JOT

F A L L 2 0 1 3 V O L U ME 8 N U MB E R 4 WWW. I I J O T . C O M
The Voices of Influence | iijournals.com
Cover_mw_Layout 1 9/27/13 1:48 PM Page 1
THE JOURNAL OF TRADING 9 FALL 2013
Optimal Trading Algorithm
Selection and Utilization:
Traders Consensus versus
Reality
JINGLE LIU AND KAPIL PHADNIS
JINGLE LIU
is a quantitative researcher
for algorithmic trading
at Bloomberg Tradebook
LLC in New York, NY.
[email protected]
KAPIL PHADNIS
is a quantitative researcher
for algorithmic trading
at Bloomberg Tradebook
LLC in New York, NY.
[email protected]
T
he market impact of orders cannot
be directly observed and is typi-
cally inferred from the data. Aca-
demic literature proposes various
models to describe market impact cost curves
and shows how real-world data can be fitted
into these models. Market impact curves
versus order size have been observed as con-
cave, indicating lower proportional impact
for larger orders. The market structure has
changed quite a bit in the last decade and is
constantly evolving. The U.S. equities market
is extremely fragmented, with 14 exchanges
and numerous dark pools. Market impact
models have traditionally looked at macro
factors, including average daily volume,
average spread, and volatility, among others,
to explain the cost of a trade. As algorithmic
tradings popularity increases, algorithm
types and their parameters become very
relevant factors affecting an orders market
impact. Institutional buy-side traders use
broker algorithms to execute orders in the
markets. These algorithms and their effect on
execution costs are the focus of this article.
Trading algorithms slice big orders into
smaller individual suborders in various styles
and route them to different venues using smart
order routing. Algorithms vary by a traders
specific objectives, which could be to trade at
a particular percentage of volume, to follow
volume-weighted average price (VWAP), to
use volume prof ile, and so on. Algorithms
also try to minimize their footprints while
extracting available liquidity eff iciently by
reacting to real-time market activities. For
buy-side traders, however, using trading
algorithms alone does not guarantee better
order execution performance. Understanding
the characteristics and performance of various
types of algorithms provided by sell-side bro-
kers is crucial to selecting optimal algorithms
for certain orders under certain market con-
ditions to better achieve investment objec-
tives. Past studies compare the performance
of different types of algorithms (Domowitz
and Yegerman [2006]; Kissell [2007]).
In this article, we investigate execution
performance of four types of commonly used
trading algorithms provided by brokers. Dis-
tribution of trade cost (average traded price
versus midpoint price at arrival of the order) is
used to quantitatively evaluate the execution
performance. We present results of a study
on traders consensus patterns for selecting
algorithms and compare those results with
actual algorithm performance. Furthermore,
we explore the dependence of performance
on order size, participation rate, limit prices,
and algorithm type. These results can be
useful to buy-side traders, helping them to
further improve their algorithmic decision-
making process and select optimal algorithm
parameters.
The following section introduces the
dataset for this study and describes termi-
JOT-LIU.indd 9 9/18/13 7:31:00 AM
THE JOURNAL OF TRADING 9 FALL 2013
Optimal Trading Algorithm
Selection and Utilization:
Traders Consensus versus
Reality
JINGLE LIU AND KAPIL PHADNIS
JINGLE LIU
is a quantitative researcher
for algorithmic trading
at Bloomberg Tradebook
LLC in New York, NY.
[email protected]
KAPIL PHADNIS
is a quantitative researcher
for algorithmic trading
at Bloomberg Tradebook
LLC in New York, NY.
