 

 

LIBRARY
Michigan State
University

 

 

 

PLACE IN RETURN BOX to remove this checkout from your record.
TO AVOID FINES Mum on or before date duo.

DATE DUE DATE DUE DATE DUE

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

AN ANALYSIS OF COMPETITION IN THE SECURITIES
INDUSTRY: THE MARKET FOR DUALLY-LISTED NYSE STOCKS

By

Matthew Jerome Knittel

A DISSERTATION

Submitted to
Michigan State University
in partial fulﬁllment of the requirements
for the degree of

DOCTOR OF PHILOSOPHY

Department of Economics

1996

ABSTRACT

AN ANALYSIS OF COMPETITION IN THE SECURITIES INDUSTRY: THE
MARKET FOR DUALLY-LISTED NYSE STOCKS

By

Matthew Jerome Knittel

Prior empirical research regarding competition within the securities industry established two
apparently conﬂicting results: the presence of statistically and economically signiﬁcant price
differentials based on trade location in an environment that appears to be characterized by
effective competition. This paper attempts to reconcile these results. Recent data are used
to replicate earlier studies, shorten time ﬁ'ames used in earlier studies and test a local market
hypothesis. In addition, this paper uses a liquidity premium measure, as opposed to the
spread, as the price of liquidity services and examines the impact of competition across all
exchanges, not just the NYSE. The paper concludes with regressions that explain price
differentials across exchanges. Evidence of selective market making or cream skimming
behavior is found on the part of non-NYSE exchanges. Non-NYSE exchanges tend to offer
superior execution when the NYSE is active (in terms of the total number of trades) and in
smaller trades. Compared to activity measures, the typical competition measure used in
earlier research has relatively weak explanatory power regarding price differentials when
time frames are reduced to allow for the shifting competitive position of non-NYSE

exchanges.

Dedicated to my Mother and Father, in appreciation of their love, kindness and patience.

iii

ACKNOWLEDGEMENTS

I would like to thank Professors Bruce Allen and Norman Obst for their input and
helpful comments in the completion of this thesis. Special thanks are extended to the chair
of my dissertation committee, Wally Mullin, for his insightfulness and timely responses
which allowed me to shape this paper into a ﬁnal product.

While writing this dissertation, I was employed at the Michigan Department of
Treasury under Mark Haas. Mark provided encouragement and understanding during this
time. In addition, I would like to thank David Zin for his helpful comments regarding both
the analysis and programming necessary to complete this dissertation.

Finally, two close friends have tolerated my, at times, unpredictable behavior during
the writing of this thesis: Jocelyn Samuel and Michael Ahmad. Thanks to both for seeing

me through to the end.

iv

TABLE OF CONTENTS

Page

LIST OF TABLES .................................................... vii

LIST OF EXHIBITS ................................................... ix

INTRODUCTION ...................................................... 1
CHAPTER 1 AN OVERVIEW OF COMPETITION IN THE

SECURITIES INDUSTRY ................................... 3

Attempts to Introduce Competition: Increasing Order Flow Fragmentation . . . 5

Costs of Fragmentation ....................................... 8

Beneﬁts of Competition ...................................... 10

Caveats ................................................... 11

Statement of Purpose and Layout of Paper ............................. 13

CHAPTER 2 ECONOMIC AND MEASUREMENT ISSUES ................. 16

Competition Measures ............................................. 18

The Measurement of Competition .............................. 18

Effectiveness of Competition Measures ......................... 19

The Price of Liquidity Services: Spread or Liquidity Premium? ............ 21

Aggregation Across Trade Size Classes ............................... 24

Unobservable Variables ............................................ 25

CHAPTER 3 LITERATURE REVIEW ................................... 29

The Effect of Competition on NYSE Spreads and Liquidity Premiums ....... 29

Does Trade Location Affect Price Execution? ........................... 32

Summary ....................................................... 35

Do Empirical Results Conﬂict? ................................ 36

Possible Explanations for Empirical Results ...................... 38

CHAPTER 4 COMPETITION AND PRICE DIFFERENCES ................ 39

The Simple Model: The NYSE as the Dominant Exchange ................ 39

Data ........................................................... 43

Replication of Empirical Results: Liquidity Premium
Difference Calculations ...................................... 46

Bag:

Replication of Empirical Results: The Effect of Competition on

Average Daily Spreads and Liquidity Premiums ................... 52
Summary of Replication Results ..................................... 55
CHAPTER 5 PRICE DISPERSION IN THE SECURITIES INDUSTRY ....... 58
The Law of One Price ............................................. 59
Structural Characteristics Three and Four: Complete Information
and Many Buyers and Sellers; The Local Market Hypothesis ......... 60
The Effect of Shorter Time Frames ............................. 63
Summary of Regressions Using Shorter Time Frames .............. 71
The Local Market Hypothesis ................................. 72
Summary of Local Market Hypothesis .......................... 81
CHAPTER 6 CREAM SKIMMING BEHAVIOR .......................... 83
Capturing Trade Activity: New Calculation of Liquidity
Premium Differentials ....................................... 83
Comparison of Different Matching Techniques ......................... 88
The Cream Skimming Hypothesis .................................... 92
Splitting of Data ............................................ 94
Cream Skimming Regressions ................................ 102
Conclusions .................................................... 1 1 1

vi

Table 1a
Table 1b
Table 2

Table 3

Table 4

Table 5

Table 6a
Table 6b
Table 6c
Table 6d
Table 6e
Table 6f
Table 6g
Table 7a
Table 7b
Table 7c

Table 7d

LIST OF TABLES

Bag:
Average Excess Liquidity Premium Paid for 1988 Off-Board Trades ...... 34
Average Excess Liquidity Premium Paid for 1989 Off-Board Trades ...... 34
Exchange Volume for 1992 Sample ................................ 40
1992 Sample .................................................. 44
Average Liquidity Premium Differences, 1992 ....................... 51
Daily Time Frame Regressions .................................... 54
NYSE OLS Regression Results ................................... 64
Boston OLS Regression Results ................................... 65
Cincinnati OLS Regression Results ................................ 66
Midwest OLS Regression Results ................................. 67
Nasdaq OLS Regression Results ................................... 68
Paciﬁc OLS Regression Results ................................... 69
Philadelphia OLS Regression Results .............................. 70
Regional Market Test, Boston .................................... 74
Regional Market Test, Cincinnati .................................. 75
Regional Market Test, Midwest ................................... 76
Regional Market Test, Nasdaq .................................... 77

vii

1332:

Table 7e Regional Market Test, Paciﬁc .................................... 78
Table 7f Regional Market Test, Philadelphia ............................... 79
Table 8 Average Liquidity Premium Differences, Trading Intensity Allowed ..... 90
Table 9 Average Liquidity Premium Differences ............................ 91

Table 10a Average Values for Liquidity Premium Difference Regressions, Boston . . . 96
Table 10b Average Values for Liquidty Premium Difference Regressions, Cincinnati . 97
Table 100 Average Values for Liquidity Premium Difference Regressions, Midwest . . 98
Table 10d Average Values for Liquidity Premium Difference Regressions, Nasdaq . . 99
Table 10c Average Values for Liquidity Premium Difference Regressions, Paciﬁc . . 100

Table 10f Average Values for Liquidity Premium Difference Regressions,

Philadelphia .............................................. 101
Table 11a Liquidity Premium Difference Regressions, Boston .................. 104
Table 11b Liquidity Premium Difference Regressions, Cincinnati ............... 105
Table 11c Liquidity Premium Difference Regressions, Midwest ................. 106
Table 11d Liquidity Premium Difference Regressions, Nasdaq .................. 107
Table 1 1e Liquidity Premium Difference Regressions, Paciﬁc .................. 108
Table 11f Liquidity Premium Difference Regressions, Philadelphia ............. 109
Table 12 Summary of Cream Skimming Regressions ........................ 110

viii

LIST OF EXIHBITS

Page
Figure 1 Data Manipulation .............................................. 47
Figure 2 Liquidity Premium Difference Calculation ........................... 49
Figure 3 Liquidity Premium Differences Using Trading Intensities ............... 86

ix

INTRODUCTION

Over the past decade, the securities industry has been the focus of considerable
academic research. A predominant issue has been the nature of competitive relationships
that exist on and between stock exchanges that trade the same asset. Recent allegations
leveled against the Nasdaq demonstrate the timeliness of this issue. The Securities and
Exchange Commission (SEC) alleged in a draft complaint (summer 1996) that the NASD
failed to force dealers to report promptly stock trades and failed to discipline some dealers
from "backing away," or refusing to honor investors' orders for stock at favorable prices.l
Prior to these allegations, controversy over competitive practices still existed due to the
relatively recent practice of payment for order ﬂow (which began in earnest in 1988),
particularly by Nasdaq dealers.2

These allegations are troublesome given efforts by the securities industry to ensure
a high degree of integration between exchanges. Improved technology has been used to

reduce information lags and search costs, theoretically allowing exchanges to be "in the same

 

1"SEC Seeks Settlement With NASD," The Wall Street Journal, June 18, 1996, Section
C, p. 1. Under the settlement of the draﬁ complaint, the NASD agreed to be censured, but not
ﬁned.

2The Cincinnati exchange has also come under recent scrutiny to end a practice known
as "preferencing." Under the Cincinnati exchange's rules, a broker can also act as a specialist,
selling his own shares for a mark-up at very low risk. NYSE supporters contend that the investor
loses a one-in-four chance to get a better price on the NYSE when a dealer does business on the
Cincinnati.

2

market" more often. That is, the price (for liquidity services) on one exchange should
competitively constrain prices on competing exchanges. The allegations made by the SEC
seem to contradict this line of reasoning. Do ﬁnancial exchanges effectively compete? This

is the central question this analysis attempts to address.

CHAPTER 1
AN OVERVIEW OF COMPETITION IN THE SECURITIES INDUSTRY

Frequently, trading in NYSE-listed stocks takes place simultaneously (or within a
relatively short period of time) on two different exchanges. Stocks that trade on more than
one exchange are known as dually-listed stocks. Possible trade locations include the NYSE,
Nasdaq, Boston, Cincinnati, Midwest (Chicago), Paciﬁc and Philadelphia exchanges; the
latter ﬁve known generally as regional exchanges.3 4 All exchanges that trade NYSE-listed
stocks are linked by the Interrnarket Trading System (ITS), which displays current quotes
and recent trade information. The ITS allows traders to route quickly a trade to the dealer
offering the best possible price. For 1993, the ITS encompassed more than 2,000 dually-
listed NYSE stocks that traded on the regional and Nasdaq exchanges.5

A broker trading stock on behalf of the public has a ﬁduciary responsibility to obtain
execution at the best bid or offer (BBO) displayed by the ITS. However, as noted by Cohen
et al., the existence of the ITS does not ensure that this ﬁduciary responsibility is always

fulﬁlled: the "ITS as it currently exists is far from perfectly operating and it is unlikely soon

 

3Trades in NYSE-listed stocks may also take place on computer-based trading systems
such as the Instinct, Posit, Lattice or Arizona stock exchange. These smaller exchanges are not
included in the analysis.

“Most trading on the regional exchanges (at times more than 95 percent) is in NYSE-
listed stocks; the remainder are local stocks. Engel and Hecht (1994), p. 144.

5Engel and Hecht (1994), p. 107. Throughout the remainder of this analysis, non-NYSE
(including Nasdaq) exchanges are referred to generally as regional exchanges.

3

4

to be. Therefore, price priority rules are sometimes violated."6 In essence, the ITS is simply
a quotation display device, not an execution system. A trader that normally deals on the
NYSE may see a better regional quote and attempt to secure it but fail to because the trade
was "hit" before he or she could secure the trade, because the quote was "stale" (old) or the
quote pertained to a smaller trade. If the trader then returns to the NYSE, the market may
have turned against him and a worse price will be obtained compared to if he had simply
submitted the order to the NYSE. Thus (especially when trade activity is hectic) "traders do
ignore seemingly better quotes available elsewhere, and execute orders at their most
convenient and reliable market . . . . Clearly this results in violations of price priority.7

It is important to note that the ﬁduciary responsibility of the broker does not mean
that all exchanges offer the same price. Specialists not offering the BBQ are not obligated
to trade at these prices. In addition, discounting (trading inside the best bid or ask) is fairly
common. Blume and Goldstein (1992) found that 12 to 31 percent of trades in their sample
occurred inside the BBQ while Petersen and F ialkowski (1993) found price improvement for
35 percent of trades in their sample. Petersen and F ialkowski also found that 19 percent of
orders in their sample were sent to an exchange offering inferior quotes.

Previous research used two terms to describe the trading of dually-listed stocks: the
ﬁagmentation of order ﬂow and competition between exchanges. Fragmentation occurs any
time the same stock trades on two different exchanges. Conceptually, the terms competition
and fragmentation may be used interchangeably because they refer to the same phenomenon:

the splitting of order ﬂow between two or more exchanges. However, using a more strict

 

5Cohen et a1. (1982), p. 133.

71bid.

5

interpretation, these two terms have different implications. Competition refers to the positive
aspects arising ﬁom the splitting of order ﬂow while ﬁagmentation refers to negative effects.
While earlier research recognized both positive and negative effects associated with
the splitting of order ﬂow, competition has implicitly remained the focus. The SEC started
this trend two decades ago by implementing a number of rule changes with the stated
purpose of fostering competition; negative aspects were largely ignored. A consensus has
yet to be reached regarding the net effect of the rule changes. A brief consideration of these
changes follows in order to provide some background regarding the controversy over efforts
to enhance competition within the securities industry.
A. Attempts to Introduce Competition: Increasing Order Flow Fragmentation
Ever since the US. Congress ﬁrst passed the Security Amendments Act of 1975, the
chief objectives of the SEC have been (1) to increase competition in the securities industry
and (2) the achievement of a National Market System (NMS). The goal of the NMS was to
promote the efficient execution of transactions, encourage fair competition between
brokers/dealers and markets and to make readily available information regarding quotes and
transactions.8 Hamilton (1987) noted the effect of linking the exchanges as follows: "The
ITS is the cornerstone of the NMS . . . . The survival of so much off-board (non-NYSE)
trading is the keystone of the NMS . . . . The NMS has increased competition and, as
intended, has pushed the NYSE to improve its performance somewhat in order to prevent the

off-board market from capturing greater market share."9

 

8Cohen and Conroy (1990), p. 278.

