Liquidity Beyond the Inside Spread:

Measuring and Using Information in the Limit Order Book

Paul Irvine, George Benston

Goizueta Business School
Emory University
Atlanta, GA, 30322

and

Eugene Kandel
Graduate School of Business
and Department of Economics
Hebrew University
Mount Scopus, Jerusalem, 919051

July, 2000

Abstract
World equity markets increasingly convert to electronic trading, in many cases adopting the format of a
pure electronic order book without intermediaries. A distinguishing feature of this format is that a high
proportion of available liquidity is committed (displayed) rather than implicit or hidden. We examine
the properties of a measure of liquidity, the Cost of Round Trip trade (CRT), which aggregates the
status of the limit order book at any moment in time for a specific transaction size. CRT, which
measures the ex ante committed liquidity immediately available in the market, complements the
effective spread, which measures the ex post combination of the committed and hidden liquidity
available over a period of time. We use data from the Toronto Stock Exchange to compare CRT to
the quoted and effective spreads, and estimate its ability to predict the subsequent trading activity.
While we propose CRT as a research tool, we also advocate its use by exchanges to indicate to
investors the level of committed liquidity.

1
We would like to thank Hendrick Bessembinder, Pina DeSantis, Michael Goldstein, Larry Harris, Shehzad Mian,
Tavy Ronen, Wendy Rudd, G. William Schwert, Matthew Spiegel, Sunil Wahal, Zvi Wiener, and seminar
participants at Hebrew University, IDC, and HEC for their helpful discussions. An earlier version of this paper
was presented at the 1997 NYSE Global Equity Trading Conference, and at the 1998 INFORMS Conference. We
are very grateful to the Toronto Stock Exchange for providing the data. Corresponding author: Paul Irvine,
Goizueta Business School, Emory University, Atlanta, GA. 30322.
 Liquidity Beyond the Inside Spread:
Measuring and Using Information in the Limit Order Book

Abstract

World equity markets increasingly convert to electronic trading, in many cases adopting the format
of a pure electronic order book without intermediaries. A distinguishing feature of this format is that
a high proportion of available liquidity is committed (displayed) rather than implicit or hidden. We
examine the properties of a measure of liquidity, the Cost of Round Trip trade (CRT), which
aggregates the status of the limit order book at any moment in time for a specific transaction size.
CRT, which measures the ex ante committed liquidity immediately available in the market,
complements the effective spread, which measures the ex post combination of the committed and
hidden liquidity available over a period of time. We use data from the Toronto Stock Exchange to
compare CRT to the quoted and effective spreads, and estimate its ability to predict the subsequent
trading activity. While we propose CRT as a research tool, we also advocate its use by exchanges
to indicate to investors the level of committed liquidity.
 2

1. Introduction

Liquidity is valuable to market participants, yet is notoriously hard to define. The consensus in
the literature is that liquidity cannot be represented by a single variable. Most empirical studies use
several alternative measures, such as quoted and effective bid-ask spreads and depth at the best bid
and offer quotations, to obtain a robust picture of a market’s liquidity. Each of these measures
captures different aspects of liquidity. The goal of this paper is first to classify these aspects and
then examine an alternative measure of liquidity, which is a scalar representation of the committed
liquidity supply. This new measure has theoretical justification, is intuitive, and allows comparisons
across markets.

As a benchmark for our analyses, we use a hypothetical perfectly liquid capital market. Every
investor is assumed to be a price taker, facing perfectly elastic demand and supply of shares at the
same market-clearing price. We propose to evaluate the liquidity of an actual market according to
its proximity to this hypothetical perfectly liquid market: the closer the resemblance, the more liquid
is the actual market. As a metric of comparison we propose the Cost of Round Trip (CRT) for
transactions of various sizes. The advantage of CRT is that it aggregates the state of the entire limit
order book into a single number for each transaction size. We show that CRT is the best single
predictor of subsequent trading activity. Consequently, we advocate its use as yet another measure
of liquidity in research and industry.

Improved access to limit order book data has been increasingly useful in market microstructure
research. Several papers have found it necessary to summarize the depth in the entire limit order
book to properly answer research questions. Benston, Irvine and Kandel (1997) examine the limit
order book on the Toronto Stock Exchange after decimalization on that exchange. Barclay,
Christie, Harris, Kandel and Shultz (1999) summarize the limit order book to study the impact of
reforms on the Nasdaq market. Goldstein and Kavajecz (2000) use the limit order book to show
how limit order traders and specialists interacted following the reduction in tick size on the NYSE.
All of these papers use essentially the same measure as we discuss in this paper. It is likely that the
usefulness of the limit order book will continue to improve as data becomes available. However, to
date, the measures that summarize the limit order book have not been formally defined nor have the
properties of the CRT been examined and contrasted with other measures of liquidity.

To illustrate the use of CRT, we analyze an extensive data set obtained from the Toronto Stock
 3

Exchange (TSE). The TSE’s Order File allows us to construct the prevailing limit order book for
263 stocks in the TSE 300 Index. We use the limit order book and trade-by-trade data to calculate
CRT and standard liquidity measures, and compare their abilities to predict subsequent market
activity. Although our tests are TSE-based, we see no reason to believe that the results are specific
to the TSE; CRT aggregates the committed liquidity in any market, and thus should be useful for
investors. It is particularly useful in markets where a high proportion of available liquidity is
committed rather than hidden. As world equity markets increasingly adopt a pure electronic order
book architecture, the applications where CRT would be useful should increase.

The paper proceeds as follows. Section 2 presents the proposed framework of analysis and
discusses CRT’s relation to other measures of liquidity. Section 3 describes the TSE trading
procedures, the data and our sample selection process. Section 4 presents the results and Section 5
concludes the paper.

2. Framework

The literature classifies agents active in speculative markets into three broad categories.
Liquidity traders participate in the market because of liquidity constraints or for portfolio rebalancing
reasons. Informed investors trade to profit from their superior information. Market makers or
dealers (broadly defined) try to make money by smoothing order flow imbalances. The two former
groups are usually considered demanders of liquidity; however some trade speed of execution for
price improvement and submit limit orders, i.e. supply liquidity. The latter group is almost always
assumed to be a liquidity supplier. In what follows, we analyze the state of liquidity supply from the
viewpoint of traders who submit market orders.

While the preferences of liquidity traders may differ from those of the informed traders, both
types prefer lower trading costs and smaller temporary price deviations from fundamental values.2
The aversion to the latter increases with the trader’s risk aversion, while the trading costs affect all
traders regardless of their risk preferences. In this paper we concentrate on the cost of trading
rather than on the temporary price deviations from the fundamental value.

Before proceeding further, we would like to classify various aspects of liquidity to focus the

2
We would like to thank Larry Harris for a discussion of this distinction.
 4

subsequent discussion and emphasize our contribution.3 First, we need to define the period of time
over which liquidity is needed and available.4 Some investors may demand instantaneous liquidity to
achieve immediate execution of an order; for example, arbitrageurs between the stock and option
markets, for whom the speed of execution is imperative. Others, such as institutions rebalancing
their portfolio, may need liquidity over a longer period of time (say a day or a week) to execute a
large order. By the same token, a market may offer little liquidity per unit of time on average, but
offer occasional pockets of liquidity, which allow patient investors to execute their orders
inexpensively. We will refer to immediate liquidity supply as liquidity available at a point in time,
and distinguish it from the liquidity supply over time.

Second, we need to define the level of commitment behind the liquidity supply. We refer to
liquidity arising from an outstanding option to buy or sell a certain amount of security, e.g. a limit
order or a quote, as committed or displayed liquidity. Committed liquidity is contrasted with
hidden liquidity, which stems from hidden limit orders, or from potential price improvements offered
on demand by a dealer, floor trader, or specialist. For example, Bacidore, Battalio and Jennings
(2000) propose using the committed liquidity in the limit order book as a benchmark with which to
illustrate hidden liquidity on the NYSE.

Third, it is well understood that liquidity is a transaction-size-specific concept. A liquid market

for small orders may be very illiquid for large ones, or vice versa. For example, Jones and Lipson
(2000) report that liquidity (measured by spreads and depths) for institutional trades worsened
following NYSE’s reduction of the tick size, while liquidity for small trades improved.

Finally, the measures of liquidity can be classified as either ex ante, which capture the liquidity
as it becomes available, and ex post, which capture past liquidity. An ex ante measure of liquidity
is useful to investors, because it indicates the cost at which a trade can be immediately executed.
Consequently, we conjecture that an ex ante measure can predict future investor behavior, such as
order flow. An ex post liquidity measure can only be constructed if a trade has been executed.
Investors can use it to measure the efficacy of their broker’s execution, but it does not capture what
Wagner (1993) refers to as the opportunity cost of not trading. Ex post measures are inherently
limited in their ability to predict future liquidity. They are, however, valuable to researchers trying to

3
Some of the discussion here follows O’Hara (1994) and Harris (1996).
4
In both cases there is a transaction cost that compensates for the immediacy of transaction; as in Demsetz
 5

describe the characteristics of a market, but, perhaps, less useful in predicting future order flow.
The quoted spread is an example of an ex ante liquidity measure, while the effective spread is an
example of an ex post measure. In this paper we focus our attention on developing an ex ante
measure of immediately available committed liquidity.

