Académique Documents
Professionnel Documents
Culture Documents
Chaiyakorn Yingsaeree
Most equity and derivative exchanges around the world are nowadays organized as order-driven
markets where traders execute their trades by submitting either market orders or limit orders. A
market order provides an immediate execution at the best price available in a limit order book
upon the order arrival. With a limit order, a trader can improve his execution price relative
to the market order price but the execution is neither immediate nor certain. Determining an
appropriate order type of a particular trade is ultimately a fundamental problem faced everyday
by all investors in such markets, and a prerequisite for any approach for making such decision is a
model of limit order execution times and the associated execution probability. However, research
into how to model such probability is very limited, and primarily focused on the probability that
a newly submitted order will be executed in a specified period of time. In addition, these models
may not be appropriate to develop a dynamic trading strategy since only information about newly
submitted order can be inferred from those models.
The objective of this research is to develop a computational model of limit-order executions
in an order driven market that can be used to predict the probability that a giving limit order
will be executed in a specified period of time as well as investigate the optimal way to utilize the
developed model to make order placement and trading decision in algorithmic trading systems.
Unlike traditional models that focus on only newly submitted limit orders, we focus on modeling
the execution probability of all orders in the order book with the aim to gain more insight on the
determinants of the execution probability and how this probability change over time. The insight
gained from this model is then utilized to solve order placement decision problem faced by an
investor as well as to derive optimal trading strategies for several decision problems in algorithmic
trading system.
During the first year of my study, I produced one review paper that presents the overview of
the emerging field of Computational Finance. I also designed a virtual hedge fund system and
supervised a group of MSc student to develop it. For the works related to my research, I developed
a program to collect real time order book information from Reuters data feed as well as developed
a program to calculate the empirical distribution of execution probability from trade information.
The future work includes developing simulation model of order driven markets to generate data
for control experiments, developing new probability model of limit-order execution by utilizing
non-parametric techniques from machine learning community (e.g. Bayesian neural network and
relevance vector machine), developing a platform to develop an automated trading system, and
developing an order submission strategy that utilizes the developed execution probability model.
This work is in collaboration with Deutsch Bank, and I expected to spend a period of internship
at Deutsch Bank to test the developed model with the real environments.
1
Contents
1 Introduction 4
1.1 Motivations from the literature and industry . . . . . . . . . . . . . . . . . . . . . 4
1.2 Objectives of this research . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3 Outline of this report . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2 Background 8
2.1 Market architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.1.1 Limit order markets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.1.2 Dealers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.1.3 Auctions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2 The limit order book . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.3 Algorithmic trading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.4 Trade execution strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.4.1 Choice of trading venue . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.4.2 Choice of trade schedule . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.4.3 Choice of order type . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3 Literature Review 16
3.1 Order submission strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.1.1 Static order submission strategies . . . . . . . . . . . . . . . . . . . . . . . . 17
3.1.2 Dynamic order submission strategies . . . . . . . . . . . . . . . . . . . . . . 20
3.1.3 Framework for order submission strategy . . . . . . . . . . . . . . . . . . . 21
3.2 Limit-order execution models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.2.1 Execution probability models . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.2.2 Execution time models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
4 Research Proposal 31
4.1 Hypothesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.2 Problem statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.3 Expected contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
4.4 Scope of the research . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
4.5 Validation of contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2
CONTENTS 3
5 Research Plan 33
5.1 Tools for real-time data collection . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
5.2 Simulation models of pure limit order book markets . . . . . . . . . . . . . . . . . 34
5.3 Platform for algorithmic trading system development . . . . . . . . . . . . . . . . . 34
5.4 Computational model of limit-order execution . . . . . . . . . . . . . . . . . . . . . 35
5.5 Order submission strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
5.6 Timetable . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
6 Work to Date 37
6.1 Banking science virtual trading platform . . . . . . . . . . . . . . . . . . . . . . . . 37
6.2 Empirical distribution of execution probability . . . . . . . . . . . . . . . . . . . . 38
Chapter 1
Introduction
Advances in telecommunication and computing technologies during the past decades have created
a trend toward the automation of financial markets which are now becoming increasingly global,
dynamic and, thus, complex. As a result, it has become essential for their participants to have the
most innovative technology in order to deliver critical solutions and improve productivity so as to
enhance their profits and competitiveness. The advent of electronic trading venues together with
the ability to store a huge amount of data enables us to capture a detailed record of all events
occurring in the market. This detailed record are gradually becoming available to researches
and, thus, enables us to analyze the characteristic of the markets, such as intraday order flows
and imbalances in supply and demand, as well as the behaviors of their trader, such as trader’s
order submission strategy, as a function of contemporaneous and past states of the market. The
insights gained from these scientific investigations can then be applied to devise new ways of profit
making and reducing the cost of trading by optimizing the timing, size, and other characteristics
of submitted order.
4
CHAPTER 1. INTRODUCTION 5
order may sometimes be executed only partially or not at all. In addition, a trader who submits
limit orders may also offset by picking-off risk2 if he does not monitor the market continuously.
Moreover, if the submitted limit order is not executed, the trader must decide when, and how, to
resubmit the order. Thus, determining an appropriate order submission strategy for a particular
trade is ultimately a fundamental problem faced everyday by all traders participating in such
markets, and the solution to this problem is of paramount significance not only to all traders,
particularly to institutional investors who frequently trade large volumes of shares representing
a quarter or more of the whole market volume, but also to market microstructure literature that
analyze the rational for, and the profitability of, limit order trading as well as the characteristics
and dynamic behaviors of the limit order market. In addition, a methodology to solve such
problem can also be utilized as a building block to solve many decision problems in algorithmic
trading literature.
The desire to understand traders’ order submission strategies has inspired a wide range of
theoretical and empirical research. On the theoretical side, many order submission models have
been proposed and examined to analyze the rational for, and the profitability of, limit order
trading as well as the characteristics and the dynamic behaviors, of the limit order market (e.g.
[11, 22, 39, 16, 21, 17]). Empirical approaches, on the other hand, analyze history of trades and
quotes that occur in the exchange to achieve the same goals. Although recent empirical studies
[7, 19, 41, 44, 6, 33, 20, 18, 10, 34] indicate that trader’s decision of when and which order type
to submit is significantly influenced by the state of the order book (e.g. the queue volume, the
market depth, and the inside spread) as well as its dynamic (i.e. recent changes to the order
book), little attention has been paid to develop an order submission strategy that utilizes this
information to optimize trade execution. Notable exceptions are Nevmyvaka et al. [37, 36] who
propose a quantitative method that allows traders to optimally price their limit orders with the
aim to minimize trading costs based on the state of the order book.
While their results indicate that incorporating market conditions into the decision can greatly
improve the execution result, their works are loosely related to traditional models since they utilize
reinforcement learning to directly optimize order execution without any consideration about the
tradeoff suggested by theoretical order submission models. As a result, the main drawback of their
approach is that, when traders’ trading objectives change, new reinforcement learning models have
to be constructed and trained to obtain an appropriate strategy. To avoid this inconvenience,
it might be better to incorporate information about market conditions into traditional order
submission models so that, after the model is calibrated, traders can utilize the model regardless
of their objectives.
Although traders’ order submission decisions can be explained by several factors, theoretical
models generally view these decisions as a tradeoff between the expected profit and the free-
trading option. The expected profit depends on the execution price and the execution probability
of a limit order, while the value of free trading depends on the arrival probability of adverse
information which may move the price through the submitted limit order. Undoubtedly, one of
millisecond, and our execution price may be affected by the submission of market orders from competing traders
as well the revision of the price and volume at the best quote.