9Hamilton (1987), pp. 1,335 and 1,334.

6

Rule changes have advanced the SEC's ﬁrst objective. Originally, due to NYSE Rule
390, NYSE members could not trade as third market (Nasdaq) dealers. That is, NYSE
members had to trade shares on the exchange ﬂoor. However, Rule 19c-3 (effective July 18,
1980) permitted NYSE members to make off-board trades. This enabled brokers to act as
dealers (market makers) by allowing them to match buy and sell orders in-house without
sending each order directly to the ﬂoor of the exchange.10 Essentially, Rule 19c-3 was an
attempt to increase market making competition in the hopes of reducing spreads and
attracting more order ﬂow.

NYSE Rules 97 and 127 may also have contributed to off-board trading. These rules
allowed smaller orders, limit orders and the NYSE specialist for a particular stock to take
precedence over some part of block trades (greater than 10,000 shares). In order not to
disrupt a pre-arranged block trade, a market maker may cross the trade on a regional
exchange.“

However, the net effect of these rule changes, particularly Rule 19c-3, has been, and
continues to be, the subject of much controversy. The initial effect of 19c-3 was favorable
as large brokerage houses began to trade in—house. Yet, by mid-1983, most large NYSE
ﬁrms had ceased trading off-board. Some market analysts attributed this to coercion by
exchange specialists.” The SEC noted that inefficiencies and frustration with the ITS

combined with a general dissatisfaction of the Rule 19c-3 environment may have contributed

 

10Cohen and Conroy (1990), p. 278. Due to the uncertainty of the effects, the SEC
restricted off-board trading to only those issues newly listed on the exchange after April 26,
1979.

11Hamilton (1987), p. 1,333.

12Cohen and Conroy (1990), p. 278.

to this outcome.13

In a report issued in 1981, the SEC stated that it was "not possible at this time to
reach any conclusions regarding effects of an active OTC (over-the-counter or Nasdaq)
market in Rule 19c-3 common stocks."M More recently, Cohen and Conroy (1990) noted
that "no consensus on the results of the 19c-3 experiment has yet emerged" and that "our
empirical results reﬂect the basic problems associated with increased competition in a
manner that also results in increased market fragmentation.“5 On the one hand, greater
competition decreased percentage spreads while on the other, higher return variances seemed
to result from the fragmentation of order ﬂow. Hence, the debate over the net effect of
increased competition (fragmentation) continues.

Yet, most market analysts agree that attempts to enhance competition have indeed
increased the attractiveness of off-board trading for many traders. Off-board transactions can
differ from NYSE transactions for a number of reasons. Often, off-board trades can be
executed more readily or information the trade may contain can temporarily be kept secret. ‘6
In a more recent development, dealers on non-NYSE exchanges (particularly Nasdaq) have
started to pay for order ﬂow, typically one cent per share. In addition, the NYSE auction

market may be ill-suited for block trades. Marketability of block trades depends on the

 

13SEC, Request for Comment on Off-Board Trading Pursuant to Rule l9c-3, Release No.
20074, 6 (1983).

1“SEC, A Monitoring Report on the Operation and Effects of Rule 19c-3 under the
Securities and Exchange Act of 1934, Release No. 18062 (1981). These results were criticized
by the NYSE. As with many other studies on the topic, the NYSE felt that the report did not
address the possible adverse effects associated with Rule 19c—3.

15Cohen and Conroy (1990), pp. 279 and 304.

“Hamilton (1987), p. 1,333.

8

broker actively searching among known large investors as NYSE specialists are precluded
ﬁ'om searching for offsetting block trade by NYSE Rule 113. "Upstairs marketmakers" may
negotiate these block transactions. Some are not NYSE members and may go off-board to
complete the trade.17

When combined, the improved technology linking exchanges, rule changes and other
factors such as payment for order ﬂow have translated into an increasing percentage of trades
involving NYSE-listed securities taking place off-board. While most studies have found that
the positive effects associated with this competition dominate negative effects from
fragmentation, undesirable effects do exist. They will tend to increase the cost to transact
and/or decrease execution quality. These effects are considered brieﬂy now.

1. Costs of Fragmentation

While the fragmentation of order ﬂow may result in poorer execution for both
dealers/traders and investors for a variety of reasons, there are essentially two underlying
causes. First, a market clearing price must now be reached on several exchanges instead of
just one. Assuming that information is not perfect and information time lags exist, prices
may not be the same at a given instance on two exchanges due to different supply or demand
conditions or information asymmetry. The price on one of the larger exchanges (NYSE) may
represent (or more closely approximate) the true market clearing price while a stock that
trades on a regional exchange may suffer from information adjustment lags or may be more
susceptible to local demand or supply shocks.

Second, fragmentation will tend to reduce the general "thickness" of the market. As

a result, investors may forego possible price improvement as the order is not exposed to all

 

1"’Ibid.

9

possible offsetting trades. In addition, traders will most likely have to wait at least as long
to execute a trade compared to fully consolidated order ﬂow. Fragmentation implies that
trades are exposed to fewer possible matching trades on any given exchanges. This means
that dealers must search and this involves costs. To offset search costs, dealers/brokers may
increase their price for liquidity services to offset the longer waiting time to trade
(particularly if they hold an undesirable inventory position).

Third, the uncertainty surrounding the equilibrium price of an asset (particularly for
the smaller exchanges) may have two effects compared to the case of consolidated order
ﬂow. First, market makers or specialists may increase their bid/ask spread for protection
against price uncertainty. This is referred to generally as satellite or price risk cost: prices
on regional exchanges tend to depend heavily on prior NYSE prices. As noted by Hamilton
"(i)f investors and marketrnakers are risk averse the probability of deviating from the price
in the dominant (NYSE) marketplace is a cost of off-board trading."18 Second, transaction
prices will become more volatile as new information reaches the exchanges and prices adjust.
Less stable transaction prices increase the price variance of an asset and serve to strengthen
the ﬁrst effect of wider bid\ask spreads.

Extemality considerations may also be used to argue against the fragmentation of
order ﬂow. Some exchanges (regionals) may free ride ﬁ'om others (NYSE) who regulate to
ensure that their exchange operates in a smooth and timely manner. Campbell et a1. noted
that "by listing a new contract, an exchange also provides a central place where integrated
traders meet, making it possible to trade off-ﬂoor at minimal search cost and to free ride on

available competitive price information to lower negotiation cost. Hence, off-ﬂoor trades

 

18Ibid, p. 1,334.

10

based on the bid/ask spread may be cost-efficient for certain traders who compliment each
other through reputation or block-trading frequency." ‘9 This extemality argument is
strengthened if economies of scale for regulatory activities exist.

Finally, increased fragmentation may lead to in-house markets for stocks (19c-3
stocks) creating submarkets that do not interact across ﬁagments. Submarkets are created
because brokers are allowed to act as dealers, matching buy and sell orders in-house without
sending each order directly to the ﬂoor of the exchange. This leads to violations in price
priority rules, inferior execution prices (due to unexposed orders), more public orders
executing against dealers instead of other public orders and violations of secondary priority
rules.20 As before, the net effect is to reduce the quality of the market.

To sum up, the pooling of orders onto one exchange should serve to increase the
effective thickness of the market (the number of traders willing to buy/sell at any given
price). In turn, a thicker market should lead to lower price variances, a shorter waiting time
to transact, shorter price adjustment delays and lower spreads.”

2. Beneﬁts of Competition

Competition, as used in this analysis, refers to competitive pressures faced by the
NYSE specialist, the dominant player in the market for NYSE-listed stocks. Competition
may originate ﬁ'om other specialists on the regional exchanges, private crossing networks,
Nasdaq marketrnakers or ﬂoor traders on the NYSE. Given guaranteed execution at the

BBC, specialists tend to compete by offering greater price improvement to the broker (and

 

19Campbell et a1. ( 1991), p. 497.
2"Cohen and Conroy (1990), p. 281.

21Cohen et a1. (1982), p. 119.

ll

ultimately the customer).22 Different transaction rules which affect the brokers choice of an
exchange may also affect competition.23 In addition, exchanges may offer inducements such
as payment for order ﬂow.

Generally, there are two main beneﬁts to competition between exchanges. First,
competition may serve to reduce the power of a monopolist specialist, lowering the cost to
transact by reducing bid/ask spreads.24 Competition eliminates any excess proﬁtability and
results in lower spreads because the specialist (or other party) is forced to accept more risk
for a given trade. Second, dually-listing stocks allows diverse markets to exist
simultaneously (e.g. specialist and competitive broker). Different markets may serve the
diverse needs of heterogeneous traders and allow order ﬂow to migrate to the most efﬁcient
market structure for that type of ﬁnancial asset.

3. Caveats

When the NYSE and OTC markets trade the same security, different market
structures compete. Each type of structure has its own advantages. Though an interesting
question, this paper does not explore the implications of this fact. Instead, its primary focus
is the measurement and effect of competition on the price of liquidity services. To simplify
this task, this analysis views exchanges as entities that compete against one another for order
ﬂow based on the price execution they offer for trades. In reality, the separate exchanges
merely offer alternative locations to complete a transaction; the specialists (NYSE and

regionals) or competing market makers (Nasdaq) are really the parties quoting stocks and

 

22Neal and Reiﬂ‘en (1994), p. 4.
23lbid.

24Demsetz (1968), p. 43.

12

competing for order ﬂow.

While it is true that parties trading NYSE-listed stocks can choose among the
different exchanges, certain parties may tend to deal only with a particular exchange,
possibly resulting in inefficiencies. If this "preferred" exchange is a non-NYSE exchange,
then the investor forgoes the opportunity of exposing their trade to the open outcry system
on the NYSE, where the market for the stock is typically most active. As a result, traders
may forego possible price improvement. However, by completing the trade off-board,
brokerage ﬁrms can often bypass an accumulation of orders on the NYSE specialist's book.
Therefore, the exchanges are able to offer different types of execution quality, which may
or may not beneﬁt the investor. The convention of competing exchanges is thought to
capture these differences and allows the analysis to simplify the institutional details of the
industry in order to focus on the measurement and effect of competition between exchanges.

Following the convention established by Lee (1993), this analysis assumes that all
trades are initiated by outside public traders who pay for liquidity, because the identity of the
trade initiator is not known. The analysis uses the price of liquidity services as the measure
of execution quality. However, this is only one aspect of trade quality. The speed of
execution, the amount of guaranteed depth and the reliability of trade settlement also affect
execution quality.” Therefore, the price of liquidity services is only one criterion that should
be used in the evaluation of exchange performance. This is particularly relevant if trades

migrate off-board to avoid queuing on the NYSE. In this instance, broker's may obtain

 

25Lee (1993), p. 1,012.

13

"better" execution for their client, though at a higher price.26
B. Statement of Purpose and Layout of Paper

The purpose of this paper is to provide insight regarding the competitive relationship
between stock exchanges and in the process help to resolve the controversy surrounding
attempts to foster a more competitive exchange environment. Prior research on the topic of
competition between stock exchanges has demonstrated that competitive effects seem to
dominate fragmentation effects, reducing spreads on the NYSE. These results were
statistically signiﬁcant and appeared to demonstrate that competition was indeed effective
in reducing NYSE liquidity prices.

However, more recent research demonstrated that there appears to be non-trivial price
differences for a given stock based on trade location. These results seem to contradict one
another. If competition is effective, price differences should theoretically not exist because
the transaction costs in this industry are most likely negligible, even for the smallest trades
(100 shares). Given these empirical results, a number of pertinent questions arise. Do both
results continue to hold using more recent data? Are the competition measures capturing
effective competition or some other phenomenon? Finally, is it possible to reconcile these
two results? These are the questions that this analysis attempts to address.

The paper proceeds in the following manner. Chapter 2 considers a number of

economic and measurement issues that may cause earlier studies to miss the true nature and

 

26Using the trade price as the sole criterion of trade quality may be misleading if
regional exchanges tend to offer worse execution when the market for the stock is very active
and the queuing of order ﬂow becomes a relevant factor in determining trade quality. However,
this analysis ﬁnds that non-NYSE exchanges are able to offer relatively better execution at these
times. Therefore, it is assumed that tradeoffs between price and speed of execution factors do
not signiﬁcantly alter the interpretation of the results.

l4

effect of competition. Issues include (1) the deﬁnition of the competition variable, (2) the
use of a spread instead of a liquidity premium measure, (3) the use of daily aggregated
measures (across time frames and trade size classes) and (4) unobservable variables.

Chapter 3 presents the formal literature review. Included are studies using regression
analysis to quantify competitive effects and two studies comparing transaction prices across
exchanges. These studies serve as a foundation for the analysis in Chapter 4.

Chapter 4 introduces the formal methodology of this paper. It is based on a simple
dominant exchange model where the NYSE dominates other exchanges due to informational
advantages. It then replicates the methodologies used in earlier studies to examine whether
more recent data conﬁrm their results. Once the results are replicated, the analysis shortens
the time frames used to derive variables used in the analysis. This may affect the
interpretation of certain key variables, particularly volume and competition. In addition,
competitive effects are measured on all exchanges, not just the NYSE.

