Page 1

Institutional Trading, Trading Volume, and Spread

Malay K. Dey

University of Massachusetts at Amherst

and

Northern Illinois University

B. Radhakrishna

University of Minnesota

[This Draft: March 2001]

We are grateful to NYSE for providing us with the TORQ data and to Ben Branch,

Shanta Hegde, and Robert Miller for their comments and numerous helpful suggestions.

All comments are welcome.

Please address correspondence to: Malay K. Dey, Dept. of Finance, Northern Illinois

University, Dekalb, IL 60115, Tel. (815) 753-6395, Email: mdey@niu.edu, OR B.

Radhakrishna, Department of Accounting, Carlson School of Management, University of

Minnesota, 321, 19th Avenue S., #3-122, Minneapolis, MN 55455, Tel. (612) 624-0302,

Fax (612) 626-1335, Email: bradha@umn.edu.

Page 2

1

Institutional trading, Trading Volume, and Spread

Abstract

Besides its academic interest, the effect of institutional trading on the bid-ask

spread is of interest to regulators and market makers. It is often (casually) argued that

greater institutional participation results in increased volatility in the market. On the

other hand, some argue that greater liquidity trading by institutions reduces spread.

There is no direct empirical evidence and little theoretical knowledge to suggest a

convincing relation between institutional trading and spread. In this paper, we present

some evidence on the nature and effect of institutional trading on spreads. We argue that

institutional trading is not completely information driven, part of it is liquidity trading in

nature. We find evidence that information induced institutional trading increases the

adverse selection component. However, large volume (liquidity) trading reduces the

order processing costs. We find the net effect of institutional trading on spread is

consistently negative. Moreover, institutional buys have differential information from

sells. Institutional trades per se reduce spreads, but only sells increase the adverse

selection component. Both effective and relative spreads impound the differential nature

of institutional buys and sells.

Keywords: Institutional Trading, Spread decomposition, information, liquidity.

JEL Classification: G19.

Page 3

2

Institutional trading, Trading Volume, and Spread

I. Introduction

Since Demsetz (1968) bid-ask spread is recognized as the price of liquidity

provided by the dealers in an equity market. A number of studies have investigated what

determines spread (Branch and Freed [1977], McInish and Wood [1992], Klock and

McCormick [1999], Heflin and Shaw [2000]). Some of the significant determinants of

spread found in the literature are order size, number of trades, competition in the dealers’

market, ownership structure, and the native characteristics of a stock e.g., price, and

volatility. Trading rules and mechanics of trading that proxy for information flow are

also found to affect the spread.

There is very little empirical evidence on institutional trading and spread and their

interrelationship. Keim and Madhavan (1997) find execution costs for institutional trades

are different between listed and NASDAQ stocks. Conrad, Johnson and Wahal (2001)

report an asymmetric relation between institutional buys and sells and soft-dollar

execution. However, there is some evidence of the effect of institutional trading on

securities prices. Empirical studies using order size or trading volume as a proxy for

institutional trading1 suggest an increased price effect associated with institutional

trading. Using proprietary data on institutional trading, Chan and Lakonishok (1993) find

the average price effect to be small for institutional trades, but the price effect for buys

and sells to be asymmetric; and Sias and Starks (1997) find that institutional trading

contributes to serial correlation of returns.

1 Lakonishok and Maberly (1990) use block trades as a proxy for institutional trades, and odd lot trades

plus non-margin account trades as a proxy for individual trades.

Page 4

3

The role of institutional trading in the determination of spread is interesting since

it is often argued (casually) that increased institutional participation in the U.S. equity

market during the past decade has led to an increase in the volatility, and has widened the

bid-ask spread in the equity market. On the other hand, some argue that institutional

trades provide liquidity, and hence decrease spread. Bertisimas and Lo (1998) show how

optimal trading strategies may be devised by execution cost minimizing investment

mangers. In any case, there is very little empirical evidence or theoretical knowledge to

conclude how institutional trading affects spread. Further since spread is considered to

be a sum of two different components, adverse selection, and order processing2 it is

unclear how institutional trading affects the individual components of the spread.