[email protected]
T
he market impact of orders cannot
be directly observed and is typi-
cally inferred from the data. Aca-
demic literature proposes various
models to describe market impact cost curves
and shows how real-world data can be fitted
into these models. Market impact curves
versus order size have been observed as con-
cave, indicating lower proportional impact
for larger orders. The market structure has
changed quite a bit in the last decade and is
constantly evolving. The U.S. equities market
is extremely fragmented, with 14 exchanges
and numerous dark pools. Market impact
models have traditionally looked at macro
factors, including average daily volume,
average spread, and volatility, among others,
to explain the cost of a trade. As algorithmic
tradings popularity increases, algorithm
types and their parameters become very
relevant factors affecting an orders market
impact. Institutional buy-side traders use
broker algorithms to execute orders in the
markets. These algorithms and their effect on
execution costs are the focus of this article.
Trading algorithms slice big orders into
smaller individual suborders in various styles
and route them to different venues using smart
order routing. Algorithms vary by a traders
specific objectives, which could be to trade at
a particular percentage of volume, to follow
volume-weighted average price (VWAP), to
use volume prof ile, and so on. Algorithms
also try to minimize their footprints while
extracting available liquidity eff iciently by
reacting to real-time market activities. For
buy-side traders, however, using trading
algorithms alone does not guarantee better
order execution performance. Understanding
the characteristics and performance of various
types of algorithms provided by sell-side bro-
kers is crucial to selecting optimal algorithms
for certain orders under certain market con-
ditions to better achieve investment objec-
tives. Past studies compare the performance
of different types of algorithms (Domowitz
and Yegerman [2006]; Kissell [2007]).
In this article, we investigate execution
performance of four types of commonly used
trading algorithms provided by brokers. Dis-
tribution of trade cost (average traded price
versus midpoint price at arrival of the order) is
used to quantitatively evaluate the execution
performance. We present results of a study
on traders consensus patterns for selecting
algorithms and compare those results with
actual algorithm performance. Furthermore,
we explore the dependence of performance
on order size, participation rate, limit prices,
and algorithm type. These results can be
useful to buy-side traders, helping them to
further improve their algorithmic decision-
making process and select optimal algorithm
parameters.
The following section introduces the
dataset for this study and describes termi-
JOT-LIU.indd 9 9/18/13 7:31:00 AM
10 OPTIMAL TRADING ALGORITHM SELECTION AND UTILIZATION FALL 2013
nology used in the rest of the article for results around sta-
tistics and algorithms. Next, we analyze the main results
of our study and present data supporting our hypothesis
for optimizing algorithm usage. The final section sum-
marizes the article and highlights the main conclusions.
TRADES CONSENSUS
AND ACTUAL PERFORMANCE
Data Summary of Algorithm Characteristics
Our dataset consists of more than 270,000 buy-side
orders executed using trading algorithms provided by
Bloomberg Tradebook in the U.S. equity market. The
dataset encompasses a time frame large enough to incor-
porate a variety of market conditions. All algorithms
were executed during regular market hours and exclude
all stocks listed as pink sheets and the OTC Bulletin
Board stocks. The dataset consists of more than 4,500
tickers distributed across all market capitalizations and
sectors.
We categorize the algorithms into four groups:
scheduled, participation rate, dark, and implementa-
tion shortfall. Scheduled algorithms are also known as
VWAP algorithms because their goal is to achieve an
average trade price as close as possible to the VWAP for
the duration of the order. Participation rate algorithms
aim to participate in the market at specified rate with
respect to the market volume. Dark algorithms restrict
participation to venues known as dark pools. Dark
pools are reported to execute 14.26% of consolidated
U.S. volume as of February 2013 (Rosenblatt [2013]).
Implementation shortfall algorithms aim to be oppor-
tunistic with respect to arrival price benchmark while
minimizing market impact.
We summarize percentage distribution of orders in
each category as shown in Exhibit 1, grouped by average
daily volume (ADV, 30 day), order size as percentage
of ADV, participation rate based on fill quantity over
market volume in price, and algorithm duration. We
group the properties mentioned above into three buckets
each, where each bucket specifies a range that is deter-
mined based on data properties. We describe the table
data by grouped properties below.
We def ine low (0 to 1 million shares), medium
(1 million to 10 million shares), and high (greater than
10 million shares) ADV levels as shown in Exhibit 1.
