + λ)∆ > δ and µ > δ
∆− r+ρ+λ
, X(α∅ ) is decreasing and has a negative limit. Then
∅
there is a unique a0 such that X(α ) > 0 for α∅ < a0 . And we have

(r + 2λ)∆ − δ 1 (r + ρ + λ)∆ − δ
µ r+ρ+λ = −(δ − (r + 2ρ)∆) + µ
2(r + ρ + λ) (1 + 2a0 ) ρ+λ r+ρ+λ

then " ρ+λ #

 r+ρ+λ
1 (r + ρ + λ)∆ − δ 1 (r + 2λ)∆ − δ
r+ρ+λ <2 ⇔ a0 > −1 >0
(1 + 2a0 ) ρ+λ (r + 2λ)∆ − δ 2 2[(r + ρ + λ)∆ − δ]
One possible sufficient conditions for the existence of an equilibrium:
" ρ+λ
 r+ρ+λ #
1 (r + 2λ)∆ − δ 1
−1 >
2 2[(r + ρ + λ)∆ − δ] 4

but it does not hold if λ is too big.

148
 We can extend the domain of validity of the equilibrium by looking for conditions to verify that G(a0 ) > ρ2 .
Let’s first notice that

ρ+λ
µ r δ−r∆ − ρ∆ 1 µ (λ + ρ) ωr − ρ∆ 1
f (α∅ ) = 2
r − r
∆ 2(r + ρ + λ) (1 + 2α∅ ) ρ+λ ∆ 2(r + ρ + λ) (1 + 2α∅ ) ρ+λ +1
ρ+λ
µ r δ−r∆2 − ρ∆ 1 µ (λ + ρ) ωr − ρ∆ 1
> r
+1
− r
∆ 2(r + ρ + λ) (1 + 2α ) ρ+λ
∅ ∆ 2(r + ρ + λ) (1 + 2α∅ ) ρ+λ +1
 
µ 1 λ+ρ δ − r∆ 1
>− ω− r
∆ 2(r + ρ + λ) r 2 (1 + 2α∅ ) ρ+λ +1

then

µ ρ + λ ω − 2δ + r ∆
  
∅ 2 ∅ δ µ δ − r∆
G(α ) > +α × −r+ −∆ +
∆ r 2(r + ρ + λ) ∆ ∆ r+ρ+λ
 
µ 1 λ+ρ δ − r∆ 1
− ω− r
∆ 2(r + ρ + λ) r 2 (1 + 2α∅ ) ρ+λ +1

assuming that
 
δ µ δ − r∆
−r+ −∆ + >0
∆ ∆ r+ρ+λ

we obtain that

µ ρ + λ ω − 2δ + r ∆
 
2 µ 1 ρ+λ δ − r∆ 1
G(a0 ) > − ω− r
∆ r 2(r + ρ + λ) ∆ 2(r + ρ + λ) r 2 (1 + 2a0 ) ρ+λ +1
µ ρ + λ ω − 2δ + r ∆
2
>
∆ r 2(r + ρ + λ)
ρ + λ ω − 2δ + r ∆
 
2 δ − (r + 2ρ)∆ µ (r + ρ + λ)∆ − δ
− − +
r (r + 2λ)∆ − δ ∆ ∆ r+ρ+λ
ρ + λ ω − 2δ + r ∆
2 δ − (r + 2ρ)∆
>
r (r + 2λ)∆ − δ ∆
δ ∆
µ ρ+λω− 2 +r2
 
1 (r + ρ + λ)∆ − δ
+ × −
∆ r r+ρ+λ 2 (r + 2λ)∆ − δ

ω− δ2 +r ∆
2 δ−(r+2ρ)∆
the first term decreases w.r.t λ and converges towards 2r∆ ∆ > ρ2 . The second term is posi-
tive.

We can also check that for µ big enough it works without this assumption. Let’s call

µ ρ + λ ω − 2δ + r ∆
  
δ µ δ − r∆
g(α∅ ) = 2
+ α∅ × −r+ −∆ +
∆ r 2(r + ρ + λ) ∆ ∆ r+ρ+λ
 
µ 1 λ+ρ δ − r∆ 1
− ω− r
∆ 2(r + ρ + λ) r 2 (1 + 2α∅ ) ρ+λ +1

1
and look at the limit µ → ∞ with α = µ →0

149
 !
µ ρ + λ ω − 2δ + r ∆
 
1 1 δ − r∆ 2 1
g( ) ≈ −∆ + + × 1− r
µ ∆ r+ρ+λ ∆ r 2(r + ρ + λ) (1 + µ2 ) ρ+λ +1
µ ρ + λ ω − 2δ + r ∆
 
∼ 1 δ − r∆ 2 r+ρ+λ 2
= −∆ + + ×
∆ r+ρ+λ ∆ r 2(r + ρ + λ) ρ+λ µ
" #
δ ∆
∼ 1 ω− 2 +r2 δ − r∆
= −∆+
∆ r r+ρ+λ
" #
1 ω − 2δ + r ∆2 ρ
> −∆ >
∆ r 2
" ρ+λ
#
  r+ρ+λ
1 (r+2λ)∆−δ
And because a0 > 2 2[(r+ρ+λ)∆−δ] − 1 > 0, for µ big enough the equilibrium α is less than a0 .

∅
This also proves that limµ→∞ αeq =0

Now if
 
δ µ δ − r∆
−r+ −∆ + <0
∆ ∆ r+ρ+λ

Since

µ ρ + λ ω − 2δ + r ∆
  
δ µ δ − r∆
G(α∅ ) > 2
+ α∅ × −r+ −∆ +
∆ r 2(r + ρ + λ) ∆ ∆ r+ρ+λ
 
µ 1 λ+ρ δ − r∆ 1
− ω− r
∆ 2(r + ρ + λ) r 2 (1 + 2α∅ ) ρ+λ +1

we obtain for the equilibrium α∅

h i
δ−r∆
µ 1 µ ρ∆ 1 µ ∆− r+ρ+λ − (δ − r∆)
r
+1
> − × λ+ρ δ−r∆

− α∅ × λ+ρ δ−r∆


2(r + ρ + λ) (1 + 2α )
∅ ρ+λ 2(r + ρ + λ) 2 r ω− 2 r ω− 2

then

δ − (r + ρ + λ)∆
X(α∅ ) > δ − (r + 2ρ)∆ + µ
r+ρ+λ
 h i 
δ−r∆
 µ ρ∆ 1 µ ∆− r+ρ+λ − (δ − r∆) 
+ [(r + 2λ)∆ − δ] × − × λ+ρ δ−r∆

− α∅ × λ+ρ
ω − δ−r∆


 2(r + ρ + λ) 2 ω− 
r 2 r 2

which is equivalent to

1 ρ∆ [(r + 2λ)∆ − δ]
X(α∅ ) > (δ − (r + 2ρ)∆)(1 + µ )− × λ+ρ
ω − δ−r∆


2(r + ρ + λ) 2 r 2
h i
δ−r∆
µ ∆ − r+ρ+λ − (δ − r∆)
∅
− α × [(r + 2λ)∆ − δ] λ+ρ
ω − δ−r∆


r 2

Since this expression is decreasing w.r.t λ, letting α∅ unchanged.

150
 ρ∆ 2r∆ µ∆ − (δ − r∆)
X(α∅ ) > δ − (r + 2ρ)∆ − × δ−r∆
− α∅ × 2r∆
2 ω− 2 ω − δ−r∆
2
 
2r∆ ∅ ρ∆ ∅ µ∆
> δ − (r + 2ρ)∆ + α (δ − r∆) − − α × 2r∆
ω − δ−r∆
2
2 ω − δ−r∆
2
2r∆ ρ∆ µ∆
> δ − (r + 2ρ)∆ − − α∅ × 2r∆
ω − δ−r∆
2
2 ω − δ−r∆2

Now we need to give a bound for α∅ × µ and show that it is small enough.

µ ρ + λ ω − 2δ + r ∆
  
δ µ δ − r∆
G(α∅ ) = 2
+ α∅ × −r+ −∆ +
∆ r 2(r + ρ + λ) ∆ ∆ r+ρ+λ
µ ρ + λ ω − 2δ + r ∆
2 1 µ (λ + ρ) ωr − ρ∆ α∅
− r + r
∆ r 2(r + ρ + λ) (1 + 2α∅ ) ρ+λ ∆ r + ρ + λ (1 + 2α∅ ) ρ+λ +1
µ (λ + ρ) ωr − ρ∆ α∅
 
µ δ − r∆
> α∅ −∆ + +
∆ r+ρ+λ ∆ r + ρ + λ (3/2)2
 
µ 1 ω ω − ρ∆ 1 δ − r∆
> α∅ − ∆ + +
∆ (3/2)2 r r + ρ + λ (3/2)2 r+ρ+λ
 
µ 1 ω
> α∅ −∆
∆ (3/2)2 r

Then at equilibrium
 
ρ ∅ µ 1 ω
>α −∆
2 ∆ (3/2)2 r

and
∅ 2r∆ ρ∆ ∆ ρ∆
X(αeq ) > δ − (r + 2ρ)∆ − δ−r∆ 2
− 2r∆ δ−r∆
× h i
ω− 2 ω− 2 2 (3/2)2 ωr − ∆
1

∅
Given the difference of magnitudes we assume it implies that X(αeq ) > 0.

