Electronic Trading in Order-Driven Markets: Efficient Execution

Yuriy Nevmyvaka
Carnegie Mellon
University
[email protected]

Michael Kearns
University of
Pennsylvania
[email protected]

Abstract.
In this paper, we address the importance of efficient
execution in electronic markets. Due to intense
competition for profit opportunities, trading costs can
represent a significant portion of overall return. They
must be taken into account both when a specific trade is
being executed, and when a general investment strategy is
being designed. We empirically demonstrate that by
combining market orders (which offer immediate
execution regardless of price) and limit orders (which
offer uncertain execution at a specified price), we are
able to obtain a superior average price than by using
market orders only. Our analysis highlights the trade-off
between expected price improvement from limit orders
and the risk of non-execution. We show how to determine
the optimal limit order price in a simplified setting and
suggest how this approach can be generalized to a
complete solution. All of our experimental results are
obtained on an extensive collection of NASDAQ limit
order data.

I. Introduction.
Execution considerations permeate every aspect of
investment activity from the development of high-level
trading strategies to post-trading performance attribution
because the actual prices at which trades occur are a
direct consequence of the execution mechanism in place.
Every market participant should be concerned with
transaction costs regardless of their investment goals and
style. Ideally, one would like to have a comprehensive
execution system, which aims to minimize the costs of
trading. Classifying and quantifying various trading costs
is a complex task (see [Kissell and Glantz, 2003]), and a
comprehensive trade optimization system must solve a
challenging multi-dimensional problem.
The goal of this paper is not to offer the ultimate
solution to efficient execution, but rather to develop a
simple building block, which can be easily implemented,
quantified, and then extended into a more comprehensive
system. In our study, we concentrate on a single but
important aspect of the overall problem: the immediate
price received or paid for a transaction of a fixed size over
a fixed period of time. This makes our setup stylized and
yet practical. More specifically, we ask the following
question: how should one buy (respectively, sell) V shares
of a given stock over T seconds while spending the least
(respectively, receiving the most) cash? In the framework

Amy Papandreou
Lehman Brothers
[email protected]

Katia Sycara
Carnegie Mellon
University
[email protected]

we consider, a trader has only three options: (1) submit a

market order at time 0 for V shares, (2) submit a market
order at time T for V shares, or (3) submit a limit order at
time 0 for V shares and a market order for the unexecuted
shares (if any) at time T. In other words, we can use a
market order at the beginning of the time period, a market
order at the end of the time period, or a limit order
combined with a market order for residual shares. In all
three cases, we will always end up with the same number
of shares, but will have spent different amounts of cash
entering the position. By formulating the execution
problem in this fashion, we can quantitatively answer a
number of important questions. Which of the three
options is the most effective? How should one price a
limit order to expect the most favorable execution? How
can we quantify the risks of non-execution and
unfavorable price movements?
We suggest a precise analytical methodology to reason
about efficient execution in modern electronic markets:
for every possible limit order price, we compute the ratio
of the execution price to the mid-spread price at the start
of the trading period; resulting return curves allow us to
pinpoint which pricing strategy produces the most
favorable expected price. We similarly build out the risk
curves, defining risk as the standard deviation of the
outcomes of each strategy. Risk curves elucidate the
relationship between limit price and execution
uncertainty. We then merge the two curves into a single
function to demonstrate the trade-off between expected
returns and risk. We conclude that it is only optimal to
operate on a specific portion of the risk-return function,
which we call efficient pricing frontier. Thus the output
of our analysis allows the trader to select some tolerable
risk threshold and to determine a strategy which delivers
the most favorable execution price for that level of risk.
Finally, we identify the relative importance of different
market microstructure variables (order size, execution
window, market activity, etc.) in the optimization process,
and explain how to expand our basic methodology to the
multi-period dynamic execution.
To be able to assign specific numbers to the expected
price improvements and risk associated with each
strategy, we perform what-if simulations within
historical limit order books. We simulate what would
have happened to a hypothetical order in the real-world
market, and we record the execution price at the end of
each trial. (Details of this experimental methodology are

provided in Section III). In our approach, we rely on the

pre-committed liquidity, represented by resting limit
orders, which is another aspect that sets our work apart
from similar studies.
This paper is structured as follows: in the next section
we will discuss related work making an emphasis on the
novelty of using limit order books in our analysis. We
then describe the experimental setup we use and point out
its strengths and shortcomings. In Section IV, we
introduce our analytical framework: how to quantify
returns, risk, and the combination of the two. Section V
contains a summary of our empirical results a study
across various stocks and trading conditions. We
generalize our findings and suggest how more complex
execution systems can be built upon them in Section VI.