We use a hypothetical perfectly liquid capital market without intermediaries as a benchmark for
our analysis. In such a market every investor is a price taker, which means that they can buy or sell
any number of shares at the prevailing price. We evaluate the liquidity of an actual market by how
closely it resembles this imaginary perfectly liquid market: the closer the resemblance, the more liquid
is the actual market.

To specify the metric of comparison, consider an investor who wants to buy and sell the same
number of shares at the same time (make a round-trip trade) by submitting market orders. Such an
investor demands immediate liquidity on both sides of the market. We do not question the
motivation for such a transaction, but rather study its cost, i.e. the cost of liquidity. In the perfectly
liquid market, such a transaction would have zero cost. We use the percentage cost of a round-trip
trade as a measure of market liquidity: for a particular trade size, smaller cost indicates a more liquid
market.5 The inside quoted spread, which is frequently used as a proxy for liquidity, is a special
case of our metric for trades that are smaller than the inside depth. For trades larger than the inside
depth, our measure reflects the entire structure of the limit order book, since it takes into account the
liquidity beyond the inside spread and depth. We refer to our measure as the Cost of a Round Trip
trade of D dollars, CRT(D).

In the table below we relate CRT(D) and the conventional measures of liquidity, Quoted Inside
Spread (QIS), Quoted Inside Depth (QID), and Effective Spread (ES), to the aspects of liquidity
outlined above. Given conventional market designs, committed liquidity is always immediate; on the
other hand, the liquidity over a period of time is hidden ex ante and not observable until the
transaction is completed. This implies that certain categories of liquidity cannot be measured
separately.

QIS and QID jointly represent the committed immediate liquidity for small trades. However,
these measures provide little information about the committed liquidity for larger trades. ES(D),

(1968); we vary the degree of immediacy.

5
Later we split this measure into the sell-side and the buy-side measures.
 6

calculated by trade size, is an ex post measure of a combination of committed and hidden liquidity
over a previous time period. It captures the immediate total liquidity for small trades, and liquidity
over time for larger trades. It is useful for evaluating the average cost of executed transactions over
a period of time; however, it cannot be used by an investor to gauge committed liquidity before
submitting a trade.6 This is where CRT(D) is most useful; it is an ex ante measure of immediate
committed liquidity for a given trade size, which shows the investor the upper bound of transaction
cost for immediate execution of a trade of that size.

Commitment Level Immediate Liquidity Liquidity over time

(measured)
Small Orders Large Small Orders Large Orders
Orders
Committed/Displayed QIS, QID --- ---
(ex ante) CRT(D) CRT(D)

Hidden --- --- --- ---

Hidden and Committed ES(D) ES(D)
(ex post)

While, conceptually, CRT(D) is appropriate for any market, it is most informative in markets
where market orders are mostly executed against a limit order book, and least accurate in markets
where most trades are negotiated. This is because the precision with which CRT(D) describes the
market rises with an increase in the proportion of committed liquidity relative to the total liquidity
supply. In a pure limit order market without intermediaries and hidden limit orders, all immediate
liquidity is committed. In this situation, CRT(D) is the most informative. On the other hand, in a
market with a large proportion of negotiated trades, such as Nasdaq prior to 1997, committed
liquidity is only a small portion of the total liquidity supply; thus CRT(D) is less representative. Even
in dealer markets committed liquidity provides some information for comparing the performance of
the same market over time.

6
Roll’s (1984) measure has some of the same features as the effective spread. However, because it cannot be
calculated by trade size, its interpretation with respect to the trade size is unclear.
 7

The dramatic increase in the role of ECNs (electronic limit order books) in the United States,
and the transition of many exchanges to an electronic format (e.g. AMEX, Chicago, Frankfurt,
Paris, Tel Aviv, Toronto) suggests that the proportion of committed liquidity continues to rise in the
equity markets.7 Consequently, the precision and usefulness of CRT(D) as a measure of liquidity
increases as well. While our main goal is to introduce a research tool, we argue that CRT(D) has
practical applications as well as a convenient way of communicating the immediate committed
liquidity of a market to investors. Thus, exchanges should consider adopting it alongside their other
reported liquidity measures.

2.1 Calculating CRT(D)

We describe the calculation of the CRT(D) at a given point in time, omitting the time subscript
for simplicity. Let us denote the inside bid price by P-0, and the inside ask price by P0.
Consequently, all the lower bid prices and higher ask prices in the limit order book are denoted by
P-1 > P-2 > P-3 > P-4 ..., and by P1 < P2 < P3 < P4 ..., respectively. The quantities of shares offered or
requested at these prices are represented by a vector

Q = [... Q-2, Q-1, Q-0, Q0, Q1, Q2, ...].

First let us calculate the number of shares that corresponds to the dollar amount D:

T(D) = 2D / (P-0 + P0)

This is the number of shares one can buy/sell at the midpoint price. Next, we define two indicator
variables, I-k and Ik,, which refer to sell and buy (respectively) orders of D dollars:

1 if T(D) > Σ i=-0,-k Qi

I-k = (T(D) - Σ i=-0,-k+1 Qi ) / Q-k if T(D) > Σ i=-0,-k+1 Qi and T(D) < Σ i=-0,-k Qi

0 otherwise.

And

7
See Glosten (1994) for a discussion of the benefits of the electronic format.
 8

1 if T(D) > Σ i=0,k Qi

Ik = (T(D) - Σ i=0,k Qi ) / Qk if T(D) > Σ i=0,k-1 Qi and T(D) < Σ i=0,k Qi

0 otherwise.

Then the per dollar cost of a roundtrip trade of D dollars, CRT (D), is:

CRT(D) = ( Σk=0,infty Ik Pk Qk − Σ k=0,infty I-k P-k Q-k ) / D

Based on the discussion above, we claim that market i offers higher liquidity for transactions of
size D than market j if CRT i(D) < CRT j(D).

As stated above, CRT(D) is an ex ante measure representing the committed immediate

liquidity. We argue that this measure is useful because it recognizes that the depth of the market at
quotes other than the inside bid and ask may have an impact on liquidity. To illustrate this point
consider, for example, the effect of a decline in the tick size on the Toronto Stock Exchange.
Bacidore (1997) and Porter and Weaver (1997) have stated that liquidity improved following the
decline in the minimum tick size from one-eighth of a dollar to 5 cents. They base this conclusion on
the fact that, on average, effective spreads declined. This decline in the ex post measure of total
liquidity, which includes the hidden orders and the willingness of market makers to provide liquidity,
does not provide evidence on ex ante immediate committed liquidity. The same authors also find
that the depth at the inside quotes declined. Consequently, as Harris (1994) points out, they should
not conclude that liquidity increased, since investors are no longer able to execute large orders at the
inside quotes; therefore, the overall cost of trading did not necessarily decline. Similarly, analysis of
effective spreads by trade size does not provide a definitive answer of whether a smaller tick size
improves liquidity. Huson, Kim and Mehrotra (1997) examine changes in effective spreads for eight
different trade size categories before and after decimalization on the TSE. Their results are
ambiguous. Although effective spreads decline for 500 -1000 share trades, they increase for trades
smaller than 500 shares. Likewise, Ronen and Weaver (1999) find that effective spreads decline for
small trades, but do not decline for larger trades following the reduction in the tick size on AMEX in
1997. Goldstein and Kavajecz (2000) using limit order book data from the NYSE, find support for
Harris’ (1994) arguments in that the liquidity supplied by limit order traders generally decreases
following the reduction in tick size on that market.
 9

Harris’ (1994) argument is further reinforced by the fact that the limit order book (LOB) is not
always full. Biais, Hillion and Spatt (1995) first observed this phenomenon when they have found
holes (prices on the grid where limit orders could be placed, but were not) in the LOB of the Paris
Bourse. These holes are also common in the TSE LOB. We found that the average distance
between the best bid or ask and the next offered price is 12.4 cents, compared with the minimum
tick size of 5 cents. The existence of holes provides an additional support to our claim that the state
of the book is complex and cannot be inferred from the inside spread and depth measures (QIS and
QID).

We reiterate that CRT(D) is not designed to replace the effective spread as the measure of
liquidity. Effective spreads capture ex post trading costs, which is important for many questions.
However, it ignores the fact that trades may occur because ex ante liquidity is high. At these times
the expected cost of a trade, in terms of greater execution speed and certainty, is particularly low.
The opportunity cost of delaying a trade even though the immediate execution cost may be high is
ignored in the calculation of effective spreads. CRT(D), which represents ex ante committed
liquidity, complements the effective spread in that respect.

2.2 Empirical Hypotheses

Following the above discussion we hypothesize that:

CRT(D) is a measure of liquidity, and thus should exhibit characteristics similar to other liquidity
measures in terms of its relation to trading volume and intraday behavior.

On the other hand, CRT(D) provides information about the state of the book that is not
contained in QIS, QID, or ES; thus it should not be highly correlated with these measures, except
when QIS is a special case of CRT(D).

When observed by investors, CRT(D) provides an observable upper bound of the cost of
immediate execution. We hypothesize that lower CRT(D) indicates lower expected cost of
execution for trades of a particular size, which should encourage subsequent submission of market
 10

orders of that size.8 This is the most important test of the proposed measure.

We test these hypotheses using data from the TSE. The next section describes the TSE
procedures, data and sample selection process. The results are presented in Section 4.

