s our search engine crawled the Web.
The web site itself may have changed. You can check the current page or check for previous versions at the Internet Archive. Yahoo! is not affiliated with the authors of this page or responsible for its content.
How Effective Are Effective Spreads? An Evaluation of Trade Side Classification Algorithms How Effective Are Effective Spreads?
An Evaluation of Trade Side Classification Algorithms
Ananth Madhavan
Kewei Ming
Vesna Straser
Yingchuan Wang*
Current Version: November 20, 2002
Abstract
The validity of the most commonly used measure of market quality using public
data, the effective spread, depends on the accuracy of trade side classification and the
benchmark quote. Using a combination of NYSEs Trade and Quote Data (TAQ) and
proprietary order data, we compare the performance and the consequent effective spread
biases of two widely accepted algorithms for samples of Nasdaq and NYSE stocks in the
post-decimalization environment. Results reveal much higher sensitivity of the effective
spread estimates to the choice of benchmark quote in the current trading environment
than previously reported. This is due to high discrepancy between the quotes recorded at
the order submission and the order execution time resulting from high frequency of trade
and quote updates. The magnitude of the bias is similar on both markets and can be
reduced by lagging benchmark quotes. Overall our results demonstrate significant
difficulty in the quest for the actual transaction cost measure and at the same time call for
more detailed trade reporting/collecting rules.
* All authors are employed by ITG Inc., 380 Madison Avenue, New York, NY 10017.
The information contained herein is for informational purposes only and has been
compiled from sources, which we deem reliable. ITG Inc. does not guarantee the
accuracy or completeness or make any warranties regarding the results from usage. All
information, terms and pricing set forth herein is indicative based upon, among other
things, market conditions at the time of this writing and is subject to change without
notice. Additional and supporting information is available upon request. ITG Inc. is a
member of NASD, SIPC.
© 2002 ITG Inc. All rights reserved.
110502-90312 2
1

Introduction
Transaction costs and execution quality are topics of great interest from both a
practical and academic viewpoint. Numerous studies have attempted to quantify
differences in market quality across various market structures (Madhavan (2000) presents
a survey). One of the key metrics for market quality is the effective spread, defined as
the signed difference between trade price and the bid-ask midpoint prevailing at the time
of order submission. Indeed, in January 2001, the U.S. Securities and Exchange
Commission adopted Rule 11Ac1-5, which requires all market centers to publish their
trading quality on a monthly basis. Such required trading quality measures include
effective spread, the rate of price improvement, execution speed, and others. Accurate
computation of effective spread requires two crucial pieces of information, the trade side
and the prevailing quote. Commonly available transaction data sets, such as the NYSEs
Trade and Quote (TAQ) data or Nastraq, provide neither, raising the possibility of serious
biases in intermarket comparisons. As such the calculation of effective spread relies
heavily on the accuracy of the underlined trade direction algorithm and the choice of the
benchmark quote. Errors in either can result in significant overstatement (or
understatement) of the actual effective spread and in turn lead to erroneous conclusions
regarding relative performance of alternative market centers.
The objective of this paper is twofold. First, we want to evaluate the accuracy of
the effective spread estimates calculated based on publicly available data as compared to
the true spreads obtained from data that includes actual trade side and order submission
time. Second, we want to assess the relative magnitude of biases due to errors in trade
classification and due to incorrect assignment of the benchmark quote.
Several algorithms have been proposed to identify whether a trade was initiated as
a buy or sell order. These algorithms include the quote rule, the tick rule, the Lee and
Ready (LR) (1991) rule, and the latest addition, the Ellis, Michaely and OHara (EMO)
(2000) rule. We chose to compare performance and consequent effective spread biases of
the most frequently used algorithm, the LR rule, and its latest improvement, the EMO
classification algorithm. Our analysis is conducted on samples of Nasdaq and NYSE 3
listed stocks, respectively, trading in a decimal tick-size environment. We use a
combination of NYSEs Trade and Quote Data (TAQ) and actual trade direction data
proprietary of ITG Inc. (ITG) from June 2002. Effective spread and relative effective
spread are computed based on true side data, LR-based side, and EMO-based side. We
evaluate different levels of quote lags in applying the LR and EMO algorithms.
Evaluation of trade classification and consequent biases of the competing
algorithms has already received much attention in recent financial literature.
1
However,
these analyses were performed using data prior to the reductions in tick size beginning in
1997.
2
Previous research has shown that biases tend to increase with higher frequency of
quote updates and trading volume and a smaller tick size. We revisit the issue as market
conditions have changed considerably since previous analyses. In comparison to 1997,
the average trading volumes have more than doubled and tick size has decreased to
decimal increments.
3
Indeed, our results demonstrate that effective spread estimates are increasingly
dependent on the specifics of the algorithm used for their calculation. Specifically, in the
current trading environment, the accuracy of the effective spread estimates is far more
sensitive to the choice of the benchmark quote than previously believed. Analysis shows
that this finding is due to high discrepancy between the quotes recorded at the order
submission and order execution time, a consequence of high frequency of quote updates
in a smaller tick environment. We find that the EMO algorithm based effective spread
provides smaller bias resulting from the incorrect trade side classification. However, due
to large differences between benchmark quote recorded at order submission and trade
execution time, it is the LR algorithm based effective spread that is significantly closer to
the true spread. When correcting for this time bias, we find that for NYSE stocks the