2
Picking-off risk, also known as adverse selection cost and winner’s curse problem, is associated with the well-
known concept that limit orders are free options to other traders [12] and these options will become mispriced
as soon as the fundamental of the instrument is changed. Hence, traders who submit limit orders may expose
potentially large losses if they do not constantly update their orders to reflect these changes.
CHAPTER 1. INTRODUCTION 6
the most important factors in valuing such tradeoff is a model of limit order execution times and
the associated execution probability [11, 39, 16, 21, 42, 1, 26]. The main reason for this is that
the expected profit of traders who decide to trade via limit orders is an increasing function of the
execution probability. The larger the execution probability, the shorter the expected waiting time,
and thus the smaller the expected adverse selection cost. In addition, recent empirical findings
indicate that there is a strong relationship between this probability and the state of the order
book. In particular, Omura et al. [38] report that the probability of execution of limit orders
on Tokyo Stock Exchange (TSE) is lower when there are open ticks between the bid-ask spread
and when the depth, the quantities of limit orders at the best price, of the same side is high.
Conversely, the execution probability is higher when the depth of the opposite side of the book
is high. This is in accordance with Biais et al. [7] who indicate that, for the Paris Bourse CAC
system, order flow is concentrated near the best quotes with depth somewhat larger at nearby
quotes. When depth at the best quote is large, traders rapidly place limit orders within the
spread, while traders place market orders when the spread is small. All of these suggest that it is
sensible to model the execution probability of limit orders using the state of the order book and
utilize this model to derive the optimal order submission strategy. However, to the best of our
knowledge, no research effort on this topic has been reported in the literature before.
to derive the optimal order submission strategy for several trading problems faced in algorithmic
trading systems.
Background
This chapter presents background information on a number of key concepts in the areas that
this research spans. Particularly, the information about trading mechanisms frequently used in
financial markets is presented with the main emphasis on the limit order book markets which is
the main market studied in this research. The broad area of algorithmic trading is also reviewed
to place our work in a well defined context and to outline a rich picture of business reality for
which the extended version of this research could be considered in the future work.
8
CHAPTER 2. BACKGROUND 9
if two orders have the same limit price, the one which was entered into the system first will has
the priority usually has second priority. Other priorities such as order quantity are also used in
some markets (e.g. LIFFE). A list of unexecuted limit orders stored in the system constitutes a
limit order book. Since limit orders can be modified or canceled at any time, the order book is
dynamic and sometimes it can change rapidly especially in active markets.
Instead of using limit orders, a trader may desire that an order must be executed at the
market, i.e., at the best available price. To achieve this, the trader can submit a market order, an
unpriced order which will be executed immediately against the best order, to the market. If the
order quantity is larger than the quantity available at the single best price on the book, the order
will walk the book resulting in partial executions at progressively worse prices until the order is
fully filled.
A market might have multiple limit order books, each managed by different broker. Limit
order books might also be used in conjunction with other mechanisms. When all trading occurred
through a single book, the market is said to be organized as a consolidated limit order book
(CLOB) which is used for actively traded stocks in most Asian and European markets.
2.1.2 Dealers
Dealer markets
A dealer is basically an intermediary who is willing to act as a counterparty for the trades of his
customers. A trade in a dealer market, such as FX market, usually starts with a customer calling
a dealer to ask for its price quotes (i.e. the dealer’s bid and ask price), and, then, the customer
may buy at the dealer’s ask, sell at the dealer’s bid, or do nothing. Unlike limit order markets
where a buyer who thinks the best price in the book is unreasonable can place his or her own
bid, a buyer in dealer markets does not have a possibility to do that. Dealer markets are also
usually characterized by low transparency since dealers usually provide quotes only in response
to customer inquiries and these are not publicly visible.
In addition to dealer-customer interactions, interdealer trading is also important for conducting
dealer business. Since the incoming buy and sell orders that a particular dealer sees are usually
imbalanced, accommodating these customer needs may leave the dealer with an undesired long
or short position. In such case, the dealer may attempt to sell or buy in the interdealer market
to balance its position. Nowadays, some of these interdealer markets, FX market, are conducted
via a limit order book, such as Electronic Broking Services (EBS) and Reuters Dealing 3000 Spot
Matching (D2).
possibly the NYSE specialist who has many roles and responsibilities but an important one is to
maintain a two-sided market when there is nothing on the limit order book and on one else on
the floor bidding or offering.
With the advent of electronic order management system, the competitive position of dealers
and other intermediaries has weakened since, nowadays, customers can update and revise their
limit orders rapidly enough to respond to market conditions. Hence, they can quickly supply
liquidity when it is profitable to do so and quickly withdraw their bids and offers when markets
are volatile. As a result, the presence of a dealer to maintain a two-sided market is considerably
diminished from most of financial markets. However, dealers today serve another useful function
in facilitating large (block) trades especially in the block market, also called the upstairs market.
When an institution contact a dealer to fill a large order, the dealer can act as a counterparty, try to
locate a counterparty for the full amount, work the order over time, or some combination of these.
The dealer’s advantage here thus lies in access to capital, knowledge of potential counterparties,
and expertise in executing large order overtimes, also known algorithmic systems.
2.1.3 Auctions
When multiple buyers and sellers are concentrated in one venue at one time, trades may not need
to be coordinated since agents can contact each other sequentially to strike bilateral bargains.
However, the result obtained from such approaches may not be economically efficient since many
participants will execute their trades at prices worse than the best price realized over the entire
set of trades. To avoid this problem, a single-price clearing, which is generally implemented with
a single-price double-sided auction should, could be employed. In this mechanism, supply and
demand curves are constructed by ranking bids and offers from all participants, and the clearing
price is usually determined by maximizing the feasible trading volume. The double-sided auction
is widely used in security markets especially for low activity securities. The Euronext markets, for
instance, conduct auctions once or twice per day depending on the level of interest. Double-side
auctions are also usually used to open continuous trading sessions (e.g. Euronext, Tokyo Stock
Exchange, and NYSE) and, also, at the close of continuous trading sessions.
Although most auctions in secondary markets are double-sided, single-sided auctions are
widely used in primary markets. These include the U.S. Treasury debt markets, and most U.S.
municipal bond offerings. They are also used, though not as often, for initial issues of equity.
In summary, financial markets have various architectures. Some of the main characteristics distin-
guish them are presence or lack of intermediation and continuous or periodic trading. Intermedi-
ated markets employ market markers, dealers, or specialist, who determine price quotes and act as
a counterparty in each trade. Consequently, these markets are usually referred to as quote-driven
markets due to the quote setting function of the dealers. In non-intermediated markets, trading
does not involve intermediates but submitted orders are stored, matched, and executed via the
limit order book. These markets are usually referred to as order-driven markets since the whole
trading process is determined by submitted orders. The second characteristic determines if trades
are executed continuously during a trading session or only at certain points in time. These two
modes are called a continuous double auction and a periodic auction respectively.
CHAPTER 2. BACKGROUND 11
• The systems that automate the first step of the trading process, namely the trading signal
generation. Thus, human intervention is required for the last two tasks of the trading
process, which are the trading decision and the execution of the trade.
• The systems that automate the trade execution which is last step of the trading process.
The aim of the execution algorithm is often focuses on placing and managing orders in the
market in order to minimize the trading cost. Using execution algorithms leaves the first
two steps to the human trader.
• The systems that combine the first two categories but leaving the trading decision to the
human trader.
• The fully automated systems, often referred to as black-box trading, that automate all steps
in the trading process.