3. Data

3.1. Overview of TSE Procedures

The Toronto Stock Exchange (TSE) is one of the ten largest equity markets in the world. In
1995 total trading value was C$207.7 billion, representing 15.7 billion shares exchanged. In
Canada, the TSE accounted for 81.0% of the value of shares exchanged and 57.7% of the share
volume. These trading figures ranked the TSE as the ninth largest exchange in the world as
measured by domestic market capitalization, and the tenth largest as measured by equity trading
volume.

In 1996, the TSE was in the process of converting to a completely computerized trading
system. The architecture for this system has been in place since the Computer Assisted Trading
System (CATS) was first introduced in 1977. In 1996, CATS traded about one-half of the issues
on the TSE. However, most of the CATS issues have lower average daily volume than issues
traded on the TSE Floor. The trading day on the CATS system begins with a pre-opening period,
during which orders may be entered into the system. The CATS opening price is established as the
value that maximizes the total number of shares traded. During the trading day CATS operates as a
continuous market that stores orders on an electronic order book. Traders can post orders in the
book or cross orders by matching orders in upstairs trading and then sending the transaction to the
CATS system.

In 1996, most of the more active issues on TSE traded on the exchange Floor. Each exchange
member may have as many as six floor traders active on the Floor for every exchange seat they
own. These traders execute orders for clients and for their own account. Certain traders are
designated as Registered Traders (RTs). Although RTs are allowed the privilege of trading in their

8
On the TSE, the Paris Bourse and the Tel Aviv Stock Exchange investors can observe up to five quotes away
from the inside on each side, which means that they can calculate CRT(D) for many transaction sizes. The
Instinet ECN reveals their entire order book. On other exchanges only the inside quote is observed; thus,
investors cannot calculate the CRT(D).
 11

stocks for their firms’ account, their primary responsibility is to maintain the integrity of the market
by serving as a market maker in certain stocks, both Floor and CATS. To this end their trading
patterns are monitored and their trading is subject to certain rules. For example, the TSE requires
that at least 70 percent of a RT’s trades must be stabilizing in nature. The most important aspect of
their duties is to maintain a steady quoted spread in these securities, subject to a spread goal (a
time-weighted average spread target). The spread goal for a particular stock is negotiated between
the RT and the exchange staff.

The TSE uses two electronic systems that complement the activities of the RTs. The Market
Order System of Trading (MOST) automatically executes smaller orders from member firms.
Market orders on the Floor of less than the minimum guaranteed fill size, usually 599 shares, are
routed into the MOST system and guaranteed execution; if an offsetting order is not available, the
RT must guarantee execution of the order. Orders that do not qualify for an immediate fill through
MOST are routed to the Limit Order Trading System (LOTS). The LOTS book is an electronic file
of orders at prices other than the current inside bid and ask quotations. A portion of the LOTS
book, for both Floor and CATS issues, is available to investors through the TSE market-by-price
information service. Investors can observe five best bids and offers for each stock and the number
of orders and share volume at each of those prices.

The TSE is not a completely transparent market, because there is hidden liquidity that is not
recorded in the LOB. For example, the automatic obligation of the RT to complete MOST eligible
trades is a form of hidden liquidity. RTs do not necessarily disclose their entire supply of committed
liquidity. Floor traders, who trade on the Floor for their own accounts are a second source of
hidden liquidity. The upstairs markets of TSE member firms provide a third source of hidden
liquidity, because member firms are allowed to internalize trades. Specifically, they have 15 minutes
in which to decide whether to send the trade to the TSE to execute against the LOB, cross the trade
against another internal trade, or send the trade to a competing exchange. The principle effect of this
hidden liquidity will be observed in ES that are lower than CRT(D). In this sense, the ratio of
CRT(D) to the effective spread can be a measure of the transparency of a market, providing a way
to quantify the importance of hidden liquidity in a market.

On April 15th, 1996 the Toronto Stock Exchange introduced decimal trading on all stocks.
The main change was that the tick size for stocks trading at prices exceeding C$5 declined from
 12

12.5 cents to 5 cents. This is a significant change in the minimum tick size, since most of TSE volume
occurs in higher-priced stocks price: the average trade price in 1995 was C$13.16. To avoid
endogenous tick size changes (when stock price crosses the C$5.00 level), we restrict our sample
to firms with a high bid of at least C$5.00.

3.2 Data Description

We obtained a new data set, the TSE Order File, which contains information on every order
submitted to the exchange. The TSE began producing the Order File in electronic form in mid-May,
1996. Therefore, the first available months of this file are June and July 1996. To accurately
classify orders on the Order File as buys or sells it is necessary to know which orders are eligible for
execution on MOST. The minimum guaranteed fill (MGF), discussed above, is the crucial piece of
information required to check MOST eligibility. We collected MGF information from the TSE
Review for all the stocks in the TSE 300, the TSE’s major index, which constitute our initial sample.

From these data we construct a LOB for June 1 through July 31, 1996 by examining every
order, subtracting the number of shares filled at the time the order was submitted, and posting the
remaining number, if any, to the LOB at the specified price.9 The LOB can be described in term of
quotes and depths. There are, of course, two sides to the LOB, the bid side and the ask side. The
lowest ask quote, P0, and the associated depth, Q0, form the inside ask quote, the highest bid, P-0,
and associated depth Q-0, form the inside bid. Orders submitted away from the inside bid and ask
quotes accumulate at each tick size (5 cents). An aggregation of all orders at each bid and ask price
yields the LOB. By examining the entire LOB we contend that we can estimate the impact of
committed liquidity more precisely than can be deduced from the inside spreads and quotes alone.

Orders are entered into the Order File continuously throughout the trading day. It is necessary
to summarize the LOB in a way that is not sensitive to order submission and execution. Biais, Hillion
and Spatt (1995), who examine the LOB on the Paris Bourse, faced a similar problem. They
summarize the Paris Bourse LOB by first calculating the time-series mean of the book for each
stock and then calculating the average LOB by taking the cross-sectional average of the each

9
Griffiths, Smith, Turnbull and White (2000) provide further details on the use of the TSE’s Order File.
 13

individual stock’s time-series mean LOB.10 Because we are interested in measuring time-varying
liquidity, we chose to construct each stock’s LOB at discrete intervals during the trading day.

For each stock-day in the sample period we construct the LOB immediately before the 9:30
AM open , again immediately after the open, at hourly intervals during the trading day, and again at
the 4.00 PM close of trading. This results in a sample of nine LOBs for each stock-day, two at the
open, six during the trading day, and one at the close. We found that approximately 10% of the
opening spreads were equal to zero. These represent orders submitted during the night that cross in
a batch at the open. As such, they do not represent the spread available to a trader who brings an
order to the market after the open. Immediately thereafter the zero spreads disappear and the
distribution of the inside spreads remains relatively constant throughout the rest of the trading day.
Therefore, we measure liquidity in the opening period immediately after the batch processing of
overnight trades is completed. Because we are interested in the effect of liquidity in the LOB on
subsequent trading, we also omit the LOB constructed at the close. This leaves us with seven daily
LOBs for each stock, one calculated immediately after the open, and six more beginning at 10:30
AM and calculated hourly through 3:30 PM.

Two sample LOBs are presented in Figure 1. At 1:30, the inside quoted spread in Barrick
Gold is 15 cents and considerable depth exists on both the bid and ask sides of the LOB one tick
away from the inside spread. On the ask side a particularly large order is present at C$37.50. One
hour later, the inside spread has decreased to 5 cents, and the large sell order at C$37.50 is still in
the LOB. However, the large amount of committed liquidity available at C$37.50 is farther away
from the highest bid price than it was at 1:30, indicating that CRT(D) has increased, at least for large
buy orders.

To construct our final sample we eliminate two groups of stocks in the TSE 300 Index. First,
five stocks with extremely light volume, which averaged less than 2 trades per day over the sample
period, were eliminated. We felt that these stocks would not provide a large enough sample to
estimate the qualities of CRT(D) with any precision. Second, as discussed in the previous section,
we restricted the sample to stock-periods where the highest bid quote was at least C$5.00. Thirty-

10
Our study has a different emphasis and is based on a more extensive sample. Our sample covers more stocks
(263 vs. 40) over a longer period (40 days vs. 19 days). Moreover, it contains all outstanding orders, whereas
Biais et. al. had access only to the five best bid and ask quotes.
 14

two firms in the TSE 300 Index never traded above $5.00. The final sample consists of the
remaining 263 firms.

These firms are a diverse group. The TSE constructs the index to reflect the contribution of
different industries to the market capitalization of the TSE. Some of the firms in the sample are
among the most active on the TSE, but others are comparatively light traders that do not always
have a deep set of limit orders.

Our data set consists of LOBs constructed for these 263 firms. To minimize the influence of
outliers on our results we further restrict the LOBs to have an inside spread of less than $3.00.11 As
discussed above, LOBs where stocks traded below $5.00 were excluded, so that we would have a
constant tick size of five cents. In addition, there are times when the LOB is empty, as Coppejans
and Domowitz (1999) found for a Swedish Futures contract, and Kandel, Lauterbach and Tkach
(1999) found for 122 most active stocks on the Tel Aviv Stock Exchange. Our final data set
consists of 61,239 LOBs. By comparison, 263 stocks sampled 7 times a day for the 40 trading
days in our sample would have produced 73,640 LOBs if the LOB was never empty.