Scheduled, participation rate, and implementation short-
fall algorithms are more popular for trading stocks that
trade 1 million to 10 million shares a day. In contrast,
more than 55% of time dark algorithms are being used
for equities of low ADV. This observation aligns with
the common belief that traders are more inclined to
trade less liquid equities in dark pools.
Order size is normalized using the stocks 30-day
ADV to make the statistics across different stocks more
comparable. This category is also broken into three
groups: small orders (0% to 1%), medium orders (1%
E X H I B I T 1
Mean of Stock ADV, Order Size, Participation Rate, and Duration of Four Classes of Algorithms
Note: Percentage normalized within each property and algorithm type.
JOT-LIU.indd 10 9/18/13 7:31:00 AM
THE JOURNAL OF TRADING 11 FALL 2013
to 10%), and large orders (10% to 100%). Among all
algorithms, more than 60% of all the orders have an
order size smaller than 1% of ADV, and more than 95%
of all the orders have an order size smaller than 10% of
ADV. This result suggests that cautious and risk-averse
traders inherently chop parent orders into smaller pieces
to avoid leaking their intentions to the market during
the execution of the parent order.
To further confirm this argument, we counted the
number of suborders within traders parent orders. Parent
orders were constructed by combining all the suborders
sent to a broker together based on same trader, ticker,
and side during the same day. Our finding shows that
parent orders larger than 10% of ADV are typically split
into two or three smaller orders by tradersconsistent
with the argument herein.
More than 90% of implementation shortfall algo-
rithms were traded on orders smaller than 1% ADV.
Thus, for small orders, traders prefer to have the order
done quickly by extracting maximum liquidity via
implementation shortfall algorithms. In the meantime,
for very large orders (>10% of ADV), traders rely on
scheduled algorithms to spread out the trade over the
day to minimize the market impact or rely on dark algo-
rithms to take advantage of block liquidity periodically
available in dark pools.
The participation rate is the ratio of algorithms
filled quantity to the market volume within the limit
price during that period of the execution. Participation
rate is strongly related to order size and duration. It
consists of three groups: low (0% to 5%), medium (5%
to 20%), and high (20% to 100%). Dark algorithms,
at 62.9%, have the highest percentage of orders falling
in the highest participation rate category, followed
by implementation shortfall, participation rate, and
scheduled algorithms. This finding could indicate that
traders are finding a high percentage of the liquidity in
dark pools regardless of the average percentage of dark
volume. We surmise that dark liquidity begets more dark
liquidity, and the average dark liquidity might not be the
entire story. We are working on a study to determine if
dark liquidity tends to be autocorrelated in nature, thus
leading to an understating on averages. Participation
rate algorithms are more likely to be used by traders
for targeting a certain level of participation; hence, the
majority fall in the 5%20% participation rate bucket.
Algorithm duration is the time between the algo-
rithms initialization and the time the order is completed
or canceled. Scheduled algorithms typically are used to
work through relatively long periods, with 67.4% of
orders lasting longer than 30 minutes. In contrast, 65.1%
of orders using a dynamic algorithm are finished within
five minutes because of the relatively smaller order size
and higher participation rate.
Data Summary of Trade Cost
Using a meaningful yardstick is important for
algorithmic performance analysis. In this study, we use
implementation shortfall (IS) as a trade cost measure
because practitioners use it widely as a benchmark to
evaluate execution performance. IS is measured as the
difference between the assets prevailing market price
at the time of the investment decision and the realized
average trade price. The normalized trade cost, TC, can
be expressed as

=
( )
avg arrival
arrival
TC S

P

(1)
where P
arrival
is the midpoint quote (average of bid/ask
price) at the point of order entry and P
avg
is the average
traded price of the algorithm. S is the order side and has
a value of +1 for buy orders and 1 for sell orders. All
values are in basis points. We will refer to this statistic
as trade cost for the remainder of the article.