Magnitude Assumptions. These following conditions are sufficient for the equilibrium to exist.
     
1ρ+λ δ 1ρ+λ 1 1ρ+λ 1
ω× + × − + + r∆ × − − >0
6 r 2 6 r 2 6 r 2

2r∆ ρ∆ ∆ ρ∆
δ − (r + 2ρ)∆ − δ−r∆ 2
− 2r∆ δ−r∆
× h i >0
ω− 2 ω− 2 2 (3/2)2 ωr − ∆
1

1 ω
−∆>0
(3/2)2 r

One can check that the following conditions are also sufficient so that the conditions above hold:
 
2r + ρ
ω > 3δ × max 1, and δ > (r + 4ρ)∆
2ρ

151
A.5.4 Proof of Proposition A.12
∅ ∅
We know that ∂G
∂α (αeq ) > 0. Since we know that when µ is big then αeq is close to zero, we must show that
∂G
∂λ around α = 0.

µ ρ + λ ω − 2δ + r ∆
  
∅ δ µ δ − r∆
G(α ) = 2
+ α∅ × −r+ −∆ +
∆ r 2(r + ρ + λ) ∆ ∆ r+ρ+λ
µ ρ + λ ω − 2δ + r ∆
2 1 µ (λ + ρ) ωr − ρ∆ α∅
− r + r
∆ r 2(r + ρ + λ) (1 + 2α∅ ) ρ+λ ∆ r + ρ + λ (1 + 2α∅ ) ρ+λ +1

Then for α → 0

µ ρ + λ ω − 2δ + r ∆
 
r
G(α∅ ) ∼ 2
× 1−1+ 2α∅
∆ r 2(r + ρ + λ) ρ+λ
 
δ µ δ − r∆
+ α∅ × −r−µ+
∆ ∆r+ρ+λ
µ (λ + ρ) ωr − ρ∆ ∅
+ α
∆ r+ρ+λ
which gives

 
δ µω µ δ − (r + 2ρ)∆
G(α∅ ) ∼ α∅ × −r−µ+ +
∆ ∆r ∆ 2(r + ρ + λ)

which is decreasing with respect to λ.

For the two last derivative this is sufficient to show that what is in factor of µ or ω in function G is
positive. For µ:

µ ρ + λ ω − 2δ + r ∆ µ (λ + ρ) ωr − ρ∆
   
2 1 ∅ 1 δ − r∆
1− r + α × r − ∆ +
∆ r 2(r + ρ + λ) (1 + 2α∅ ) ρ+λ ∆ r + ρ + λ (1 + 2α∅ ) ρ+λ +1 r+ρ+λ

which works also for ω.

152
A.6 Comparative statics
The execution rate depends on 3 important parameter: the depth α, the monitoring rate λ∗ of the investor
and the monitoring rate λ0 of other investors with the constraint that λ∗ = λ0 = λ. Writing l in function of
these 3 parameters allows for disentangling where the effect of λ from. l depends on

u u d d
U1 + ∆ = Vln−out (0) + A0 − Vlo−A 0
0 (0) = Vho−out (0) − B − Vhn−B 0 (0)
Z T ∗
δ ∆ D(t)e−(r+λ −λ0 )t
= −[ω − + r ] ds
2 2 0 α
Z T
δ ∆ ∗
= −[ω − + r ] h(t)e−(r+ρ+λ )t dt
2 2 0
d d u u
U2 = Vln−out (0) + B 0 − Vlo−A 0
0 (0) = Vho−out (0) − A − Vhn−B 0 (0)

ω + 2δ − r ∆
2 (λ∗ + ρ) ωr − ρ∆ −(r+ρ+λ∗ )T
= + e
r + ρ + λ∗ r + ρ + λ∗
−(λ0 +ρ)t
D(t) 1 (1−e ) 1 1 (λ0 +ρ)t ln(1+2α)
with h(t) = αe−(λ0 +ρ)t
= 1− 2 αe−(λ0 +ρ)t = 1+ 2α − 2α e and T = ρ+λ0 , noticing that
h(T ) = 0.

δ − (r + 2ρ)∆
l(α, λ∗ , λ0 ) =
"2∆ Z T #
1 δ ∆ −(r+ρ+λ∗ )t
+ µ −∆ − (ω − + r ) h(t)e dt
2∆ 2 2 0
" #
∗
1 ω + 2δ − r ∆ (λ + ρ) ω
− ρ∆ ∗
+ µ 2
+ r
e−(r+ρ+λ )T
2∆ r + ρ + λ∗ r + ρ + λ∗
RT ∗
The derivative of 0
h(t)e−(r+ρ+λ )t
dt with respect to α is
Z T Z T
∂T ∗ ∂h ∗ ∂h ∗
h(T )e−(r+ρ+λ )T + (t)e−(r+ρ+λ )t dt = (t)e−(r+ρ+λ )t dt > 0
∂α 0 ∂α 0 ∂α

then
∂l
<0
∂α
Decomposition: Obviously we have
∂U2 ∂U1
> 0 and >0
∂λ0 ∂λ∗
Moreover
Z T
∂U1 δ ∆ ∂h ∗
= −[ω − + r ] (t)e−(r+ρ+λ )t dt > 0
∂λ0 2 2 0 ∂λ0
∂l ∂U1
And since ∂λ∗ < 0 (appendix section J.) and ∂λ∗ > 0 we must have

∂U2
<0
∂λ∗

153
A.7 Empirical Implications
A.7.1 Liquidity supply in the ”pre-signal” phase
Before the asset fundamental value switch to a different high or low level the liquidity supply on both side
∅
of the order book is equal to αeq L.

A.7.2 Information integration speed

The duration between the change in the asset fundamental and the adjustment of transaction prices in the
limit order book is the duration of the transition phase.

1
T = ln(1 + 2α∅ )
ρ+λ

Corollary A.3. For high enough values of λ and µ an increase of the monitoring rate (λ) decreases the
duration of the transition phase T . An increase of the volatility of the asset fundamental (µ or ω) decreases
as well this duration.

A.7.3 Order flow decomposition of the price adjustment (transi-

tion phase)
Corollary A.4. In the transition phase the amount of market order executed and limit order cancelled are
" #
∅ ∅
ln(1 + 2αeq ) ∅
ln(1 + 2αeq )
MO = L, LOC = αeq − L
2 2

Moreover the ratio of limit order cancellation over executed market order is increasing with respect to
∅
αeq :

∂ LOC
∅ MO
>0
∂αeq

A.7.4 Risk of being picked-off

Corollary A.5. Conditionally on having a limit order on the ”wrong side” of the limit order book when the
value of the utility flow change, the risk of having this limit order executed during the transition phase is

∅
ln(1 + 2αeq ) ∂Ppo
Ppo = ∅
, ∅
<0
2αeq ∂αeq

When an investor send a limit order during the initial phase the unconditional probability for this limit order
to be picked-off in the transition phase is

∅
µ µ ln(1 + 2αeq )
× Ppo = ∅ + ρ
µ + ρ + λ + l∅ 2(µ + λ)αeq 2

The effect of the monitoring rate and the fundamental volatility on the unconditional probability is
unclear and depends on the parameters value.

154
A.7.5 Proof of Corollary A.4
The amount of cancellation is given by
Z T Z T
1 1
(ρ + λ)D(t)dt = (ρ + λ)(− L + [α∅ + ]Le−(ρ+λ)t )dt
0 0 2 2
1 1
= (− (ρ + λ)T + [α + ](1 − e−(ρ+λ)T ))L
∅
2 2
1 ∅ 1 1
= (− ln(1 + 2α) + [α + ](1 − ))L
2 2 1 + 2α
The second part of the corollary comes from the the fact that ln(1 + x)/x is increasing.