Occasionally, the LOB does not have sufficient depth to calculate CRT (D); that is, the book is
not full. The percentage of all stock-periods during which the book is not full increases with trade
size, a mechanical relation, and decreases with average daily volume. For stocks matched on
average daily volume, the book is full less often on the Floor, where more liquidity is hidden, than in
CATS, where more liquidity is committed. When the book is not full, CRT (D) is calculated by
assuming that there is a perfectly elastic supply of and demand for shares at 50 cent (10 ticks) above
and below the best bid and ask orders.12

In the next section we use the TSE data to compare the various liquidity measures in our
sample and test the hypotheses presented above.

11
This truncation reduced the number of LOBs by only 202. Most of these observations appear to be legitimate
quotes on very high-price stocks, but a few were observed for prices below C$100.
12
We experimented with prices further away from the inside. While this changes the size of CRT (D), the
regression and correlation results presented later in the paper are not affected.
 15

4. Results

Table 1 presents summary statistics on trading activity, depth and spreads for the 263 sample
firms from June 1 through July 31, 1996. These statistics were computed first for each stock, then
cross-sectional averages were calculated: means and standard errors of these averages are shown.
The first column presents the cross-sectional averages for all 263 firms. These data show an
average daily volume of 98.1 thousand shares for a stock with an average price of C$21.39.
Moving to the book, we find that there are a substantially greater number of shares in the LOB than
are revealed at the inside quote. The quoted inside depth (QID) constitutes only 25% of the total
depth in the LOB (total number of shares bid plus ask). On average, total inside depth is 9.0
thousand shares compared to total ask depth of 31.2 thousand shares and total bid depth of 20.3
thousand shares. Average LOB depth is higher on the ask side than the bid side, which suggests
that the direction of trades may be important for measuring liquidity, or it could be just a sample-
specific artifact.

The other columns in Table 1 present the same summary statistics for nine low, medium and
high combinations of price and volume (stock types). We constructed these stock-type sub-
samples by ranking the 263 stocks by average share price and average daily volume (number of
shares) over the period, wherein one-third of the observations are classified as low, medium, or
high. The stock types are grouped into low-volume, low-price (LL) if they rank in the lower third of
both attributes, and so forth. The table reveals considerable cross-sectional variation of stock types
in the sample, with several empirical regularities. Higher volume and lower prices have a strong
effect on depth and spreads. Lower volume and higher priced stocks have significantly smaller
depth (QID, ask, and bid) and lower spreads (QIS and ES), with the exception of high-volume,
high-price stocks (HH) that had an average bid depth marginally larger than that of high-volume
medium-price stocks (HM). Consistent with results observed in other markets, spreads decline as
volume rises and increases as price rises. By any conventional measure of liquidity, high volume
stocks would be considered more liquid.

The results in Table 1 indicate that much of the available liquidity in the market is outside of the
depth offered at the inside bid and ask quotes. Generally, the greater the average daily volume in a
stock, the lower the fraction of limit orders at the inside quotes, but for no stock-type does inside
depth surpass 33 percent of total depth. This supports our assertion that the inside quoted spread
 16

and the inside depth alone are not sufficient to measure the supply of immediate committed liquidity;
an aggregate measure is required. We proceed to calculate CRT(D) using information contained in
the LOB and compare it with the conventional measures of liquidity.

4.1. Liquidity as a Function of Trade Size.

Since liquidity measures have to be trade size specific, we must decide which transaction sizes
to use for our measure of liquidity. We chose five Canadian dollar trade size ranges: $0-$10,000,
$10,001-$25,000, $25,001-$50,000, $50,001-$100,000 and $100,000-$200,000. We choose
ranges of trades rather than exact trade sizes to increase the number of observed transactions in
each group. These five trade sizes are inclusive; they encompass more than 98 percent of all trades
on the TSE. The smallest dollar trade size group, C$0-$10,000, contains the largest number of
trades, 214,500, with an average of 218 shares per trade. The number of trades declines
monotonically with trade size; the C$100,000 - C$200,000 trade size group contains 15,592
trades, with an average of 5,392 shares per trade.

Table 2 presents data on depth and spreads for these five trade sizes and for Floor and CATS,
separately. In addition, we distinguish among the daily-volume and share price groupings (low,
medium, and high) of the stock types presented in Table 1. In Table 2 and hereafter we present
only five of the nine stock types: medium-volume medium price (MM) is the base case, and four
other cases when only one of the categories is changed: low and high volume (LM, HM) and low
and high price (ML, MH).

Two conventional measures of liquidity for which trade size is irrelevant are presented first: QIS
and QID. Then, ES(D), and CRT(D) are presented for each of the five trade sizes. ES(D) is
calculated from trades within each dollar trade size range. Reported ES(D) is the cross-sectional
time-series average of each hours’ effective spreads. CRT(D) is calculated at the middle of each
trade size range (C$5,000, C$17,500, C$37,500, C$75,000 and C$150,000).

Notice that CRT(D) nearly equals QIS for C$5000 trades. Because it is calculated from the
LOB, CRT(D) can never be less than QIS. For this trade size, only rarely is it necessary to go
outside the inside depth to find committed liquidity. As trade size increases, so does CRT(D). This
is a mechanical relation, since committed liquidity offered for a larger order has to be more
 17

expensive (on a per share basis) than committed liquidity for a smaller order. However, this is in
marked contrast with the effective spread, which is non-monotonic in trade size. Larger trades
frequently have smaller effective spreads, which indicates that the order-initiating trader shopped
around for the best execution, rather than demanding immediate liquidity. CRT(D) increases with
price and declines with volume, just like the conventional liquidity measure, QIS. CRT(D) increases
most rapidly for low volume stocks, indicating that these stocks have thinner LOBs than more
heavily traded stocks.

The difference between Floor and CATS is illuminating. Except for a few cases, CRT(D) for
CATS stocks is significantly smaller than CRT(D) for Floor stocks. Recall that CATS, which is an
electronic limit order book, has a larger proportion of committed liquidity, which is captured by
CRT(D). The Floor venue, populated by professional floor traders, offers more hidden liquidity,
which acts as a substitute for the committed liquidity. This phenomenon is reflected in the ratio of
CRT(D) to ES, CRT/ES (presented in Table 2), which generally rises faster with dollar trade size on
the Floor than on the CATS system. This result is consistent with the findings of Sofianos and
Werner (1997), who show that floor brokers provide a significant part of NYSE liquidity, especially
for larger trades.

Occasionally, LOB does not have sufficient depth to calculate CRT(D); that is the book is not
full. Table 2 reports the percentage of all stock-periods during which the book was full (BOOK
FULL %). Empty books occur infrequently in the smallest trade size; however the percentage of
time during which the book is not full rises rapidly with the trade size. For stocks matched on
volume and price, the book tends to be full more often in CATS, consistent with the notion that
higher percentage of liquidity there is committed, as opposed to the Floor, where more of it is
hidden.

If we ignore the observations when the book is not full CRT(D) is subject to a sample selection
bias. To illustrate this point we compared ES in the hour after calculating CRT(D) for cases in
which the book was full and those when it was not (results are not reported for brevity). Depending
on the dollar trade size, ES is on average 60-100 percent higher when the book is not full. This
suggests that the selection bias is potentially quite severe. To mitigate this potential source of bias,
when the book is not full we calculate CRT(D) by assuming that there is a perfectly elastic supply of
and demand for shares at 50 cents (10 ticks) above and below the best ask and bid.
 18

Table 3 presents Pearson correlations between the standard liquidity measures (QID, QIS, and
ES(D), forthcoming effective spreads, and CRT(D) by trade size, stock-type and trading system.
We also include effective spread on average transactions in the previous hour (LAG ES(D)),
because, when observed, it also can be used to estimate current market liquidity. 13 We examine the
correlations to see how CRT(D) varies with these standard measures and to determine whether
CRT(D) contains different information.

CRT(D) is significantly correlated with the standard measures of liquidity: positively with QIS,
ES(D), and LAG ES(D), and negatively with QID. As expected, the highest correlation is with QIS
for the small trade sizes, since the inside spread is a special case of CRT(D). The correlation
between CRT(D) and QIS declines with the dollar trade size, as the inside quote becomes less
representative; this holds for all stock groups. The correlations between CRT(D) and ES(D), and
LAG ES(D) also generally declines with trade size.

We show that the correlations among the liquidity measures for Floor and CATS are similar,
except for CRT(D) with ES(D) for larger trade sizes. The correlation generally is higher on the
CATS system, which is consistent with the absence of floor traders therein. As discussed above,
Floor traders provide an additional source of liquidity by occasionally offering improvement to the
committed LOB prices.

Finally, the correlation of CRT(D) with ES(D) is generally greater than with LAG ES(D), which
suggests that CRT(D) may be a better predictor of the subsequent trading activity than LAG ES(D).
However, this will be tested in the next section.

Finally, we would like to evaluate intra-day behavior of CRT(D) to verify that, being a measure
of liquidity, it exhibits general pattern consistent with other liquidity measures. Figure 2 presents the
average intraday values of CRT(D): it is the highest at the open and declines throughout the day.
This is similar to the intraday pattern of QIS and ES in this sample, and is consistent with findings
from similar markets.