II. Related Work.

We believe that the main innovation in our approach is
bringing together the subjects of efficient execution and
limit order trading. While each area has received attention
in academic literature, we are not aware of other studies
that cast the execution problem within the context of an
order-driven marketplace.
Both academics and professionals have long
acknowledged the importance of execution optimization.
An overview of a number of quantitative execution
methods is presented in [Kissell and Glantz, 2003]. This
book describes a comprehensive top-down approach to
evaluating trading costs and optimizing execution. Other
important works on the general trade execution are
[Almgren and Chriss, 1999] and [Almgren and Chriss,
2003]. [Bertsimas and Lo, 1998] suggest a dynamic
programming approach to solving the execution problem,
which can be further extended to the portfolio setting. The
majority of studies in this area concentrate on the market
impact of trading, which is the most important aspect of
execution. Their approach to quantifying this effect,
however, is fundamentally different from ours: other
studies look at the post-trade price changes, while we
analyze the ex-ante pre-committed liquidity that we see
on the order books.
Analyzing and modeling limit order trading is a part of
the broader discipline of market microstructure, which
studies interactions among market participants and the
process of price formation. The extensive overview of the
subject is [Harris, 2002], which includes the introduction
to the order-driven market mechanisms. One of the first
inquiries into the real-world limit order markets is [Bias et
al., 1995], which develops a theoretical model of limit
order submissions and conducts a comprehensive
empirical study of the Paris Bourse. A number of papers
try to explain the rationale behind using limit orders,
modeling behavior of different types of traders: [Lo et al.,
2000] and [Hollifield et al., 2003].

Perhaps the closest in spirit to our work are the studies

that perform empirical analyses of various order
submission strategies. [Hasbrouck and Harris, 1996] and
[Handa and Schwartz, 1996] compare market order and
limit order trading strategies in different markets and
confirm that limit orders can indeed enhance returns. It is
this exact approach that we are applying to efficient
execution.

III. Experimental Setup.

In our study, we used historical records from the Island
Electronic Communication Network (ECN). Island
(recently acquired by INET) is essentially an electronic
and completely automated stock exchange which accounts
for a significant volume of trading in NASDAQ stocks. In
the Island files we can see every event that happened to a
stock through a trading day all order submissions,
cancellations, and executions. Every action is timestamped, and since all transactions are recorded
electronically, the exact sequence of order flow is
unambiguous. These features allow us to precisely
reconstruct buy and sell limit order books at any point in
time.
We go through the record of order arrivals,
cancellations, and executions, updating the state of limit
order books after every event; at designated times, we
insert into the order flow artificial orders that represent
various trades we wish to investigate. We then perform all
the executions and maintain order priorities in the same
way it is done in the real ECN. Such a setup allows us to
run what-if simulations in historical order books. In all
simulations, we use one-and-a-half years of data from
January 2003 to June 2004 , which is available for every
stock traded on Island. While we have obtained results
for a broad cross-section of stocks, for the purpose of
brevity, we concentrate here on a single security
Microsoft Corp. common stock (MSFT). For an extended
version of our results, see [Nevmyvaka et al., 2004].
We investigate efficient execution by examining the
pre-committed liquidity, which is precisely what limit
order books represent. When a trader submits a limit
order, he provides an option to the rest of the market
participants to transact at a pre-specified price up to the
orders size i.e. other traders can lift off liquidity,
while knowing ex-ante how much they are paying. This is
crucial for our analysis transaction costs in our model
come from two sources: the bid-ask spread, and price
concessions as payment for liquidity. When a trader
submits a market order, he has to first step over the
spread, and then pay increasingly disadvantageous prices
the deeper in the opposing book he needs to reach to
satisfy his liquidity demands. On the other hand, when he
submits a limit order, he risks having the price move
away from his order and then being forced to demand
liquidity at the end of the interval. Historical limit order

books allow us to quantify and compare these two

dimensions of order submissions.