A.7.6 Proof of Corollary A.5

At t a limit order is cancelled if the owner monitor the market, with probability (ρ + λ)dt, and it executed
(ρ+λ) L
with probability D(t)
2
dt. The probability to have the stale limit order executed between t and t + dt is
" Z #
(ρ + λ) L2 t
(ρ + λ) L2
Pt = dt × exp − (ρ + λ) + ds
D(t) 0 D(s)

Noticing that
∂D L
= −(ρ + λ)D(t) − (ρ + λ)
∂t 2
this probability is equal to
"Z #
(ρ + λ) L2 t ∂D
∂t (ρ + λ) L2 ρ+λ
Pt = dt × exp + ds = dt = dt
D(t) 0 D(s) D(0) 2α

Then the overall probability of is

Z T
ρ+λ ln(1 + 2α)
Pt dt = T =
0 2α 2α

155
A.8 Extension to the problem with 2 monitoring rates,
λ1 and λ2, in the symmetric case, ρ+ = ρ−, s = 21
A.8.1 steady-state populations
We consider the case where the population is heterogeneous. A mass L1 of agents monitor the market with
intensity λ1 and a mass L2 of agents monitor the market with intensity λ2 . We assume that the second type
is the intensive monitoring type, λ2 > λ1 .
Both populations have the same law of motion for their private value of the asset. So in steady state we
must have

1 i
Liho + Lihn = Lilo + Liln =
L
2
Moreover on aggregate we must have that for the asset supply condition
1
L1ho + L1lo + L2ho + L2lo = L
2
which imply that at aggregate there remains only one degree of liberty α as in the case with only one
monitoring intensity,
L1lo + L2lo = L1hn + L2hn = αL
Let’s call Lilo = αoi Li and Lihn = αni Li . It could be the case that αni 6= αoi which would introduce an
asymmetry in each of the populations whereas at the aggregate level αn1 L1 + αn2 L2 = αo1 L1 + αo2 L2 = αL.

Assumption A.4. Here we restrict ourselves to the case where αni = αoi = αi . In the perfectly symmetric
case of the game with uncertainty it will be the case because the terms of the trade-off are the same for lo’s
and hn’s.

A.8.2 Limit order in steady state without fundamental uncertainty

In this case the terms of the trade-off between market orders and limit orders is the same for agents in L1 or
in L2 because the monitoring intensity does not play a role in the steady state value function.
The equilibrium value functions must be the same as in the case with a unique monitoring rate as well
as the equilibrium values α, lA and lB .

Let’s look at the micro-dynamic of the limit order book

Ask Side: At time t, on the ask side of the market the depth is constantly equal to DA = L1lo + L2lo and
the order flows sustaining this steady state are
• Outflow due to execution: a share lA (L1lo + L2lo ).dt is executed. This share is equal to the flow of
buy market order

lA (L1lo + L2lo ).dt = m1A (λ1 L1hn + ρL1ln ).dt + m2A (λ2 L2hn + ρL2ln ).dt

• Outflow due cancellation: people switching from lo to ho, ρ(L1lo + L2lo ).dt, lo type cancelling their
sell limit order to send a sell market order, m1B λ1 L1lo .dt + m2B λ2 L2lo .dt
• Inflow: people switching from ho to lo type sending a sell limit order, (1 − m1B )ρL1ho .dt + (1 −
m2B )ρL2ho .dt
The steady state condition is then:

ρ(L1lo + L2lo ) + lA (L1lo + L2lo ) + lB (L1hn + L2hn ) = ρ(L1lo + L2lo )

Which leads to following equation:

156
 ρ
αL = α1 L1 + α2 L2 = L
2(2ρ + lA + lB )
which is the same as in the case with a unique monitoring rate.

We need to add another steady state conditions for the populations L1 and L2 ,

dLilo
= ρ(1 − miB )Liho − lA Lilo − ρLilo − miB λi Lilo
dt
1
= ρ(1 − miB )( − αi )Li − lA αi Li − ραi Li − miB λi αi Li
2
ρ i 1
= L − (lA + 2ρ)αi Li − miB (λi αi Li + ρ( − αi ))Li
2 2
A possible equilibrium is the one where
1
miB (λi αi Li + ρ( − αi ))Li = lB αi
2
in this case
ρ
αi = =α
2(2ρ + lA + lB )
and we can check that 0 < miB < 1. miA can be defined the same way.

With this equilibrium, this is easy to replicate the equilibrium strategies in the case where we converge
i
to a steady state equilibrium. For the ask side, if we call DA the share of the depth of the order book due to
i
the population of L , we would get the dynamic
i
dDA
= ρLiho (t) − ρDA
i i
(t) − lA DA i
(t) − lB DB (t) + λi (Lilo (t) − DA
i
(t))
dt
and then the same dynamic for αi (t) as in the case with a unique monitoring rate.

A.9 Transition phase

We assume that the strategies are the same than in the unique monitoring
Let’s take the case of the transition phase when the fundamental value switches from v0 to v0 + ω.
At t = 0, the beginning of the transition phase the depth at price A in the limit order book is equal to
1 2
DA (0) = DA (0) + DA (0). During the transition phase the flow of market orders that hit the limit orders at
price A is equal to

u(t).dt = (λ1 + ρ)(L1hn (t) + L1ln (t)).dt + (λ2 + ρ)(L2hn (t) + L2ln (t)).dt
The evolution of the population Li with limit orders at price A is given by
i
dDA i Di (t)
= −(λi + ρ)DA (t) − u(t) 1 A 2
dt DA (t) + DA (t)
We have to solve for a system of equation of type
(
f
f 0 =−af − u(t) f +g
g
g 0 =−bg − u(t) f +g

We can assume that g ≥ 0 without generality because one of the functions has to be strictly positive. It
gives that

157
 gf
g(f 0 + af ) = f (g 0 + bg) = −u(t)
f +g
and then
 0
f 0 g − g0 f f f
= = (b − a)
g2 g g
and finally
f f0
= exp[(b − a)t]
g g0
or equivalently, there is a function h such that

f (t) = f0 e−at h(t), and g(t) = g0 e−bt h(t)

In our case, we know that there is a function hA such that hA (0) = 1 and
i
for i ∈ {1, 2}, DA i
(t) = DA (0)e−(λi +ρ)t hA (t)
Then we obtain that
d 1
[(DA (0)e−(λ1 +ρ)t + DA 2
(0)e−(λ2 +ρ)t )hA (t)] = − (λ1 + ρ)DA
1
(0)e−(λ1 +ρ)t hA (t)
dt
−(λ2 + ρ)DA 2
(0)e−(λ2 +ρ)t hA (t) − u(t)
dhA
⇐⇒ (DA 1
(0)e−(λ1 +ρ)t + DA 2
(0)e−(λ2 +ρ)t ) = −u(t)
dt
We can use this first result to define the intensity of limit order for execution, kA (t). We know that
u(t).dt is the flow of market order. Then we mus have u(t) = kA (t)DA (t).
1
DA (t) = DA 2
(t) + DA 1
(t) = DA (0)e−(λ1 +ρ)t hA (t) + DA
2
(0)e−(λ2 +ρ)t hA (t)
The execution intensity is equal to
ḣA
kA (t) = −
hA
It is easy to see that the strategies are the same. Indeed in the value functions in the transition phase, the
dynamic of the transition phase affects these value functions only through the execution intensities kA or kB
and through the duration of the transition phase, T u or T d , defined here by hA (T u ) = 0 and hB (T d ) = 0. But
to prove that deviations from the conjectured strategies are optimal we don’t need to look at the particular
form of k or T .

Now we can look more precisely at the function u(t). The law of motion for the population of hn and ln
are the following

dLihn
= −(λi + ρ)Lihn (t)
dt
dLiln Di (t)
= −(λi + ρ)Liln (t) + u(t) 1 A 2
dt DA (t) + DA (t)

then we obtain that

d i
(L (t) + Liln (t) + DA
i
(t)) = −(λi + ρ)(Lihn (t) + Liln (t) + DA
i
(t))
dt hn
then
Lihn (t) + Liln (t) = (Lihn (0) + Liln (0) + DA
i
(0))e−(λi +ρ)t − DA
i
(t)

158
 The differential equation for hA is then

d 1
[(DA (0)e−(λ1 +ρ)t + DA2
(0)e−(λ2 +ρ)t )hA (t)]
dt
= − (λ1 + ρ)(L1hn (0) + L1ln (0) + DA1
(0))e−(λ1 +ρ)t
− (λ1 + ρ)(L2hn (0) + L2ln (0) + DA
2
(0))e−(λ2 +ρ)t

which gives

1
(DA (0)e−(λ1 +ρ)t + DA
2
(0)e−(λ2 +ρ)t )hA (t) − (DA
1 2
(0) + DA (0))
=(L1hn (0) + L1ln (0) + DA
1
(0))e−(λ1 +ρ)t + (L2hn (0) + L2ln (0) + DA
2
(0))e−(λ2 +ρ)t
− (L1hn (0) + L1ln (0) + DA
1
(0) + L2hn (0) + L2ln (0) + DA
2
(0))

and finally

(L1hn (0) + L1ln (0))(1 − e−(λ1 +ρ)t ) + (L2hn (0) + L2ln (0))(1 − e−(λ2 +ρ)t )
hA (t) = 1 − 1 (0)e−(λ1 +ρ)t + D 2 (0)e−(λ2 +ρ)t
DA A
u u
In the transition phase, an important difference is value function is Vln−out (t)+A0 −Vlo−A0 (t) that appears

in the value of the execution intensity that makes agents indifferent between limit and market orders. the
formulation of this expression is very similar to the unique monitoring intensity case.
u u
Let’s call X(t) = Vln−out (t) + A0 − Vlo−A 0 (t)