In this paper, we investigate if institutional trading has any information content

beyond what has been documented as a size or volume effect. In a multivariate, panel

regression framework, we determine if there exists a relation between bid-ask spread and

institutional trading after adjusting for size and price effects. Most studies on the

determinants of spread focus on the supply side of dealership market i.e., competition in

the dealer market, and use a cross sectional regression approach. Our approach is

different from previous studies in two important ways. First, we focus on the demand

side (investors’ characteristics) to determine the relation between spread and institutional

trading3; and second, we use a panel data approach accounting specifically for both serial

and contemporaneous correlation in the error terms4. We report regression results using

2 Arguably there is also inventory effect in spread. In this paper, we use a decomposition technique that

ignores inventory effect.

3 Recently, some other papers have studied the demand side, particularly the ownership structure. For

example, Heflin and Shaw (2000) document a relation between block holding and spread, and Chung and

Charoenwong (1998) look at insider holding as a determinant of spread.

4 Dey (2000) introduces a similar but reduced form panel data regression model for effective spread.

Page 5

4

both effective ($) and relative spread as the dependent variable in a set of regression

equations.

Further we decompose the spread into order processing and adverse selection

components and investigate how those components vary with changes in trading volume,

net order flow (buy vs. sell), and institutional trading. We assume contemporaneous

correlation between the disturbances and use an SUR (Seemingly Unrelated Regression)

analysis to find the significant determinants of the adverse selection and the order

processing components of the spread for our sample firms. We use a unique data set

(TORQ) that identifies institutional trading. Prior studies proxy institutional participation

by using measures based on trade size that are subject to measurement error.

Our results show that institutional trading proportion is inversely related to both

effective and relative spreads. We also find that the negative slope (suggestive of the

inverse relation) is not constant and flattens out at higher concentrations of institutional

trading. We find that this negative slope is provided by both institutional buys and sells

alike. Results from a SUR analysis show that the adverse selection costs tend to increase

and the order processing costs tend to decrease with increases in institutional trades.

The rest of this paper is organized as follows. In section II, we present the

motivation for this study. In section III, we describe an empirical model for spread using

panel regressions, provide data description, and explain the results. Further, we describe

the decomposition of the spread into order processing and adverse selection costs

components and report results from a SUR analysis of the determinants of those

components. Section IV concludes the paper.

Page 6

5

II. Motivation

II.A. Relation between Spread and Institutional trading

Schwartz (1988) identifies four classes of variables, namely, activity, risk,

information, and competition as determinants of spread. Existing literature find trade

size, number of trades, ownership structure, and extent of market power in the dealership

market to be the key determinants of bid-ask spread (McInish and Wood [1992], Laux

[1993, 1995], Klock and McCormick [2000], Heflin and Shaw [2000]). Dealer market

competition represents the supply side of the market for liquidity services5. On the other

hand, trade size, ownership structure, and frequency of trading measure the activity in

securities markets and represent the demand side of the market for liquidity services.

Prior research suggests an inverse relationship between spread and trading activity

measured by order size, and number of trades (McInish and Wood [1992]). Institutions

trade large sizes, and also trade frequently.6 Thus institutional trading will induce low

spread. However, trading activity also contributes to both information and risk associated

with a security7. Hasbrouck (1991) provides evidence that large trades contain more

information than small trades and cause spreads to widen. Lin, Sanger and Booth –

hereafter LSB (1995) find evidence of an increasing (although not continuously), non-

linear relation between spread and trade size. Jones, Kaul, and Lipson (1994) report that

5 In a securities market, dealers provide liquidity (both buying and selling of securities) services, while

investors demand those services.

6 Lee and Radhakrishna (2000) confirm earlier evidence that institutions tend to trade in larger volumes.

They present a technique of classifying trades into institutional and non-institutional based on size in a way

that reduces the error in classification to statistically manageable levels.

7 See, for example, the summary of the empirical evidence on the price effect of block trades in Dey (2000).

Page 7

6

the number of trades captures the essence of volatility in financial markets even in the

presence of volume and trade size.8

Seppi (1990) argues that institutions use trade size strategically; they trade large

orders when they can signal to the market that their trades are not information motivated

and hence large institutional trades may not have information content. Dey and Kazemi

(2000) distinguish between large, pure information trades and large institutional trades

and argue that institutional trades are driven by both “pure information” and liquidity

needs. Dey and Kazemi (2000) predict the “pure information” component of equilibrium

spread to be an increasing, while the liquidity component of the spread to be a decreasing

function of institutional trading.