Hence, most algorithmic trading systems consist of two main parts: determining when to trade
and how to trade. Determining when to trade is the analytic part of the strategy which revolves
CHAPTER 2. BACKGROUND 13
around watching the changing market data and detecting opportunities within the market. For
example, consider a pair trading strategy that examines pairs of financial instruments that are
known to be statistically correlated. Normally statistically correlated instruments are likely to
move together. When these instruments break correlation, the trader may buy on and sell the
other at premium with the hope to gain profit when both instruments become correlated again. In
this case, the algorithm involves monitoring for any changes in the price of both instruments and
then recalculating various analytics to detect a break in correlation. Another example is market
making strategy which tries to place a limit order to sell above the current market price or buy
a limit order below the current price in order to benefit from the bid-ask spread. In addition,
any sort of pattern recognition or predictive model can also be used to initiate the trade. Neural
networks and genetic programming have been extensively used to create these models.
Determining how to trade focuses on placing and managing orders in the market. At the lowest
level, this involves with determining the suitable choice of order type (i.e. limit and market order)
for each trade. This choice is no simple matter and requires some sophistications since market
orders are executed immediately but incur substantial price impact while limit orders incur no
price impact but may not be executed immediately, if at all. A higher level problem involves
with breaking up a large order into smaller orders and placing them into the market over time.
The benefit of this is that large orders have a major impact in moving the market while smaller
orders are more likely to flow under the market’s radar, and subsequently have less impact to the
market. In addition, when an interested instrument is traded in multiple exchanges, an execution
strategy also needs to determine where the order should be submitted to. Since this research is
more related to this issue, more detailed information about it will be discussed in the next section.
Normally, the trader may want to submit the order to the market whose characteristics suit his
requirements most. Some of the most important characteristics the trader usually considers are
liquidity, trading mechanism and degree of trader’s anonymity.
A financial instrument in a particular market is considered liquid if the volume of trades and
orders of that instrument is large. Liquidity is important because high liquid market is usually
associated with fast trade execution and low transaction costs. Thus, all other things being equal,
the trader would prefer to submit his orders to the market with the most liquidity.
A trading mechanism employed in the market is also an important characteristic the trader
usually considers before making trade execution decision since each mechanism has it own ad-
vantages and disadvantages. As discussed in section 2.1, trades in a continuous double auction
market are executed continuously during a trading session, while trades in a periodic auction are
executed only at certain points in time. As a result, it is more appropriate to trade in a continuous
double auction market when immediacy is required. However, trades in the periodic auction have
lower price volatility when compares to trades in the continuous double auction [13].
In the case that trader’s order is too large to be executed instantaneously without an unwanted
price impact, the trader’s action will be influenced by his trading motivation. If trading is infor-
mation motivated, it might be more appropriate to carry out in a market that offers anonymity.
In addition, trader may also break up the large order into a sequence of smaller orders and submit
them to the market over a period of time with the aim to reduce price impact by hiding from
other participants the fact that all those orders were originated by the same trader. On the other
hand, a liquidity-motivated trader whose motivation is not information related is not necessary
to do that and may submit his order to an upstairs market directly.
prices for the day). The second step is to specify the degree of risk-aversion (i.e. how much we
penalize variance relative to expected cost) which indicates the level of trading aggressiveness or
passiveness. Aggressive trading is associated with higher cost and less risk while passive trading
is associated with lower market impact and higher risk. Market impact, the degree to which an
order affects the market price, consists of permanent impact cost due to information leakage of
the order and temporary impact cost due to the liquidity and immediacy needs of the investor.
These market impacts are usually approximated by fitting some parametric functions (e.g. linear
and power laws function) using historical data. In addition, these functions can be both time
dependent and time independent. To specify the dynamics of future market prices, arithmetic
random walk is the most popular model. Giving specifications of all these factors, an optimal
trading strategy for a specific trading objective may be obtained by solving the corresponding
stochastic dynamic optimization problem.
1
The uncertainty of execution price is usually caused by rapid changes to the limit order book during a period
between order submission and trade execution. In an electronic market, multiple events can happen within a
millisecond, and our execution price may be affected by the submission of market orders from competing traders
as well the revision of the price and volume at the best quote.
2
Picking-off risk, also known as adverse selection cost and winner’s curse problem, is associated with the well-
known concept that limit orders are free options to other traders and these options will become mispriced as soon
as the fundamental of the instrument is changed. Hence, traders who submit limit orders may expose potentially
large losses if they do not constantly update their orders to reflect these changes.
Chapter 3
Literature Review
In this section, we present and analyze the current state of research of order submission strategies
and limit-order execution models. In section 3.1, the order submission problems frequently faced
by traders are presented and formalized. The existing techniques for solving those problems are
also presented in this section. The method for modeling the execution probability, which is one of
the most important information to make order submission decision, are presented in section 3.2.
Finally, our conclusion remarks about the existing techniques and the way to improve them are
given in section 3.3.
16
CHAPTER 3. LITERATURE REVIEW 17
who decide to trade via limit orders is an increasing function of the execution probability.
In reality, an order submission strategy that a trader selects normally depends on the trading
problem he tries to solve. As suggested by Harris [24], three main trading problems frequently
faced by traders are: (i) the liquidity trader problem considering how a liquidity trader who must
fill his order before some deadlines should trade, (ii) the informed trader problem considering
how an informed trader who receives a single signal about asset value should trade before his
information becomes obsolete and (iii) the value-motivated trader problem considering how a
trader who continuously estimates security value should trade. Specifically, liquidity traders must
fill their order before some deadlines which may arise when they need to invest or disinvest their
cash flow. The main objective of these traders is to obtain the best price for their trades by
carefully choosing their order submission strategies. On the other hand, informed traders, who
have private information about the underlying value of the asset, want to profitably trade on
their information. Although informed traders do have a trading deadline, which is the time
their information becomes obsolete, they did not have to fill their orders before the deadline like
liquidity traders. In fact, informed traders will trade only when it is profitably to do so. Like
informed traders, value-motivated traders also have private information about the values of asset.
However, unlike informed traders, they do not have a specific deadline and are assumed to trade
repeatedly in the market since they receive continuous information about the values.
Order submission strategies previously proposed in the literature can be classified into two
main categories: static strategies and dynamic strategies. Static order submission strategies view
this problem as a one-shot game where traders can make their order decision only once. If they
decide to submit a limit order, no additional change can be made to the order and it will stay
in the order book until it is executed or the end of the trading period is reached. Conversely,
dynamic order submission strategies allow traders to cancel or make changes to their orders before
the order expires or is executed [24, 43]. Empirically, traders change their order submission as
market conditions change. They continuously monitor the market and make appropriate changes
to their orders whenever necessary. For example, to reduce the execution risk, they may convert
their limit order to market order when the demand for immediacy increases. They may also
reprice or cancel their limit orders when the underlying value of the asset changes to manage
the adverse selection cost. Hence, it is more appropriate to model this decision with dynamic
strategies than with static strategies.
This section briefly describes related work in order submission strategy. Static order submis-
sion strategies are presented in section 3.1.1, while dynamic strategies are discussed in section
3.1.2. A general framework that can be utilized to describe theoretical order submission strate-
gies will be described in section 3.1.3. The overview of all models discussed in this section is
summarized in table 3.1.
Table 3.1: Overview of order submission strategies reviewed in this section. Each model is char-
acterized by the trading problem it tries to solve, whether it is static or dynamic strategy and the
market variables that it utilizes.
that liquidity traders have to fill their order before the deadline; thus, if liquidity traders decide
to submit a limit order and their orders are not executed, they have to submit market orders to
execute the trade when their deadlines are approached. On the other hand, informed traders will
submit market orders to fill the traders only when it is still profitably to do so.
their limit orders to minimize trading costs and control corresponding risks. The importance of a
number of microstructure variables (e.g. order size, time window and liquidity) is also highlighted.
current best quote by large amounts, when the order arrival rate is low or when the proportion of
patient traders is larger. As a result, markets with a high proportion of patient traders or a small
order arrival rate are more resilient. Also, a reduction in the tick size reduces market resiliency,
and, in some case, increases the average spread. Their analysis also yields several testable pre-
dictions: (i) a positive relationship between inter-trader durations and market resiliency, (ii) a
negative relationship between the order arrival rate and market resiliency, (iii) a joint decline of
limit order aggressiveness and market resiliency at the end of the trading session and (iv) limit
order traders submit more (less) aggressive orders when the spread is large if patient (impatient)
traders dominate the trading population.