Previous studies indicate that an orders size and
aggressiveness play important roles in driving trade
cost performance (Almgren and Neil [2001]). Smaller
orders tend to outperform larger orders, and participa-
tion exhibits a complicated inf luence on implementation
shortfall. To make a fair comparison of implementation
shortfall performance for the four common algorithms
listed here, we compare them by each pair of order size
bucket and participation rate bucket. Therefore, we
have, in total, 3 3 = 9 categories if we divide order
size and participation rate into three buckets, respec-
tively, as shown in Exhibit 2.
Order size increases from left to right, and partici-
pation rate increases from bottom to top. At each order
size and participation rate category, we calculate usage-
based percentages, which are the ratios of order counts
for each algorithm to the total count for all algorithms,
along with the mean and standard deviation of trade cost.
Exhibit 2 summarizes those numbers. Mean of trade cost
measures expected algorithm performance against arrival
JOT-LIU.indd 11 9/18/13 7:31:01 AM
12 OPTIMAL TRADING ALGORITHM SELECTION AND UTILIZATION FALL 2013
price, and standard deviation of trade cost is an indicator
of pricing risk, or the probability that the average price
is away from the expected value. Both should be consid-
ered for performance evaluation and comparison.
For small orders (<1% of ADV), traders tend to use
scheduled algorithms with a low aggressive level (low
participation rate) or use participation algorithms with a
mid/high aggressive level (mid/high participation rate)
while choosing dark algorithms to search for large fills
and quickly finish the order.
For medium-size orders (1% to 10% of ADV), traders
prefer scheduled or participation algorithms with low or
mid aggressiveness (low/mid participation rate) while
preferring to leverage block liquidity by using dark algo-
rithms to achieve a decent trade rate.
For large orders (>10% of ADV), traders rely on
the scheduled algorithms (34.5% for high participation
rate and 60.9% for medium participation rate) or block
trade opportunities from the dark pools to minimize
their footprints and stay away from predatory (48.1% for
high participation rate). Large orders with low participa-
tion rates have few data points because, naturally, most
of those large orders have >5% participation rate in the
market to get the order done by the end of day.
The usage percentage statistics in Exhibit 2 represent
a consensus of traders opinions about which algorithm(s)
to use based on nature of order itself and urgency level.
Interestingly, the results of trade cost performance do not
necessarily agree with the consensus of algorithm usage.
For example, 77.2% of small orders with low participation
level were done using scheduled algorithms even though
implementation shortfall algorithms provide better average
performance and smaller pricing risk. Compared with the
planned allocation scheme used by scheduled algorithms,
implementation shortfall algorithms can better adjust sub-
order allocation and aggressiveness based on real-time
market price movement and liquidity availability and,
therefore, can achieve lower trading cost, on average.
Implementation shortfall algorithms can immediately
extract continuous liquidity from lit exchanges, but dark
algorithms depend on the discrete liquidity in dark pools.
The continuous filling profile leads to a small variation
of the trade cost of implementation shortfall algorithms.
After finding the matching opposite sides in the dark
pools, however, dark algorithms can trade at a higher par-
ticipation rate without moving the market compared with
algorithms relying on lit exchanges. This result might
explain why dark algorithms are able to achieve lower
average trade cost at mid/high participation rates but not
at the smallest variance.
For medium-size orders (1% to 10% of ADV), dark
algorithms perform the best, while scheduled algo-
rithms perform the worst for mid/high participation
rate; participation algorithms perform the worst for low
participation rate. Best performance of algorithm over-
laps with most-used algorithms for high participation
rate but not for mid/low participation rate.
For large orders (>10% of ADV), dark algorithms
have the highest trade cost performance and smallest
standard deviation, while scheduled algorithms turn in
E X H I B I T 2
Trade Cost Performance for Four Groups of Algorithms for Various Order Sizes and Participation Rates
Notes: Algorithm groups in each category with sampling size of less than 200 are marked in gray and are not taken into consideration for comparison.