Chan and Lakonishok (1993), and Keim and Madhavan (1994) find that the price

effect and cost implications of institutional buys and sells are not symmetric. Koski and

Michaely (2000) find that buys and sells provide different information for different trade

sizes. Saar (2000) provides a theoretical framework based on a dynamic portfolio

rebalancing process of institutions to explain the documented asymmetry in the price

effect of institutional buys and sells.

In this paper, in a multivariate regression framework, we determine whether

institutional trading has information content beyond that provided by trade size denoted

by trading volume, and number of trades. Specifically, we hypothesize that the variation

in bid-ask spread can be explained by trading volume, number of trades, price, and

institutional trading. Further, we hypothesize that institutional trading per se and not the

8 Recently, Chan and Fong (1999) find that after adjusting for order imbalance, trade size is important for

the volume-volatility relation.

Page 8

7

direction of institutional trading - buy or sell affects the bid-ask spread. Stated in

alternate form, we hypothesize:

H1: The bid-ask spread should vary significantly with institutional trading

after controlling for number of trades, price, and trading volume. Further,

institutional trades per se affect bid ask spread, and thus institutional buys

and sells do not have any differential effect on the bid-ask spread after

controlling for number of trades, price, and trading volume.

II.B. Components of the spread

We extend our analysis of the relation between institutional trading and spread by

decomposing the spread into its order processing and adverse selection components and

investigating the effect of institutional trading on the individual components. We use a

technique from LSB (1995) to decompose the spread into order processing and adverse

selection components and hypothesize a relation between the individual components and

institutional trading.

Sias (1996) reports positive correlation between institutional activity and market

volatility; however, Cohen et al (1987) suggest that institutions trade frequently because

of their low order processing costs. Bertisimas and Lo (1998) derive optimal trading

strategies of institutions based on minimum execution cost. LSB (1995) find the order

processing cost to be decreasing and adverse selection cost to be increasing in trade size.

We conjecture that if institutional trading is a mix of information and liquidity trading

then the information effect should increase the adverse selection component and the

liquidity effect should decrease the order processing costs. We also recognize that while

Page 9

8

gross volume is important for the determination of order processing cost, trade direction

or net volume (buy volume - sell volume), is important in the determination of adverse

selection costs. We therefore include log (buy/sell) as a variable in the regression model

for the adverse selection component.

We determine through a set of simultaneous equations how institutional trading

affects the order processing and the adverse selection components of the spread after

controlling for number of trades, volume, and trade direction. The simultaneous

equations approach uses the cross correlation between the two regression equations to

improve the estimates. Further we determine how the asymmetric information content

and the liquidity motive in institutional buys and institutional sells affects the adverse

selection (information) component of the spread. Stated in alternate form, our hypotheses

are:

H2a. The adverse-selection component should increase with institutional

trading, and the order-processing component should decrease with

institutional trading.

H2b. Institutional buys should have a differential effect on the adverse

selection component of the spread from institutional sells.

III. Regression Models, Data Description, and Results

III.A.1. Panel Data Regression Model

We use a multivariate, panel (time series – cross sectional) regression framework

to investigate the effect of institutional trading on the bid-ask spread. We use a unique

data set (TORQ) that allows us to identify the order origination for a trade as institutional

Page 10

9

or otherwise. Most studies on the determinants of spread use pooled OLS estimates of

the parameters of a regression model. OLS estimates ignore the covariance structure of

the error term both across firms and over time.

We assume disturbances are both serially and contemporaneously correlated.

Specifically, we assume an AR(1) process with contemporaneous correlation for the

disturbance term. In our model for the spread, the serial correlation may be due to lagged

spread or lagged values of the independent variables or their interactions. Kim and

Ogden (1996) find higher order serial correlation for the spread, and Peles (1992) report

contemporaneous correlation among equity trading of institutional investors. Parks

(1967) provide consistent and efficient estimates of the parameters when disturbances

follow a first order auto regressive process - AR(1) with contemporaneous correlation.