We have examined the levels (Table 2) and correlations (Table 3) of CRT(D) to show the
relation between CRT(D) and conventional measures of market liquidity. CRT(D), as expected, is
negatively related to average trading volume. CRT(D) declines during the day as do other liquidity

13
Correlations between CRT(D) and ES(D) or LAG ES(D) can be calculated only when a trade has occurred in a
 19

measures in most markets and follows a similar intraday pattern as QIS, with the highest CRT(D)
and QIS occurring at the open. CRT(D) is positively correlated with QIS and ES, but the
correlations are less than 1.0, particularly for higher trade sizes. This suggests that CRT(D) carries
different information than QIS and ES(D), but it remains to be determined whether the incremental
information in CRT(D) is useful to investors. In the next subsection, we consider whether CRT(D)
provides valuable information on the current state of market liquidity over that already provided by
QIS and ES(D).

4.2. Liquidity and Subsequent Transactions: A Regression Analysis

If CRT(D) is a useful proxy for investors, it should help predict order arrival and subsequent
transaction frequency. In particular, if CRT(D) is low (high) at a given point in time, ceteris
paribus, we anticipate a higher (lower) probability of the subsequent submission of a market order.
A similar prediction can be made for QIS and LAG ES(D). We include QID in both sets of
regressions, because this traditional measure of liquidity usually is presented together with QIS and
ES(D). To test whether CRT(D) contains information incremental to QIS or LAG ES(D) about
market liquidity, we test the ability of CRT(D) at the beginning of a trading hour to predict the next
hour’s transaction frequency against a model that uses QIS or LAG ES(D), also measured at the
beginning of or during the previous trading hour. If CRT(D) predicts subsequent transaction
frequency better than QIS or LAG ES(D), we will conclude that CRT(D), which captures
information in the LOB beyond that of these traditional measures, is a useful measure of market
liquidity.

To test our hypothesis we assume a Poisson arrival rate for transactions and use a Poisson
regression to estimate the impact of CRT(D), QIS, and LAG ES(D) on transaction frequency. 14
The expected arrival rate, λ, is assumed to be a log-linear function of the explanatory variables:
(λ) = βX.
Ln(λ) β In this specification, the estimated coefficient can be interpreted as a partial
elasticity, the percentage change in the rate of transactions due to a one-unit change in the
explanatory variable.

particular period.
14
Transaction frequency is a dependent variable of counts (non-negative integers). The appropriateness of the
Poisson specification to models of count data is addressed in Hausman, Hall and Griliches (1984) and Cameron
and Trivedi (1996).
 20

One of the drawbacks of the Poisson specification is that, theoretically, the variance of its
distribution should equal its mean. Empirically, however, count data generally exhibit
overdispersion; where the variance of the dependent variable is greater than its mean. The
overdispersion in λ is modest; the variance ranges from 1.1 to 1.8 times the mean, depending on
trade size. Nevertheless, because failure to correct for overdispersion results in coefficient standard
errors that are smaller than they should be (thereby inflating the significance of the coefficients), we
correct the standard errors by multiplying the Poisson maximum likelihood standard errors by the

square root of the Pearson χ2 statistic divided by the degrees of freedom. 15 In addition to
controlling for overdispersion we include controls for intraday effects and day of the week effects
with dummy variables.16

The ability of CRT(D) to predict transaction frequency is compared with the ability of QIS to
predict transaction frequency using a goodness of fit measure. Because the Poisson regression is a
non-linear specification, standard goodness of fit measures, such as R2, are not appropriate.
Consequently, we present the pseudo-R2 suggested by Cameron and Windmeijer (1993) which is
commonly used in the Poisson framework (Greene 1997).

Table 4 presents the results of Poisson regression estimation of transaction frequency. The first
four columns present the statistics for Floor and CATS where CRT(D) is compared with QIS. The
next four columns present comparisons of CRT(D) with LAG ES(D) for Floor and CATS. Each
model is computed for each of the five trade sizes.

As expected, and CRT(D) and QIS have similar explanatory power for the $5,000 trade size.
Recall from Table 3 that CRT(D) and QIS are very highly correlated for $5,000 trades, reflecting
the fact that these trades usually can be executed at the inside spread, because the inside depth is
generally larger than the trade size. The correlations decrease as trade size increases, reflecting the
fact that as trade size increases, it becomes more likely that we must go outside the spread to
calculate CRT(D). The interesting question is whether the information in CRT(D) that is not related
to QIS is useful or superfluous. We can determine the answer to this question by examining how the

15
This procedure assumes that the overdispersion in the data is multiplicative. An alternative assumption that is
sometimes used is that overdispersion is quadratic in the mean. This alternative is estimated with a negative
binomial specification. We estimated the Poisson regressions with this specification. All of the results reported
in the paper are robust to the alternative assumption . Cameron and Trivedi (1996) provide a detailed discussion
of this issue.
16
The coefficients of these control variables are not reported in Table 4 to save space.
 21

CRT(D) regressions perform relative to the QIS regressions as trade size increases. If information
beyond the inside spread is useful to investors, we should see the performance of the CRT(D)
regression relative to QIS improve with trade size.

The additional information obtained from this calculating CRT(D) is apparently useful to traders.
For trade sizes above C$5,000, the goodness of fit measure shows that the CRT(D) regression
specification outperforms the QIS specification on both trading systems. Although the differences
are small at lower trade size levels, they increase substantially as trade size increases. For our
largest trade size group, C$100,000 – C$200,000 CRT(D) explains considerably more of the
variation in future transaction frequency than the inside spread: pseudo-R2 equals 24.7% for
CRT(D) on the Floor relative to 15.4% for QIS.

Regressions with CRT(D) and QID as explanatory variables count the inside depth twice, since
it is also imbedded in the calculation of CRT(D). We would like to know how much additional
information is contained in QID relative to CRT(D). Table 4 also presents the pseudo-R2 for the
same regressions with CRT(D), but excluding QID (lines in italics below each regression). We
conclude immediately that the majority of predictive power comes from the CRT(D), and, except
for the lowest trade sizes, the CRT(D) alone outperforms the combination of the inside spread, QIS,
and the inside depths, QID.

In general, an increase in the trade size makes the likelihood of trade submission more sensitive
to our measure of committed liquidity. The coefficient of CRT(D) for Floor trades ranges between -
0.514 and -0.586 for trades of $75,000 or less, but is -0.663 for $150,000 trades. For CATS
trades, the coefficients go from -0.311 to -0.435. These results indicate that a 1 percent increase in
CRT(D) results in a 0.311 to 0.663 percent decrease in transaction frequency. We argue that the
results support our assertion that the inside quoted spread, and the inside quoted depth do not tell a
complete story about the liquidity available in the market, especially for the larger trade sizes. We
conclude that CRT(D) provides additional information about liquidity, which is apparently useful to
investors, and thus can be used to predict order submission activity in the market.

Lower ex post trading costs, ES(D), could also predict increased transaction frequency if it
signals a subsequent period of high liquidity. To test this hypothesis we estimate transaction
frequency with LAG ES(D) and compare the performance in predicting future transaction frequency
to that of CRT(D), using a sample of observations wherein LAG ES(D) can be observed. The
 22

results of these regressions are presented in the last four columns of Table 4. It should be noted that
the ability of LAG ES(D) to predict transaction frequency is severely constrained by the fact that
LAG ES(D) must have been recently observed to be used. Many of the hourly periods, particularly
for low volume stocks, have no trades of a given trade size in the preceding hour. Even when LAG
ES(D) can be calculated from recent transactions, it has inferior explanatory power relative to
CRT(D). The coefficient on LAG ES(D) is negative and generally significant, indicating that higher
effective spreads reduce the number of transactions in the following period. However, as trade size
increases, CRT(D) substantially outperforms LAG ES(D), particularly on the Floor.

We examined the robustness of these results in two ways. First, we repeated all of our
reported regressions using Hausman et al (1984) fixed effects Poisson specification. We find that
our results our robust to the fixed effects alternative. Second, we estimate all of our regressions using
CRT(D) only when the book is full (treating all other observations as missing) and find similar results.

We argue that these results support our assertion that the inside quoted spread, and the
effective spread do not tell a complete story about the liquidity available in the market. We
conclude that CRT(D) provides additional information about liquidity, which is apparently useful to
investors, because it predicts their transactions in a following trading hour better than the traditional
liquidity measures.

4.3. Buy-side and Sell-side Liquidity

When submitting a buy order of size D, the investor may be more interested in the cost of that
trade than in the CRT(D), which is the cost of a roundtrip trade. To examine this conjecture, we
decompose CRT(D) into buy-side and sell-side costs, where CST(D) is the Cost of Sell trade, and
CBT(D) is the Cost of Buy trade. Using the previously defined indicator variables, Ik and I-k, we
calculate CST(D) and CBT(D) as deviations from the midpoint of the quoted spread:

CST(D) = [ 0.5*(P 0 + P-0) - Σ k=0,infty I-k P-k Q-k ] / D

CBT(D) = [Σk=0,infty Ik Pk Qk - 0.5*(P 0 + P-0) ] / D

Given the definition of the two variables it is easy to see that:

 23

CRT(D) = CST(D) + CBT(D)

Ahn, Bae and Chan (2000) use data from the Stock Exchange of Hong Kong to examine the
relations between buy and sell trading and bid and ask depth. They suggest that looking at each side
of the book separately can yield interesting insights. To understand how each side of the book
affects one-way transaction frequency we replicate the CRT(D) regressions in Table 4, separating
transactions into buy orders and sell orders and CRT(D) into CBT(D) and CST(D).