Highest-usage-based percentage and best performance in each category are marked in bold.
JOT-LIU.indd 12 9/18/13 7:31:01 AM
THE JOURNAL OF TRADING 13 FALL 2013
the worst mean performance. The poor performance of
scheduled algorithms for large, highly aggressive orders
is likely attributable to the fact that the imposed schedule
reduces the chance of participating in the liquidity event
while increasing the chance of trading too much during
inactive market periods.
High-frequency trading activity in the frag-
mented markets is also considered to play a role in the
performance of a brokerages algorithms. Such preda-
tory strategies estimate the probability of forthcoming
big orders or those under way by watching the tape
and front-running those orders to amplify the market
impact. Scheduled algorithms, which trade in a predict-
able fashion, tend to be more easily spotted by preda-
tory gamers and, therefore, become more susceptible
to technological adverse selection (Agatonovic et al.
[2012]). Opportunistic and dynamic algorithms are
less susceptible to predatory strategies because of the
nature of high dynamics and unpredictability, which
might partly explain the lower market impact of dark
and implementation shortfall algorithms compared with
scheduled and participation algorithms.
ALGORITHMIC PERFORMANCE FACTORS
Order Size
The persistence of order f low is overwhelmingly
attributable to order splitting (Toth et al., [2011]). Order
E X H I B I T 3
Trade Cost Dependence on Order Size
E X H I B I T 3 (Continued)
Note: In Panel C, the vertical line indicates the mean of the distribution.
JOT-LIU.indd 13 9/18/13 7:31:02 AM
14 OPTIMAL TRADING ALGORITHM SELECTION AND UTILIZATION FALL 2013
size is a major factor contributing to trade costs because
the market impact generated by the executed shares
during the early part of the order might affect the later
part of the order execution. Other factors mentioned
earlier have a more complex dependency on order size.
For example, duration depending on order size can affect
trade costs but is also algorithm-dependent.
In this section, we study the conditional distribu-
tion of trade cost with respect to order size as well as the
type of algorithm. Panels A and B of Exhibit 3 show the
mean and standard deviation of trade cost against order
size as a percentage of 30-day average daily volume.
We plot trade cost versus grouped percentage order
size in six intervals given by [0%, 1%], [1%, 3%], [3%,
10%], [10%, 20%], [20%, 40%], and [40%, 100%]. The
boundaries of intervals are selected based on having a
relatively uniform distribution of data density for each
group as well as enough data resolution to show cost sen-
sitivity in relation to order size. We find that the order
size dependence can be fitted by an exponent function
mean(TC) O

, where O is the order size and a is the

exponent. The result of least square regression based on
empirical data is = 0.4761, which agrees with previous
studies that show market impact to be a concave function
of order size (Almgren et al. [2005]). Pricing riskthat
is, the standard deviation of the trade costis also fit
by an exponent function sd(TC) O

, where the best

fit value for is 0.21.
We also provide a unique way to look at the distri-
bution of trade cost with varying order sizes. Exhibit 3,
Panel C illustrates the distribution profiles of the trade
cost for order size intervals. From bottom to top, each
density represents the trade cost for an interval of
order sizes. The vertical line indicates the distributions
meanwe can clearly see it shift higher with order size.
We also observe that the density distribution becomes
less peaked as we go from bottom to top or as percentage
of order size increases. This observation indicates, and
we verify, a substantial change in the third and fourth
momentsthat is, a decrease in kurtosis and an increase
in the positive skew of the distribution along with an
increase in standard deviation. This dynamic indicates
that trading costs become quite uncertain as the order
size goes up. We confirm by estimating standard errors
for these statistics that are robust estimates. The distri-
bution profile changes quickly as order size increases to
20% of ADV and changes less significantly as the order
size rises above 20%.