Lillo [31] considers the problem of the optimal limit order price for a financial asset in the
framework of utility maximization. The analytical solution of the problem gives insight on the
origin of the recently empirically observed power law distribution of limit order prices. In the
framework of the model, the most likely proximate cause of this power law is a power law hetero-
geneity of traders’ investment time horizons.
are very small. If monitoring of open orders is expensive or if the trader is risk averse, distant
order placement strategies will not be optimal. The only exception to this rule is for traders who
believe that prices are mean-reverting. They may place limit orders far from the market to benefit
if prices move far from fundamental values.
Nevmyvaka, Feng and Kearns [36] present the first large-scale empirical application of re-
inforcement learning to the problem of trade execution. In their problem, the goal is to sell
(respectively, buy) V shares of a given stock within a fixed period of time in a manner that
maximizes the revenue received (respectively, minimizes the capital spent). Their results indi-
cate that introducing market variables into the model can greatly improve the execution result
and reinforcement learning can indeed result in significant improvement over simpler forms of a
single-period optimization model.
Wang and Zhang [45] present a dynamic focus strategies that incorporate a series of market
orders of different volume into the limit order strategy and dynamically adjusts their volume
by monitoring state variable such as inventory and order book imbalance in real-time. The
sigmoid function is suggested as the quantitative model to represent the relationship between the
state variables and the volume to be adjusted. The empirical results indicate that the dynamic
focus strategies can outperform the limit order strategy, which does not adopt dynamic volume
adjustment.
Slive [43] derives the optimal dynamic order submission strategies of a trader in a limit order
market who has the ability to actively monitor his order and use cancellations and order changes
to mitigate the adverse selection and execution risks inherent in limit orders. His results suggest
that the ability to implement a dynamic strategy has a large impact on the payoffs to submit
limit orders and on limit order submission strategy of a trader. After calibrating the parameters
to a stock on the Vancouver Stock Exchange, profits from limit order submission are 48% higher
when implementing a dynamic strategy compared to a one-shot strategy. Cancellations and order
changes are used to avoid adverse selection by moving orders when the underlying value changes.
Order changes are used to mitigate execution risk by converting to a market order when the
probability of execution declines.
Although recent empirical studies [7, 19, 41, 44, 6, 33, 20, 18, 10, 34] indicate that trader’s
decision of when and which order type to submit is significantly influenced by the state of the
order book (e.g. the queue volume, the market depth, and the inside spread) as well as its dynamic
(i.e. recent changes to the order book), little attention has been paid to develop a dynamic order
submission strategy that utilizes this information to optimize trade execution. Notable exceptions
are Nevmyvaka et al. [36] and Wang and Zhang [45] who propose quantitative methods that allow
traders to optimally price their limit orders with the aim to minimize trading costs based on the
state of the order book. While their results indicate that incorporating market conditions into the
decision can greatly improve the execution result, their works can only be applied to solve trading
problem of liquidity traders. Thus, all of these suggest us to develop a new order submission
model that incorporate this information and can be utilized to solve all three trading problems
as in Harris [24].
ered by theoretical order submission model. A single-period model that can be utilized to derive
static order submission strategy is firstly presented. Then, a multi-period model for deriving
dynamic strategy is discussed.
Single-period Model
In single-period setting, the order submission decision is viewed as a one-shot game where traders
can make their decision only once. If he decides to submit a limit order, no additional change
can be made to his order and it will stay in the order book until it is executed or the end of
the trading period. Assume that a trader wants to transect q shares of a particular financial
instrument within a specific trading horizon T . At the time of decision t, the investor is faced
with three possible choices:
1. Do nothing.
2. Submit a market order to immediately transact q shares at the current market price pM
t .
To model these choices, the trader needs to specify the objective of the trading by defining two
payoff functions: a function fe (p, q) that defines the payoff he will get when he transact q shares
at a price p, and a function fnt (q) that defines the cost he needs to pay if he is not able to trade
q shares of that instrument. By specifying the form of these two functions, all three problems
mentioned in the beginning of this section can be formalized in the same framework. The liquidity
traders who has a responsibility to fill the trade before the deadline can set fnt (q) = −∞ so that
the strategy should always fill the trade when the deadline is reached. Informed traders may set
fe (p, q) to be the difference between the execution price and their private value of the asset. For
valued-motivated traders, these two functions will be time dependent.
If the trader chooses to do nothing, he will pay a cost of not trading which is equal to fnt (q).
If the trader chooses to submit a market order, his order will be immediately executed and he will
receive a payoff of fe (pMt , q). If the trader chooses to submit a limit order, one of three events can
occur: the limit order may be fully executed, partially executed, or not executed at all. When
the order is not fully executed, he will have a choice either to submit a market order to transect
the unexecuted shares or to do nothing depended on his trading objective. If the limit order is
executed for q e shares, the payoff the trader get will be:
fl (pL e L e M e e
t , q, q ) = fe (pt , q ) + max{fe (pT , q − q ), fnt (q − q )} (3.1)
This payoff is a random variable since its value depends on the number of executed share q e and
the future market price pMT , whose values are not known beforehand, and this expected payoff
can be calculated by taking expectation over q e and pM
t as in the following equation:
Eqe ,pM
t
[fl (pL e L e
t , q, q )] = Eq e [fe (pt , q )] + max{Eq e ,pM
t
[fe (pM e e
T , q − q )], Eq e [fnt (q − q )]} (3.2)
If the payoff function fe (p, q) and fnt (q) are linear in q, which is usually the case, the expected
payoff the trader gets from submitting a limit order will be:
CHAPTER 3. LITERATURE REVIEW 23
Eqe ,pM
t
[fl (pL e L e M e e
t , q, q )] = fe (pt , Eq e [q ]) + max{EpM [fe (pT , q − Eq e [q ]]), fnt (q − Eq e [q ])} (3.3)
T
To simplify this computation, previous works usually assume that when a limit order is executed,
it is satisfied fully at the stated price (except for [27]). Given a vector of variables describing
market conditions Xt , the expected payoff in this case is reduced to:
UT,Xt (pL L e L L L L
t ) = Eq e [fl (pt , q, q )] = P (pt , q, T, Xt )fe (pt , q) + (1 − P (pt , q, T, Xt ))fne (pt , q) (3.4)
where P (pLt , q, T, Xt ) is the execution probability that a limit order to transact q shares at price
L
pt will be fully executed before the end of the trading horizon T under the market condition Xt ,
and fne (pL
t , q) = max{EpM [fe (pMT , q)], fnt (q)}.
T
To determine the optimal order choice, the trader will select the choice that maximizes their
expected payoff. Although there are three possible choices (i.e. do nothing, submit a market order,
and submit a limit order), the expected payoff of all these choices can be represented by using the
same expected payoff equation UT,Xt (pL L
t ) with different value of the limit price pt . In the case of
buying, the do nothing choice can be represented by very low limit price so that the probability
of execution is zero. Accordingly, the market order choice can be represented by setting limit
price to be higher than current market price pM t so that the probability of execution is equal to
one. Hence, the optimal choice for the trader is the value of pL L
t that maximizes UT,Xt (pt ) , i.e.,
pbL
t ≡ arg maxpL t
[UT,Xt (pL L
t )]. In order to explicitly derive the functional form of UT,Xt (pt ) we need
to find the expression for P (pL t , q, T, Xt ) which will be the main subject of the next section.