ḣA ∂X
(r + ρ + λ − )X(t) = − [v u − δ − rA0 ]
hA ∂t
and at the end we find that
Tu
hA (s)e−(r+ρ+λi )s
Z
u 0 u u 0
Vln−out (t) +A − Vlo−A0 (t) = −[v − δ − rA ] ds
t hA (t)e−(r+ρ+λi )t
and for the same reason we would find

u u v u − rA0 (λi + ρ)(Au − A0 ) − ρ∆ −(r+ρ+λi )T u

Vho−out (t) − A0 − Vhn−B 0 (t) = + e × e(r+ρ+λi )t
r + ρ + λi r + ρ + λi

A.10 Steady-state with uncertainty in the perfectly sym-

metric case
In the perfectly symmetric case the populations are symmetric

Lihn = Lilo = αi Li

i i
DA = DB = γi αi Li
We don’t assume that all the lo’s or hn’s are at the best price in the limit order book. Some of them
could send their order at Au and B d . It could be optimal with the monitoring rate heterogeneity.
In the perfectly symmetric case, the duration of the transition phase are the same as well as the execution
intensity during the transition phase.

T u = T d = T, kA = kB = k, hA = hB = h

159
 Then for h we have
1 (1 − e−(λ1 +ρ)t )L1 + (1 − e−(λ2 +ρ)t )L2
h(t) = 1 −
2 γ1 α1 L1 e−(λ1 +ρ)t + γ2 α2 L2 e−(λ2 +ρ)t
Then T is defined by

(1 + 2γ1 α1 )L1 e−(λ1 +ρ)T + (1 + γ2 α2 )L2 e−(λ2 +ρ)T = L1 + L2

Using the calculation of the unique monitoring rate case, we obtain the value of the execution rate that
makes an agent indifferent between a limit order at the best price and a market order

δ − (r + 2ρ)∆
li =
"2∆ Z T #
1 δ ∆ −(r+ρ+λi )t
+ µ −∆ − (ω − + r ) h(t)e dt
2∆ 2 2 0
" #
1 ω + 2δ − r ∆
2 (λi + ρ) ωr − ρ∆ −(r+ρ+λi )T
+ µ + e
2∆ r + ρ + λi r + ρ + λi

we can rewrite

Z T
δ − (r + 2ρ)∆ µ 1 δ ∆
li = − + µ(ω − + r ) c(t)e−(r+ρ+λi )t dt
2∆ 2 2∆ 2 2 0
" #
1 δ ∆ 1 − e−(r+ρ+λi )T ω + 2δ − r ∆ (λ i + ρ) ω
− ρ∆
+ µ −(ω − + r ) + 2
+ r
e−(r+ρ+λi )T
2∆ 2 2 r + ρ + λi r + ρ + λi r + ρ + λi

−(λ1 +ρ)t 1 −(λ2 +ρ)t 2

with c(t) = 12 γ(1−e )L +(1−e )L
1 −(λ1 +ρ)t +γ α L2 e−(λ2 +ρ)t . The first line of the expression of li is obviously decreasing
1 α1 L e 2 2
with respect to λi because c(t) > 0 for all t ∈ [0, T ]. For the second line, we can rewrite it as

δ ∆ 1 − e−(r+ρ+λi )T ω + 2δ − r ∆
2 (λi + ρ) ωr − ρ∆ −(r+ρ+λi )T
− (ω − +r ) + + e
2 2 r + ρ + λi r + ρ + λi r + ρ + λi
δ − r∆ ω ω − 2δ + r ∆
2 − (ω + ρ∆) −(r+ρ+λi )T
= + e−(r+ρ+λi )T + e
r + ρ + λi r r + ρ + λi
δ−(2ρ−r)∆
δ − r∆ ω
= + e−(r+ρ+λi )T − 2
e−(r+ρ+λi )T
r + ρ + λi r r + ρ + λi
The derivative of this expression with respect to λi is
δ−(2ρ−r)∆
" δ−(2ρ−r)∆
#
δ − r∆ ω
− + 2
e−(r+ρ+λi )T − T e−(r+ρ+λi )T − 2
(r + ρ + λi )2 (r + ρ + λi )2 r r + ρ + λi
which is clearly negative. It implies that
l1 > l2
To be indifferent between a competitive limit order and a market order, an investor with a low monitoring
intensity λ1 requires a higher rate of limit order execution than an investor with a high monitoring intensity
λ2

160
Appendix B

Appendix to chapter 3

B.1 Mini flash crash events

Mini flash crashes Dates

United Airlines stock, from $12 to $3 in 15 min and recovered the next day September 8, 2008
Progress energy stock, 90% drop in a few seconds and recovered in few minutes September 27, 2010
ACOR stock, lost 11% of it’s value in under 5 seconds then quickly recovered May 24, 2011
STBC stock, from $12.51 to $10.28 in 3 sec before quickly recovering May 26, 2011
TLT stock, from $96.63 to $97.90 and then dropped back down in less than 1 sec July 14, 2011
AMJ stock, from $34.90 to $32.61 and then recovered, all in just under 4 seconds October 11, 2011
Brocade (BRCD) stock, dropped 5.5% and then recovered in about 1.5 seconds. August 17, 2012
LTXC stock, drops 10% in 1 second, recovers in 2. August 28, 2012
Perion Network (PERI) stock, dropped 7.6% in 1/3 of a second.
33 seconds later, the price rocketed 5.6% January 8, 2013
GeoEye, Inc (GEOY) stock, raced 8.5% higher, stayed up for about 1/2 second,
then returned near where it started a second later. January 9, 2013
Dow Jones index, lost 200 basis points and quickly recovered, after hacked tweet:
«Breaking: Two Explosions in the White House and Barack Obama is injured» April 23, 2103

Figure B.1: sample of mini flash crashes (see http://www.nanex.net/FlashCrash/OngoingResearch.html).

B.2 Proofs
Proof of proposition 3.2.
P[Q̃1 = q|V = 1]P[V = 1]
P[V = 1|Q̃1 = q] =
P[Q̃1 = q]

and P[V = 1] = 1/2,

P[Q̃1 = q|V = 1] = (P[U = 1] + P[U = 0,  = 1])P[˜l1 = q − α] + P[U = 0,  = 0]P[˜l1 = q + α]

1−δ 1−δ
= (δ + )φ(q − α) + φ(q + α),
2 2

161
 P[Q̃1 = q] = P[S = 1]P[˜l1 = q − α] + P[S = 0]P[˜l1 = q + α]
1 1
= φ(q − α) + φ(q + α).
2 2

Proof of proposition 3.3. At period 1 if the signal is S = 1, the profit for a participating strategic
trader is
 
(1 + δ)φ(l) + (1 − δ)φ(l + 2α) 1
Z
1 1+δ
π (α, S = 1) = − × φ(l)dl
[−Q,Q] 2 φ(l) + φ(l + 2α) 2
Z
δφ(l + 2α)
= φ(l)dl
[−Q,Q] φ(l) + φ(l + 2α)
Z
δφ(l + α)
= φ(l − α)dl
[−Q+α,Q+α] φ(l − α) + φ(l + α)
Z
δφ(l + α)
= φ(l − α)dl
[−Q+α,Q−α] φ(l − α) + φ(l + α)
Q−α
Z
δ δ
= dl = ×
[−Q+α,Q−α] 2 2 Q

because φ(l + α) = 0 for l > Q − α.