Exhibit 4 shows the relationship between order size
and estimated mean and standard deviation of trade cost
distribution by algorithm type. Scheduled, participa-
tion rate, dark, and implementation shortfall algorithms
exhibit monotonically decreasing performance with
order size at different rates. Among these algorithms,
scheduled algorithms have the worst performance and
their trade cost curve falls off the fastest compared with
E X H I B I T 4
Trade Cost Dependence on Order Size for Each
Algorithm Class
JOT-LIU.indd 14 9/18/13 7:31:03 AM
THE JOURNAL OF TRADING 15 FALL 2013
the other algorithms. In part, this occurs because
scheduled algorithms typically have more fixed share-
allocation schedules, which does not allow for f lexibility
to adjust the participation rate over the course of trading
and, therefore, cannot capture the occasional excessive
surging liquidity. Another reason is that traders typi-
cally set schedule algorithms to longer periods, leading
to longer algorithm duration and higher pricing risk.
Therefore, for medium and large orders, traders should
be discouraged from using scheduled algorithms when
trade costs are measured with respect to arrival price.
Participation rate algorithm trade cost falls off
before that of dark and implementation shortfall algo-
rithms. Using this curve, we can also estimate a favor-
able order size to use for participation rate algorithms on
the impact curve. Implementation shortfall algorithms
exhibit greater ability to incur lower trade costs because
of the allocation design, which balances market impact
and pricing risk.
Finally, dark algorithms are much less dependent
on order size than the other three and exhibit consis-
tently lower trade cost across the whole range of order
size. This result suggests that trading in dark pools could
be very benef icialespecially for very large orders,
which would not incur the large trade cost of other
algorithms. If certain equity has a decent number of
shares traded in dark historically, traders should strongly
consider using the liquidities in dark pools before or
at the same time that they trade in a lit exchange. We
caution that this result applies only for Tradebooks dark
algorithms. Because dark algorithms are not standard-
ized across brokersthat is, their logic and access to
dark pools is highly broker-dependentresults may vary
across broker/dealers.
The standard deviation of the trade cost of sched-
uled and participation rate algorithms increases much
faster than trade costs incurred when using dark and
implementation shortfall algorithms. For order sizes
larger than 20% of ADV, the standard deviations of the
trade cost of dark and implementation shortfall algo-
rithms are more than 30% lower than that of scheduled
and participation rate algorithms, as shown in Exhibit 4,
Panel B. Because of the relatively fixed time schedule
or participation rate, schedule and participation rate
algorithms tend to last longer and have more pricing
uncertaintythat is, the effect of trade cost standard
deviation is conditional on other factors.
Participation Rate
Participation rate is another important factor that
affects algorithm performance. Higher participation rate
takes away more liquidity from the market, leading to
a larger market impact that will affect the execution
price of the orders remaining shares. Exhibit 5 shows
the trade cost distribution at different participation rate
ranges for all the algorithms. The profile width decreases
as the participation rate increases from 0% to 20% and
then stays relatively the same at above 20% participation
rate. Kurtosis (peakedness) increases as the participation
rate increases.
Exhibit 6 illustrates mean and standard deviation
of execution performance for each individual algorithm
type. Scheduled algorithms performance shows much
E X H I B I T 5
Distribution of Trade Cost Performance against
Arrival Price for Various Participation Rate Ranges
JOT-LIU.indd 15 9/18/13 7:31:03 AM
16 OPTIMAL TRADING ALGORITHM SELECTION AND UTILIZATION FALL 2013
greater sensitivity to participation rate than that of the
other algorithms. Again, this high sensitivity can be
attributed to the rigid share-allocation schedule imposed
by the scheduled algorithm. Interestingly, no significant
pattern of relationship appears between participation rate
and variance of the performance. The participation rate
dependence of performance standard deviation is rela-
tively f lat for all of the algorithms.