The results of these regressions are presented in Table 5. The most interesting result is that the
size of the coefficients on CBT(D) and CST(D) are larger than the CRT(D) coefficients in Table 4.
On the Floor the coefficients of CBT(D) range from -1.051 to -1.406 and the coefficients on
CST(D) range from -0.664 to -0.894. These results indicate that the elasticity of buy (sell)
transaction frequency is greater for buy (sell) trades than the elasticity of total transactions with
respect to the entire book (CRT(D)). When determining when to submit a particular trade, the side
of the book against which the trader plans to execute apparently is more important than the opposite
side of the book.

On the other hand, the extant theory suggests that liquidity on both sides of the market may
determine which type of order (limit or market) is submitted. The prediction is that higher CBT(D)
is likely to encourage a prospective seller to submit a limit rather than a market order. To test this
idea we estimate one way transaction frequency using both CBT(D) and CST(D) as explanatory
variables. The explanatory power of this specification increases relative to the specification that only
includes one side of the book. The coefficients on the opposite side of the book are negative and
significant, as predicted, despite the fact that CBT(D) and CST(D) are generally fairly highly
correlated.17 This suggests that the liquidity on the opposite side of the book plays a role in the
investor order-submission decisions, as conjectured.

In Table 4 we claim that CRT(D) is a superior measure of liquidity because it explains

considerably more of the upcoming transaction frequency than standard measures of liquidity. To
test the robustness of this result when both sides of the LOB are calculated separately, we estimate
two additional models. In the first model, the specification is identical to the regressions presented in
Table 5 except QID is omitted from the specification. As in Table 4, the most important
explanatory variables are those calculated using liquidity beyond the inside spread, CST(D) and
 24

CBT(D). The additional explanatory power of QID is relatively small and declines with trade size.
Our second model estimates the frequency of the buy and sell trades separately using only QIS and
QID. When we compare the explanatory power of QIS and QID to that of CST(D) and CBT(D)
we find that, except for the smallest trade size, CST(D) and CBT(D) outperform the combination of
the inside depth and spread. The size of the differences in pseudo-R2 is comparable to the
differences reported in Table 4, and is occasionally higher. These robustness results (not presented
to save space) provide additional evidence on the validity of our measure.

5. Summary

This paper examines a liquidity measure based on the ex-ante level of committed liquidity in a
stock and related to the size of a trade – the Cost of a Round Trip Trade of D dollars, with missing
values in the Limit Order Book filled in, or CRT(D). The measure can be used as a benchmark
against a theoretical perfectly liquid market, which makes it useful for comparing liquidity shifts in a
particular market or liquidity across different markets. We claim that market i offers higher liquidity
for transactions of D dollars than market j if CRTi(D) < CRTj(D). Using data from the Toronto
Stock Exchange, we show that CRT(D) is correlated with other common measures of liquidity, but
provides additional information. We document the relevance of our measure by showing its ability
to predict the number of trades of that size that will occur in the forthcoming period.

Although we test our liquidity measure using TSE data, there is little reason to believe that the
results are specific to that exchange. We believe that CRT(D) will be particularly useful in any
market where a high proportion of available liquidity is committed rather than hidden. As world
equity markets increasingly adopt a pure electronic order book architecture, the applications of
CRT(D) should increase. For example, CRT(D) should be useful in comparing the committed
liquidity of competing ECNs trading the same stock at the same time. Consequently, we suggest
that it will be an increasingly useful tool for researchers and investors alike and urge exchanges to
report it.

17
The Pearson correlations for CBTF and CSTF range from 0.60 to 0.81, depending on trade size.
 25

References

Ahn, H., Cao, C., Hyuk C., 1997. Decimalization and Competition Among Exchanges: Evidence
from the Toronto Stock Exchange Cross-Listed Securities. manuscript, The Pennsylvania State
University.

Ahn, H., Bae, K., Chan, K., 2000. Limit Orders, Depth and Volatility: Evidence from the Stock
Exchange of Hong Kong. forthcoming Journal of Finance.

Bacidore, J., Battalio, R., Jennings, R., 2000. Depth Improvement and Adjusted Price
Improvement on the NYSE, NYSE Working Paper #2000-04.

Barclay, M., Christie, W., Harris, J., Kandel, E., Schultz, P., 1999. The Effect of Market Reform
on the Trading Costs and Depths of Nasadaq Stocks. Journal of Finance, 54.2, 1-34.

Benston, G., Irvine, P., Kandel, E., 1997. Liqidity Revisited, Working Paper, Emory University and
Hebrew University.

Biais, B., Hillion, P., Spatt, C., 1995. An Empirical Analysis of the Limit Order Book and the Order
Flow in the Paris Bourse. Journal of Finance, 50.5, 1655-1689.

Bacidore, J. M., 1997. The Impact of Decimalization on Market Quality: An Empirical Investigation
of the Toronto Stock Exchange. Journal of Financial Intermediation, 92-100.

Cameron, C. and Windmeijer, F., 1993. R-Squared Measures for Count Data Regression Models
with Applications to Health Care Utilization Working Paper No. 93-24, Department of Economics,
University of California, Davis.

Cameron, C. and Trivedi, K., 1996. Count Data Models for Financial Data In: Maddala, G., Rao,
C., (Eds.), Statistical Methods in Finance, North Holland, 363-392.

Coppejans, M., Domowitz, I., 1999. Stock and Flow Information as Inputs to Limit Order Book
Trading Activity. Manuscript, Duke University and The Pennsylvania State University.

Demsetz, H., 1968. The Cost of Transacting. Quarterly Journal of Economics, 82, 33-53.

Goldstein, M., Kavajecz, K., 2000. Eighths, Sixteenths, and Market Depth: Changes in Tick Size
and Liquidity Provision on the NYSE. Journal of Financial Economics, 56.1, 125-149.

Greene, W., 1997. Econometric Analysis. Third Edition, Prentice Hall.

Griffiths, M., Turnbull, A., Smith, B., White, R., 2000. The Costs and Determinants of Order
Aggressiveness. Journal of Financial Economics, 56.1, 65-88.

Glosten L., 1994. Is the Electronic Open Limit Order Book Inevitable? Journal of Finance, 49.4,
 26

1127-1161.

Harris, L. E., 1994. Minimum Price Variations, Discrete Bid-Ask Spreads, and Quotation Sizes.
Review of Financial Studies 7, 149-178.

Harris, L. E., 1996. Does a Large Minimum Price Variation Encourage Order Exposure? NYSE
Working Paper #96-05.

Hausmann, J., Hall B., Griliches Z., 1984. Econometric Models for Count Data with an Application
to the Patents-R&D Relationship. Econometrica, 52, 909-938.

Huang, R. Stoll H., 1996. Dealer Versus Auction Markets: A Paired Comparison of Execution
Costs on Nasdaq and the NYSE. Journal of Financial Economics 41, 313-357.

Huson, M., Kim, Y., Mehrotra, V., 1997. Decimal Quotes, Market Quality and Competition for
Order Flow: Evidence from the Toronto Stock Exchange. Manuscript, University of Alberta.

Jones, C., Lipson, M., 2000. Sixteenths: Direct Evidence on Institutional Execution Costs.
forthcoming, Journal of Financial Economics.

Kandel, E., Lauterbach, B., Tkach, I., 2000. Tel Aviv Continuous Trading, working paper, Hebrew
University, April.

O’Hara, M., 1994. Market Microstructure Theory. Blackwell.

Ronen, T, Weaver, D., 1999. Teenies Anyone? Manuscript, Baruch University.

Porter D., Weaver, D., 1997. Tick Size and Market Quality. Financial Management, 26, 5-26.

Roll, R., 1984. A Simple Implicit Measure of the Effective Bid-Ask Spread in an Efficient Market.
Journal of Finance 39.4, 1127-1139.

Stoll, H. R., 1989. Inferring the Components of the Bid-Ask Spread: Theory and Empirical Tests,
Journal of Finance 44, 115-134.

Sofianos G., Werner, I., 1997. The Trades of NYSE Floor Brokers. NYSE Working Paper #97-
04.

Wagner, W., 1993. Defining and Measuring Trading Costs. In: Execution Techniques, True Trading
Costs, and the Microstructure of Markets. AIMR conference proceedings.
 27

Figure 1

Sample limit order book for Barrick Gold, symbol ABX, on June 26th. Solid bars represent buy orders and the
open bars represent sell orders. The quoted spread and the midpoint of the quoted spread changed significantly
over the hour. Notice that the large sell order at $37.50 remains, but is much farther from the midpoint of the
quoted spread at 2:30 than it was as 1:30. As a consequence CRT, at least for large buy orders, has increased
over the hour.

ABX Barrick Gold

June 26th 1:30
24,000
22,000
20,000
18,000
16,000
14,000
12,000
10,000
8,000
6,000
4,000
2,000
0
36.90 37.00 37.10 37.20 37.30 37.40 37.50 37.60 37.70 37.80

ABX Barrick Gold

June 26th 2:30

24,000
22,000
20,000
18,000
16,000
14,000
12,000
10,000
8,000
6,000
4,000
2,000
0
36.70 36.80 36.90 37.00 37.10 37.20 37.30 37.40 37.50
 28

Figure 2

This figure presents intraday behavior of CRT(D), which is the average percentage Cost of a Round Trip trade at
the beginning of each hour in the trading day. CRT(D) is presented for each period by five dollar trade sizes,
C$5000, C$17,500, C$37,500, C$75,000 and C$150,000. All averages are calculated from orders provided to the
Toronto Stock Exchange (TSE) Order File and trades recorded by the TSE from June 1, 1996 through July 31,1996.