Chapter 5 investigates the price dispersion phenomenon and presents a discussion of .
the Law of One Price and the characteristics needed for it to hold in a particular market. An
analysis is performed to test whether or not the regional exchanges are local markets. If they
are local markets, the Law of One Price need not take effect, thus price differentials may
exist between the exchanges.

Chapter 6 completes the analysis. It attempts to explain price differentials between
the NYSE and regional exchanges and the conditions under which the regional exchanges
compete. In order to examine the possibility that the regional exchanges cream skim (i.e.,
offer selective market making) and do not offer effective competition all of the time,

explanatory variables are derived that characterize the trading environment. A cream

15

skimming explanation permits the existence of an effective competition measure in an
environment characterized by non-trivial price differences. This analysis ﬁnds evidence of
this behavior. Regional trade activity and relevant price execution are found to be a function
of NYSE activity. Regional exchanges tend to trade or "compete" when the trading
environment is active (thus free-riding on the NYSE price discovery process, minimizing
price risk costs and inventory adjustment costs) and tend to target smaller trades (thus

avoiding information traders).

CHAPTER 2
ECONOMIC AND MEASUREMENT ISSUES

Numerous studies have examined the effect competition has on the price of NYSE
liquidity services. Typically, studies have used regression analysis to determine whether
competition, as measured by volume dispersion or the number of exchanges actively trading
a particular stock, signiﬁcantly reduced NYSE spreads. The theories underlying most
regressions were based on one of two models: (1) a transaction or inventory cost model
where the spread compensates the dealer or specialist for providing services of immediacy
or (2) an adverse selection model where the spread serves as compensation for the existence
of individuals who have superior information regarding a trade (information traders).27

In both models, the NYSE spread is typically modelled as a function of least ﬁve
independent variables: trade activity, price volatility, price, information and competition.
Trade activity is measured using daily trade (total) volume or the number of transactions
throughout the day. The more active a market, the shorter the waiting time to execute a
particular transaction. This should reduce spreads as it reduces the risk associated with
carrying a particular inventory. In general, dealers participate in trading activities to the
extent required by temporary imbalances in the inﬂow of orders. Other things being equal,

the probability that there is an imbalance varies inversely with the time rate of transactions.28

 

27Mann and Seijas (1991), p. 54.
28Tinic and West (1972), p. 1,709.

16

17
The price volatility of a stock is measured using a price variance measure over the
trade day. Higher price variances increase uncertainty and risk and therefore increase
spreads. The stock's price also increases spreads because "the price directly affects the total
borrowing costs or the opportunity cost of capital tied up when the specialist takes a position

in an issue."29

Information has been measured using the volume or percentage of large or
block trades over the trade day. Block trades originate from institutional investors who may
have superior information. Spreads will increase as a form of protection against information
traders. Competition has typically been measured using a Herfmdahl-Hirschman Index
(HHI) calculated using exchange-speciﬁc trade volumes over the trade day or a measure
representing the number of exchanges that were active during the day. If competitive effects
dominate ﬁagmentation effects, then both measures should reduce the quoted spread.
Typically, all ﬁve independent variables have been found to have signiﬁcant
explanatory power and had the hypothesized effect on NYSE spreads. However, while
earlier research seems to have reached a consensus regarding the effect of these variables,
they may also contain similar ﬂaws. Certain economic and measurement issues may not
have been addressed due to the nature of the data. In most analyses (except the most recent),
daily average measures have been used because transaction data were not available. The use
of a transactions database allows for improved speciﬁcation. While this holds true for all of

the variables mentioned, only economic and measurement issues affecting the competition

and volume variables are considered as they are the focus of this analysis.

 

29Tinic (1972), p. 80.

18

A. Competition Measures

Although two competition measures have been used with success in the past, the
speciﬁcation of this variable is important because while the measures are similar, they are
not equivalent. Two distinct issues arise regarding this variable. First, which competition
measure is more accurate? Second, does either measure effectively represent the competitive
environment?

1. The Measurement of Competition

Because both competition variables were found to have the hypothesized effect on
the price of NYSE liquidity services, this analysis utilizes both measures. It will investigate
the robustness of each measure on both NYSE and regional liquidity prices moving ﬁom
daily to hourly time ﬁ'ames. It is possible that the two measures capture different phenomena
and therefore have disparate impacts on regional and NYSE liquidity prices. If true,
competition coefficients may differ markedly depending on the exchange and the measure
used in the regression. For example, the overall dispersion of volume (between the NYSE
and other exchanges) may be irrelevant to the smaller (regional) exchanges. However, a
variable representing the number of exchanges that are active during a particular time frame
may be relevant for these exchanges because if one regional exchange ﬁnds conditions
suitable for trading, most likely the conditions are suitable for others to "compete" as well.
In addition, relatively small increases in regional trade volume will not greatly affect the HHI
variable, but will affect the number of exchanges that are active.

Conversely, for a large exchange such as the NYSE, the dispersion of total volume
(NYSE versus off-board volume) may be more important. Hamilton noted that "the division

of off-NYSE trading between regional exchanges and the third market may have no

19

discemable effect on NYSE spreads; perhaps only the total off-board trading matters."30 In
this case, the location of volume would not matter so much as the fact that trading took place
on an exchange other than the NYSE.

2. Effectiveness of Competition Measures

The second issue regarding the competition variable is whether either measure truly
represents the presence of effective competition. Prior research also raised this concern.
Demsetz (1968) questioned the accuracy of an activity competition measure. Using the
number of exchanges trading a particular stock as a competition variable, Demsetz noted that
"an enumeration of the forces of competition is not, by itself, convincing evidence of

"3‘ In reference to Demsetz’s paper, Tinic (1972) echoed this concern:

competition.
"(e)specially in the case of issues experiencing heavily concentrated trading activity in the
NYSE, the mere presence of external markets (as measured by the number of multiple
listings) need not indicate the degree of effective competitive pressure on the specialist."32

Having noted this weakness, Tinic used an HHI measure in his analysis. This
measure was not without acknowledged shortcomings. This occurs because the index might
simply act as a proxy for long-run trading activity; lower HHI values may simply reﬂect the
inﬂuence of heavy trading volume. Mulherin (1996) noted that the reported association
between a Herﬁndahl index and spreads may be driven in whole or in part by the incidence

of limit orders and heated ﬂoor trading activity on the NYSE. Heated ﬂoor competition

keeps specialists "honest" by forcing them to maintain narrow spreads. This internal

 

30Hamilton (1979), p. 180.
31Demsetz (1968), p. 44.

”Tinic (1972), p. 88.

20

competition measure may be correlated with an HHI measure.

Therefore, both measures may not represent effective inter-exchange competition.
Essentially, the issue is the exogeniety of the competition variable: competition, as it has
been measured in the past, may simply be a function of market activity (volume) or even the
spread of an asset.33 If competition is merely a function of trading activity, then the inclusion
of volume and competition variables in a regression not only reduces the explanatory power
of both, but may also spuriously credit the competition variable with reducing spreads.

Given this relationship, improved speciﬁcation may result by reducing the correlation
between volume and competition. This is particularly important if non-NYSE exchanges
engage in cream skimming activity because these exchanges may become active or
"compete" only under certain conditions. Coffee (1996) noted that these exchanges typically
pay for order ﬂow (nonprice competition) only in active NYSE stocks, trade
disproportionately in smaller trades (avoiding information traders) and often match quotes
on the NYSE and ﬁ'ee-ride on the NYSE price discovery process (which is easier when
volume is heavier). McInish and Wood (1992) found that regionals and Nasdaq ﬁ'ee-ride at
least 90 percent of the time on trading in NYSE-listed securities. By isolating trading
environments that are more conducive to cream skimming and reducing the correlation
between volume and competition, this analysis can more accurately measure competitive
effects relative to other market characteristics.

To test the robustness of the competition measure, this paper utilizes three

approaches. First, three time frames are used to quantify competition's impact on the cost

 

33McInish and Wood (1992) noted that the direction of causality for the inverse
relationship between NYSE spreads and regional trading volume is uncertain. Tight NYSE
spreads may result in greater regional trading.

21

to transact: daily, half-day and hourly. It is postulated that this breakdown will increase the
size and signiﬁcance of the volume and competition coefﬁcients because shorter time ﬂames
are able to separate volume effects ﬂom competitive effects more accurately; there should
be less correlation in the variables using shorter average time ﬂames. This is particularly
true for non-NYSE exchanges where trading may be sporadic.

Second, the analysis in the ﬁnal section of this paper uses competition measures (in
addition to other variables) to explain price differentials between the regional exchanges and
the NYSE. To be truly effective, competition should not only reduce the spread or price on
the NYSE, but it should also reduce price differentials between exchanges.

Finally, the competition variable measuring the number of exchanges that are
actively trading a stock is modiﬁed to include a threshold value. This measure does not
recognize an exchange as "active" unless a certain threshold of trades were executed on the
exchange during the relevant time frame.34 In many instances, only one or two trades occur
on an exchange during the time ﬂame. It is questionable whether these exchanges should
truly be considered "active" and included in the competition variable.

B. The Price of Liquidity Services: Spread or Liquidity Premium?

Earlier studies used the NYSE spread as a measure of the price for liquidity services.
The spread is deﬁned as the difference between the bid and ask price a dealer or specialist
quotes. The bid is the price at which the dealer or specialist is willing to accept a trade while
the ask is the price at which the dealer is willing to trade stock. However, more recent

literature has noted that using the spread as the dependent variable is inappropriate because

 

34Analyses were also performed without the threshold, but threshold values greatly
increased the explanatory power of this variable. This is attributed to a reduction in correlation
between the volume and competition variables.

22

transactions often occur "inside" the quoted spread. To trade inside the spread means price
improvement for the investor (a transaction price greater than the bid or less than the ask).
Blume and Goldstein (1992) found that between 12 to 31 percent of the trades in their sample
occurred inside the quoted spread.

Transactions may occur inside the quoted spread for a number of reasons. Discounts
may be offered due to competitive pressures. Alternatively, dealers or specialists may offer
selected discounts for particular trades in order to balance inventory or because of temporary
information advantages. Transactions may also occur inside the spread if specialists do not
display the best limit orders or if market orders are matched.” As a result, using the posted
spread as a price measure will overstate the cost to transact.

Recent literature supports this contention. Reiss and Werner (1996) and Peterson and
Fialkowski (1994) noted that researchers should not use average spreads to measure the
efﬁciency of a market: "(e)mpirical tests of the sources of the spread must use the spread
actually paid by investors."36 Peterson and Fialkowski went on to state that using the spread
as a price measure may not be a problem if spreads serve as a proxy for actual transaction
prices. However, in their analysis the authors found that "(t)he effective spread faced by

investors in our samples is signiﬁcantly smaller than the posted spread."37 The possible

 

35Specialists need only post quotes that are "representative" of orders in the limit order
book. In particular, limit orders which are better than the current quote are not automatically
posted as new quotes. McInish and Wood (1995) refer to undisplayed limit orders that are at
better prices than those of limit orders actually displayed as the "hidden limit order book."
Fictitious price improvement may occur if market orders execute against hidden limit orders.
The authors feel that if true spreads were revealed by the NYSE, the spreads would be matched
by the regionals and the illusory price improvement would disappear.

36Peterson and F ialkowski (1994), p. 290.

”Ibid, p. 270.

23

rnisspeciﬁcation is compounded if the two measures are not highly correlated. The authors
derived this result: "(n)ot only do we ﬁnd the size of the posted spread and the effective
spread differ in magnitude, but their correlation is also small."38

A more appropriate measure of the cost to transact is the liquidity premium. The
liquidity premium is deﬁned as the difference between the transaction price and the
equilibrium price of a transaction. Although the true equilibrium price is not known with
certainty, it is assumed to equal the mid-point of the spread. This is a strong assumption.
However, it is maintained that the liquidity premium is a superior measure to the spread
because it captures the essential elements of the transaction: the spread and the actual
transaction price.

As a simple example of how the liquidity premium is calculated, assume a bid equal
to 45.0 and an ask equal to 45.5. Given a transaction at the ask, the equilibrium price would
then be calculated as (45.0 + 45.5)/2 or 45.25 and the liquidity premium would equal .25
(45.5 - 45.25). The strength of this approach is that the liquidity premium measure allows
an analysis to use observable information that is pertinent (transaction price) and captures
discounts.

In order to calculate liquidity premiums, this analysis uses the midpoint of NYSE
spreads as an estimate of the equilibrium price of a stock. Regional quotes are not used
because they are unreliable; often they are not meaningﬁrl. Petersen and F ialkowski noted
that "price improvement relative to an exchange's own quote can be deceiving."39 That is,

regional quotes may be irrelevant; however, while their quotes may be inferior, they tend to

 

38Ibid, p. 271.

39Petersen and F ialkowski (1994), p. 275.

24
match best quotes even if they do not display them.40
C. Aggregation Across Trade Size Classes

Because prior research did not control for the size of trades, regressions treated trades
of different sizes as homogeneous products. However, it is questionable whether a trade of
5,000 shares should be considered the same product as a trade of 100 shares. Trades that are
not the same size may be heterogeneous products because the information content of the
trades differ, non-NYSE exchanges are less willing to deal (compete) in larger trades (as they
have a greater impact on inventory and are often infonnationally motivated) and often larger
trades beneﬁt ﬂom discounting. Reiss and Werner noted this discounting behavior: "(o)ur
estimates reveal that medium to large trades on average receive discounts ﬂom the touch
(inside) spread."‘“

Recent literature has demonstrated the need to control for the composition of volume.
McInish and Wood (1992) found that trade size had signiﬁcant explanatory regarding
relative bid/ask spreads in their analysis. Lin, Sanger and Booth (1995) also conﬁrmed the
importance of trade size. The authors noted that "(t)he average effective spread is found to
increase with trade size. . . . The increase in the effective spread associated with trade size
is largely the result of an increase in adverse information revealed ﬂom large trades."42

To control for volume composition, this analysis separates trades into ﬁve size

categories when comparing liquidity premiums across exchanges. Individual size classes

 

4°Ibid, p. 289.