At period 1 if the signal is S = 1, the profit for a participating strategic trader is
 
(1 + δ)φ(l − 2α) + (1 − δ)φ(l) 1 1 − δ
Z
π 1 (α, S = 0) = × − φ(l)dl
[−Q,Q] φ(l − 2α) + φ(l) 2 2
δφ(l − 2α)
Z
= φ(l)dl
[−Q,Q] φ(l − 2α) + φ(l)
δφ(l − α)
Z
= φ(l + α)dl
[−Q−α,Q−α] φ(l − α) + φ(l + α)
δφ(l − α)
Z
= φ(l + α)dl
[−Q+α,Q−α] φ(l − α) + φ(l + α)
Q−α
Z
δ δ
= dl = ×
[−Q+α,Q−α] 2 2 Q

because φ(l − α) = 0 for l < −Q − α. Finally,

1 1 1
π 1 (α) = π (α, S = 1) + π 1 (α, S = 0)
2 2

Proofs of propositions 3.4 and 3.5. At t = 2, if U = 1 the proof is similar to the period 1 case.
If U = 0, the trading decision of strategic traders is not obvious. Then we take as unknown the mass of
trading
Z β
M0 = Xi (U = 0)di, −β ≤ M0 ≤ β.
0
The value of M0 at equilibrium depends on the profit that can be achieve at period 2, knowing that the
expected value of the asset is 1/2. This profit is directly linked to the average spread between 1/2 and the
transaction price at period 2, when the strategy associated to the state U = 0 is played. This spread is
Z
1
Σ2, (q1 , M0 ) = [P2 (q2 , q1 ) − ]φ(q2 − M0 )dq2
[−Q−β,Q+β] 2

Using this function we can distinguish between three types of trading outcomes at equilibrium associated to
trading decisions of strategic traders

162
 • the all-selling outcome in which all event traders sell, M0 = −β, is an equilibrium strategy if and only
if Σ2, (q1 , −β) ≥ 0
• the all-buying outcome in which all event traders buy, M0 = β, is an equilibrium strategy if and only
if Σ2, (q1 , β) ≤ 0
• A ”mixed” outcome in which event can behave differently, which generates an order flow −β < M0 < β,
is an equilibrium strategy if and only if Σ2, (q1 , M0 ) = 0. Otherwise there would be an incentive to
unilaterally deviate to take advantage of the non zero spread Σ2,
To figure out the equilibrium strategies, we need to have the pricing policy of the market maker given a
trading schedule M0 (q1 ).

P[Q̃2 = q2 , V = 1|Q̃1 = q1 ] P[Q̃2 = q2 , V = 1, Q̃1 = q1 ]

P[V = 1|Q̃2 = q2 , Q̃1 = q1 ] = =
P[Q̃2 = q2 |Q̃1 = q1 ]] P[Q̃2 = q2 , Q̃1 = q1 ]]
and

P[Q̃2 = q2 , V = 1, Q̃1 = q1 ] = P[U = 1, Q̃2 = q2 , V = 1, Q̃1 = q1 ]

+ P[U = 0,  = 1, Q̃2 = q2 , V = 1, Q̃1 = q1 ]
+ P[U = 0,  = 0, Q̃2 = q2 , V = 1, Q̃1 = q1 ]
with

1
P[U = 1, Q̃2 = q2 , V = 1, Q̃1 = q1 ] = φ(q2 − β)φ(q1 − α) δ
2
11
P[U = 0,  = 1, Q̃2 = q2 , V = 1, Q̃1 = q1 ] = φ(q2 − M0 )φ(q1 − α) (1 − δ)
22
11
P[U = 0,  = 0, Q̃2 = q2 , V = 1, Q̃1 = q1 ] = φ(q2 − M0 )φ(q1 + α) (1 − δ)
22
and

P[Q̃2 = q2 Q̃1 = q1 ] = P[U = 1, V = 1, Q̃2 = q2 , Q̃1 = q1 ]

+ P[U = 1, V = 0, Q̃2 = q2 , Q̃1 = q1 ]
+ P[U = 0,  = 1, Q̃2 = q2 , Q̃1 = q1 ]
+ P[U = 0,  = 0, Q̃2 = q2 , Q̃1 = q1 ]
with

1
P[U = 1, V = 1, Q̃2 = q2 , Q̃1 = q1 ] = φ(q2 − β)φ(q1 − α) δ
2
1
P[U = 1, V = 0, Q̃2 = q2 , Q̃1 = q1 ] = φ(q2 + β)φ(q1 + α) δ
2
1
P[U = 0,  = 1, Q̃2 = q2 , Q̃1 = q1 ] = φ(q2 − M0 )φ(q1 − α) (1 − δ)
2
1
P[U = 0,  = 0, Q̃2 = q2 , Q̃1 = q1 ] = φ(q2 − M0 )φ(q1 + α) (1 − δ)
2
Lemma B.1. At period 2, when U = 0, the trading strategies and order flow outcome of event traders are
• for q1 ∈ [−Q − α, −Q + α) the all-buying outcome, M0 = β is the only equilibrium outcome
• for q1 ∈ (Q − α, Q + α] the all-selling outcome, M0 = −β is the only equilibrium outcome
• for q1 ∈ [−Q + α, Q − α] all values M0 ∈ [−β, β] can correspond to an equilibrium.
However in this case we set M0 = 0 which is the only value robust to a small trading cost and the
most natural because there is no coordination issue between strategic traders, they trade on the ”non-
information” event only if there is mispricing.

163
Proof of lemma B.1. knowing that
P2 (q2 , q1 )) = P[V = 1|Q̃2 = q2 , Q̃1 = q1 ] =
δφ(q1 − α)φ(q2 − β) + 1−δ
2 [φ(q1 − α) + φ(q1 + α)]φ(q2 − M0 )
δ[φ(q1 − α)φ(q2 − β) + φ(q1 + α)φ(q2 + β)] + (1 − δ)[φ(q1 − α) + φ(q1 + α)]φ(q2 − M0 )
The proof of this lemma corresponds to the analysis of the function
Σ2, (q1 , M0 ) =
δ[φ(q1 − α)φ(q2 − β) − φ(q1 + α)φ(q2 + β)]φ(q2 − M0 )
Z
1
dq2
2 δ[φ(q1 − α)φ(q2 − β) + φ(q1 + α)φ(q2 + β)] + (1 − δ)[φ(q1 − α) + φ(q1 + α)]φ(q2 − M0 )
which is equal to Z
2, 1 1 N (q1 , q2 )
Σ (q1 , M0 ) = dq2
2Q 2 D(q1 , q2 )
with
N (q1 , q2 ) = δ[I{q1 ∈[−Q+α,Q+α]} I{q2 ∈[−Q+β,Q+β]} − I{q1 ∈[−Q−α,Q−α]} I{q2 ∈[−Q−β,Q−β]} ]I{q2 ∈[−Q+M0 ,Q+M0 ]}
and
D(q1 , q2 ) = δ[I{q1 ∈[−Q+α,Q+α]} I{q2 ∈[−Q+β,Q+β]} + I{q1 ∈[−Q−α,Q−α]} I{q2 ∈[−Q−β,Q−β]} ]
+ (1 − δ)[I{q1 ∈[−Q+α,Q+α]} + I{q1 ∈[−Q−α,Q−α]} ]I{q2 ∈[−Q+M0 ,Q+M0 ]}
If q1 ∈ [−Q − α, −Q + α), necessarily the integration is on q2 ∈ [−Q − β, Q + M0 ] because in this case
the period 1 signal is negative and at period 2 strategic traders either sell because the true value of the asset
is 0 or ”trade M0 ” to correct the mispricing. Then

0 if q2 ∈ [−Q − β, −Q + M0 )
N (q1 , q2 ) 
= −1 if q2 ∈ [−Q + M0 , Q − β]
D(q1 , q2 ) 
0 if q2 ∈ (Q − β, Q + M0 ]


Then
Q−β β − M0
Σ2, (q1 , M0 ) = −δ −δ <0
2Q 4Q
In this case, we set M0 = β and Σ2, (q1 , β) = −δ Q−β
2Q

If q1 ∈ [−Q + α, Q − α], the integration is on q2 ∈ [−Q − β, Q + β]. Then




 0 if q2 ∈ [−Q − β, −Q + M0 [
1
if q2 ∈ [−Q + M0 , −Q + β[


N (q1 , q2 )  2−δ

= 0 if q2 ∈ [−Q + β, Q − β]
D(q1 , q2 )  1
if q2 ∈]Q − β, Q + M0 ]




 2−δ
0 if q2 ∈]Q + M0 , Q + β]
Then
Σ2, (q1 , M0 ) = 0
In this case, M0 can take all possible value in [−β, β].