IV. Analytical Framework.

In this main section we present the essence of our
approach to execution: first, we describe how to
determine which limit order price results in the most
advantageous execution price; second, we introduce risk
into our analysis; and finally, we will combine the two to
derive the efficient pricing frontier.

A.

Expected Execution Price.

Let us re-visit the basic setting for our execution

problem: a high-level investment strategy issues a
directive to acquire V shares of some stock, and this
position must be entered within T seconds.
This task can be executed using the following trading
strategies:
(1) submit a market order for the entire amount
immediately guarantees both the execution and the
amount of cash paid (respectively, received), but has to
pay for liquidity demanded;
(2) wait until the end of the time period, hoping for a
favorable price move, and then go to the market with the
entire amount can achieve price improvement, but has
exposure to price volatility and still has to pay transaction
costs;
(3) submit a limit order at the beginning of the time
period; this order may execute completely, partially, or
not at all; then submit a market order for the remainder of
shares (if any) at the end of the interval can avoid
paying transaction costs, otherwise becomes the worst
case of Strategy 2.
All these strategies end up with the same position V
after T seconds, but will have spent different amounts of
cash. Therefore, if we plot levels of cash that each
strategy spent (respectively, generated), we can find
which one is the most efficient and if there is some
general relationship between various strategies
performance.
It is important to understand first what it is exactly that
we are measuring. Although we are interested in
comparing levels of cash, we cannot simply express our
results in dollars. For example, if a stock traded at $5 in
January 2003 and at $50 in June 2004, cost savings of
$0.02 will have very different implications in the two
cases. Therefore, the variable that we chose to optimize is
the difference between the execution price (weighted by
volume) and the mid-spread price at the beginning of the
time period, expressed as a fraction of the initial midspread price. Say the market at the beginning of the time
period has a bid of $24.12 and an ask of $24.18, yielding
a mid-spread price of $24.15. If we execute a buy order at
the ask, our price differential is ($24.15 - $24.18)/$24.15
= 0.001242. We express these numbers in basis points (12 bp in this example) and for the lack of better

terminology refer to them as returns; however, note that

they are not returns in the conventional sense of the term.
If during the execution the trader steps over the spread
and demands liquidity from the other side of the market,
returns are negative, thus representing transaction costs. If
the traders limit order is priced below the initial midspread and is later executed, then returns are positive
(transaction savings).
We submit orders from high-priced to low-priced and
record average returns for each strategy. As expected,
these returns tend to peak around a certain price level,
which consequently represents the optimal pricing level to
achieve the most advantageous execution price. A typical
order price-return curve is shown below in Figure 1.

Figure 1. The peak in the curve represents pricing

strategy that produces the most favorable expected
execution price.
This is a realistic curve for many situations, but for
concreteness, the above graph is a plot of a scenario
where the task is to acquire 10,000 shares of MSFT
within one hour.
On the x-axis, we plot where the limit order stands
relative to its own side of the market. For example, x = 70
means that the order is submitted at a limit price of best
bid minus 70 cents (in other words, it is 70 cents deep
within the buy book); x = -10 means that the limit price is
best bid + 10 i.e. the order is submitted either within the
bid-ask spread, or it is a marketable order, which
transacts with the sell book until the specified number of
shares is bought or the limit price is reached. In the later
case, the shares that are not executed immediately get
placed at the top of the buy book. On the y-axis, we plot
the aforementioned returns from our execution
strategies, expressed in basis points. They represent
transaction costs of trading a given block and thus are
negative.
By this construction, the parts of the curve on the
extreme left and right of the graph correspond to Strategy
1 and Strategy 2 respectively: on the left, we have orders
with high prices that get executed immediately and
completely; on the right, orders are priced so far away
from the inside market that they never get executed, thus
forcing a market order at the end of the trial. The peaked

shape in the middle of the graph supports our main thesis

superior execution price can be achieved by using
carefully tuned limit orders. The maximum of the curve
represents the lowest possible transaction cost that can be
achieved, and the corresponding x-value is the optimal
limit order price. Therefore, the main message of Figure 1
should be interpreted as follows: if you want to acquire
10,000 shares over 60 minutes and seek the most
attractive expected execution price, you should submit a
limit order at the best bid plus 5 cents. We discuss the
shape and behavior of these curves in great detail in
Section V.
Notice that the entirety of our return curve is below
zero. This means that transaction costs are always present
i.e. limit orders help to lose less money, as opposed to
generating profit opportunities on their own. While this
may be somewhat counterintuitive, one has to remember
that our results are averaged over many trials; therefore,
in many cases limit orders end up not being executed, and
the trader is forced to incur all the regular costs of a
market order at the end of the time period.