Multi-period Model
In multi-period setting, the trading horizon T is further divided into a smaller period indexed by
t ∈ {0, ..., T }. At any specific time t, the investor observes the current state of the market Xt ,
and he then chooses an action θt ∈ Θ(Xt ) which can be either resubmitting an order at price pL t
or canceling the previously submitted order. The action space of the investor can be described as
Θ(Xt ) = {pL L
t } ∪ {C}, where C denoted cancellation. If θt 6= pt−1 , the trader will amend his order
price and his order goes to the back of the time priority queue at the new price. If the trader set
his new price such that the price is better than the opposite side best price, he has converted his
order to a market order that will transact immediately.
To determine the optimal order submission strategy, the trader chooses his action at each time
to maximize his expected payoff over the trading horizon. This problem can be formulated as
a stochastic dynamic optimization problem with a finite horizon and a finite action space. The
formal description of the valuation function consists of two parts.
ut (Xt , qt ) = max {Et,qte [fe (θt , qte ) + ut+1 (Xt+1 , qt − qte )]} (3.5)
θt ∈Θ(Xt )
subject to the boundary condition uT (XT , qT ) = max{fe (pM T , qT ), fnt (qT )} depending on the
trading problem the trader wants to solve. Like in the single-period model, most of the previous
works usually simplify this calculation by assuming that when a limit order is executed, it is
satisfied fully at the stated price. In this case, the valuation function is reduced to:
CHAPTER 3. LITERATURE REVIEW 24
ut (Xt , qt ) = max {P (θt , q, Xt )fe (θt , q) + (1 − P (θt , q, Xt ))Et,qte [ut+1 (Xt+1 , q)]} (3.6)
θt ∈Θ(Xt )
where P (θt , q, Xt ) is the one-period execution probability of a limit order to transect q shares at
price θt .
Theoretical models
Since the execution probability depends on future traders’ order submission, an order submission
model that describes traders’ behavior can be utilized to determine the execution probability.
Given the model of the market together with the distribution of traders and their valuation, the
execution probability in an equilibrium1 can be derived (See [39, 16, 17] for examples). Specifically,
assuming that all traders in the market are rational and use the same optimal order submission
strategy, the execution probability of the limit order can be estimated by calculating the proba-
bility that other traders will submit an order to trigger our limit order.
For example, Foucault [16] studies a market for a single risky asset whose trading day is
divided into discrete time intervals denoted t = 1, 2, ..., T , where T is a random stopping time,
and, at each t, the probability that trading will terminate is (1−ρ). When the trading terminates,
the payoff of the asset, whose value is constant and equal to v, will be realized. At each time
t, a trader who is characterized by the reservation price, Rt = v + yt , arrives. Yt can take two
value yh = +L or yl = −L with probability k and (1 − k), respectively. Thus, in this market,
there are two types of traders: the yh type only place buy orders and the yl type who only place
sell order. Consider a trader of type yh who arrive at time t. Let B be the bid price this trader
choose if he posts a buy limit order. This order will be executed only if (i) the game does not
stop before the arrival of the next trader (with probability ρ), (ii) the next trader is type yl (with
probability (1 − k)), and (iii) the next trader submit a market order. The probability of the last
event is endogenous and depends on the bid price. In particular, if the bid price is too low, the
yl -type trader will be better off posting a sell limit order. However, if the bid price is greater than
the price that yl -type trader is indifferent between market order and a limit order, this buy limit
order will be executed with probability ρ(1 − k). Foucault illustrates that this threshold price is
1−ρ(1−k)
θh = v − L + 1−ρ 2 k(1−k) (2L).
Another example is an equilibrium model of Parlour [39]. In her model, there is an asset
that are traded on day 1 and pays off a certain amount V on day 2. On day 1, at each times
t = 1, 2, ..., T , a randomly drawn agent arrives at the market. Agents are characterized as potential
buyers or sellers with probability πb and πs respectively. A potential buyer has an endowment of
cash that can use to purchase one share. On the other hand, a potential seller holds one share
which can be sold for cash. Agent t’s preferences are given by a utility function U (C1 , C2 ; βt ) =
C1 + βt C2 , where C1 is agent’s consumption on day 1, C2 is agent’s consumption on day 2 and
βt is a trader’s personal trade-off between C1 and C2 . The parameter βt determines an agent’s
willingness to trade and is assumed to be randomly distributed over an interval (β, β) with some
continuous distribution F (.). Normally, a potential seller with a low value of βt will be eager to
sell, while a potential buyer with a low value of βt will be disinclined to buy. All agents have
only one opportunity to submit orders, and, once they submit an order, the order cannot be
withdrawn. The market, in this setting, is characterized by the designated bid and ask price B, A
1
In an equilibrium, optimal order submission strategies are determined based on the execution probability which
is computed by assuming that all traders follow the same strategy
CHAPTER 3. LITERATURE REVIEW 26
and a limit order book bt . The limit order book is described by the number of shares on the
bid and ask sides, bB A
t and bt , immediately prior to the arrival of agent t. The optimal strategy
for each agent in this market can be determined by recursively working backward from time T .
Particularly, the agent at time T has two options whether to trade using limit order or to do
nothing. If the agent is a potential seller, he will compare the utility of doing nothing (βT V )
with the utility of selling at the bid (B). If βT V < B, then he will enter a market sell order.
Giving a particular distribution for βT , the probability that the agent T will enter a market sell
order can be computed. A similar computation can be utilized to compute the probability of a
market buy order at time T . For the agent arrive at time T − 1, if he is a buyer and bB T −1 = 0,
he can enter a buy limit order which will be executed if agent T enter a market sell order, the
probability of which we just computed. If bB T −1 ≥ 1, agent T − 1’s limit buy will not be first in
the execution queue, and, thus, cannot be executed in the one remaining period. Given agent
T − 1’s direction, the limit order book bT −1 , and the limit order execution probability, Parlour
Buy
illustrates that there are cutoffs β Buy
Limit
and β Limit such that if βT −1 < β Buy
Limit
, agent T − 1 will
Buy Buy
do nothing; if β Buy
Limit
< βT −1 < β Limit , he enters limit buy order; and if βT −1 > β Limit , he will
enter a market buy order. Given a distribution of βT −1 , the probability of these events can be
computed and utilized to define the execution probability for time T − 2 which defines agent T − 2
optimal strategies, and so on.
Although the results obtained from these analytical studies may not be appropriated for real
situations since they depend on the assumption about the market model which is usually a lot
simpler than the real market, these results provide us more understanding about the relation
between the execution probability and other related variables.
Empirical models
Apart from using analytical methods, we can also utilize historical data on trades and quotes to
estimate the execution probability. Given a specific limit of time, the execution probability can
be defined as a ratio of the number of limit orders that are executed within that limit to the total
number of orders submitted to the market. This definition has been utilized by several authors
including Omura et al. [38] and Hollifield et al. [27]. To estimate the execution probability, the
main task of this model is to determine whether the limit order is executed in the specified period
of time or not. If every action relating to each limit orders is time-stamped in our historical
data, this task will be very easy to achieve. However, in real situations, we can observe only
the change in quantities of limit orders in the limit order book and information about the trade
(i.e. execution price and its volume). Thus, some approximation methods should be employed to
estimate the time that the order is submitted to the market as well as the time that the order is
executed.