“Reiss and Werner (1996), p. 161. This result pertains to trades on the London Stock
Exchange's electronic quotation system, known as SEAQ.

42Lin, Sanger and Booth (1995), p. 1,180.

25

include trades between 100 and 400 shares, 500 and 900, 1,000 and 1,900, 2,000 and 4,900
and trades greater than 5,000 shares. This breakdown follows the convention established by
Lee (1993), which is summarized in the literature review. In addition, regressions that are
used to explain liquidity premiums on speciﬁc exchanges include a size variable to capture
any effects attributable to the size of the trade.
D. Unobservable Variables

The ﬁnal issue is the possible bias resulting ﬂom relevant unobservable variables.
While these variables are observable to the market participant, they are not observable to the
econometrician viewing the data set. There are three general categories of unobservable
variables: (1) the queuing of order ﬂow on the NYSE specialist's book (after price, orders
are handled based on the time they are placed); (2) the payment for order ﬂow; and (3) a
trader's taste for a particular non-NYSE exchange, more generally described as a principal-
agent phenomenon. Principal-agent problems may arise in the industry as brokers and
traders forego their ﬁduciary responsibility to seek best execution for investors. The search
for best execution may be costly to the brokers and traders not only in terms of time
commitment, but also because "gentlemen's agreements" exist. In addition, certain brokers
may limit their trading to a particular exchange.

All unobservable variables may be regarded as "non-price" competition variables.
They are methods by which the regional exchanges compete for order ﬂow with the NYSE.
Non-price factors increase regional volume at the expense of the NYSE, making competition
look relatively more active. In particular, the queuing of order ﬂow on the NYSE may cause
trades to migrate off-board for reasons not related to the price of the stock. This may bias

the competition coefﬁcient.

26

The direction of the bias hinges upon (1) the effect non-price competition factors
(omitted) have on liquidity premiums (dependent variable) and (2) the correlation between
the omitted variables and the variables included in the analysis. For example, let X,
represent a "typical" competition variable, X2 an unobserved non-price competition variable

and Y the average liquidity premium. If the real regression is:

(1) Y=a+Ble+BZX2+8
and the estimated equation is

(2) Y=a+131X1+8

then the expected value of the competition coefﬁcient becomes

(3) 1943,1213) = B. + 13, * 10.2)

where f(r,2) is the simple correlation coefﬁcient between X, and X2. A bias will exist as long
as the true [52 does not equal zero (i.e., X2 is relevant) or the simple correlation coefﬁcient
does not equal zero.

As an example of potential bias, consider the HIH competition variable. We would
expect the HHI coefficient to have a positive impact on liquidity premiums; greater volume
dispersion (lower HHI) represents greater competition (lower liquidity premiums).
Therefore, B, should be positive. The correlation between omitted non-price competition
factors and the HHI variable should be negative. When the regional exchanges compete
using non-price factors, volume dispersion increases, ceterius paribus, reducing the value of

the HHI variable. Therefore f(r,2) is negative.

27

This relationship is well-grounded in empirical research. Harris noted that it may
become proﬁtable to supply liquidity when volume is high and spreads equal their minimum
price variation (a spread of 1/8): "supplying liquidity will be highly . . . perhaps excessively
. . . proﬁtable."43 In general, the OTC and regional exchanges will ﬁnd it more attractive to
"compete" at higher volume levels. McInish and Wood stated that "competition will vary
with volume . . .the crowd is more active as volume levels rise.“4 Coffee (1996) noted that
traders tend to pay for order ﬂow only in stocks that are active. Because regionals cannot
compete on price terms at minimum spreads, they may compete using non-price factors.

Hence, the direction of the bias (if any), depends on the sign of [32, the effect of non-
price factors on the dependent variable. If it is equal to zero, no bias will exist. However,
if [32 is positive, then the HHI competition coefﬁcient will be biased downward (less
positive), moving the coefficient away ﬂom its hypothesized value. Most likely, if dealers
had to resort to discounting (and were able to discount, given the spread) or reducing the
outstanding spread to attract order ﬂow, they would have, assuming it was proﬁtable.
Therefore, it is asserted that non-price competition factors cause liquidity premiums to
remain at higher levels or increase because traders are willing to accept higher costs to
transact (which are passed on to the investor) in exchange for speedier transactions, payment
for order ﬂow or the maintenance of certain trade relationships. This biases an HHI
competition coefficient downward (less positive) and an activity competition coefﬁcient
upward (less negative), reducing both of their true impact on liquidity premiums. This

occurs because volume leaves the NYSE for non-price reasons.

 

“Harris (1992), p. 14.

4“McInish and Wood (1992), p. 754.

28

Unfortunately, given the data, there exists no acceptable method to control for this
type of bias. This remains a challenge for future research. However, if a bias does exist,
shortening time ﬂames used to measure variables should limit any bias to fewer

observations, hopefully allowing the competition measure to be modelled more accurately.

CHAPTER 3
LITERATURE REVIEW

The literature reviewed in this chapter is separated into two general categories. The
ﬁrst group of analyses use regressions to quantify competition's effect on NYSE spreads and
liquidity premiums. The second group compares the price of liquidity premiums and
effective spreads across exchanges that trade dually-listed NYSE stocks. Both
methodologies are replicated in Chapter 4.

A. The Effect of Competition on NYSE Spreads and Liquidity Premiums

Demsetz (1968) published the ﬁrst recognized paper analyzing the effect of
competition on NYSE spreads. The analysis used a random sample of 192 stocks over two
days in 1965; only competitive effects ﬂom regional exchanges were considered. Two
simple regressions demonstrated that while the competition coeffrcient (the number of
exchanges that listed the stock) had the expected sign, the results were not statistically
signiﬁcant. Demsetz attributed the weak explanatory power of the competition variable to
measurement difﬁculties.

Tinic (1972) improved Demsetz's speciﬁcation through the use of an HHI variable
to measure competition and the inclusion of a number of other relevant explanatory
variables. Data were averaged over nineteen trading days in 1969 for 80 stocks. Tinic found
that competition had a statistically signiﬁcant effect on the average NYSE spread. Other

coefficients included in the regression were also found to have a statistically signiﬁcant

29

30

effect on NYSE spreads.

Tinic and West ( 1972) used the average number of dealers quoting spreads as a
competition variable to measure the effect on Nasdaq spreads. Two regressions using
separate data sets were performed. The ﬁrst regression used data on 68 stocks for a single
day in 1962; in this instance competition was found to have poor explanatory power. The
authors attributed this result to the fact that a single day of data was used. In addition, the
authors noted that the “low t-ratios with V (volume) and Nq (competition) result in part ﬂom
the high intercorrelation between the two variables.”45 This intercorrelation “makes the
partitioning of explanatory power between the two variables virtually impossible.”46

The second regression used 1971 data for 300 Nasdaq stocks averaged over the ﬁrst
ﬁve days in November. The authors felt that the ﬁve day averaging procedure would make
the data more stable. Using this data set, Tinic and West found signiﬁcant competitive and
volume effects. However, though they found signiﬁcant competitive effects, the authors
noted that “(t)rading activity still remains a dominant factor in determining the number of
dealers to participate in the market-making process.”47 Tinic and West attributed the
signiﬁcance of their results (compared to Demsetz's analysis) to their different measurement
of competition and the different markets (NYSE versus Nasdaq). The authors termed the
regional competition used in Demsetz's analysis as "possibly irrelevant."

Branch and Freed (1977) reconﬁrmed the statistically negative effect competition has

on the average NYSE spread. The authors used data ﬂom 1,732 stocks over one day in 1974.

 

45Tinic and West (1972), p.1,713.
“Ibid, p. 1,714.

“Ibid, p. 1,720.

3 1
Competition was measured using the number of different exchanges on which the security
traded (including the third market); it was found to signiﬁcantly reduce spreads, though the
effect was small.

Hamilton (1979) analyzed the determinants of the NYSE spread using cross-sectional
data on 315 NYSE-listed stocks ﬂom 1975. Variables were derived by averaging over four
trading days. Using the number of marketrnakers that provide quotes for a stock as a
competition measure, Hamilton found that off-exchange trading of NYSE-listed stocks can
result in smaller and less volatile spreads if effects attributable to increased competition
dominate ﬂagmentation effects.

Cohen and Conroy (1990) examined whether Rule l9(c)-3 led to (1) higher levels of
market making activity, (2) increased competition among market makers, and (3) increased
return variance. Rule 19(c)-3's objective was to introduce competition by allowing NYSE
members to act as third market dealers for a limited group of stocks. Using residual analysis,
Cohen and Conroy found that 19(c)-3 stocks beneﬁted ﬂom the attempt to increase
competition: "Overall, it is apparent that, on an ex post basis, the OTC market making
activity of exchange members did have a favorable effect on bid/ask spreads, at least for
some groups of stocks."48

McInish and Wood (1992) used a linear regression model to identify time-of-day
effects for NYSE spreads. The data covered all quotes for by NYSE specialists for the ﬁrst
six months of 1989. The authors constructed a minute-by-minute time series of time
weighted percentage NYSE bid/ask spreads (BAS) over the trading day. Competition was

deﬁned as the square root of volume traded off-board divided by NYSE volume. It was

 

48Cohen and Conroy (1990), p. 294.

32
found to have a (highly) statistically signiﬁcant effect on the BAS.

Harris (1992) found unusual results in his analysis. Using data ﬂom the ﬁrst and
second quarter of 1989, Harris found that less volume dispersion (as measured by the log of
the ratio of NYSE volume over total volume) decreased the average relative spread. He
attributed this result to the fact that trades migrate to the primary exchange (NYSE) when
spreads are tight: "(t)he (competition) variable coefﬁcient is negative. If large (NYSE)
market share is a proxy for lack of competition among dealers, this result would be
surprising. Note, however, that if spreads in the primary market are tight, order ﬂow will
migrate to the primary market and primary market share will be large, which is consistent
with the observed result."“9

Using a linear programming model, McInish and Wood (1994) found that
competitive effects dominate ﬂagmentation effects. The analysis used data ﬂom 198 8-1 990
and separated stocks into 5 portfolio classes, which maximized the degree of ﬂagmentation
between them but minimized differences in other explanatory variables. Both the spread and
liquidity premium were found to decrease with ﬂagmentation.

B. Does Trade Location Affect Price Execution?

Lee (1993) attempted to measure the degree of integration between the NYSE and
Nasdaq/regional exchanges by comparing liquidity premiums of similar orders across
exchanges. He found that the price execution (using liquidity premiums) of trades differed
systematically based on trade location. These interrnarket differences were found to be a

function of trade size, with the smallest and largest trades exhibiting the greatest liquidity

premium differences.

 

”Harris (1992), p. 11.

33

Lee's hypothesis was that the ITS linking the exchanges did not effectively
communicate intermarket liquidity differences. If it did, then identical orders would have
equal opportunity for best price execution regardless of their initial routing location. Lee
noted that even though dealers and traders are required to use the best ITS quote to satisfy
their ﬁduciary responsibility, price differences based on the location of execution can still
exist due to hidden liquidity that is not reﬂected by these quotes. As noted, many trades take
place inside these quotes.

Lee's analysis was performed in the following manner. Trades on the NYSE were
matched to regional trades based on the size of the trade and a two minute time interval.50
Liquidity premiums were then calculated for all matched trades. If there was more than one
NYSE trade to match, then the liquidity premiums on the NYSE were averaged for all
matched trades. The "excess cost" of the trade on the regional exchange was then deﬁned
as the difference between the liquidity premium on the regional exchange minus the liquidity
premium on the NYSE, which may be positive or negative.“

All trading days in 1988 and 1989, except October 25, 1989, were used in the
analysis. Tables 1a and 1b list liquidity premium differences calculated by Lee for 1988 and
1989 (a positive value indicates that execution costs were higher on the regional/Nasdaq

exchange):

 

50This time frame was changed to intervals of one, ﬁve and ten minutes, with no
appreciable change in the results.

51It was noted that a comparison of intermarket average liquidity premiums for a given
size may not be accurate. Regional dealers may "skim the cream" by making markets only in
more liquid stocks, meaning that they would have lower liquidity premiums than the NYSE. In
addition, effective spreads are known to display pronounced intraday patterns.

34

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Table la
Average Excess Liquidity Premium Paid for 1988 Off-Board Trades
in Time Matched Sample

Charles Lee (1993)

Trade Size (Shares)
Exchange 100 - 400 500 - 900 1000 - 1900 2000 - 4900 5000 - 9900
Boston 1.4* 0.71 0.67* -0.03* 2.11
Cincinnati 1.36* -1.21* -1.36* -0.020 0.73
Midwest 0.68“ -0.70* -0.95* -0.9* -0.28*
Paciﬁc 1.10* -0.23* -0.l8* -0.5* 071*
Philadelphia 156* 0.5* -0.5 -0.41 0.32
Nasdaq 1.51* 0.76* 0.59* 0.21 * 2.61 *
All 107* -0.22* -0.42* -0.49* 0.31
*Denotes statistical signiﬁcance at the one-percent level.