B.

Risk.

This brings us to the second major point returns

alone do not tell the entire story. While it may be
tempting to just adopt the previous conclusion that the
optimal order should be submitted exactly at the price
where returns peak, we have to remember that higher
returns come with higher risks. In our case, we are mostly
concerned with the risk of non-execution and being forced
to transact at a later time at an inferior price. And while
risk of non-execution, mid-spread volatility, and volume
volatility are all slightly different concepts, we study them
jointly by defining risk as the standard deviation of
returns.

Figure 2. Returns become more uncertain the further

we move from the inside market.
A risk profile that corresponds to the returns curve
from Figure 1 is shown in Figure 2. We simply plotted the
standard deviation of returns (y-axis) which are
averaged in Figure 1 for every limit order price (x-axis).
A couple of observations about the shape of the curve.
First, it generally slopes upwards from left to right, which
means that the deeper you hide your order in the book, the

less likely it is to execute before the end of the allotted

time interval, and the higher is the uncertainty around the
final price.
This shape is partly a result of our choice of the
scoring system. Since we "mark" our position to the
beginning of the time period, the further in time we get
from the starting point (time to execution is proportional
to the distance from the inside market for limit orders),
the wider becomes the distribution of our "returns". Thus
the general upward trend in our risk profile.
The more curious aspect of the graph is the "dip" in
risk between the pure market orders and the nonmarketable limit orders. This is partly an artifact of our
simulation setup, but it is also grounded in reality. Market
orders (e.g. limit orders with prices of bid plus several
dollars) sweep the sell book for the entire size at once;
therefore they trade through multiple price levels with
volume getting smaller and more volatile as we move
away from the inside market. The upside of this strategy
is that execution is guaranteed. On the other hand,
"marketable" limit orders (not as highly priced, with limit
prices of bid plus a few cents) transact with the top of the
sell book where volume is the highest and then leave the
residual shares sitting on top of the buy book. The
execution is still quasi-guaranteed, but the transaction
price is now capped at some more reasonable level.

C.

Efficient Pricing Frontier.

Now that we have described both returns and risk

profiles, we need a method for combining the two
measures so that we can optimize them together to derive
the actual optimal limit order price.
In order to perform a meaningful comparison among
alternative strategies, we borrowed a popular tool from
the classic Finance Theory the Markowitz efficient
frontier [Markowitz, 1952]. This methodology was
developed to show the trade-off between the risk and
return in an investment portfolio: in order to achieve
higher returns, investor has to assume more risk. The
same holds true for our domain to get price
improvement the trader has to employ a riskier strategy.
To plot a risk-return profile, we place every possible
execution strategy on a two-dimensional graph, where xaxis represents its standard deviation, and the y-axis its
returns. By connecting all the strategies together, we get
the plot presented in Figure 3. This profile is a
combination of results from Figures 1 and 2. The semicircle shape of the graph can be explained by the dip in
the risk function described above.
This shape has one important implication: many
trading strategies from our setup are sub-optimal. Only
the top part of the risk-return profile where the increase in
risk results in higher expected returns should be
considered in the strategy selection process. This is what
we call an efficient pricing frontier (a similar concept is
used in [Kissell and Glantz, 2003]). In this example, this

is the upward-sloping section of the curve, which

connects the point of minimum risk (8, -18) to the point of
maximum returns (29, -9). For any other point along the
curve, we can always find either less risk for the same
expected return, or higher expected return for the same
level of risk, or both. In terms of actual trading strategies,
we conclude that it only makes sense to price limit orders
in the interval [best bid +5, best bid +11].

done within one-hour period; solid line represents 1,000

shares, dashed line 5,000 shares, dash-dot line 10,000
shares. Trading a smaller volume is clearly less expensive
than a larger quantity. Returns and frontier curves are
therefore stacked horizontally without intersecting.
Furthermore, trading larger orders is riskier, as we can see
from the second graph in Figure 4 larger orders put a
larger dollar value at the risk of an adverse price
movement during the execution period, thus again making
large-order risk curves dominate those of smaller orders.
From the position of the peaks in return curves and
from the shape of the pricing frontier (they are shifting to
the left with increasing size), we can conclude that the
trader has to price his orders more aggressively for larger
quantities. While it is clear that trading large volumes is
costly, it is difficult to propose a definite remedy most
of the time acquiring or selling of a significant block of
securities is a necessity. One way to address this issue is
to split a large order into several pieces and transact them
sequentially.