We run the following regression model for our panel data:

1234

_Pr

nt ntntntnt

Spread

where:

NtradePctDly AvgiceInstprop

abbbbe

=+++++

(1)

n = 1…N; number of firms in sample, t = 1…T; number of trading days

Spreadnt = Effective or Relative spread for the nth firm on day t

NtradePctnt = Number of trades on day t for firm n expressed as a proportion of

average number of daily trades over the entire sample period

Dly_Avgnt = Volume of trade on day t for firm n expressed as a proportion of

average daily trading volume over the entire sample period

Pricent = Closing price on day t for firm n

Instpropnt = Institutional trading (number of trades) as a proportion of total

trading for the nth firm on day t.

Further, the error structure is assumed as follows:

E

ntnn

es

,

E

itjtij

e es

di=

where

e r em

ntn n tnt

=+

−

, 1

mf

ntnn

N

~

g,

2c h=

injn

∈∈

and ,(due to contemporaneous correlation),

(disturbance term is AR(1)),

, 0 b

Page 11

10

E

i t

, −

jt

i=

i=

em

=

1

0

d

d

d

i

Ets

it js

mm

≠

0 where (cross correlation is zero).

E

it jtij

m mf

(contemporaneous correlation),

III.A.2. Data

To estimate the parameters of our regression model, we use data from the TORQ

data set. The TORQ files released by NYSE were prepared under the supervision of

Professor Joel Hasbrouck during his tenure as a Visiting Economist to the NYSE. This

dataset contains trades, quotes, order processing, and audit trail data for a sample of 144

NYSE stocks for the three months (63 trading days) from November 1990 through

January 1991.9 These firms represent a size stratified random sample of firms in the

NYSE and thus cover the broad spectrum of NYSE firms.

As noted by Lee and Radhakrishna (2000), the marginal contribution of TORQ

data over ISSM or TAQ data is in providing identification for traders’ classes, as

institutions, individuals, and dealers. Most studies using other trades/quotes databases

use size as a proxy for institutional trades.

We impose a restriction that is common among studies that study effective

spreads or the components of the spread to ensure adequacy of data in estimation. We

select all firms in the data set that have on average 20 trades per day or more during the

sample period. Further, in classifying the trades, whenever there are executions of

multiple orders on the active side of a trade, we take the trader class of the largest order

9For more details on the TORQ database and on trading procedures at the NYSE, see Hasbrouck (1992).

Page 12

11

on the active side as the initiator of that trade. The active side of the trade is determined

using the Lee-Ready (1991) algorithm.10 This reduces our sample to 65 firms.

Of the 65 firms that survived our initial cutoff, 14 firms have one or more days of

missing observations or days with trading halts. We leave out the firms with trading halt

days from our study since the effect of trading halts on spreads and the price discovery

process is unique and beyond the realm of this paper. We chose to omit the firms with

missing observations, since there are questions about the reliability of estimates using an

unbalanced panel11. Thus we have a balanced panel of 51 firms with 63 days data that we

use in our panel regression.

The panel data set includes daily data for the firms in the data set. For each firm,

we calculate the mean daily effective bid-ask spread, and the proportion of buy and sell

orders initiated by institutions for each day. We compute the effective spread for each

trade defined as twice the absolute value of the difference between trade price and the

prevailing mid-quote. The mean effective spread is the average of effective spreads

across all trades in a day. To determine the mean proportion of trades by institutions in

that firm, we calculate, for each day, the proportion of trades by institutions on the buy

and sell side.

Besides proportion of institutional trading, there are three other independent

variables in the panel regression. The “number of trades” variable is computed as the

number of trades each day divided by the average number of trades for the firm over the

10 Lee and Radhakrishna (2000) show the effectiveness of Lee-Ready algorithm for classifying trades into

buys and sells.

11 We explored another alternative i.e., to consider all the firms (65) but for fewer days (61). However,

there was a problem in deciding which 2 days to delete for firms that have full 63 days data. Since any

choice on this was likely to be arbitrary, we chose to use a balanced panel of 51 firms over 63 days.