If q1 ∈]Q − α, Q + α], necessarily the integration is on q2 ∈ [−Q + M0 , Q + β] because in this case the
period 1 signal is positive and at period 2 strategic traders either buy because the true value of the asset is
1 or ”trade M0 ” to correct the mispricing. Then

0 if q2 ∈ [−Q + M0 , −Q + β)
N (q1 , q2 ) 
= 1 if q2 ∈ [−Q + β, Q + M0 ]
D(q1 , q2 ) 
0 if q2 ∈ (Q + M0 , Q + β]


164
Then
Q−β β + M0
Σ2, (q1 , M0 ) = δ +δ >0
2Q 4Q
In this case, we set M0 = −β and Σ2, (q1 , β) = δ Q−β
2Q

Proof of proposition 3.6. To compute the profit at period 2, we can notice that the symmetry of
the problem implies that this profit conditionned on the value of the signal at period 1 is the same in the
two cases S = 0 and S = 1. As these two cases are equally likely the unconditionnal value of the profit is
equal to the value of one of these conditionnal profits.
Let’s focus on the case where S = 1. Then the order flow at period 1 is necessarily such that q1 ∈ [−Q +
α, Q + α].
If q1 ∈ [−Q + α, Q − α], at period 2 strategic traders buy if U = 1 (q2 ∈ [−Q + β, Q + β]) and do not
participate if U = 0. Then the corresponding share of the profit is equal to

P[q1 ∈ [−Q + α, Q − α]] × δ×

1 1
[P[q2 ∈ [−Q + β, Q − β]] × (1 − ) + P[q2 ∈ [Q − β, Q]] × (1 − ) + P[q2 ∈ [Q, Q + β]] × 0]
2 2−δ
Q−α Q−β 1 β 1−δ
= ×δ×[ × + × ]
Q Q 2 2Q 2 − δ

If q1 ∈ [Q − α, Q + α], at period 2 strategic traders buy if U = 1 (q2 ∈ [−Q + β, Q + β]) and sell if U = 0
(q2 ∈ [−Q − β, Q − β]). Then the corresponding share of the profit is equal to

P[q1 ∈ [Q − α, Q + α]]×
1+δ
{δ × [P[q2 ∈ [−Q + β, Q − β]] × (1 − ) + P[q2 ∈ [Q − β, Q + β]] × 0]+
2
1+δ 1
(1 − δ) × [P[q2 ∈ [−Q − β, −Q + β]] × 0 + P[q2 ∈ [−Q + β, Q − β]] × ( − )]}
2 2
α Q−β 1−δ Q−β δ
= × {δ × [ × ] + (1 − δ) × [ × ]}
Q Q 2 Q 2

Proof of proposition 3.8.

∂π 2 (α, β)
 
δ 1 β β
= −(1 − ) + (1 − δ)(1 − )
∂α 2Q 2−δQ Q
 
δ 1 β
=− δ + (1 − δ − )
2Q 2−δ Q

1 ∂π 2 (α,β)
If δ < δ0 then 1 − δ − 2−δ > 0 and ∂α < 0.
If δ > δ0
∂π 2 (α, β)
   
δ 1 δ 1−δ
< −δ + ( − (1 − δ)) = − <0
∂α 2Q 2−δ 2Q 2 − δ
A δ δ ∂π
Then, for β = A, if C ∈ [(1 − Q)2, 2] we have ∂C > 0.
2
For any fixed α, ∂π ∂β
(α,β)
< 0, then if C > (1 − A δ ∂π
Q ) 2 , ∂A < 0 because αe q does not depend on A in this
interval under the assumption 2.
∂π 1 (α) ∂π 2 (α,β) ∂π 2 (α,β) A δ
Because ∂α < 0, ∂β < 0 and ∂α < 0 then, for C < (1 − Q)2

∂π ∂ 2
= [π (A, A) + π 1 (A)] < 0
∂A ∂A

165
Proof of proposition 3.10. For this proof we study the sign of π2 (0, A) − π2 (A, A) − π1 (A).

 
2 2 1 δ 1 A A A 1 A A A
π (0, A)−π (A, A)−π (A) = Q× 1 − − (1 − ) − (1 − ) × (1 − ) − (1 − δ) (1 − )
2 2−δQ Q Q 2−δQ Q Q

We can rewrite it as

  2
A 1 A A
U( ) = 1 − δ − 2
+ (1 + δ) − 1
Q 2−δ Q Q

1
U (0) = −1 and U (1) = 1 − 2−δ ≥ 0. Because this is a 2nd degree polynom there is a unique value between
0 and 1 for which U = 0.

Proof of proposition 3.12. To compute this measure at period 1 we need the following probabilities:

P[Q1 ∈ [−Q − α, −Q + α)] = P[S = 1] × P[l1 + α ∈ [−Q − α, −Q + α)]

+ P[S = 0] × P[l1 − α ∈ [−Q − α, −Q + α)]
1 1α
= 0 + P[l1 ∈ [−Q, −Q + 2α)] =
2 2Q
1α
P[Q1 ∈ (Q − α, Q + α]] =
2Q
P[Q1 ∈ [−Q + α, Q − α]] = P[S = 1] × P[l1 + α ∈ [−Q + α, Q − α]]
+ P[S = 0] × P[l1 − α ∈ [−Q + α, Q − α]]
1 1
= × P[l1 ∈ [−Q, Q − 2α]] + × P[l1 − α ∈ [−Q + 2α, Q]]
2 2
Q−α
=
Q

Then

1α 1−δ1+δ Q−α 1 1 α 1+δ1−δ

E[(Ṽ − P1 )2 ] = × + × + ×
2Q 2 2 Q 4 2Q 2 2
   
1 α Q − α 1 α
= (1 − δ 2 ) + = 1 − δ2
4 Q Q 4 Q

To compute this measure at period 2 we need the following probabilities:

166
 P[Q2 ∈ [−Q − β, −Q + β), Q1 ∈ [−Q − α, −Q + α)]
=P[Q1 ∈ [−Q − α, −Q + α)|U = 1, V = 1] × P[U = 1, V = 1] × P[l2 + β ∈ [−Q − β, −Q + β)]
+P[Q1 ∈ [−Q − α, −Q + α)|U = 1, V = 0] × P[U = 1, V = 0] × P[l2 − β ∈ [−Q − β, −Q + β)]
+P[Q1 ∈ [−Q − α, −Q + α)|U = 0] × P[U = 0] × P[l2 + β ∈ [−Q − β, −Q + β)]
α 1β α 1β
=0 + δ +0= δ
Q 2Q Q 2Q

P[Q2 ∈ [−Q + β, Q − β], Q1 ∈ [−Q − α, −Q + α)]

=P[Q1 ∈ [−Q − α, −Q + α)|U = 1, V = 1] × P[U = 1, V = 1] × P[l2 + β ∈ [−Q + β, Q − β]]
+P[Q1 ∈ [−Q − α, −Q + α)|U = 1, V = 0] × P[U = 1, V = 0] × P[l2 − β ∈ [−Q + β, Q − β]]
+P[Q1 ∈ [−Q − α, −Q + α)|U = 0] × P[U = 0] × P[l2 + β ∈ [−Q + β, Q − β]]
α 1Q−β 1α Q−β
=0+ δ + (1 − δ)
Q 2 Q 2Q Q
1 α Q−β
=
2Q Q

P[Q2 ∈ (Q − β, Q + β], Q1 ∈ [−Q − α, −Q + α)]

=P[Q1 ∈ [−Q − α, −Q + α)|U = 1, V = 1] × P[U = 1, V = 1] × P[l2 + β ∈ (Q − β, Q + β]]
+P[Q1 ∈ [−Q − α, −Q + α)|U = 1, V = 0] × P[U = 1, V = 0] × P[l2 − β ∈ (Q − β, Q + β]]
+P[Q1 ∈ [−Q − α, −Q + α)|U = 0] × P[U = 0] × P[l2 + β ∈ (Q − β, Q + β]]
1α β
=0 + 0 + (1 − δ)
2Q Q

Symmetrically,

1α β
P[Q2 ∈ [−Q − β, −Q + β), Q1 ∈ (Q − α, Q + α]] = (1 − δ)
2Q Q
1 α Q−β
P[Q2 ∈ [−Q + β, Q − β], Q1 ∈ (Q − α, Q + α]] =
2Q Q
α 1β
P[Q2 ∈ (Q − β, Q + β], Q1 ∈ (Q − α, Q + α]] = δ
Q 2Q

167
and

P[Q2 ∈ [−Q − β, −Q), Q1 ∈ [−Q + α, Q − α]]

=P[Q1 ∈ [−Q + α, Q − α]|U = 1, V = 1] × P[U = 1, V = 1] × P[l2 + β ∈ [−Q − β, −Q)]
+P[Q1 ∈ [−Q + α, Q − α]|U = 1, V = 0] × P[U = 1, V = 0] × P[l2 − β ∈ [−Q − β, −Q)]
+P[Q1 ∈ [−Q + α, Q − α]|U = 0] × P[U = 0] × P[l2 ∈ [−Q − β, −Q)]
Q−α 1 β
=0 + δ +0
Q 2 2Q