B.
Figure 3. Trade-off between risk and return. Only the
top portion where higher risk results in higher returns
should be used for execution.
This is the pivotal part of our analysis: we run
historical simulations to construct return and risk curves,
combine the two into the risk-return profile, and
ultimately extract the efficient pricing frontier. Using this
frontier, the trader can do one of two things pick a target
level of returns (price improvement) and find a strategy
that will deliver these returns in expectation with minimal
risk; or he can select a level of risk he is comfortable with
and get the strategy which will deliver the highest
expected return for that level. Obviously, after picking a
point on the frontier, we need to refer back to the returns
and risk graphs to determine the corresponding limit
price.

V. Results.
In this section we present a summary of our results.
Our goal here is two-fold: to show practical applications
of the suggested model, and to point out various
microstructure variables that must be taken into account
during the analysis. We first examine the effects of
modifying the inputs of the execution strategy order
size, execution window, and time of the day, and then
explore the real strength of our approach: conditioning the
execution on the state of the market trading volume and
book depth in this case. For every variable we examine
we provide plots of returns, risk, and pricing frontiers.

A.

Order Size.

Perhaps the most straightforward parameter of the

execution strategy is the order size. Everything else kept
constant, it is more expensive to trade larger orders.
Figure 4 (last page) illustrates this point. All trading is

Time Window.

If we are to divide a large order into multiple small

orders, we must reduce the execution window for each
smaller piece. This effect is explored in Figure 5. Trading
here is performed over 60 minutes (solid line), 10 minutes
(dashed line), and 1 minute (dash-dot line); every
transaction is for 1,000 shares. Not surprisingly, return
graphs show that it is more expensive to transact over
shorter time intervals limit orders remain in the book
only briefly thus making it less likely that transaction
price will reach the limit level, forcing the trader to
submit market orders and incur price impact. The case of
the time window is not as clear-cut, however, as that of
the order size. While transacting on longer time scale is
less expensive, it is also more risky the solid line
dominates the others in both return and risk plots. This
means that none of the three strategies is strictly superior,
and therefore the choice of the time window depends on
the traders attitude towards risk and return.
The efficient frontier plot illustrates this point: if the
trader picks the target risk level of 15, then he should buy
1,000 shares over 10 minutes (the dashed line corresponds
to highest expected returns for that level of risk), whereas
if the he can tolerate the risk level of 20, then he should
transact over 60 minutes and expect higher returns
(smaller transaction costs). Final observation: similar to
the order size, shorter execution time necessitates more
aggressive order pricing.

C.

Time of the Day.

One other variable that the trader can potentially

control is the time of the day when the execution is
performed. Temporal liquidity patterns are well
documented: there is more volume right after the open
and before the close than in the middle of the day. We are
trying to answer a slightly different question: should we

consider time of the day as a separate input variable,

which can influence the outcome of our analysis? In other
words, do curves differ significantly in the morning and in
the afternoon? Figure 6 seems to suggest that time of the
day indeed makes a difference. There we transact 1,000
shares over 60 minute intervals starting at 11 am (solid
line), 12 pm (dashed), and 2 pm (dash-dot). We avoided
open and close on purpose, since it is widely believed that
the price formation process and liquidity dynamics are
different during those times. While Figure 6 does show
that curves can differ significantly from one time period
to another, it is impossible to make meaningful
generalizations. It does not appear that trading at a
particular time of the day is more profitable than at some
other time.
When we trade during a 10-minute window, however,
all curves are much closer together, which makes time of
the day effect much less pronounced. Therefore, if the
trader is planning on transacting over a long time period,
he should take the time of trade into account; otherwise,
this variable can be disregarded.

D.

Market Conditions.