1
They include Lee and Radhakrishna (2000), Odders-While (2000), Peterson and Sirri (2002), Finucane (2000), Ellis,
Michaely and O'Hara (2000), and Piwowar and Wei (2001).
2
The most recent analysis by Piwowar and Wei (2001) uses July 1999 Nastraq data. Since their data does not include
the information about the order submission, their study is not directly comparable to ours. Peterson and Sirri (2002)
study, which is the closest to our study in terms of data availability and questions addressed, uses transaction data from
June 1997.
3
For example, in 1997, trading volumes on NYSE and Nasdaq markets were 132.7 and 149.9 billion shares
respectively. In 2001, the volumes jump to 301.0 billion shares on NYSE and 456.3 billion shares on Nasdaq. As of the
end of September 2002, we have observed 263.7 and 314.5 billion shares on NYSE and Nasdaq markets. 4
smallest bias is achieved using EMO algorithm while for Nasdaq stocks the smallest bias
is still achieved by LR algorithm.
While quote lags can effectively reduce the time delay bias, the actual reporting
lag information is not readily available in public data such as TAQ or Nastraq. In turn, a
recommendation on what is the optimal algorithm for accurate estimation of the effective
spread becomes very difficult. Based on our data, we conclude that the optimal quote lags
for the determination of trade side is 2 seconds for both algorithms. The optimal time lag
for determination of the benchmark quote, however, is much more ambiguous. If using
the LR method, no time lag in assignment of the benchmark quote results in the most
accurate estimate for NYSE market while for Nasdaq stocks the best lag is 2 seconds. If
the EMO algorithm is used, we recommend a quote lag of 10 seconds for NYSE stocks
and 2 seconds for Nasdaq stocks. While the latter procedure provides the least amount of
bias, the optimal trade delay may vary depending on the liquidity of the sample used in
the analysis and may also change with time. We conclude that considerable caution
should be used in the interpretation of the estimated effective spreads depending on the
algorithm and time period of data used in the analysis.
The study proceeds as follows. In Section 2, we provide a brief description of trade
classification algorithms and discuss how our study relates to previous work on the topic.
In Section 3, we present sample selection and data considerations. In Section 4, we
present and discuss the results. We conclude in Section 5.
2

Background and Previous Research
Four major trade classification algorithms are widely used: the quote rule, the tick
test rule, the LR rule, and the EMO rule. While the quote method infers trade dir