P[Q2 ∈ [−Q, −Q + β), Q1 ∈ [−Q + α, Q − α]]

=P[Q1 ∈ [−Q + α, Q − α]|U = 1, V = 1] × P[U = 1, V = 1] × P[l2 + β ∈ [−Q, −Q + β)]
+P[Q1 ∈ [−Q + α, Q − α]|U = 1, V = 0] × P[U = 1, V = 0] × P[l2 − β ∈ [−Q, −Q + β)]
+P[Q1 ∈ [−Q + α, Q − α]|U = 0] × P[U = 0] × P[l2 ∈ [−Q, −Q + β)]
Q−α 1 β Q−α β
=0 + δ + (1 − δ)
Q 2 2Q Q 2Q
Q−α2−δ β
=
Q 2 2Q

P[Q2 ∈ [−Q + β, Q − β], Q1 ∈ [−Q + α, Q − α]]

=P[Q1 ∈ [−Q + α, Q − α]|U = 1, V = 1] × P[U = 1, V = 1] × P[l2 + β ∈ [−Q + β, Q − β]]
+P[Q1 ∈ [−Q + α, Q − α]|U = 1, V = 0] × P[U = 1, V = 0] × P[l2 − β ∈ [−Q + β, Q − β]]
+P[Q1 ∈ [−Q + α, Q − α]|U = 0] × P[U = 0] × P[l2 ∈ [−Q + β, Q − β]]
Q−α 1Q−β Q−α 1Q−β Q−α Q−β
= δ + δ + (1 − δ)
Q 2 Q Q 2 Q Q Q
Q−αQ−β
=
Q Q
by symmetry again
Q−α2−δ β
P[Q2 ∈ (Q − β, Q], Q1 ∈ [−Q + α, Q − α]] =
Q 2 2Q
Q−α 1 β
P[Q2 ∈ (Q, Q + β], Q1 ∈ [−Q + α, Q − α]] = δ
Q 2 2Q

Then
α 1β 1 α Q−β 1−δ1+δ 1α β 1
E[(Ṽ − P2 )2 ] =δ ×0+ + (1 − δ) ×
Q 2Q 2Q Q 2 2 2Q Q 4
1α β 1 1 α Q−β 1−δ1+δ α 1β
+ (1 − δ) × + + δ ×0
2Q Q 4 2Q Q 2 2 Q 2Q
Q−α 1 β Q−α2−δ β 1−δ 1
+ δ ×0+ ×
Q 2 2Q Q 2 2Q 2 − δ 2 − δ
Q−αQ−β 1
+
Q Q 4
Q−α2−δ β 1−δ 1 Q−α 1 β
+ × + δ ×0
Q 2 2Q 2 − δ 2 − δ Q 2 2Q
 
1α Q−β β
= (1 − δ)(1 + δ) + (1 − δ)
4Q Q Q
 
1Q−α 1−δ β Q−β
+ 2 +
4 Q 2−δQ Q

168
 2
Finally the fact that E[(Ṽ − P1 )2 ] = 2δ π 1 (α) + 1−δ
4 is obvious and to obtain the relation between E[(Ṽ − P2 )2 ]
2
and π (α, β), we just have to develop the two expressions:

δ α 1 β α β
π 2 (α, β) = × [(1 − ) × (1 − ) + (1 − δ) (1 − )]
2 Q 2−δQ Q Q
δ α 1 β 1 α β
= × [1 − δ − +( − (1 − δ)) ]
2 Q 2−δQ 2−δ QQ
and
   
2 1α β 1 α δ β
E[(Ṽ − P2 ) ] = (1 − δ) (1 + δ) − δ + (1 − ) 1 −
4Q Q 4 Q 2−δQ
1 1 α δ β 1 α β
= + [−δ 2 − + δ( − (1 − δ)) ]
4 4 Q 2−δQ 2−δ QQ

169
170
Bibliography

[1] Back K. and Baruch S. (2007), Working orders in limit order markets and floor ex-
changes, Journal of Finance 62, 1589-1621.

[2] Biais B., Foucault T. and Moinas S. (2013), Equilibrium Fast Trading, Working Paper.

[3] Biais B., Hillion P. and Spatt C. (1995), An empirical analysis of the limit order book
and the order flow in the Paris Bourse, Journal of Finance 50, 1655-1689.

[4] Biais B., Martimort D. and Rochet J-C. (2000), Competing mechanisms in a common
value environment, Econometrica 68, 799-838.

[5] Biais B. and Weill P.-O. (2009), Liquidity shocks and order book dynamics, Working
Paper.

[6] Biais B., Hombert J. and Weill P.-O. (2012), Pricing and Liquidity with Sticky Trading
Plans, Working Paper.

[7] O. Brandouy, Barneto P. and Leger L.A. (2003), Asymmetric Information, Imitative
Behaviour and Communication : Price Formation in an Experimental Asset Market,
European Journal of Finance, 9, 393-419.

[8] Brogaard J., Hendershott T. and Riordan R. (2012), High Frequency Trading and Price
Discovery, Working Paper.

[9] Brunnermeier M. (2005), Information leakage and market efficiency, Review of Financial
Studies 18, 417-457.

171
[10] Carvalho C., Klagge N. and Moench E., The Persistent effects of a false news shock,
Federal Reserve Bank of New York Staff Reports.

[11] Chaboud A., Chiquoine B., Hjalmarsson E. and Vega C. (2009), Rise of the machines:
algorithmic trading in the foreign exchange market, Working Paper.

[12] Chordia T., Roll R. and Subrahmanyam A. (2000), Commonality in Liquidity, Journal
of Financial Economics 56, 3-28.

[13] Chordia T., Roll R. and Subrahmanyam A. (2001), Market Liquidity and Trading Ac-
tivity, Journal of Finance 56, 501-530.

[14] Colliard J-E. and Foucault T. (2012), Trading fees and efficiency in limit order markets,
Review of Financial Studies 25, 3389-3421.

[15] Daniel K., Hirshleifer D. and Subrahmanyam A. (1998), Investor psychology and security
market under- and overreactions, Journal of Finance 53, 1839-1885.

[16] Della Vigna S. and Pollet J. M. (2009), Investor inattention and friday earnings an-
nouncements, Journal of Finance 64, 709-749.

[17] Dow J., Goldstein I. and Guembel A. (2011), Incentives for information production in
markets where prices affect real investment, Working Paper.

[18] Duffie D., Garleanu N. and Pedersen L. (2005), Over-the-counter markets, Econometrica
73, 1815-1847.

[19] Duffie D., Garleanu N. and Pedersen L. (2007), Valuation in over-the-counter markets,
Review of Financial Studies 20, 1865-1900.

[20] Duffie D. (2010), Presidential address: asset price dynamics and slow moving capital,
Journal of Finance 65, 1237-1267.

[21] Ederington L.H., and Lee J. (1995), The Short-run dynamics of the price adjustment to
new information, Journal of Financial and Quantitative Analysis 30, 117-134.

172
[22] Engle R., Fleming M., Ghysels E and Nguyen G. (2011), Liquidity and Volatility in the
U.S. Treasury Market: Evidence From A New Class of Dynamic Order Book Models,
Working Paper.

[23] Fleming M. J. and Remolona E. M. (1999), Price formation and liquidity in the U.S.
treasury market: the response to public information, Journal of Finance 54, 1901-1915.

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

[25] Foucault T. (2010), Limit Order Markets, Encyclopedia of Quantitative Finance, Rama
Cont editor, John Wiley & Sons Limited.

[26] Foucault T., Kadan O. and Kandel E. (2005), Limit order book as a market for liquidity,
Review of Financial Studies 18, 1171-1217.

[27] Foucault T., Hombert J. and Rosu I. (2012), News trading and speed, Working Paper.

[28] Foucault T., Kadan O. and Kandel E. (2013), Liquidity cycles and make/take Fees in
electronic markets, Journal of Finance 68, 299-341.

[29] Foucault T. and Menkveld A. (2008), Competition for order flow and smart order routing
systems, Journal of Finance 63, 119-158.

[30] Foucault T., Roell A. and Sandas P. (2003), Market Making with Costly Monitoring:
An Analysis of the SOES Controversy, Review of Financial Studies 16, 345-384.

[31] Foucault T., Pagano M. and Roell A. (2013), Market Liquidity: Theory, Evidence, and
Policy. Oxford University Press.

[32] Froot K., Scharfstein D. and Stein J. (1992), Herd on the street: informational efficiencies
in a market with short-term speculation, Journal of Finance 47, 1461-1484.

[33] Gajewski J.-F. and Gresse C. (2007), Centralised Order Books versus Hybrid Order
Books: A Paired Comparison of Trading Costs on NSC (Euronext Paris) and SETS
(London Stock Exchange), Journal of Banking and Finance 31, 2906–2924.