Intraday CRT Pattern

6.00

5.00

4.00

3.00

2.00

1.00

0.00
9:30 AM 10:30 AM 11:30 AM 12:30 PM 1:30 PM 2:30 PM 3:30 PM

CRTF $5,000 CRTF $17,500 CRTF $37,500

CRTF $75,000 CRTF $150,000
 29

Table 1

Trading statistics, depth and spreads per dollar means (first row) and standard errors (second row) by
volume and price groups averaged over stocks, June 1 - July 31, 1996 [note 1] Low (L), Medium (M) and
High (H) triads (e.g., LL = lowest third of volume and lowest third of price)

Low volume Medium volume High volume

Variable All LL LM LH ML MM MH HL HM HH

Trading Statistics
VOLUME (thousands 98.1 8.0 7.3 7.6 42.6 41.5 45.7 226.7 189.9 296.3
of shares) 8.8 1.2 0.9 1.1 2.8 3.5 3.7 28.7 22.0 34.8

PRICE (per share) 21.39 8.41 16.85 45.61 8.94 16.72 34.74 9.47 16.78 35.55
1.13 0.48 0.45 7.21 0.35 0.55 2.47 0.40 0.54 2.36

STD DEV RETURN .017 .017 .013 .011 .022 .016 .022 .020 .015 .014
[note 2] .001 .002 .001 .001 .005 .003 .004 .002 .002 .002

Depth
QID (quoted inside 9.0 6.6 4.5 3.8 9.2 8.1 5.2 16.3 13.2 13.8
depth) [note 3] 0.5 0.6 0.4 0.6 0.9 0.5 0.3 1.6 1.6 2.1

ASK DEPTH 31.2 23.1 9.9 6.7 31.1 22.6 13.1 78.8 52.1 42.7
[note 3] 1.9 4.6 1.8 0.8 5.3 3.5 1.3 8.5 7.0 5.0

BID DEPTH 20.3 14.0 8.2 7.1 23.5 13.7 10.8 37.0 32.4 35.2
[note 3] 1.3 2.0 1.3 1.1 4.8 1.6 0.9 3.8 5.3 5.3
QID / TOTAL (BID 25% 25% 33% 31% 22% 29% 27% 15% 18% 21%
+ ASK) DEPTH 1% 3% 3% 3% 2% 3% 2% 1% 2% 2%

Spreads per Dollar

QIS (quoted inside 1.41 2.74 2.10 1.33 1.84 1.33 0.95 1.23 0.76 0.44
spread %) [note 4] 0.05 0.21 0.14 0.12 0.14 0.11 0.08 0.07 0.06 0.03

ES (effective 0.94 1.87 1.35 0.98 1.23 0.83 0.65 0.80 0.48 0.27
spread %) 0.04 0.17 0.11 0.10 0.10 0.07 0.07 0.05 0.04 0.44

Number of stocks 263 23 38 26 31 27 30 33 23 32

Note 1: for each stock over the period, all variables except effective spread (ES) were measured at seven hourly daily intervals;
ES were measured for each transaction

Note 2: measured with daily closing bid prices

Note 3: thousands of shares

Note 4: per dollar spreads are calculated from the midpoint of the inside bid and ask prices
 30
Table 2
Means (first row) and standard errors (second row) of Quoted Inside Depth and Spreads per Dollar (percentage) for Floor
and CATS trades. By stock types -- Volume and Price triads: E.g., MM = medium volume, medium price (base group)

Floor CATS
Trade Liquidity base volume: L & H price: L & H base volume: L & H price: L & H
Size (C$) Measure MM LM HM ML MH MM LM HM ML MH

QID (Quoted Inside 7,518 5,624 13,836 8,369 5,285 9,892 2,945 11,178 10,413 5,205
Depth) 99 91 195 128 67 246 44 395 208 121

QIS (Quoted Inside 1.30 2.06 0.76 1.91 0.99 1.26 2.00 0.67 1.62 0.69
Spread) 0.02 0.03 0.01 0.03 0.01 0.03 0.03 0.02 0.03 0.01

0 - 10,000 ES (effective spread) 0.70 1.28 0.45 1.12 0.63 0.79 0.97 0.50 1.14 0.50
0.01 0.04 0.01 0.02 0.01 0.02 0.02 0.01 0.02 0.01
CRT (cost round trip) 1.43 2.32 0.81 2.26 1.05 1.35 2.38 0.72 1.99 0.72
0.02 0.03 0.01 0.03 0.01 0.03 0.03 0.02 0.04 0.02

CRT / ES 2.04 1.81 1.80 2.02 1.67 1.71 2.45 1.44 1.75 1.44
BOOK FULL % 98.1 98.1 99.8 98.6 99.1 100.0 94.7 99.4 98.4 100.0

10,001 - ES (effective spread) 0.62 1.19 0.39 1.02 0.52 0.70 0.81 0.41 0.95 0.44
25,000 0.02 0.07 0.01 0.03 0.01 0.03 0.03 0.02 0.03 0.01
CRT (cost round trip) 1.88 3.10 1.02 3.35 1.28 1.78 3.43 0.89 2.92 0.86
0.02 0.04 0.01 0.04 0.02 0.05 0.04 0.03 0.06 0.02

CRT / ES 3.03 2.61 2.62 3.28 2.46 2.54 4.23 2.17 3.07 1.95
BOOK FULL % 91.0 86.0 98.5 91.0 96.4 95.7 82.1 98.9 92.0 98.4

25,001 - ES (effective spread) 0.60 1.04 0.38 1.05 0.51 0.67 0.79 0.39 0.98 0.43
50,000 0.03 0.07 0.01 0.05 0.02 0.03 0.07 0.02 0.04 0.02
CRT (cost round trip) 2.52 4.03 1.34 4.74 1.59 2.25 4.50 1.24 4.14 1.10
0.03 0.05 0.02 0.06 0.02 0.05 0.05 0.04 0.07 0.02
CRT / ES 4.20 3.88 3.53 4.51 3.12 3.36 5.70 3.18 4.22 2.56
BOOK FULL % 79.5 70.5 95.4 77.1 89.1 88.8 59.8 96.4 82.3 96.4

50,000- ES (effective spread) 0.65 1.14 0.39 0.86 0.47 0.65 0.97 0.38 0.86 0.42
100,000 0.04 0.10 0.01 0.06 0.02 0.04 0.10 0.02 0.04 0.02
CRT (cost round trip) 3.41 5.23 1.83 7.05 2.01 2.98 5.73 1.81 5.90 1.45
0.03 0.05 0.03 0.07 0.02 0.07 0.05 0.06 0.09 0.03

CRT / ES 5.25 4.59 4.69 8.20 4.28 4.58 5.91 4.76 6.86 3.45
BOOK FULL % 60.2 43.0 89.7 50.6 78.1 79.0 36.3 87.2 65.6 92.5

100,000 ES (effective spread) 0.62 1.00 0.34 1.16 0.35 0.63 0.91 0.35 1.07 0.42
-
200,000 0.05 0.12 0.01 0.17 0.03 0.05 0.18 0.02 0.10 0.02
CRT (cost round trip) 4.80 6.70 2.72 9.87 2.60 4.05 6.91 2.71 8.23 2.01
0.04 0.05 0.08 0.06 0.03 0.07 0.04 0.08 0.10 0.03

CRT / ES 7.74 6.70 8.00 8.51 7.43 6.43 7.59 7.74 7.69 4.79
BOOK FULL % 32.0 12.9 74.5 21.2 56.4 64.2 21.6 70.0 44.3 75.9
QID and QIS are measured at hourly intervals; QIS per Dollar at the midpoint of the bid and ask prices
ES(D) is measured for each trade within the trade dollar size group (unweighted)
CRT measured at the midpoint of each trade dollar size group (e.g., $5,000 for the 0 - $10,000 trade size, $17,500 for the $10,000-$25,000 trade
size)
BOOK FULL % is equal to the percentage of stock-periods where the enough committed liquidity exists (on each side of the book) to
execute a trade size equal to the midpoint of each trade dollar size group.
 31
Table 3
Pearson correlations of CRT(D) (Cost of Round Trip per Dollar) with: QID (Quoted Inside Depth), QIS (Quoted Inside
Spread per Dollar), ES(D) (Effective Spread per Dollar), and LAG ES(D) (lagged Effective Spread per Dollar). Floor and
CATS trades By stock types -- Volume and Price triads: E.g., MM = medium volume, medium price (base group)

Trade Floor CAT

S
Size (C$) base volume: L & H price: L & H base volume: L & H price: L & H
MM LM HM ML MH MM LM HM ML MH

0-10,000
QID -.13 -.15 -.25 -.17 -.13 -.15 -.12 -.24 -.29 -.09
QIS .91 .93 .93 .88 .97 .98 .92 .94 .93 .98
ES .47 .59 .45 .43 .61 .59 .54 .58 .56 .47
LAG ES .37 .51 .36 .34 .49 .48 .47 .46 .48 .27