One example of such approximation is the work of Omura et al. [38] who analyze the execution
probability of limit orders on the Tokyo stock exchange. Their dataset contain information about
any changes in quantities of the limit order at the best bid and the best ask, and to an execution
price and its volume. They reconstruct the market order flow using the following rule. Suppose
that at time t an execution is recorded. Let Pt be the execution price at time t, Vt be the executed
amount in share, Askt− be the ask price, and Bidt− be the bid price just before the execution
occurs. If they observe P t = Askt− , then the execution is carried out by an incoming market buy
order of the size Vt . If Pt = Bidt− , then there is a market sell order of the size Vt . To reconstruct
CHAPTER 3. LITERATURE REVIEW 27
limit order flow, they trace any changes that occurred in the limit order book. Suppose that no
execution occurs at time t. If at time t − 1 an execution occurs, then the change in the book from
t − 1 to t is solely caused by the execution at t. They will disregard this t record for the purpose of
constructing incoming limit order flow. If instead, there is no execution at time t − 1, the change
in the book must be caused by either a submission or a cancellation of limit orders. Let ALmtt
be a placed sell limit order at time t, AskVt and AskVt−1 be the amount of sell side of the book
at time t and t − 1, respectively. When the ask prices Askt and Askt−1 at the consecutive times
are the same, then the amount of orders changed is:
where T indicates the trading period, and AM kts , t < s ≤ T , indicates the amount of the buy
market order with an execution price higher than or equal to the limit price Askt that arrives
after the submission of the sell limit order of the limit sell order ALmtt . Similarly, a buy limit
order BLmtt is executed if
T
X
BBookt− + BLmtt ≤ BM kts
s>t
where BM kts , t < s ≤ T , is the amount of the sell market order with an execution price lower
than or equal to the buy limit price of the limit order BLmtt .
To estimate the execution probability, they assume that a limit order remains on the book until
the closing of the day it is submitted. Thus, they ignored all cancellations, and this may cause
underestimate of the execution probability especially when the preceding limit orders, ABookt−
or BBookt− are cancelled. In such case, the limit order concerned may be actually executed, but
the method may fail to recognize it.
Although we can utilize this approach to estimate the probability of execution, the main
drawback of this approach is that it depends on the specific time period. If the trading period of
the trader changes, the execution probability must be recalculated.
CHAPTER 3. LITERATURE REVIEW 28
where is the tick size of the asset. By modeling the property of the first-passage time, one can
utilize it to estimate the property of time-to-first-fill and time-to-fill of an interested asset. For
example, one can use the first-passage time as an approximation of the time-to-fill and use it to
estimate the probability of execution using the following equation:
However, the probability obtained from this estimation usually overestimates the actual execution
probability since the estimated first-passage time is typically lower than the actual time-to-fill2 .
The property of the first-passage time of a particular financial instrument can be modeled
both by theoretical and empirical approaches. On the theoretical side, the statistical property of
the first-passage time can be explicitly derived if a stochastic property of the asset price process
is given. For example, if the dynamic of the asset price S(t) is given by a Brownian motion with
diffusion rate σ, the probability distribution of the first-passage time will be [15]:
∆ ∆2
f (∆, t) = √ exp √ (3.12)
2πσ 2 t3 2σ 2 t
2
The reason why the estimated first-passage time is typically lower than the actual time-to-fill is that when the
asset price first reaches the limit price the order will be executed only when it is the first order in the queue. Thus,
the time-to-fill is usually longer than the first-passage time.
CHAPTER 3. LITERATURE REVIEW 29
Using equation (3.11), the probability that the limit order is fully executed during a specific
trading period T can be approximated as (see [31] for example):
Z T
∆
P r(F ullyExecuted|T, ∆, σ) ≈ P r(F P T ≤ T |∆, σ) = f (∆, t)dt = erfc √ (3.13)
0 2σ 2 T
Although many stochastic processes (e.g. geometric Brownian motion and Markov processes)
can be applied in this context, the estimated result is largely depended on the asset price model
specification. If this specification is not appropriate, the estimation error can be incredibly large.
For example, if the asset price exhibits a short-term mean revision but a geometric Brownian
motion is utilized to model the execution time, the predicted execution time will largely underes-
timate the real execution time as reported in [32]. To amend the problem, one can use empirical
approaches that directly model the first-passage time using historical time series of transactions
data (see [5, 22] for example). The primary advantages of such approach to statistical models is
that if the stochastic process for the asset prices exhibits mean revision or more complex forms of
temporal dependence and heterogeneity, this will be automatically incorporated into the empirical
model. While empirical approach can solve model specification problem, the first-passage time
model still suffers from several other important limitations. The most obvious weakness is the
assumption that the order is executed when the limit price is first attained; hence, the model
can not be easily modified to handle the variation of limit order sizes as well as the distinction
between time-to-first-fill and time-to-fill. In addition, it cannot easily incorporate the effects of
explanatory variables such as market conditions and the state of the order book. Thus, although
the first-passage time model is a natural theoretical framework for modeling the order executions,
it leaves much to be desired from a practical point of view.
time distribution. The advantage of such approach is that information from unexecuted orders can
be easily and correctly accommodated. The idea behind this is that although we can not calculate
the exact execution time of unexecuted orders, we know that if these orders were executed their
execution time should be at least as long as their lifetime. In addition, survival analysis also
enables us to estimate the execution time distribution as a function of dependent variables such
as information about the order (e.g. limit price and order size), state of the order book and
market conditions. This property is very important since recent empirical research suggests that
the execution probability is largely depended on these variables [38], and, without the ability to
incorporate these variables into the model, the result may not be accurate.
Although empirical execution time models based on survival analysis seem to be a good can-
didate to model the execution probability since it can solve all the mentioned problems, they still
have some limitations concerning its use in developing an order submission strategy. The most
important limitation is the inability to estimate the expectation over the filled quantity mentioned
in equation 3.2 and 3.5. Although they can be used to estimate the expectation of the simplified
models mentioned in equation 3.4 and 3.6, these simplified models are based on the assumption
that, when a limit order is executed, it is satisfied fully at the stated price, which usually does
not hold especially for a large limit order.
3.3 Summary
This section presents an overview of order submission strategies together with it relation to
models of limit-order execution. Although recent empirical studies indicate that a trader’s order
submission strategy is largely influenced by the state of the order book and current market
conditions, little attention have be paid to develop a theoretical model to optimize such strategy.
Most of the theoretical models consider this problem as a trade-off between the free-trading
option and the expected profit, which is largely depended on the execution price and the execution
probability. Thus, one of the most important factors to make such decision is a model of execution
probability. Consequently, a natural choice to develop an order submission strategy that utilizes
this information is to incorporate it into the model of execution probability. However, all of the
previous methods to model such probability still have limitations and need further development.
Specifically, an ideal limit-order execution model should be able to incorporate the information
about market conditions and the state of the order book as well as to estimate the expected profit
of limit order trading without the assumption that when limit order is executed it is satisfied fully
at the stated price.
Chapter 4
Research Proposal
4.1 Hypothesis
Traditional order submission models usually do not use information about market conditions
(i.e. the state of the order book and its dynamic) to make trading decision. However, recent
empirical investigations suggest that traders’ order submission decision is largely depended on
this information and incorporating it into the decision model can greatly improve the execution
result. One way to achieve this is to develop a model of limit-order execution that utilize this
information as a determinant to determine the execution probability of limit orders and utilize
the resulted execution probability to derive the optimal order submission strategy.
1. Develop a model of limit-order execution that utilizes information about market conditions
to estimate the execution probability of limit orders. To achieve this, several limitations of
previous models will be examined and improved in the following directions:
• The developed model will incorporate the explanatory variables (i.e. market conditions
and the state of the order book) to estimate the execution probability.
• The developed model will consider all orders in the order book rather than consider
only newly submitted orders as in previous models so that the developed model can
be utilized to analyze the dynamic of the execution probability with the aim to gain
more insight on the determinants of the execution probability and how this probability
changes over time.
• The developed model will be able to handle the variation of order size as well as to
estimate the expected fill quantity of a limit order.