Table 1b
Average Excess Liquidity Premium Paid for 1989 Off-Board Trades
in Time Matched Sample

Charles Lee (1993)

Trade Size (Shares)
Exchange 100 - 400 500 - 900 1000 - 1900 2000 - 4900 5000 - 9900
Boston 104* 1 0.55* 0.33 0.35 0.88
Cincinnati 2.65* 0.44 -0.44* 1 .49 1.32
Midwest 0.87* -0.03* -0.27* -0.32* -0.06
Paciﬁc 1.19* 0.38 0.23 -0.02* -0.42
Philadelphia 182* 1.21 * 0.89* 0.46 1.26
Nasdaq 1.58* 1.23* 0.89* 1.13“ 2.26*
All 122* 0.45* 0.14 022* 0.51

 

 

 

 

 

 

 

*Denotes statistical signiﬁcance at the one-percent level.

 

 

 

35

For the smallest trades (100 - 400 shares), the NYSE was found to offer better
execution. The Midwest and occasionally the Cincinnati and Paciﬁc exchanges offered
superior price execution for midsize trades (500 - 4900). For large trades, the results were
mixed. However, a consistent result was that "(o)nly the Nasdaq performs consistently
worse than the NYSE in all size categories."52

Petersen and F ialkowski (1994) preformed a similar analysis. The authors found that
there was signiﬁcantly less price improvement for orders sent to regional exchanges. On
average, orders routed to regional exchanges received 3.1 cents less price improvement
compared to the NYSE. This result was not attributable to worse regional quotes. Regionals
tended to match the best quote when they executed an order even if they were not currently
displaying it. The authors attributed price differences due to liquidity that was not displayed,
particularly on the NYSE.

C. Summary

Prior empirical research appears to have settled a number of issues regarding the
determinants of NYSE bid-ask spreads. All studies found that volume signiﬁcantly reduces
spreads. There was a positive and signiﬁcant relation between spreads and the price and
price volatility of an asset. The effect of information appears to depend on the measure one
uses, though in most cases it proved to have signiﬁcant explanatory power. McInish and
Wood (1992) demonstrated that the variables listed above have a similar impact on liquidity
premiums and that trade size has signiﬁcant explanatory power regarding liquidity
premiums. Regarding the competition variable, research demonstrated that competition

reduces spreads, and in one study, reduces liquidity premiums, no matter the competition

 

”Lee (1993), p. 1,011.

36

measure used in the analysis.

However, many researchers have also commented on the inherent difﬁculty of
measuring competition accurately due to the high correlation between volume and
competition. As a result, a number of authors noted that the competition measure may
capture effects that are more properly attributed to volume. Some papers even called into
question the direction of causality: Does volume reduce spreads thereby attracting dealers
to the more active market or does greater competition force down spreads? The implicit
assumption in prior research was that competition caused spreads to fall. This need not be
the case. What appears to be a more competitive environment may simply be a by-product
of higher volume that reduces spreads.

The two ﬁnal studies in the literature review demonstrated that price execution for
similar adjacent trades can differ systematically depending on the location of the trade. Lee
(1993) found that non-NYSE exchanges offered worse execution on average between one
and one and one-half cents for the smallest trades (100 - 400) while Peterson and Fialkowski
(1994) found that regionals offer signiﬁcantly less price improvement. As noted, these price
differences are most likely the result of hidden liquidity (i.e., traders willing to execute a
trade at better terms for the investor) that is not reﬂected by intermarket (ITS) quotes: many
trades take place inside these quotes.

1. Do Empirical Results Conﬂict?

The fact that (1) prices can differ systematically based on trade location while (2)
competition appears to reduce both NYSE spreads and liquidity premiums may be troubling.
It seems to imply that the securities industry is not typically characterized by a competitive

environment. That is, competition or volume dispersion may indeed be associated with

37

lower spreads and prices, but this condition is not prevalent enough to force price
differentials between exchanges down to their transaction cost, which in this case should not
differ signiﬁcantly ﬂom zero. Do statistically signiﬁcant price differentials imply that this
industry is non-competitive the majority of the time?

It may not if there is no causal relationship between these two results. However, it
is asserted that a competitive environment which reduces prices on individual exchanges
should also serve to reduce price differentials between exchanges. This follows because an
active trade environment (as measured by volume or the number of trades) also tends to
improve the competitive position of non-NYSE exchanges. When trading activity increases
(in most cases competition will as well), there is less room for price differences to exist
because non-NYSE exchanges know true prices with greater certainty, information and cost
heterogeneity between exchanges declines and price discreetness makes it costly to discount
when spreads are low (spreads tend to decrease in active trading environments). In addition,
if the market is active and spreads are reduced to their minimum level of 1/8 (a liquidity
premium of 6.25, assuming a transaction at the bid or ask), then NYSE specialists cannot
hide liquidity and price differences cannot exist because discounting is no longer an option.

Another explanation for the coexistence of these two results is that competition has
a different effect on regional exchanges compared to the NYSE. However, as will be shown
in the next chapter, competition not only has the same negative impact on regional liquidity
premiums, but the effect is similar in magnitude as well (based on the sign and size of the

competition coefficient). Hence, this explanation is dismissed.

38

2. Possible Explanations for Empirical Results

There are three possible explanations for the coexistence of the apparently conﬂicting

empirical results found in the literature:

1.

depth.

The competition results derived ﬂom regressions may not be directly comparable to
price difference calculations. While most regressions used data aggregated over the
trade day, Lee used two-minute time ﬂames to derive price differences. Regressions
may simply capture a general relation between average daily volume dispersion and
average spreads/liquidity premiums: this may not hold at shorter time ﬂames. The
broader time ﬂame may also allow the competition measure to capture volume
effects inadvertently or introduce a bias ﬂom unobserved non-price variables.

The individual exchanges are not completely integrated, but retain some local power
over price. Hence, it may be improper to pool price data across exchanges. The
exchanges may not be completely integrated due to information lags or because
certain dealers or traders prefer a particular regional exchange, foregoing possible
price improvement resulting ﬂom hidden liquidity on the NYSE.

The two results do not conﬂict, but the competition measure is ﬂawed because it
captures cream skimming activity.53 Evidence suggests that trades take place on non-
NYSE exchanges when trades are relatively small (thus avoiding information traders)
and when trade volume is relatively heavy (thus ﬂee-riding on NYSE price discovery
and minimizing price risk and inventory carrying costs). Cream skimming activity
can potentially make competition look effective while simultaneously allowing price
differences to exist.

The remainder of this paper is devoted to exploring these explanations in greater

 

53Cream skimming refers generally to selective market making by regionals dealers or

selective trade location decisions by traders.

CHAPTER 4
COMPETITION AND PRICE DIFFERENCES
The simple model outlined in this chapter provides a general ﬂarnework for viewing
the securities industry. The model also provides motivation for price difference calculations
between the NYSE and non-NYSE exchanges, time ﬂame breakdowns and trade size
breakdowns used later in the analysis. Earlier empirical studies are then replicated to ensure
that their results were not driven by speciﬁc data sets and that structural changes have not
occurred within the industry which may alter their general conclusions.
A. The Simple Model: The NYSE as the Dominant Exchange
Although many exchanges can trade dually-listed NYSE securities, trading is
unquestionably dominated by the NYSE, particularly when trades are large. In 1993, 90
percent of NYSE-listed stocks were traded on at least one other exchange.54 Of these dually-
listed stocks, the largest regional exchange (Midwest or Chicago) traded only 4 percent of
the total volume, while Nasdaq traded only 7 percent.55 Market shares were similar in this
analysis. Table 2 lists each exchange's share of total volume for stocks included in the

sample:

 

54Engel and Hecht (1994), p. 107.
5-"Ibid.

39

40

Table 2
Exchange Volume for 1992 Sample

 

 

 

 

 

 

 

 

 

 

 

 

Exchange Total Volume Percent of Total Volume
Boston 79,395,500 1.54%
Cincinnati 84,669,000 1 .64%
Midwest 170,3 12,600 3.30%

NYSE 4,416,448,900 85.46%
Nasdaq 233,409,700 4.52%

Paciﬁc 1 16,339,600 2.25%
Philadelphia 67,405,800 1 .3 0%

Total 5,167,981,100 100.00%

 

 

Given the large NYSE market share, the market for NYSE-listed securities is
modeled as being dominated or led by the NYSE, while smaller exchanges follow by
forming a competitive ﬂinge. The NYSE leads or dominates because its large market share
gives dealers and specialists on the NYSE a virtual monopoly on the price discovery process.
Due to their position, NYSE specialists have access to highly signiﬁcant information: the
number of shares and prices at which orders are awaiting execution. In other words, they
know the supply and demand for a particular stock and to some degree, can predict the
direction in which the price is headed.56 As a result, NYSE dealers and specialists know the
true state of the market with greater certainty and therefore suffer less ﬂom price risk and
inventory carrying costs. The presence of hidden limit orders on the NYSE serves to

strengthen this information advantage. Specialists need not reveal these orders because they

 

56The New York Institute of Finance (1992), p. 82.

41

do not necessarily have to display their best quote, but only a representative quote.57

Non-NYSE exchanges are modeled as a fringe competitor because their small market
share places them at an information disadvantage, making them price takers the majority of
the time. As noted, they will often ﬂee-ride ﬂom the NYSE price discovery process.58
Regionals are occasionally forced to submit orders to the NYSE to discover true prices. In
addition, the ﬂinge also competes on non-price terms, occasionally paying for order ﬂow or
providing quicker execution if the NYSE is active.59

Within the context of this model, the magnitude of NYSE information and location
advantages change throughout the day based on the volume of trading. As volume increases
and more information becomes available, parties that choose to transact on the ﬂinge reduce
their price risk because heavier volume reduces the time between trade observations.
Specialists on the ﬂinge also reduce their inventory carrying costs as market volume

increases.60 Finally, queuing on the NYSE may increase the attractiveness of ﬂinge trading

 

57McInish and Wood (1995) found that NYSE specialists failed to display about one-half
of all limit orders that were inside existing quotes for a sample of 118 NYSE-listed stocks.

58Garbade and Silber (1979) were the ﬁrst to characterize the NYSE as a dominant
competitor, with the regional and Nasdaq dealers offering satellite competition. They found
that the satellites provided price leadership only 10-20 percent of the time. McInish and Wood
(1992) found that the satellites did not meaningfully contribute to price discovery. Lee (1993)
noted that NYSE specialists may widen spreads (and increase inside-the-spread trading on the
ﬂoor) for strategic reasons; to deter nonprimary exchanges from ﬂee riding on the NYSE price
discovery mechanism.

59For this analysis, it is assumed that all non-NYSE exchanges operate under similar
information disadvantages. Hence, prices are only compared between exchanges in the two
information states.

60Inventory considerations such as these are not trivial. Most models specify inventory
carrying and asymmetric information costs as the two major determinants of the spread. Mann
and Seijas (1991) noted that "(d)ealers argue that it is not the cost of carrying inventory, but the
risk of carrying it" that determines spreads. They seemed to worry almost exclusively about
losing money by a markdown in the value of their portfolios.

42

because it potentially improves the quality (speed) of these trades. Therefore, when trading
is relatively heavy, the ﬂinge marginal cost/supply curve shifts down as information and
other general advantages of the dominant exchange are diminished. This causes the market
share of the dominant exchange to decline as competition increases. Conversely, at lower
volume levels, both price risk and inventory carrying cost risk increase, shifting the fringe
supply curve up and increasing the market share of the dominant exchange.

Because the level of trade activity changes over the trading day, the ﬂinge marginal
cost/supply curve will also not remain constant: it will shiﬁ depending on the trade
volurne.‘51 This simple modeling points to a weakness of the daily variables used in earlier
studies: a daily competition measure cannot isolate shifts in the competitive position of the
fringe that occur throughout the day. To capture the changing competitive position of the
ﬂinge given the volume of trading, this paper reduces time ﬂames used in liquidity premium
regressions ﬂom daily to half-day to hourly. Assuming competition has a negative effect on
liquidity premiums, the improved measurement resulting ﬂom smaller time ﬂames should
strengthen this effect.

A similar argument is made regarding the size of trades. The ﬂinge is in a better
position to compete for smaller trades because these trades tend to contain less information
and have a smaller impact on inventories. Because trades of different magnitude cannot be
considered a homogeneous product, they are separated into ﬁve size classes when comparing

liquidity premiums across exchanges.

 

61It is assumed that there is a positive, though decreasing, relationship between trade
volume and price information.

43

B. Data
Data used in this analysis are ﬂom the Institute for the Study of Securities Markets
(ISSM) for the entire 1992 calendar year. Each trade and quote that appears in the
transaction database is time stamped to the nearest second with the originating exchange
identiﬁed. In addition, each trade entry speciﬁes the transaction price and the size of the
trade. Due to the immense size of the transaction database, a sample of forty-two stocks was
selected. This sample was not selected at random. Given the relatively small size of the
sample, a random sample may have resulted in too many "small" or "large" stocks, based on
volume and overall price levels. Stocks were purposely selected to represent a broad
spectrum of prices, daily volume levels and industry classiﬁcations, though there is certainly
no reason to suspect that industry type would have meaningful implications for this analysis.
Table 3 lists the stocks included in the sample (along with a number of other pertinent
statistics) ﬂom smallest to largest based on the number of trades for the year. It also
provides some insight regarding the size of the individual stock databases. As shown, there
is a wide range of values. The second column of Table 3 lists the total number of trades and
quotes for the entire year for each stock on all exchanges. The third column lists the number
of "clean" trades that remained after quotes and cancelled or problem trades were removed.
Cancelled trades are ﬂagged in the original transaction database with a "-9999" in the price
ﬁeld while problem trades (inaccurate price transmitted) are ﬂagged with a "0." The fourth
column lists the number of cancelled or problem trades. Except for Reebok, these records

comprise a relatively small portion of the total number of records in the database.