The real strength of our approach is presented in

Figures 7 and 8, where we condition our optimization on
the state of the market. In all the previous experiments,
the optimal pricing frontiers that we have derived are
unconditional i.e. we use all the data available to us in
order to construct the curves. It may be more informative
and practical to condition our results on a specific state of
the market by using only those parts of data that conform
to desired conditions. In Figure 7, we create two sets of
curves for days with high and low transaction volume
(solid and dashed line correspondingly). It appears that it
is cheaper to trade on high-volume days, but it is also
riskier. High volume means more liquidity and thus
smaller market impact, but surges in volume are also
correlated with higher volatility thus making adverse
price movements more likely. Just as in the case of
different execution windows, efficient pricing frontiers
intersect in a non-trivial way, and thus the choice of an
optimal pricing strategy depends on the traders risk
tolerance. Also, orders should be priced more
aggressively on low-volume days. This is consistent with
our previous findings.
In Figure 8, we want to see how our results change
when we submit our orders into thick (solid line) and
thin (dashed line) books. (We define the depth of a
book by the total volume within 20 cents from the inside
market). Results are similar to those in the transaction
volume conditioning experiments, but somewhat less
clear. This leads us to believe that the depth of the book
may be not as significant of a variable as trading volume,
when it comes to limit order pricing. In any event, our
goal here is to demonstrate how our optimization
technique can be applied to specific market conditions.

There are many other conditioning schemes that can be

informative: low vs. high price volatility, low vs. high
volume volatility, directional market, and so on. We
explain how our model can be extended even further in
the next section.

VI. Generalization and Future Directions.

While we spent the bulk of this paper laying out our
basic methodology, it is important to remember that this
analysis is only a building block, which can be expanded
considerably. In all the experiments we have conducted,
our trading strategies remained static we submit a limit
order at the beginning of the trial interval and then do not
touch them until the last second. Such approach to trading
is certainly unrealistic, but our findings can be extended
to a more complex framework.
Say we have an hour to acquire 1000 shares of MSFT.
We can look at the results of our experiments and figure
out that, for example, the optimal strategy is to submit a
limit order at the best bid minus one cent; which is what
we do. But, as the time goes by, conditions may change:
the inside market may move far away from our standing
order, or our order may get partially executed, or the
spread may widen, etc. But as all these events happen, at
any point after the initial submission we have an option to
re-evaluate our pricing strategy. We can use the shortened
time horizon and updated order size to go through the
same analysis as before, and then come up with a new
optimal price and re-submit our order. If we perform this
operation continuously, we can ensure that at every point
in time we are trading at the lowest expected cost.
One way to envision this process is through a large
look-up table as shown in Table 1. This matrix specifies
the optimal order price for every possible combination of
order size and time to execution. The table can be filled
following the analysis from Section IV. The number of
entries in Table 1 is likely exaggerated, as it can be built
with an arbitrary resolution i.e. sampling every 10
minutes and every 100 shares probably makes more
sense.
1000 s.
999 s.

1 share
60 min
Bid-10
Bid-11

Bid-1
59 min
Bid-9
Bid-11

Bid -1

1 min
Bid+100 Bid+99

Bid+1
0 min
market
market

market
Table 1. Look-up table for optimal price updates.
While we can adopt this data-intensive approach and
run a large number of experiments to construct the lookup table exactly as described above, this is not very
practical. First, it will take a significant amount of time,
processing power, and raw data to create such table; then,
the resulting database will be very large and difficult to
maintain and update; and, finally, most information will
be repetitive and thus redundant because of the

similarities in trading characteristics across stocks.