Page 13

12

sample period. Thus this is a measure of abnormal trading in each day. Trading above

(below) mean would give this variable a value higher (lower) than one. The “trading

volume” variable is computed by dividing the daily share volume for the firm by the

average daily share volume over the sample period. Therefore, this variable also has

values above (below) one when trading volume is higher (lower) than the average daily

share volume. The price variable is the closing price of the stock.12

In Table 1, we present means of the computed statistics of the variables used in

the panel regression. We first compute the relevant statistics for the sample firms over

the sample period and then compute the means of those statistics. Thus we report the

means of the cross sections of firm means, medians, and standard deviations. The inter-

firm mean (median) spread is .126 (.12) that is about an eighth. The inter-firm mean

standard deviation is quite low at .02. The largest spread in our sample is .514,

approximately one-half, and the lowest .019 or approximately one-sixty-fourth. The

inter-firm mean (and median) institutional trading in our sample is around 30%.

Although institutions generally trade on a regular basis, in some trading days there is no

institutional trading. Of the 3,213 (51×63) firm trading days covered in our sample there

were 10 firm-days when there was no institutional trading, and one firm-day when all the

trading was initiated by institutions. There is also sufficient variability in the number of

trades and volume. The low trading (volume) in our sample was 9% (3%) of average

daily trades (volume). The high trading (volume) was 392% (1235%) of average daily

volume. In Table 2, we present Pearson correlations for variables in the regression.

12 Using average price over a day instead of closing price does not change the results.

Page 14

13

Panel A presents pooled correlations computed from 3,213 observations. In Panel B, we

present the means of correlations computed in time series for each firm.

III.A.3 Regression Results of Panel Model

Table 3 reports the regression results for the panel data regressions with effective

spread (in dollars) and relative spread as dependent variables. For our first model (1-ES),

the independent variables are number of trades, trading volume, price, and the proportion

of institutional trades. All four variables are significant in determining effective spread.

The significant coefficients show that effective spread increases as trading (number of

trade) and price increase and decreases as trading volume and institutional trading

increase. The coefficient for institutional proportion is negative (-.0111) and significant at

less than 1% level. Thus an increase in institutional proportion reduces the spread. The

R2 for this model is 22%. However, for a similar model with relative spread (1-RS) as

the dependent variable, number of trades fails to be a significant determinant of spread.

All other variables, namely average trading volume (-.0137), price (-.028), and

institutional proportion (-.024) remain significant at less than 1% level, and the R2 for the

model is 90%. The change in the effect of price on relative spread (decreasing in price)

from that on effective spread (increasing in price) is expected since relative spread is

computed as effective spread over mid quote. This change in the effect of price on

effective and relative spreads is consistent across all the four models.

For our second model, we introduce two dummy variables for high and medium

institutional proportions. The high (medium) institutional proportion variable has a value

of 1 when the level of institutional trading is in the top (middle) 33% percentile, and zero

Page 15

14

otherwise. The coefficients of the dummy variables are both positive and significant in

the regression. The high dummy has a larger coefficient (.0016) than the medium

dummy variable (.0007). Taken in conjunction with the significant negative (-0.0153)

coefficient of the institutional proportion variable, this suggests that the negative slope of

the institutional proportion variable flattens out at higher levels of institutional trading.

This conjecture is confirmed with model 3.

In model 3, the institutional proportion variable is replaced by three variables –

high, medium and low proportion. The high (medium, low) proportion variable has the

same value as institutional proportion if institutional trading proportion is in the top

(middle, bottom) 33 percentile, and zero otherwise. The coefficients of all three variables

are negative and significant in the regression, but while the coefficient for low proportion

is -.0144, that for the high proportion is significantly (14%) less at -.0122. We interpret

these results as follows. On average, there may be a mean positive effect on institutional

trading embodied in the positive intercept. However, when there is an increase in

institutional trading within a level, the spread declines, but the rate of decrease is lower at

higher levels of institutional proportion. Between low and medium levels, the rate of

decrease is similar indicating these are perhaps two discrete levels of institutional

proportions.

Finally, for our fourth model, we break up the institutional proportion into

institutional buy and institutional sell to test for a difference in their effect on the spread.

Results from the fourth model show that spread reduces as institutional trading increases,

be it buy or sell. The coefficients for buy and sell are significantly negative for both

effective spread and relative spread.