OOQO‘UIAUJN—

hbbwwwwwwwwwwwNNNNNNNNN—————~r—~—~
N—‘OONQQM-ﬁ-WN—‘O‘OOOQO‘MAUJN—OOOOQOsUI-th—OO

Slack
Honda
Nashua
Geico

Ionics
Bradlees
KLM
ChrisCraft
Quik & Reily
MGM
Johnson
Goodrich
Skyway
Safeway
Mattel
Kroger

LA Gear
Ofﬁce Depot
Rockwell
Dayton
Xerox
Quaker Oats
Ameritech
Nynex
Caterpillar
Kellogg
General Mills
Wendy's
Reebok
Gillette
Amoco

Toys R Us
Dupont
Sears
McDonald's
KMart
General Electric
Ford
Homedepot
Walrnart
GTE

Disney
General Motors
Total

44
Table 2

1992 Sample

Total Number

of Trades
ammm
10,666
1 1,077
10,599
14,198
15,406
25,459
17,247
20,502
20,943
37,266
53,415
54,502
54,502
73,237
75,982
92,729
84,847
91,577
101,5 1 5
125,81 8
97,981
96,942
103,293
123,207
132,315
124,807
128,553
133,255
196,330
222,791
219,993
279,486
244,066
25 1,502
260,622
443,797
461,909
454,334
449,961
452,280
433,471

5.12.133
6,814,520

Number of

Clean

Trails:

5,298
5,393
5,817
7,722
8,735
9,015
9,299
10,024

1 1,392
18,982
24,757
3 1,000
3 1,000
38,336
45,03 1
46,609
47,3 13
48,034
48,675
52,920
53,705
61,863
63,15 1
65,363
65,683
69,069
82,501
98,842
1 10,466
1 13,599
129,996
144,489
146,135
156,800
168,065
260,507
295,940
307,704
322,350
333,948
337,206
322.663

4,285,402

Number of
Deleted

Records

12

14

23

9

1 8

96

8

29

14

20

27
650
650
814
822
3,865
869
744
922
1,321
1,094
2,959
2,283
1,022
809
1,746
1,681
17,729
1,725
2,102
2,527
1,660
1,458
1,741
2,3 16
2,308
2,292
4,944
2,003
l ,599
3,371

72,840

45

Occasionally, data entry errors will occur in the database; these are unavoidable.
The IS SM runs a number of screening procedures to catch these errors by comparing suspect
trades to trades in the immediate time vicinity. Yet, even these procedures cannot identify
all errors. Because some errors remained, additional checks were performed when
calculating price differences. Liquidity premium differences greater than $0.75 were
individually checked by comparing the price for that trade to comparable surrounding trades.
If convincing evidence of a data entry error or stale quote existed, the observation was
removed.62 In addition, the program which calculated liquidity premium differences
automatically removed individual liquidity premium differences greater than $2. It is argued
that differences this large cannot exist and need not be checked.‘53

Prior to the presentation of results, it is instructive to describe how the original data
were manipulated in order to derive price differentials and data sets used in regressions. This
allows the reader to conceptualize the ﬁnal data sets used in the analysis. Figure 1 illustrates
how each of the 42 separate stock data sets were split for liquidity premium difference
calculations. The data set for General Motors is used as an example. For 1992, there were
512, 138 total trades and quotes for this stock. After removing cancelled and problem trades
and regional quotes, a ﬁeld was created representing the continuously updated outstanding

quote on the NYSE. The total data set was then split between the six non-NYSE exchanges.

 

62Perhaps the most commonplace error (though not directly attributable to data entry
error) was the existence of " stale " quotes in the data set. If fresh quotes are not immediately
listed for a new day, then the program calculates liquidity premiums using the previous day‘s
quote if a trade is matched. If the bid and ask, and hence the transaction price change, then the
liquidity premium will be miscalculated: it will appear larger than it really is. The screening
procedure was able to remove these instances.

63Lee also included a $2 ﬁlter.

46

For example, the split for the Boston exchange contains all Boston and NYSE trades and the
outstanding NYSE quote ﬁeld.

Next, for each of the six smaller exchange data sets, the data were separated again
into ﬁve size classes, which contain all regional and NYSE trades for that size category and
the outstanding NYSE quote ﬁeld. This results in 30 mini-data sets for each stock. At this
point, regional trades were matched to NYSE trades using the appropriate time ﬂame.
Liquidity premiums were then calculated for the NYSE or regional trades and a difference
was taken. This resulted in a total of 1,260 possible data sets (30 times 42) of price
differences for the entire sample. Finally, these data sets were then aggregated across the 42
stocks based on size class and exchange so that 30 data sets remained. A similar process
was used for regression data sets, although one less step was required because there was no
breakdown based on size class.

C. Replication of Empirical Results: Liquidity Premium Difference Calculations

To analyze liquidity premium differentials between exchanges, regional trades were
matched to their NYSE counterparts (i.e., the same size class) if the NYSE trade occurred
one minute before or one minute after the regional trade.“ Liquidity premiums were then
derived for these matched trades. Liquidity premium differences were calculated as the
difference between the regional and NYSE liquidity premiums. If more than one NYSE
trade was matched to a regional trade, then an average was taken in order to derive the

liquidity premium on the NYSE.

 

6“Five minute and one minute time frames were also used: this did not alter the
general results and generally increased or decreased the number of matched trades. This is
consistent with Lee (1993).

47

//
Neﬁppéé
25m A

83.- 88 /
82.. 821.11... £88

com- 8m
sum 88 Bob om \ 8...- 8_\ // //

awn—U oﬂm \\ l \\O<Dm<Z / wage:

was owﬁﬁxm fl: 3.350

8 88m 8.8% W 28322“

883. 88m / \\
can .2 858 //
/ waﬂgmoﬁmu

//
|V\ cocmom

5:22:52 «:5
ﬂ charm

48

As an illustration, consider Figure 2. Assume that a trade occurs on a regional
exchange at 9:31 at the ask price with an outstanding quote of $30 and $30.25. The regional
trade (R1) would be matched to NYSE trades N1, N2 and N3. The liquidity premium
difference (regional - NYSE) would equal (12.5 - (12.5 + 0 + 12.5)/3) or 12.5 - 8.33 for a
difference of 4.166 cents per share in favor of the NYSE. Assume further that the
outstanding quote shifts a short time later at 9:32:30 to $30.375 and $30125. The second
regional trade (R2) would be matched to NYSE trades N4 and NS. The liquidity premium
difference would be calculated as (0 - (12.5 +12.5)/2) or -12.5 cents per share, which favors
the regional exchange. NYSE trade N6 would not be matched.65

Table 4 reports annual average liquidity premium differences for time matched trades
in 1992. Positive (negative) values indicate that the regional exchange paid a higher (lower)
average liquidity premium for that trade size class. All values are expressed in cents per
share. An asterisk denotes statistical signiﬁcance using a one-percent conﬁdence interval
(i.e., it rejects the null hypothesis that the mean liquidity premium differential is equal to
zero) while two asterisks denote statistical signiﬁcance using a ﬁve-percent conﬁdence

interval.66

 

65Liquidity premiums for all exchanges use the outstanding mid-spread on the NYSE as
the equilibrium price because regional quotes are oﬁen unreliable. In addition, if
regional/Nasdaq spreads had been used, it may have produced inaccurate results because the
midpoint of the spread may not represent the true equilibrium price.

“Most of the price differences found in Table 4 were found to be signiﬁcant. However,
this result may be driven by the large number of observations used in the analysis.

49

:90 omU

a

 

 

 

e2 Nu

 

r .13., ..r...-il..-..i! . 1:1 :imiﬂ.ﬁim°mwnvmmv

pagans—3.
825 .5 Z

 

 

mnwdmm

--... _ 83mm

  

:dqcmwwdag
madame—me.

.3 msdmm

I cmmdmm

1 mnmdmw

 

zeta—5.50 coconuts 53.5.5 5:555

N 953m

ﬁ oomdmm

SJBIIOG

50

Using Table 4, a number of trends become apparent. First, matched regional trades
occurred disproportionately in the smallest trade size class. Execution results for this group
were mixed: three exchanges offered worse average execution while the other three offered
better average execution compared to the NYSE. Moving ﬂom the smallest to largest trade
size class, liquidity premium differences tend to exhibit a U-shape, decreasing at ﬁrst ﬂom
their original value and increasing when the 1,000 - 1,900 size class is reached: this holds
for all exchanges. Except for the Philadelphia exchange, all non-NYSE exchanges offered
better average execution for trades between 500 - 900 and 1,000 - 1,900 shares. Only in the
largest trade size class did all non-NYSE exchanges offer noticeably poorer execution than
the NYSE.

In order to understand the economic signiﬁcance of these statistics, consider the price
difference for trades on the Paciﬁc exchange for the 100 - 400 size class. The average
calculated price difference is 0.61 cents in favor of the NYSE. For this sample of stocks,
160,969 Paciﬁc trades were matched to at least one similar NYSE trade. If the average
volume of these matched trades is approximately 200 shares (more trades occur at the
smallest size of 100 shares), then the total trade volume on the Paciﬁc for these matched
trades is equal to 32,193,800 shares. Given the average liquidity premium difference, similar
trades that occurred on the Paciﬁc exchange for this sample in 1992 paid a total of $19.6
million (32.1 million shares times 0.61 cents) in excess liquidity premiums compared to
similar NYSE trades. For 1992, approximately 2,400 stocks were listed on the NYSE, most
(approximately 90 percent) were dually listed on the non-NYSE exchanges as well. If stocks
not included in the sample are similar to those included in the sample, then the total cost to

execute off-board for this size class is much greater.

51

 

 

.3283 359286 .aoueoqéé a $3: 3505:»3 30:32am museum 3..
.3385 852.38 E33908 a man: connoﬁewmm 305333 88:60 _._

 

 

 

Seed 5a .. EN: 3.6.— .. $35 9:? $3. 23. a 82.6 maﬁa”
2.2.m Gs; _._ :85 cm.» 3.. mid- 35.8 a 33.6. need.“ .. $26- coin—m
wound :k. _. Need Ema 2.. Game. 332 a 386. SEN _. 83.: acade—
wvzl whim 8.. «3.? coin _. mmﬂl- $5.3 .. 83.7 8%: _._ .826 ”26.2
amend 2m; 8.. 336. main ._ ehmvﬁ- mat: a Fund- maﬁa; 386- v2.3
2:: em... _. emote- 39m .. cal- 3&6 _. mNQT wad .. 33.7 «8.3
own—03x 3:23.: owﬂo>< 3:23.: owﬁo>< 3:3wom owﬂo>< 3aomwom ems—03‘ 3.830%
3.332 3332 3:032 @2332 3:332
33A 25.125" 83.8.: 25.2% 8.. - 8—
320 03m

«:3 .220 05m .3 869:. 62—832 2:3.
connecoua SEE—o...— bm—zav: ou§o><

v ozah

22835
332
858..
.8322
3:585
830m

gm

52

D. Replication of Empirical Results: The Eﬂ'ect of Competition on Average Daily

Spreads and Liquidity Premiums

To test competition's effect on the price of liquidity services, simple ordinary least
squares regressions were performed using the NYSE spread or average liquidity premium
for a given exchange as the dependent variable. Competition was proxied using two
variables. The ﬁrst is the HHI measure reviewed earlier. It is equal to the sum of the daily
NYSE market share (of total volume) squared plus the daily non-NYSE share squared.‘57 The
second competition variable (ACTIVEZ) is a simple count of the number of exchanges that
were active during the trading day. In order to be considered active, an exchange had to
trade at least 1,000 shares of the given stock during the trading day.

Five additional explanatory variables were also used. These include:

1. The log of the average price on a given exchange

2. The log of the average size trade on a given exchange

3. The log of daily trade volume executed on a given exchange

4. Price volatility on a given exchange ((maximum price - minimum price) /
average price)

5. Block volume (trades greater than 10,000 shares) on a given exchange

Log transformations were used because the explanatory variables were hypothesized to have
a decreasing impact on the dependent variable as they increase in value. The volatility and

block volume (information) variables did not use a log transformation.68

 

“Regressions were also performed using (1) the sum of each individual exchange's
market share squared and (2) the sum of NYSE, Nasdaq and regional exchange shares squared.
These alternative measures did not alter the general results.

68Typically, the volume variable used a log transformation in earlier studies while the
average price variable occasionally used a log transformation. McInish and Wood (1992) used a
square root transformation for the average size variable. Most studies did not transform the
volatility or information variable (block trading). Regional regressions using log transformations
of these variables did not alter the general results, although NYSE regressions (using the
liquidity premium as the dependent variable) had a noticeably better ﬁt.

53

Separate regressions were performed using each competition measure (ACTIVEZ and
HHI) with the same set of explanatory variables. For both sets of regressions, the general
results for the non-competition explanatory variables were quite similar. Therefore, in order
to conserve space, complete results are detailed only for the unshaded competition variable
in Table 5. The unshaded competition variable represents the "preferred" competition
measure for a given exchange (the coefﬁcient had the expected sign and attained greater
statistical signiﬁcance). Non-competition coefﬁcients, R-squared and Durbin-Watson
statistics associated with the regressions using the shaded competition variable are not listed.

The dependent variable for the ﬁrst regression in Table 5 is the log of the average
daily NYSE spread (measured in cents). Regressions that follow use the log of the average
daily liquidity premium on the NYSE or regional exchange as the dependent variable.
Coefﬁcient values are listed while t-statistics appear in parenthesis below the coefﬁcient
values.

Results using the average NYSE spread as the dependent variable were very
comparable to earlier studies. All coefﬁcients had the expected sign and were statistically
signiﬁcant using a one-percent conﬁdence interval. Average price, average size, price
volatility, and the number of large trades (information) increased average spreads. The
volume coefﬁcient was negative and highly signiﬁcant: both competition coefﬁcients had
considerable signiﬁcance and the expected sign.

The results changed somewhat using an average liquidity premium measure as the
dependent variable. In particular, the average price and volume coefﬁcients, though
maintaining their signiﬁcance, lost much of it. The average size coefﬁcient became negative.