31
CHAPTER 4. RESEARCH PROPOSAL 32
2. Develop an order submission model that utilize the execution probability estimated from the
developed limit-order execution model to optimize trade execution. Although recent em-
pirical investigations suggest that traders’ order submission is largely depended on market
conditions and incorporating this information to make order submission decision can greatly
improve the execution result, little attention has been paid to develop an order submission
strategy that utilizes this information to optimize trade execution. By utilizing the ex-
ecution probability from the developed limit-order execution model that utilizes market
conditions to estimate the execution probability, the developed order submission model will
accordingly utilize market conditions to make order submission decision, and, thus, an or-
der submission strategy produced by the developed model is expected to be better than the
strategy produced by traditional models that do not utilize this information.
2. A new probability model of limit-order execution that models all orders in the order book.
4. A new order submission model that utilize the developed order execution model to estimate
the execution probability
Research Plan
The research activities that need to be carried out to achieve the contributions of this research
can be categorized into five main tasks, which are:
• Task 1: Implement tools to collect real-time data for experiments and analysis,
• Task 2: Implement simulation models of a pure limit order book to generate data for
experiments and analysis,
• Task 3: Develop a virtual financial market environment which provides a testbed for subse-
quence experiments and analysis,
• Task 4: Develop new computational models of limit-order execution that incorporates vari-
ous explanatory variables and utilizes all orders in the order book to estimate the execution
probability,
• Task 5: Develop a new order submission strategy that can be utilized to make order sub-
mission decision in an algorithmic trading system.
The detailed information of each task will be summarized in the subsequence sections starting
from section 5.1 to section 5.5. To conclude the chapter, the timetable of this research will be
discussed in section 5.6.
33
CHAPTER 5. RESEARCH PLAN 34
ArcaEX)2 . The reason why we interest in these trading venues is that they are pure electronic limit-
order book markets and we have an access to their real-time order book information. The real-time
information of MCX market can be collected from Reuters data feed that Thomson Reuters spon-
sors it to UCL free of charge. The real-time information of NYSE Arca market is obtained from
Arca Web Book service (accessible at http://www.archipelago.com/marketdata/book info.asp).
Since real-time information comes from two different sources, two different systems have to be
implemented. The system for collecting information from MCX market was implemented using
Reuters System Foundation API (SFC) Java Edition3 , while the system for NYSE Arca market
is a web robot that is scheduled to retrieve a snapshot of the order book from Arca Web Book
service every a specific period of time. Note that the developments of these two systems are
already finished and they are currently running to collect real-time information every single day.
is that it obviates the need for a so-called fill model to predict from price information alone
when limit orders would be executed [29]. Specifically, if only price information is available, a
simulation model must make a decision whether a limit order is executed or not when the asset
price crosses the limit price. A typical fill model might probabilistically execute the limit order
based on historical data and the volume of the order. In contrast, in our simulation, such models
are not required since a limit order will be filled only when there is a matched with the opposing
order just as in the real exchanges.
Once the virtual financial exchange is running, any number of automated trading clients can
connect to it and trade according to their programmed strategy. The communication between
the exchange and their clients will be implemented using the Financial Information eXchange
(FIX) protocol4 . To facilitate trading client development, a framework for developing a trading
client, that abstract the detail about the FIX protocol from trading clients by providing pro-
gramming functions for basic order management (e.g. submit, amend and cancel) and commands
for retrieving current status of the market (e.g. current order book and recent trades), will be
implemented.
5.6 Timetable
The timetable of this research is illustrated in the following figure.
Work to Date
This section briefly describes all the works that I have done during the first year of my study.
During the first six months, I spend most of the time reviewing research articles in the emerging
field of Computational Finance. The result I get during this period is a review paper of this area.
After finishing the review, I found myself interest in the area of algorithmic trading which leads
me to further my research in this particular area. I also have an opportunity to design a virtual
trading platform and supervised a group of MSc student to develop it. The experiences I gain
from this opportunity help me understand more about the investment process as well as how to
apply computational methods to solve its problem. For the works that related to my research,
I developed a program to collect real time order book information from Reuters data feed (see
section 5.1) as well as developed a program to generate the empirical distribution of execution
probability from trade information.
The rest of this section is organized as follow: information about the virtual trading platform
will be described in section 6.1. Section 6.2 presents the method for generating the empirical
distribution of execution probability from trade information.
37
CHAPTER 6. WORK TO DATE 38
I supervised nine MSc students to develop this platform, each of them responsible for each
component of the system. The number of students focused in each component is summarized in
Table 6.1.
Component # of students
Market simulator 1
Virtual hedge fund 1
Trading strategy 1 4
Trading strategy 2 1
Execution strategy 2
that a newly submitted limit order will be executed; however, these models have no information
about how these probabilities are changing over time since they model only the newly submitted
order. This limitation suggests us to construct an execution probability model of all orders in the
order book. When time goes by, a position of the order in the order book will be changed, and
the execution probability of the order should be changed accordingly. Surprisingly, it seems that
only trade data is enough to model this kind of probability.
Specifically, we can use trade data to estimate the following conditional probability, essentially
the CDF of the execution time Tk of a limit order k:
P r(Tk ≤ t|Xk , Ok )
where Xk is a vector of explanatory variables that captures market conditions and other condi-
tioning information, and Ok is a vector of variables describing information about the limit order
(i.e. location in the order book, limit-order price, size, and side indicator).
Unlike previous methods that model the probability of newly submitted order, we try to
model the probability of every order in the order book. Thus, we need some way to specify a limit
order that we would like to investigate. One way to do this is to use the location in order book
measured by the amount of the order preceding it. Consequently, an order can be represented
by a quadruple Ok = (Lk , Pk , Qk , Sk ) where Lk , Pk , Qk , Sk are location of the order, limit-order
price, size (in shares) and side indication (buy or sell), respectively.
As a starting example, assume that we want to estimate the probability that a limit order
Ok will be executed (time-to-first-fill) during the next t minute. In this case, the order can be
represented by a triple Ok = (Lk , Pk , Sk ) since quantity have no effect on time-to-first-fill. The
probability P r(Tk ≤ t|Ok ) can be estimated directly by counting the event that an order having
this specification will be executed during the next t minute. Using historical trade data, giving
a particular time, we can perform ”what-if” simulation to determine whether a specified order
will be executed during the next t minute or not. If we perform this simulation throughout the
available time period, the probability P r(T k ≤ |Ok) can be estimated by the ratio between the
number of time the order is executed and the number of time that we perform the simulation.
The outline of this process is illustrated in figure 6.2. The examples of empirical distribution of
execution probability generated from such approach is illustrated in 6.3.
N, EN ← 0
for each timestamp ts in historical data
get all trades during ts and ts + t from historical data
perform simulation to determine whether there are enough trade to fill Ok
if Ok can be filled: EN ← EN + 1
N ←N +1
return EN/N
Figure 6.2: Outline of the method for generating an empirical distribution of execution probability.
CHAPTER 6. WORK TO DATE 40
Figure 6.3: Example of empirical distribution of execution probability generated from trade in-
formation.
Bibliography
[1] Hee-Joon Ahn, Kee-Hong Bae, and Kalok Chan. Limit orders, depth, and volatility: Evidence
from the stock exchange of hong kong. The Journal of Finance, 56(2):767–788, 2001.
[2] Robert Almgren and Neil Chriss. Optimal execution of portfolio transactions. Journal of
Risk, (3):5–39, 2001.
[3] Robert Almgren and Julian Lorenz. Bayesian adaptive trading with a daily cycle. Journal
of Trading, 2006.
[4] Robert Almgren and Julian Lorenz. Adaptive arrival price. Algorithmic Trading III: Preci-
sion, Control, Execution, 2007.
[5] James Angle. Limit versus market orders. Working Paper No. FINC-1377-01-293, School of
Business Administration, Georgetown University, Washington, DC, 1994.
[6] Helena Beltran-Lopez, Pierre Giot, and Joachim Gramming. Commonalities in the order
book. Working Paper 2005014, Universit catholique de Louvain, Dpartement des Sciences
Economiques, January 2005.