173
[34] Glosten L. and Milgrom P.(1989), Bid, ask and transaction prices in specialist market
with heterogeneously informed trader, Journal of Financial Economics 13, 71-100.

[35] Glosten L. (1994), Is the electronic open limit order book inevitable?, Journal of Finance
49, 1127-1161.

[36] Goettler R., Parlour C. and Rajan U. (2005), Equilibrium in a dynamic limit order
market, Journal of Finance 60, 2149-2192.

[37] Goettler R., Parlour C. and Rajan U. (2009), Informed traders and limit order markets,
Journal of Financial Economics 93, 67-87.

[38] Green T. C. (2004), Economic news and the impact of trading on bond prices, Journal
of Finance 59, 1201-1233

[39] Gresse C. (2006), The Effect of crossing-network trading on dealer market’s bid-ask
spreads, European Financial Management 12, 143-160.

[40] Gresse C. (2012), Market fragmentation in Europe: Assessment and prospects for market
quality, Foresight project on The Future of Computer Trading in Financial Markets -
Driver Review 19.

[41] Gross-Klussmann A. and Hautsch N. (2011), When machines read the news: Using
automated text analytics to quantify high frequency news-implied market reactions,
Journal of Empirical Finance 18, 321-340.

[42] Hasbrouck (2007), Empirical Market Microstructure: The Institution, Economics, and
Econometrics of Securities Trading. Oxford University Press.

[43] Hasbrouck J. and Saar G. (2009), Technology and liquidity provision: The blurring of
traditional definitions, Journal of Financial Markets 12, 143-172.

[44] Hasbrouck J. and Saar G. (2011), Low-Latency trading, Working Paper.

[45] Hasbrouck J. (2013), High frequency quoting: short-term volatility in bids and offers,
Working Paper.

174
[46] Hendershott T., Jones C. and Menkveld A. (2011), Does algorithmic trading improve
liquidity?, Journal of Finance 66, 1-33.

[47] Hollifield B., Miller R. and Sandas P. (2004), Empirical analysis of limit order markets,
Review of Economic Studies 71, 1027-1063.

[48] Hollifield B., Miller R., Sandas P. and Slive S. (2006), Estimating the gain from trade
in limit order markets, Journal of Finance 61, 2753-2804.

[49] Jovanovic B. and Menkveld A. (2010), Middlemen in limit-order markets, Working Pa-
per.

[50] Kim O. and Verrecchia R. E. (1991), Trading volume and price reactions to public
announcements, Journal of Accounting Research 29, 302-21.

[51] Kim O. and Verrecchia R. E. (1994), Market liquidity and volume around earnings
announcements, Journal of Accounting and Economics 17, 41-67.

[52] Kyle A. (1985), Continuous auctions and insider trading, Econometrica 53, 1315-1336.

[53] Lagos R. and Rocheteau G. (2009), Liquidity in asset markets with search frictions,
Econometrica 77, 403-426.

[54] Lagos R., Rocheteau G. and Weill P.-O. (2011), Crises and liquidity in over-the-counter
markets, Journal of Economic Theory 146, 2169-2205.

[55] Mankiw, N. G. and Reis R. (2002), Sticky Information Versus Sticky Prices: A Proposal
to Replace the New Keynesian Phillips Curve, Quarterly Journal of Economics 117,
1295-1328.

[56] Mondria J. (2010), Portfolio Choice, Attention Allocation, and Price Comovement, Jour-
nal of Economic Theory 145, 1837-1864.

[57] Pagnotta E. (2010), Information and liquidity trading at optimal frequencies, Working
Paper.

[58] Pagnotta E. and Philippon T. (2012), Competing on speed, Working Paper.

175
[59] Parlour C. (1998), Price dynamics in a limit order market, Review of Financial Studies
11, 789-816.

[60] Parlour C. and Seppi D. (2003), Liquidity-based competition for order flow, Review of
Financial Studies 16, 301-343.

[61] Parlour C. and Seppi D. (2008), Limit order markets: a survey, Handbook of Financial
Intermediation & Banking.

[62] Peng L. and Xiong W. (2006), Investor attention, overconfidence and category learning,
Journal of Financial Economics 80, 563-602.

[63] Rosu I. (2009), A Dynamic model of the limit order book, Review of Financial Studies
22, 4601-4641.

[64] Rosu I. (2010), Liquidity and information in order driven markets, Working Paper

[65] Securities and Exchange Commission (2010), Concept release on equity market structure,
Federal Register 75, 3594-3614.

[66] Seppi D. (1997), Liquidity provision with limit orders and a strategic specialist, Review
of Financial Studies 10, 103-150.

[67] Sims C. (2003), Implication of Rational Inattention, Journal of Monetary Economics

50, 665-690.

[68] Tetlock P. C. (2010), Does public financial news resolve asymmetric information? Review
of Financial Studies 23, 3520-3557.

[69] Van Nieuwerburgh S. and Veldkamp L. (2009), Information immobility and the home
bias puzzle, Journal of Finance 64, 1187-1215.

[70] Vayanos D. and Weill P.-O. (2008), A Search-based theory of the on-the-run phe-
nomenon, Journal of Finance 63, 1351-1389

[71] Weill P.-O. (2007), Leaning against the wind, Review of Economic Studies 74, 1329-1354.

176
[72] Weill P.-O. (2008), Liquidity premia in dynamic bargaining markets, Journal of Eco-
nomic Theory 140, 66-96.

177
 Essays in Financial Market Microstructure

This dissertation is made of three distinct chapters. In the first chapter, I show that traditional
liquidity measures, such as market depth, are not always relevant to measure investors’ welfare. I
build a limit order market model and show that a high level of liquidity supply can correspond
to poor execution conditions for liquidity providers and to a relatively low welfare. In the second
chapter, I model the speed of price adjustments to news arrival in limit order markets when in-
vestors have limited attention. Because of limited attention, investors imperfectly monitor news
arrival. Consequently prices reflect news with delay. This delay shrinks when investors’ attention
capacity increases. The price adjustment delay also decreases when the frequency of news arrival
increases. The third chapter presents a joint work with Thierry Foucault. We build a model to
explain why high frequency trading can generate mini-flash crashes (a sudden sharp change in the
price of a stock followed by a very quick reversal). Our theory is based on the idea that there is
a trade-off between speed and precision in the acquisition of information. When high frequency
traders implement strategies involving fast reaction to market events, they increase their risk to
trade on noise and thus generate mini flash crashes. Nonetheless they increase market efficiency.

Keywords: liquidity, welfare, limit order market, news, limited attention, imperfect market mon-
itoring, high frequency trading, mini flash crash, market efficiency.

Essais en Microstructure des Marchés Financiers

Cette thèse est composée de trois chapitres distincts. Dans le premier chapitre, je montre que les
mesures de liquidité traditionnelles, tels que la profondeur du marché, ne sont pas toujours perti-
nents pour mesurer le bien-être des investisseurs. Je construis un modèle de marché conduits par
les ordres et montrent qu’une offre de liquidité élevée peut correspondre à de mauvaises conditions
d’exécution pour les fournisseurs de liquidité et à un bien-être relativement faible. Dans le deuxième
chapitre, je modélise la vitesse des ajustements de prix à l’arrivée de nouvelles dans les marchés
conduits par les ordres, lorsque les investisseurs ont une capacité d’attention limitées. En raison de
leur attention limitée, les investisseurs suivent imparfaitement l’arrivée de nouvelles. Ainsi les prix
s’ajustent aux nouvelles après un certain délai. Ce délai diminue lorsque le niveau d’attention des
investisseurs augmente. Le délai d’ajustement des prix diminue également lorsque la fréquence, à
laquelle les nouvelles arrivent, augmente. Le troisième chapitre présente un travail écrit en collab-
oration avec Thierry Foucault. Nous construisons un modèle pour expliquer en quoi le trading à
haute fréquence peut générer des «mini flash crashes» (un brusque changement de prix suivie d’un
retour très rapide au niveau antérieur). Notre théorie est basée sur l’idée qu’il existe une tension en-
tre la vitesse à laquelle l’information peut être acquise et la précision de cette information. Lorsque
les traders à haute fréquence mettent en oeuvre des stratégies impliquant des réactions rapides à
des événements de marché, ils augmentent leur risque de réagir à du bruit et génèrent ainsi des
«mini flash crashes». Néanmoins, ils augmentent l’efficience informationnelle du marché.

Mots clefs: liquidité, bien-être, marché conduit par les ordres, nouvelles, attention limitée,
surveillance de marché imparfaite, trading haute fréquence, efficience informationnelle de marché.