Presumably, larger trades were able to beneﬁt from discounting, which was captured when

54

mmﬁd

vow—d

ham—.0

momd

$.06

«26

2 vmmd

53mg»

«55%
33:2

2..—

2..—

3..—

cc. _

$.—

hm.—

v.—

Sad

gag

$325

anal
”36.

A263
886.

$33
:86.

892-2
$3.?

3.2.3
:85.

$3.3
28....

e3:

...-..~w.~...w.3w
398°

-

x

s

_ .

:39.

38.0...

ﬁat...
harmed

A533.
a $36».
.233...
23.3
mmﬂd.
32.”
V Emacs...

h.

39.2
ES

:3:
2.5.

a

9%.:
3%

32.2
886

22.2
3%

Gene
2a":

98.2
385

98.2
886

83:
3mm

33 :
2.2a

«EH—6N
zoo—m

 
 

503.83.». S: 203 032:; neurons—8 392% m5? maommmocwom 6qu

63.:
885

$85
2 :3

$8.2
2.25

85.8
885

82.9
385

$32
$85

83.2
$85

38.2
885

.02...—

893
286.

$35
685

3.2a
386.

62.3
83.?

3.8.3
38.?

833
$8.?

as: E
«25.?

$33
22.?

a

own-395— we..—

nnemnmouwom 2.3.:— oEC. 3.5

m 033—.

5.3
385.

GB 3
12 3-

ASN.8
$85

83.3
83.?

misc
38.?

33.2
885

a§l
826.

RS s
285

«an

sis...

813
£85

5.3
$25

832
$2.:

€22
«33

22.:
m2:

3%.:
$25

332
825

£38
~22

HEN

.u><

8355 e:
«2.2.03.2; .mE

825$ .5
050mm .w><

SEEP—m .34
.3332 .w><

SE5 .5
H3322 .w><

8382a as
$85050 .m><

SE05 .5
aoamom .w><

83805 .34
mm; at...

82%
mm; .92

«3:3
«cannons

55

the liquidity premium measure was used. Both competition coefﬁcients maintained their
appropriate signs, though their signiﬁcance was reduced slightly.

Moving to the regional regressions, the equations generally had greater difﬁculty
explaining average liquidity premiums (as measured by the R-squared statistic), particularly
the Cincinnati exchange. One interesting result was the superior explanatory power of the
ACTIVE2 competition measure compared to HHI. For all non-NYSE exchanges, the
ACTIVE2 coefﬁcient had the anticipated sign and was highly signiﬁcant. Conversely, ﬁve
of the regional regressions derived an unanticipated sign for the HHI coefﬁcient.69

Except for the Paciﬁc exchange, the regional volume coefﬁcient always had the
anticipated sign and in three of the regressions, it was statistically signiﬁcant using a one-
percent conﬁdence interval. Regional price volatility, average price and block volume
coefﬁcients always had the anticipated sign and were usually signiﬁcant. The results for the
regional average size variable were mixed.

E. Summary of Replication Results

Based on results in this chapter, it appears that data from 1992 conﬁrm results from
earlier studies. Statistically signiﬁcant price differences were still present. In general, the
results were quite comparable to Lee, though many average liquidity premium differences

were larger in this study. This may be attributable to the smaller size of the sample. The

 

69Additional regressions were performed using both competition measures
simultaneously. This did not alter the general results; the regional competition coefﬁcients were
very similar to regressions using the measures separately. However, an F-test revealed that ﬁve
of the six regional regressions had a statistically better ﬁt using both competition variables in the
regressions, compared to using only the ACTIVE2 variable. This result appears to be driven by
the large number of observations; the differences in the sum of squared error terms was
negligible. However, there was a much larger difference between these two terms comparing
regressions using only HHI to those using both competition variables.

56

major difference between the two sets of results was that (1) data from 1992 did not ﬁnd
worse average execution for all non-NYSE exchanges for the smallest trade size class and
(2) all non-NYSE exchanges paid much higher liquidity premiums for large trades.

Competition was found to signiﬁcantly reduce NYSE liquidity prices, whether they
were measured using a spread or liquidity premium measure. In general, most regression
coefﬁcients had the expected signs. However, regressions had a considerably more difﬁcult
time explaining average liquidity premiums, presumably due to unpredictable discounting
behavior as the liquidity premium measure should essentially equal half the spread unless
discounting was involved. Regional regressions were similar to the NYSE regression,
though generally the variables had less explanatory power.

Using the average daily spread as the dependent variable, NYSE volume and
competition elasticities were quite similar to earlier studies. The NYSE volume elasticity
was equal to -.1513 while the HHI competition elasticity was equal to .41 (using the average
value of .77) and the ACTIVE2 elasticity was equal to approximately -.01 (average value of
5.02). A ten percent increase in the HHI measure will increase the dependent variable by 4.1
percent, or 0.87 cents, given the average NYSE spread of 21.2 cents. If two less exchanges
are active on average, it will increase the dependent variable by .4 percent, or .08 cents.70
For the regional regressions, the elasticity of ACTIVE2 was approximately equal to -.01.

The dominant exchange and fringe competitors seem to have different views of

competition based on the superior explanatory power of the ACTIVE2 competition measure

 

”These results are consistent with Hamilton (1979). Hamilton, using the number of
marketmakers trading the stock as a measure of competition, derived net competition elasticities
between -.02 and -.04 and a volume elasticity of -.07. Branch and Freed (1977), derived a
competition elasticity of -. l 7 (using the number of exchanges that trade the stock) and a volume
elasticity of -.02 using the relative spread as a dependent variable.

57

in the fringe regressions. Apparently, NYSE market share or overall volume dispersion does
not matter to the regionals, only that some or most of them are active. The two measures
most likely capture different phenomena. Compared to ACTIVE2, the HHI variable may not
capture the queuing and free-riding process that occurs when the market for a stock is active,
which serves to increase non-NYSE activity and "competition." Queuing and free-riding are
most likely associated with tight spreads and lower liquidity premitnns, hence the
signiﬁcance of the ACTIVE2 coefﬁcient on reducing local liquidity premiums. This ﬁnding
is consistent with the view of these exchanges as a competitive fringe. Presumably, if one
or two regional exchanges ﬁnds it attractive to compete, free-ride or beneﬁt from queuing
on the NYSE, most do.

The volume variable had noticeably less explanatory power in the regional
regressions compared to the NYSE. It is possible that sporadic regional trade activity
increases the correlation between the volume and competition measures, reducing the

explanatory power of both variables. This possibility is examined in the next chapter.

CHAPTER 5
PRICE DISPERSION IN THE SECURITIES INDUSTRY

Chapter 4 established a number of results regarding the determinants of spreads
and/or liquidity premiums in the securities industry. One of the stronger results was that
competition had a similar and signiﬁcant impact on liquidity premiums across all exchanges
(as measured by the sign, signiﬁcance and size of the ACTIVE2 competition coefﬁcient).
Most other explanatory variables also had similar impacts on liquidity premiums across
exchanges. This result is noteworthy given the large disparity in trade volume between the
exchanges. In general, it seems to imply that the exchanges are integrated to a substantial
degree.

Yet, this implication contradicts another result from Chapter 4: statistically and
economically signiﬁcant price differences. As noted in Chapter 3, in an environment where
competition seems to constrain or reduce prices on individual exchanges; competition should
also serve to reduce price differentials between exchanges. In the case of the securities
industry, price differences should in fact approach zero because technology largely
eliminates transaction or transport costs between exchanges.

Chapter 5 continues to examine the coexistence of these two results now that the
empirics from earlier studies have been conﬁrmed. The relevant issue now becomes the
proper interpretation of the empirical results. Should price dispersion be interpreted as an

indictment of the Law of One Price, or is another explanation responsible, such as a failure

58

59

to deﬁne the market properly?71 Using this framework, the nature of competition within the
securities industry is explored.
A. The Law of One Price

Researchers often use two standard assumptions to characterize industry conduct and
performance: competitive markets and the Law of One Price. In most cases, these
assumptions are not unrealistic and simplify the task of studying industry dynamics. The
competitive market assumption states that individual ﬁrms (exchanges) do not have power
over price, but must take the market price as given. The Law of One Price states that for any
product, there can exist only one market price at any point in time. Although the two
assumptions are technically independent, it is assumed for the purposes of this analysis that
a competitive market is a sufficient condition for the Law of One Price to hold in the
securities industry. That is, price dispersion can exist only if the industry is not
competitive.”

The structure of an industry determines whether or not it is competitive and can be
characterized as a single market. Speciﬁcally, four structural characteristics are necessary
to allow competition to exist: (1) a homogeneous product; (2) low transaction costs; (3)
many buyers and sellers; and (4) complete information. If the industry encompasses these
four structural characteristics, then the conditions within the industry are such that it should

be competitive and the Law of One Price can take effect. Conversely, if any one structural

 

“For the analysis in this chapter, it is implicitly assumed that the individual exchanges
comprise a single market. Following Stigler and Sherwin (1985), the market for the good is
deﬁned as "the area within which the price of a good tends to uniformity, allowance being made
for transportation costs."

72In the context of this analysis, competition implies that activity on other exchanges
effectively reduces liquidity premiums and liquidity premium differentials between exchanges.

6O

characteristic does not hold, then the industry need not be competitive, and price dispersion
may result.

The ﬁrst two structural characteristics are assumed to hold in this analysis. The
quality of matched trades was held constant using a two-minute time frame and size
breakdowns. Transaction (and transportation, assumed negligible) costs are more
troublesome because they are unobservable. There may be costs associated with the decision
to transact on exchanges or with other parties with whom the broker does not normally deal,
more generally a principal-agent phenomenon. Dealers or brokers may have reciprocal
relationships with counterparts on a particular exchange, hence they may not actively search
for best execution.

Unfortunately, there is no acceptable method to measure or address these costs,
assuming they exist in some meaningful form. However, if these costs do exist, they are
more likely to have a greater impact on smaller exchanges: these exchanges have thinner
markets, tend to deal primarily in smaller trades (hence the cost of paying higher liquidity
premiums due to transaction costs for any one trade is smaller) and most likely transact on
behalf of less sophisticated parties compared to the NYSE. Based on the empirical results
from the previous section, it does not appear that these costs systematically affect smaller
off-board trades. Therefore, transaction costs are regarded as negligible and price differences
should approach zero if competition is effective in this industry.

B. Structural Characteristics Three and Four: Complete Information and Many

Buyers and Sellers; The Local Market Hypothesis

The simple model used earlier depicted the securities industry as being led by a

dominant exchange (NYSE) followed by fringe competitors. Clearly, complete information

61

does not exist within the context of this model. Given this information asymmetry, a
pertinent question is how much asymmetry exists. Is there enough information asymmetry,
or enough of an information lag between exchanges, that non-NYSE exchanges can be
characterized as separate local markets? Alternatively, does the ITS ensure adequate
integration?

The local market hypothesis proceeds in two separate, but related steps. The ﬁrst step
uses regressions to explain average liquidity premiums on the individual exchanges. Time
frames are shortened ﬂom daily to half-day to hourly averages. The purpose of these
regressions is to (1) establish the effect that shorter time ﬂames have on coefﬁcient values,
(2) provide a benchmark for local market regressions and (3) provide a more consistent basis
(in terms of time ﬂames) for comparisons between price difference results and the effect
competition has on individual exchange liquidity premiums. It is hypothesized that reducing
time ﬂames will increase the explanatory power of the volume and competition variables,
particularly for the regional exchanges. This should occur because shorter time ﬂames (1)
reduce the correlation between these two variables, (2) allow the analysis to more accurately
identify when activity actually occurs on non-NYSE exchanges (their competitive position
may change during the day based on market activity) and (3) limit any bias in the
competition coefﬁcient attributable to non-price competition from non-NYSE exchanges.

The second step tests the formal local market hypothesis. Regressions are used once
more to explain average exchange liquidity premiums, but now a set of fringe and dominant
exchange explanatory variables are included simultaneously. Given the modeling of the
NYSE as a dominant exchange, prices on the fringe should be responsive to information and

activity at this market center. If they are not, then this provides evidence that these

62

exchanges are more properly viewed as local markets.”

A few notes regarding the methodology are in order. The competition measure
(ACTIVE2) again includes threshold values. For the daily time frame, an exchange had to
trade at least 1,000 shares in a given stock to be considered active. For the half-day time
frame the threshold was 400 shares; for hourly it was 200 shares.74 In the ﬁrst set of
regressions which establish trends in coefficients moving to smaller time ﬂames, competition
coefﬁcients appear shaded or unshaded following the convention used in the previous
chapter.

For local market regressions using two sets of explanatory variables, only the average
price on the regional was included due to the high correlation between average prices on
regional exchanges and the NYSE (a correlation of one was not uncommon). Also, these
regressions include only the ACTIVE2 competition measure because this measure seems to
have superior explanatory value (compared to the HHI measure) regarding regional liquidity
premiums. Finally, due to space constraints, Durbin-Watson statistics were not included for

either set of regressions: values typically ranged between 1.8 and 1.9.75

 

"Unfortunately, the local market hypothesis cannot be tested rigorously. Claiming that
regional conditions have a greater impact on local liquidity premiums involves a certain amount
of discretion. This is true for any analysis that attempts to measure the extent of a market.
Regarding the deﬁnition of a single market, Stigler and Sherwin (1985, p. 562) noted that "no
unique criterion exists. . . . (m)arkets can show every level of interdependence from absolute
homogeneity to complete independence." Therefore, the classiﬁcation of the separate exchanges
as a single market necessarily involves the use of some discretionary judgement.