10,001 - 25,000
QID -.23 -.29 -.30 -.29 -.21 -.22 -.25 -.31 -.36 -.16
QIS .81 .80 .81 .76 .90 .91 .82 .86 .76 .91
ES .32 .54 .30 .28 .53 .61 .46 .41 .51 .45
LAG ES .28 .50 .25 .18 .45 .44 .44 .32 .48 .35

25,001 - 50,000
QID -.33 -.36 -.36 -.37 -.27 -.27 -.30 -.37 -.41 -.22
QIS .74 .77 .74 .72 .85 .87 .79 .75 .71 .80
ES .11* .43 .32 .26 .51 .48 .40 .52 .54 .39
LAG ES .13 .42 .22 .18 .45 .44 .33 .39 .54 .38

50,000 -
100,000
QID -.41 -.41 -.42 -.41 -.32 -.32 -.30 -.43 -.44 -.31
QIS .69 .75 .70 .69 .81 .82 .78 .66 .67 .75
ES .17 .30 .27 .36 .54 .48 .26 .55 .43 .41
LAG ES .12 .36 .29 .41 .41 .42 .38 .60 .34 .28
50,000 -
100,000
QID -.41 -.38 -.48 -.36 -.34 -.37 -.21 -.50 -.45 -.36
QIS .66 .77 .67 .64 .79 .78 .80 .63 .65 .63
ES .28 .29** .32 .42 .24 .36 .23* .58 .61 .40
LAG ES .36 .29** .33 .40 .21 .34 .36 .50 .60 .23

QID and QIS measured at hourly intervals ES(D) and LAG ES(D) are measured for each trade within
the trade dollar size group (unweighted) CRT(D) measured at the midpoint of each trade dollar size
group (e.g., 5,000 for the C$ 0 - 10,000 trade size) All coefficients are significantly different from
zero with probability < .01 except those mark with * (.01<p <.05) and those marked with ** (p >
.05)
 32
Table 4

Number of Transactions in a trading hour (dependent variable) predicted by: Cost of a Round Trip (CRT(D)) vs. Quoted
Inside Spread (QIS) at the beginning of the hour or Effective Spread (LAG ES(D)) during the previous hour. Poisson
regressions coefficients (first row) and standard errors (second row) are presented by trade size group.

CRT vs. QIS CRT vs. LAG ES 1

Floor CATS Floor CATS

Trade Dollar Size CRT QIS CRT QIS CRT LAGES CRT LAGES
(C$)

0 - 10,000

QID 0.256 0.273 0.190 0.211 0.219 0.257 0.145 0.137

0.003 0.003 0.005 0.005 0.003 0.003 0.006 0.006
CRT -0.576 -0.311 -0.526 -0.217
0.005 0.006 0.005 0.007
QIS / LAG ES -0.655 -0.366 -0.209 -0.177
0.005 0.006 0.004 0.007

pseudo R-square 17.5% 17.6% 8.6% 8.8% 14.3% 9.2% 4.9% 4.4%
the same without QID 14.7% 7.1% 11.7% 4.0%

10,000 - 25,000

QID 0.094 0.179 0.083 0.174 0.036 0.132 0.055 0.120

0.005 0.005 0.008 0.008 0.006 0.006 0.012 0.012

CRT -0.572 -0.359 -0.516 -0.250 -0.030

0.006 0.009 0.009 0.014 0.007
QIS / LAG ES -0.807 -0.504 -0.116
0.009 0.012 0.007

pseudo R-square 13.3% 12.6% 8.3% 7.3% 9.1% 3.2% 3.7% 1.2%
the same without QID 13.0% 8.1% 9.1% 3.6%

25,000 - 50,000

QID 0.028 0.177 0.151 0.296 -0.039 0.087 0.052 0.148

0.006 0.006 0.014 0.013 0.007 0.007 0.022 0.022
CRT -0.514 -0.367 -0.498 -0.101 -0.249
0.006 0.011 0.012 0.009 0.021
QIS / LAG ES -0.849 -0.574 -0.014
0.011 0.021 0.014

pseudo R-square 13.1% 11.0% 10.2% 8.0% 8.6% 2.2% 4.4% 1.5%
the same without QID 13.1% 9.7% 8.5% 4.3%
 33
Table 4 (continued)

CRT vs. QIS CRT vs. LAG ES 1

Floor CATS Floor CATS

Trade Dollar Size CRT QIS CRT QIS CRT LAGES CRT LAGES
(C$)

50,000 -100,000

QID -0.047 0.185 0.201 0.400 -0.122 0.042 0.056 0.205

0.007 0.007 0.017 0.017 0.010 0.010 0.031 0.030
CRT -0.586 -0.385 -0.555 -0.410
0.007 0.012 0.013 0.032
QIS / LAG ES -1.138 -0.699 -0.156 -0.038
0.016 0.029 0.012 0.021

pseudo R-square 18.1% 13.0% 13.8% 9.7% 10.3% 2.3% 8.0% 1.9%
the same without QID 18.0% 13.2% 9.9% 8.0%

100,000 - 200,000

QID -0.030 0.276 0.218 0.474 -0.085 0.118 0.17 0.297

0.009 0.008 0.023 0.023 0.013 0.012 0.044 0.044
CRT -0.663 -0.435 -0.565 -0.284
0.008 0.015 0.017 0.039
QIS / LAG ES -1.592 -0.943 -0.069 -0.005
0.025 0.047 0.015 0.015

pseudo R-square 24.7% 15.4% 18.3% 11.1% 10.4% 1.5% 4.2 8.4% 4.3%
%
the same without QID 24.7% 17.6% 10.2% 7.7%

Note 1. Number of observations: CRT vs QIS, Floor 41,721; CATS, 19,518. CRT vs. ES:

Trade size Floor CATS

$0-$100,000 27,767 10,273
10-25,000 16,769 5,604
25-50,000 12,023 2,881
50-100,000 9,007 1,883
100-200,000 6,172 1,066
 34
Table 5

Number of Buy or Sell Transactions in a trading hour (dependent variable) predicted by: Cost of a Buy Trade, CBT(D),
for Buy Orders or Cost of a Sell Trade, CST(D), for Sell Orders Poisson regressions coefficients (first row) and
standard errors (second row) are presented by trade size group.

Trade Floor CATS

Size (C$) Buys Buys Sells Sells Buys Buys Sells Sells

0 - 10,000

QID 0.233 0.229 0.292 0.283 0.205 0.198 0.189 0.179

0.004 0.004 0.004 0.004 0.007 0.007 0.007 0.008

CBT -1.406 -1.114 -0.732 -0.779 -0.517 -0.384

0.012 0.021 0.019 0.017 0.026 0.023

CST -0.266 -0.894 -0.263 -0.275 -0.418 -0.124

0.018 0.010 0.017 0.026 0.014 0.020

pseudo R-square 14.6% 14.7% 12.8% 13.5% 7.8% 8.0% 5.2% 5.6%

10,000 - 25,000

QID 0.119 0.087 0.135 0.100 0.110 0.089 0.107 0.073

0.007 0.007 0.007 0.007 0.013 0.013 0.013 0.013

CBT -1.199 -0.817 -0.727 -0.746 -0.507 -0.460

0.019 0.022 0.020 0.025 0.030 0.028

CST -0.441 -0.837 -0.375 -0.296 -0.485 -0.226

0.018 0.015 0.016 0.024 0.020 0.021

pseudo R-square 10.4% 11.3% 8.8% 10.5% 6.8% 7.5% 5.1% 6.3%

25,000-50,000

QID 0.072 0.020 0.087 0.033 0.197 0.168 0.383 0.276

0.009 0.009 0.009 0.009 0.021 0.020 0.043 0.044

CBT -1.051 -0.720 -0.590 -0.705 -0.501 -0.723

0.021 0.021 0.019 0.033 0.036 0.080

CST -0.448 -0.664 -0.370 -0.287 -0.629 -0.281

0.017 0.015 0.015 0.029 0.063 0.058

pseudo R-square 10.2% 11.7% 8.0% 10.2% 8.2% 9.1% 11.2% 13.9%
 35
Table 5 (continued)

Trade Floor CATS

Size (C$) Buys Buys Sells Sells Buys Buys Sells Sells

50,000 - 100,000

QID 0.016 -0.049 0.033 -0.047 0.217 0.190 0.270 0.199

0.011 0.015 0.011 0.011 0.026 0.025 0.025 0.025

CBT -1.080 -0.729 -0.732 -0.778 -0.573 -0.510

0.022 0.022 0.020 0.038 0.040 0.036

CST -0.507 -0.760 -0.439 -0.293 -0.463 -0.238

0.018 0.017 0.016 0.032 0.028 0.028

pseudo R-square 12.9% 15.5% 11.5% 15.1% 11.6% 12.7% 9.0% 11.6%

100,000 - 200,000

QID 0.041 -0.023 0.061 -0.015 0.293 0.277 0.283 0.217

0.011 0.011 0.010 0.010 0.027 0.027 0.025 0.026

CBT -1.131 -0.711 -0.697 -0.771 -0.563 -0.479

0.019 0.019 0.018 0.034 0.038 0.032

CST -0.603 -0.796 -0.471 -0.294 -0.545 -0.305

0.017 0.013 0.013 0.029 0.026 0.027

pseudo R-square 19.4% 23.3% 17.4% 21.7% 16.4% 17.8% 13.5% 16.4%