Therefore, the most promising development direction is to
try to come up with a functional description of the optimal
order placement strategy.
So far, we have been deriving our optimal price by
visually determining the peak of the curve. Through our
experiments in Section V, we have already gained some
key insights: the limit price of the optimal order is directly
related to the size of the and inversely related to the time
to execution, just to mention a few dependencies. There
are other factors that clearly must be considered: bid-ask
spread, liquidity, volume, volume volatility, price
volatility, time of the day, etc. We can run a regression
using these inputs and having the optimal limit price as
output. From the technical standpoint, this essentially
amounts to finding coefficients for the following
equation:
Limit distance = 1*Volume + 2*Volatility +
3*Spread + 4*Size + 5*Time +
The benefit of this approach is that we dont have to
use the entire universe of stocks to come up with an
accurate model; and when we know the coefficients, we
can revise our estimates of the optimal price on-line. The
main challenge of doing this derivation is that it will
require even more data analysis to be performed on a very
large dataset we will have to quantify and keep track of
spreads, volumes, price volatility, volume volatility, etc.
This can be a very data-intensive and challenging process.
And, finally, to achieve even more precision and cost
savings, we can introduce conditional optimality. This is
essentially an extension of the conditioning experiments
from Section V: in practice, we want to transact
differently under different market conditions. For
example, in a market with a pronounced trend we want to
trade more aggressively than when a stock is essentially
flat. Therefore, all our optimizations should be performed
for the market conditions in which we are planning on
executing. When we are building a look-up table as in
Table 1 learning optimal prices for different times and
sizes, we can either use all the available data, thus getting
the unconditionally optimal prices, or we can use only
those pieces of data that conform to our desired
conditions. If we know beforehand that we will be
transacting in a rising market, then we should use only the
data from the prior rising markets, and thus learning
optimal limit order prices conditioned on the uptrend in
the stock. This should render our trading much more
precise and efficient.

Conclusion.
In this paper, we propose a limit order book approach
to efficient execution. We demonstrated how to estimate
return curves, risk curves, and risk-return profiles by
using historical data, and how to derive optimal pricing
frontiers. Our quantitative method allows traders to

optimally price their limit orders in order to minimize

trading costs and control corresponding risks. Through
many experiments we have studied the behavior of risk
and return in this domain, highlighting the importance of
a number of microstructure variables: order size, time
window, liquidity, etc. And, finally, we suggested several
ways how our methodology can be taken to the next level:
continuous order revision, functional description of price
curves, conditional optimization, and others.
We believe that our main contribution is making the
first step towards a very important task of optimizing
trade execution in order-driven markets. We introduce
precision in the process of order submission, and more
specifically, optimal limit order price determination. We
emphasize, however, that our method is just a building
block, which can be turned into much more significant
research projects. In this regard, we remark that we are
currently engaged in a large-scale application of the
methods of reinforcement learning to optimized execution
along the lines discussed in Section VI.

Bibliography.
1. Almgren, R. and Chriss, N., Value Under
Liquidation, Risk, 1999.
2. Almgren, R. and Chriss, N., Optimal Execution of
Portfolio Transactions, Journal of Risk, 2002.
3. Bertsimas, D. and Lo, A., Optimal Control of
Execution Costs, Journal of Financial Markets 1, 1998.
4. Bias, B., Hilton, P., and Spatt, C., An Empirical
Analysis of the Limit Order Book and the Order Flow in
the Paris Bourse, Journal of Finance, 50, 1995.
5. Handa, P. and Schwartz, R., Limit Order Trading,
Journal of Finance 51, 1835-1861, 1996.
6. Harris, L., Trading and Exchanges: Market
Microstructure for Professionals, Oxford University
Press, October 2002.
7. Hasbrouck, J., and Harris, L., Market vs. Limit
Orders: The SuperDot Evidence on Order Submission
Strategy, Journal of Financial and Quantitative Analysis,
31, 1996.
8. Hollifield, B., Miller, R., Sandas, P., Empirical
Analysis of Limit Order Markets, working paper,
Carnegie Mellon University, 2003.
9. Nevmyvaka, Y., Kearns, M., Papandreou, A., and
Sycara, K., 2004, Optimized Execution Via Order Book
Simulation, working paper, CMU.
10. Kissell, R., and Glantz, M., Optimal Trading
Strategies: Quantitative Approaches For Managing
Market Impact and Trading Risk, AMACOM, 2003.
11. Lo, A., MacKinlay, C., Zhang, J., 2000,
Econometric Models of Limit-Order Executions, Journal
of Financial Economics, 65, 2002.
12. Markowitz, H., Portfolio Selection, Journal of
Finance, 7 (1), 77-91, 1952.

Figure 4. Transacting larger orders is both more expensive and riskier.

Figure 5. Execution over shorter time periods can be more expensive, but less risky.

Figure 6. When executing over 60 minute period, time of the day should be used as one of the models inputs.

Figure 7. Transacting on a high-volume day is less expensive and less risky.

Figure 8. Thick book may be preferable to thin book.