Limit Price
Traders typically use limit prices in algorithms to
manage price risk when routes are sliced from larger
orders to brokers. In our dataset, more than 90% of
algorithms are used with a limit price. Limit price affects
both participation rate and trade costs. Our hypoth-
esis is that setting a too aggressive or marketable limit
price would make the execution faster but provide very
little price improvement. On the other hand, setting a
too passive or unmarketable limit price would lead to a
big price improvement but could not guarantee the full
fill. We analyze the trade-off between participation rate
with marketable limit price and trade cost, and we find
optimal curves for the trader to choose limit price based
on the level of order urgency.
All the algorithms with a user-set limit price are
grouped into buckets based on the normalized differ-
ence between arrival price and limit, P
lmttoarrPx
calcu-
lated as
P P
P
lm
PP
t a
P
rrPx PP
a
P
rrPx PP
where the sign of buy order is positive
and the sign of sell order is negative. The more positive
the difference, the more marketable the order. Panel A
of Exhibit 7 plots the trade cost distribution of all the
algorithms with limit price. When limit price is passive
or unmarketable, the distribution has positive average
performance with a long negative tail. When limit price
is set close to arrival price, the distribution has slightly
worse average performance with smaller standard devia-
tion and larger positive skewness, which indicates that
trade cost is tightly concentrated around arrival price.
When the limit price is aggressive or marketable, distri-
bution widens again, with a long positive tail and worse
average performance. Panel B of Exhibit 7 presents the
mean of the trade cost within each bucket. The trade cost
performance gradually decreases as P
lmttoarrPx
becomes
larger, but not at a constant rate. It increases quickly as
limit price changes from passive to neutral, and slowly
as limit price changes from neutral to aggressive.
Panel A of Exhibit 8 shows how the mean of trade
cost changes with limit price for each algorithm type.
In the unmarketable zone, all the algorithms exhibit
similar dependence. In the marketable zone, dark algo-
rithm performance decreases the fastest with limit price,
while scheduled algorithm performance decreases the
slowest. In other words, the average performance of
the dark algorithm is more sensitive to the limit price.
E X H I B I T 6
Trade Cost Dependence on Participation Rate
JOT-LIU.indd 16 9/18/13 7:31:04 AM
THE JOURNAL OF TRADING 17 FALL 2013
The standard deviation of the performance is the lowest
as limit price is set close to arrival and increases linearly
as the price moves away from arrival in either direc-
tion as shown in Exhibit 8, Panel B. Therefore, if a trader
wants to reduce the variance of his or her algorithmic
performance, setting limit price close to arrival price
will help achieve this objective.
To characterize traders behavior while setting
limit prices, and limit prices inf luence on participation
rate, we measure the ratio of missing interval volume
to volume in price, which can be calculated as
V V
V
inPx all
inPx

,
where V
all
and V
inPx
are total volume within order exe-
cution horizon and volume in price, respectively. This
measure indicates how much liquidity traders would
miss by setting a limit price on the algorithms. We call
this liquidity loss ratio. Panel C of Exhibit 7 plots the
liquidity loss ratio as a function of distance of limit price
from arrival in percentage points. A sharp increase is
observed as the limit price approaches the arrival price,
and then the liquidity loss ratio increases slowly all the
way to 0 as the limit price is set more marketable. To
achieve optimal balance between price improvement and
liquidity loss ratio, the utility objective function should
be maximized with respect to limit price, as shown in
the following equation:

( ) max( (2)
E X H I B I T 7
Trade Cost Performance against Arrival Price for
Various Limit Price Ranges
E X H I B I T 7 (Continued)
JOT-LIU.indd 17 9/18/13 7:31:06 AM
18 OPTIMAL TRADING ALGORITHM SELECTION AND UTILIZATION FALL 2013
for the implementation shortfall algorithm, the trader
is willing is to give up a lot of liquidity to get fills at his
or her price. This is less so for participation rate algo-
rithms (i.e., only when they become marketable), and
even less so for dark and scheduled algorithms. Traders
are more inclined to give up liquidity for opportunistic
algorithmsthat is, they are hopeful of getting a favor-
able price.