[7] Bruno Biais, Pierre Hillion, and Chester Spatt. An empirical analysis of the limit order book
and the order flow in the paris bourse. Journal of Finance, 50(5):1655–89, December 1995.
[8] Adam Blazejewski. Computational Models for Stock Market Order Submissions. PhD thesis,
The University of Sydney, 2005.
[9] Sergiy Butenko, Alexander Golodnikov, and Stanislav Uryasev. Optimal security liquidation
algorithms. Comput. Optim. Appl., 32(1-2):9–27, 2005.
[10] Charles Cao, Oliver Hansch, and Xiaoxin Wang. Order placement strategies in a pure limit
order book market. The Journal of Financial Research, 31(2):113–140, 2008.
[11] Kalman J. Cohen, Steven F. Maier, Robert A. Schwartz, and David K. Whitcomb. Trans-
action costs, order placement strategy, and existence of the bid-ask spread. The Journal of
Political Economy, 89(2):287–305, 1981.
[12] Thomas Copeland and Dan Galai. Information effects and the bid-ask spread. Journal of
Finance, 38:1457–1469, 1983.
[13] Nicholas Economides and Robert A Schwartz. Electronic call market trading. The Journal
of Portfolio Management, 21, 1995.
41
BIBLIOGRAPHY 42
[14] Zoltan Eisler, Janos Kertesz, and Fabrizio Lillo. The limit order book on different time scales.
Proceedings of SPIE, 6601, 2007.
[15] Zoltan Eisler, Janos Kertesz, Fabrizio Lillo, and Rosario N. Mantegna. Diffusive behavior
and the modeling of characteristic times in limit order executions, 2007.
[16] Thierry Foucault. Order flow composition and trading costs in a dynamic limit order market1.
Journal of Financial Markets, 2(2):99–134, May 1999.
[17] Thierry Foucault, Ohad Kadan, and Eugene Kandel. Limit order book as a market for
liquidity. Review of Financial Studies, 18(4):1171–1217, 2005.
[18] Joachin Grammig, Andreas Heinen, and Erick Rengifo. Trading activity and liquidity supply
in a pure limit order book market: An empirical analysis using a multivariate count data
model. In the 14th Annual Conference of the Deutsche Gesellschaft fr Finanzwirtschaft,
Dresden, Germany, September 2007.
[19] Mark D. Griffiths, Brian F. Smith, D. Alasdair S. Turnbull, and Robert W. White. The costs
and determinants of order aggressiveness. Journal of Financial Economics, 56(1):65–88, April
2000.
[20] Anthony D. Hall and Nikolaus Hautsch. Order aggressiveness and order book dynamics.
Emprical Economics, 30(4):973–1005, 2006.
[21] Puneet Handa, Robert Schwartz, and Ashish Tiwari. Quote setting and price formation in
an order driven market. Journal of Financial Markets, 6(4):461–489, August 2003.
[22] Puneet Handa and Robert A Schwartz. Limit order trading. Journal of Finance, 51(5):1835–
61, December 1996.
[23] Larry Harris. Trading and exchanges: Market microstructure for practitioners. Oxford Uni-
versity Press, New York, 2003.
[24] Lawrence Harris. Optimal dynamic order submission strategies in some stylized trading
problems. Financial Markets, Institutions and Instruments, 7(2):1–76, 1998.
[25] Joel Hasbrouck. Empirical Market Microstructure: The Institutions, Economics, and Econo-
metrics of Securities Trading. Oxford University Press, New York, Dec 2007.
[26] Burton Hollifield, Robert Miller, Patrik Sandas, and Joshua Slive. Liquidity supply and
demand: Empirical evidence from the vancouver stock exchange. GSIA Working Papers
1999-E19, Carnegie Mellon University, Tepper School of Business, March 2001.
[27] Burton Hollifield, Robert A. Miller, and Patrik Sandas. Empirical analysis of limit order
markets. Review of Economic Studies, 71(4):1027–1063, October 2004.
[28] Patrik Idvall and Conny Jonsson. Algorithmic trading: Hidden markov models on foreign
exchange data. Master’s thesis, Department of Mathematics, Linkopings Universitet, Linkop-
ings, Sweden, 2008.
BIBLIOGRAPHY 43
[29] Michael Kearns and Luis Ortiz. The penn-lehman automated trading project. IEEE Intelli-
gent Systems, 18(6):22–31, 2003.
[30] Robert Kissell and Roberto Malamut. Algorithmic decision making framework. Journal of
Trading, 2006.
[31] Fabrizio Lillo. Limit order placement as an utility maximization problem and the origin of
power law distribution of limit order prices. The European Physical Journal B, 55(4):453–459,
February 2007.
[32] Andrew W. Lo, A. Craig MacKinlay, and June Zhang. Econometric models of limit-order
executions. Journal of Financial Economics, 65(1):31–71, July 2002.
[33] Ingrid Lo and Stephen Sapp. Price aggressiveness and quantity: how are they determined in
a limit order market? In The Microstructure of Equity and Currency Markets, Oslo, Norway,
September 2005. Norges Bank/BI (Norwegian School of Management).
[34] Lukas Menkhoff, Carol L. Osler, and Maik Schmeling. Order-choice dynamics under asym-
metric information: An empirical analysis. In 12th Conference of the Swiss Society for
Financial Market Research, SWX Swiss Exchange, Zrich, April 2008.
[35] Szabolcs Mike and J. Doyne Farmer. An empirical behavioral model of liquidity and volatility.
Journal of Economic Dynamics and Control, 32(1):200–234, January 2008.
[36] Yuriy Nevmyvaka, Yi Feng, and Michael Kearns. Reinforcement learning for optimized
trade execution. In Proceedings of the 23rd international conference on Machine learning
(ICML06), pages 673–680, New York, NY, USA, 2006. ACM.
[37] Yuriy Nevmyvaka, Michael Kearns, Amy Papandreou, and Katia Sycara. Electronic trading
in order-driven markets: efficient execution. In Proceedings of the 7th IEEE International
Conference on E-Commerce Technology (CEC05), pages 190–197, 2005.
[38] Keiichi Omura, Yasuhiko Tanigawa, and Jun Uno. Execution Probability of Limit Orders on
the Tokyo Stock Exchange. SSRN eLibrary, 2000.
[39] Christine A Parlour. Price dynamics in limit order markets. Review of Financial Studies,
11(4):789–816, 1998.
[40] T. Preis, S. Golke, W. Paul, and J. J. Schneider. Multi-agent-based order book model of
financial markets. EPL (Europhysics Letters), 75(3):510–516, 2006.
[41] Angelo Ranaldo. Order aggressiveness in limit order book markets. Journal of Financial
Markets, 7(1):53–74, January 2004.
[42] Duane J. Seppi. Liquidity provision with limit orders and strategic specialist. The Review of
Financial Studies, 10(1):103–150, 1997.
[43] Joshua Slive. Dynamic strategies in a limit order market. In Annual Meetings of The Eastern
Finance Association 2008, St. Pete Beach, Florida, USA, April 2008.
BIBLIOGRAPHY 44
[44] Peter Verhoeven, Simon Ching, and Hock Guan Ng. Determinants of the decision to submit
market or limit orders on the asx. Pacific-Basin Finance Journal, 12(1):1–18, January 2004.
[45] Jiaqi Wang and Chengqi Zhang. Dynamic focus strategies for electronic trade execution
in limit order markets. In CEC-EEE ’06: Proceedings of the The 8th IEEE International
Conference on E-Commerce Technology and The 3rd IEEE International Conference on En-
terprise Computing, E-Commerce, and E-Services, page 26, Washington, DC, USA, 2006.
IEEE Computer Society.