CONCLUSION
Quantitative analysis of algorithms performance
plays an important role in algorithmic trading. We ana-
lyze trader usage of execution algorithms and extract
consensus patterns. We use these patterns to offer
insights that can help traders make optimal decisions
when selecting algorithms and determining their param-
eters to achieve reduced implicit trade costs.
Analyzing more than 150,000 trading algorithms
executed via Bloomberg Tradebook in varied market
conditions, we provide a comprehensive picture of buy-
side traders usage patterns. By looking at statistics around
algorithm and trade cost parameters, we describe the
type of algorithms and the conditions the traders select
to execute their orders. Statistics we comprehensively
study are ADV, size, participation rate, and duration,
as well as implicit trade costs for four different types of
where L(P
lmt
) is the liquidity loss ratio and is a constant
that depends on the traders risk aversion or the order
urgency level.
Panel C of Exhibit 8 plots trader behavior as a
trader compares normalized limit price (versus arrival
price) with liquidity loss ratio. Here, we characterize how
limit prices are set by traders for different algorithms
potentially giving us insight into the traders perception
of the different algorithms. Interestingly, we see that
E X H I B I T 8
Trade Cost Dependence on Limit Price for Each
Algorithm Class
E X H I B I T 8 (Continued)
JOT-LIU.indd 18 9/18/13 7:31:08 AM
THE JOURNAL OF TRADING 19 FALL 2013
algorithms. By comparing this empirical dataset of usage
patterns with implementation shortfall performance, we
offer insights on using optimal parameters for different
algorithms.
We find that our implicit trade cost curve is con-
cave, scaling as power law with an exponent of close
to 0.5 for all strategies. When examined by strategy
type, trade cost curves are conditional on strategy type.
Scheduled algorithms have the largest exponent, and
dark and implementation shortfall strategies have smaller
exponents. Standard deviation of the estimate of trade
cost also increases with order size and differs for different
strategy types.
We describe different ways of looking at trade cost
distribution versus participation rate, including using
higher moments of trade cost and conditional histo-
grams for various buckets of participation rates. Sched-
uled algorithms exhibit high trade costs as participation
rates increase, but surprisingly, other algorithms show
f latter curves as participation rates increase. We intend
to research this dynamic further, but we anticipate that
other factors are confounding elements in these curves,
including duration, volatility, and the microstructure of
individual stocks.
We construct a metric called the liquidity loss ratio
based on limit prices set on various types of algorithms.
This metric analyzes the usage of limit prices with respect
to loss in volume over the duration of the algorithm. We
will quantify this relationship in a future article, but we
point out the importance of setting limit prices and their
performance impact for different algorithms.
REFERENCES
Agatonovic, M., V. Patel, and C. Sparrow. Adverse Selection
in a High-Frequency Trading Environment. The Journal of
Trading, Vol. 7, No. 1 (2012), pp. 18-33.
Almgren, R., and C. Neil. Optimal Execution of Port-
folio Transactions. Journal of Risk, Vol. 3, No. 2 (2001), pp.
5-39.
Almgren, R., T. Chee, H. Emmanuel, and L. Hong. Equity
Market Impact. Risk, ( July 2005), pp. 57-62.
Domowitz, I., and H. Yegerman. The Cost of Algorithmic
Trading: A First Look at Comparative Performance. The
Journal of Trading, Vol. 1, No. 1 (2006), pp. 33-42.
Kissell, R. Statistical Methods to Compare Algorithmic
Performance. The Journal of Trading, Vol. 2, No. 2 (2007),
pp. 53-62.
Rosenblatt Securities. Rosenblatts Monthly Dark Liquidity
Tracker. March 2013.
Toth, B., I. Palit, F. Lillo, and J.D. Farmer. Why Is Order
Flow So Persistent? arXiv: 1108.1632 (2011). http://arxiv.
org/abs/1108.1632
To order reprints of this article, please contact Dewey Palmieri
at [email protected] or 212-224-3675.
JOT-LIU.indd 19 9/18/13 7:31:10 AM