7"The smallest possible trade is equal to 100 shares. For an exchange to be considered
"active," threshold values force exchanges to trade at an average hourly rate slightly greater than
the smallest trade. Hence, the daily threshold level, which encompasses seven hourly time
periods, is set at 1,000 and the half-day threshold level, which encompasses three to four hourly
time periods, is set at 400.

"It is possible that reducing time frames will introduce serial correlation, in turn
reducing the signiﬁcance of coefficients. This may occur if large or small values of a particular

63

1. The Effect of Shorter Time Frames

Tables 6a-g list results ﬂom the ﬁrst set of regressions. (Daily results are listed again
for comparative purposes.) T-statistics appear in parenthesis below coefﬁcient values. Table
6a lists results for NYSE regressions using both an average spread and liquidity premium
measure as the dependent variable. In general, all coefficients had the expected signs and
most did not change appreciably moving to smaller time ﬂames: all coefficients were
statistically signiﬁcant using a one-percent conﬁdence interval no matter the time ﬂame
used. However, both competition coefﬁcients tended to increase in size and signiﬁcance
moving to an hourly time ﬂame when the NYSE liquidity premium was used as a dependent
variable.

Tables 6b-g lists results for regional regressions. Many of the regional regression
results were similar to the NYSE. Again, most coefﬁcients had the expected signs. Most
regional average price, price volatility and block trade coefﬁcients remained positive and

statistically signiﬁcant moving to smaller time ﬂames.

 

variable (e.g., volume) are not independent of values in time periods that precede or follow it.
This analysis did not ﬁnd a problem with serial correlation moving to smaller time ﬂames.

64

836

Sand

womﬁo

bowed

:VNNd

hvowd

33.3
”23.

$63
88.0.

333
35.0-

R33
$3.0.

$3.3
$3.?

$23
38.?

51.8
$25

3.5
325

93:
”2.3

335
:32

$3.8
Ewe

cow:
MES

gag

833.2

92.8 23.3 5.3 933
885 £85 023. a: .o-
93: 88.8 $33 $23
3mm :85 S _ 3- 885
2:2..— oﬁﬁ. 3.5::
9:.va 93.3 23.3 SEE
3mm :86 83.? 385.
5.3 3.3 $33 an. 5
8-2 886 ”3.3- :86
«Each «BE. man—Ea:
82.: 8.0.4.3 E: 3 EEE
3mm $86 $35. 8mg-
33 a $8.3 33.3 52 3
88¢ 88... 22.? 286
:8... mm; amt; .m>< mm; ”3 . 3. «3

2:8..— oEF 2:5

Essa 8:92?”— mqo mm;

.5 oEaH

Gods
886

823
$8.9

€39
Std

A393
Same

336
8:5

328
3de

Hana
. 3&3

.boﬁaaom :E can? 22%? couuanoQ woumnm mam: mcoﬁmocwom ”202

.62.. £233
owﬂo>< we we;

82am mo wou—

éem $233
0333.. .«0 we;

32mm we we;

.82: base:
owﬁo>< .8 we;

vacuum we we:

«33
Hegemon—

65

NNZd

mag
33?.

mo_.o

3
3.2.22

woﬁo

“3%
32:2

933
$8.?

a

:33
83.?

"a

$3.3
38.?

a

9%.:
$8.?

a

£53
88.?

a

$3.3
$2.?

a

50833,“ S: 203 using :oucanOo 33% m5? mcoﬁmewom ”802

23.3 8.5 833

386 286 mad-

3 Maﬁa; «Ea—aw.
zoo—m .uom .wom .u>< .uom we‘—

oEEh 2:5. basal

88$ 22$ 8:3

886 w Sod ambod-

nan—H dang—aw «Ea—q
:3:— .uo- .wom .u>< .95 as

oEEh 25h man—Ea:

5.3 333 82.3

386 $36 wooed-

33H g «53
zoo—m .wom .uom .u>< .91 m3

252% «SE. 2:5

3:53— =22”...me mac acumen

no 939—.

3.8V
$85

«madam
.w>< wed

3.8.8
:85

“Ended
.u>< wea—

93.5
83¢

«33
.u>< we:

82:
$35

«aw—Mug
.u>< wed

5.3
an;

avg
.95. Mo‘—

63.:
:55

g
.u>< wou—

éea £23:

owaco>< we:

and;
301.595

.205 £235

owﬁo>< we;

«Eda
«gov—.23:

.55 £23:

09203.. woA

«Eda
«nonunion—

66

weed

Hun—ﬁg
8333

owned

3
83.3.2

Sued

gamma
33:2

$2.3
35.?

«a

823
3.8.0-

3

32.3
Eco-

some
$8...-

a

SSE
at?

A323
32¢.

a

@no: Gone Codi

eoood mo-mm mahod-

as 353 man—Sm
x3:— .95 .9“ .m>< .uom MS

can; 08:. .3283

e no: 33. o 83.3

eoood Noood ammod-

mug g «33
:85 do: do: .u>< as. as

2:9:— 085. hand—a:

$8.3 $3.8 $6.3

voeod wooed omood-

maven—H gnu—Hm «ﬁg
:8:— ao: a3. a3. a3. ”3

059$ 08%. .225

3:53— :emmuéuwom mAO 52.5950

3 «Eu—NH

came
_ S3

nudged
.w>< as

98.3
$3.?

g
.w>< wou—

63.3
$8.?

a? ”3

2 Se
53

Hug
.u>< med

5.:
:85

.u.>< ”3

22.3
Mum—ed

g
.w>< as

52983,.” E: can; 03%.? 5:33:80 392% main mcommmﬂwom 682

.82; £2.55

owEo>< woo

mag
nae—Eugen

.85 $23:

$803.. we‘—

«SSH—3m
Eco—Soon

.55 £233

owﬁo>< wot—

«Eda
«noun—onen—

7
6

v:.e

wanna—“mum
8:33.

eve—d

gum
8333

demo

Bung
33:2

:23
23¢.

a

$33
88....

avid
88¢.

a

@mVéd
embed-

a

$3;
82.?

$3.3
£3.

a

89$
886

38H
:83 .3:

E Ed
885

ﬁns
:8:— .3:

33$
58...

"and“
x8:— .3—

£33— 5_...8._u§ So .8322

and: 32$

mvwod Seed-
dag «Ea—Sm
.wom .u>< .9: wow—

oEauh 25H. :23:

$3: Ede

mmeod bowed-
maﬁa—Sm «Eng
.wum .w>< .wum and

can...”— uEE. mane—am

83.: 38.3

oovod oovod-
ﬂag «Ea—aw
.wom .u>< .wom we:

082m 0:5. 2:5

.5 03:9

ASN.3
2.56

g
.w>< as

883
$8.?

«aided
.w>< as

28.3
25.?

«£3
.u>< «3

33$
325

g
.w>< Moa—

£23
Sumo

g3
.u>< now—

5.3
823

g
.w>< wou—

58888 S: 803 oBmtg 83288 3an maﬁa mcommmouwom “Bo Z

.505 92%:

oweo>< we;

«Hun—«N
duo—Eamon

.805 £23:

owﬂo>< we;

ﬂag
«consonan—

.a...& £23:

owﬁo>< woq

mag
one—Eons:

68

25.0

gamma
3.3.3

9890

Bung
33:2

80— .o

33...!

3:3
28¢.

a

a? E
285.

£33
:86-

Gmm6¢
28¢.

a

83.3
$3.?

a

$33
33¢.

a

58838 E: 203 033:2, costumes coca—a mam: mcommmohwom 682

8%. c 825 956$
2-9. 28.9 $2.?
as 9.3an «3:3»
:8:— QE .8: .u>< as. ”3
2.5.:— oEF .238:
28d 83.3 :23
game 28.: Ea?
as gnu—aw «ﬁg
:85 .3. .9: .u>< .9: «3
252,.— oEﬁ. man—.2“:

2 as $8.8 GEE
3mm 32 .o 385.
as gag—aw «sad»
:8:— ﬁi .9: a: a»: no:

252...— oéh man:

338% 568..»3— mac 3132

ow 059—.

33:
~33.

“Em ﬂum
.w>< as

32.8
83...

an
.u>< 3.

ASN.8
38d

«madam
.u>< ac..—

sis
885

g
.u>< as

603
as;

g
.w>< as

833
$25

g3
.93 «3

.805 £235

owﬁo>< mag

33
Eon—anon

.505 £2.53

owﬁo>< wag

wag
«uneconom—

éoa £2.65

0353. we;

«33
Eugene:

9
6

god

nag
c8322

mde

3%
821.2

aam _.o

13533—
33:3

33$
33.0.

3

GdeV
3.36.

833
$8.?

a

$8.3
«So

a

83.1
32.?

98.3
88¢.

a

A33
886

mag
:85 .uom

839
88...

33H
:83 .3—

End
885

38H
:85 as.

82.8
385

ﬂag—aw
.wum .u><

2:th 25H 2.3.5

83.3
88.:

Q3
.uom .w><

252m 25H. ham—ﬁn:

:86
M: :3

2:8,.“ 25H. :25

:m. _ 3
$8.?

«5.3

.uem wad

38$
Sad

33%

.95. ”3

$35
$86

8:53— ..Emmouwom mac 9525

.G «Eu—NH

an”:
: :3-

£43
.w>< aca—

83 3
2.8.?

.9? mod

333
8:5-

«madam
.u>< as

5.3
8»;

mum—3
. >< as

£30
836

g
.u>< wou—

$.03
vmmmd

«madam
.w>< wow—

.58833 :E 29$ oEmth consume—Bo uocmﬁ wEm: maoﬁmemom ”802

.52; £23:

owﬁo>< we;

a
Eco—5&5

.805 £23:

owgo>< we:

g
«31:32—

éE £2.53.

owﬁo>< we;

«3%
«co—Eamon

0
7

Shed

Edna
8.8.22

Rbd

885
8.25%

owed

gm
8.8.:2

83$
:88-

a

E: E
38.8.

a

83.3
«8.8-

38.3
38.8

Rams
88.8

a

:38
388

a:

928
88.8

33H
:8:— 9:

cams
_ 88.8

«was
:8:— a»:

32. 5
8-8

ﬂan—H
:8:— .8:

2.283% E: 983 053:? 8:23:80 6285 main mcoﬁmeoM ”082

:55 $3.28 €85
m ~ cod $86. 5 mod
.woﬂ .u>< .91 as .w>< wou—
oEauh 25h £28:
an a $8.3 33.8
386 386- 386
.91 .u>< .MS— as .u>< nod
08.2.5 985. mania—u
68. a a8. 3 c _ . 3
good good- good-
.wom .u>< .uom aca— .w>< as

2.8.:— oEF b8:

8.33: 8828: 38 38.82::

8 mean:

$33
9.2.8

3543:—
.u>< as

82.88
82.8

«an;
.u>< us

$38
883

3543
.95.. as

.58: £28:

09203: we:

«Eng
«not—.289

.82: £28:

owﬁo>< we;

«HE
286.8qu

.52.: £283

owﬁo>< mo:—

«as
82.—8&5

71

However, there were also some noticeable differences between the NYSE and
regional regressions. While the NYSE average trade size coefficient was negative and
signiﬁcant at all time ﬂames, most regional average trade size coefﬁcients were positive and
signiﬁcant (except the Paciﬁc), particularly at smaller time frames. This result may occur
because larger trades are able to beneﬁt ﬂom discounting on the NYSE, but must pay a
premium on the smaller exchanges where traders prefer to deal in smaller-sized trades that
contain less information. The shorter time frames are able to capture this discounting
behavior. This result in consistent with the price difference calculations from Table 4.

Perhaps most interesting was the behavior of the volume and competition
coefﬁcients. For all regional exchanges, the volume coefﬁcient became noticeably more
signiﬁcant and (slightly) larger at shorter time ﬂames (regardless of the competition measure
used), whereas the signiﬁcance of the NYSE volume coefﬁcient did not change appreciably
(using liquidity premiums as the dependent variable). The ACTIVE2 competition measure
maintained its superiority compared to an HHI measure moving to an hourly time frame. In
four of six regional regressions, the HHI coefﬁcient did not attain the expected sign using
an hourly time frame. Conversely, both competition variables had signiﬁcant explanatory
power and the expected sign in the NYSE regressions, no matter the time frame.76
2. Summary of Regressions Using Shorter Time Frames

Reducing time ﬂames to explain average liquidity premiums does not seem to affect

greatly NYSE coefﬁcients. In particular, the volume coefﬁcient did not change appreciably,

 

76Regressions were also performed using both competition measures simultaneously.
The competition coefficients did not change appreciably compared to the regressions where they
were used separately. F-tests revealed similar ﬁndings to those performed for the daily
regressions in Table 5.

72

although both competition coefﬁcients gained size and signiﬁcance (using the liquidity
premium as the dependent variable). However, for the regional regressions, shorter time
frames had a much stronger impact, especially for the volume and average trade size
coefﬁcients. For all regional exchanges, the volume coefﬁcient became much more
signiﬁcant using shorter time ﬂames (except Nasdaq), while the average trade size
coefﬁcient tended to become positive and more signiﬁcant. The superiority of the ACTIVE2
competition measure also became clear using hourly averages.

Given these expected results, it is argued that hourly time ﬂames improve the
accuracy of coefﬁcient estimates, particularly competition, regional trade volume (due to
sporadic trade activity) and regional average trade size coefficients. Though R2 statistics are
generally lower using hourly time ﬂames, only at these time ﬂames do coefﬁcients have
similar impacts on liquidity premiums across regional exchanges.

Based on these results, the ﬁrst explanation for statistically signiﬁcant price
differences in a competitive industry environment is dismissed. When time ﬂames were
shortened ﬂom daily to hourly values, the competition coefﬁcient gained signiﬁcance.
Hence, the coexistence of these two empirical results is not attributable to the different time
frames used to derive these results.

3. The Local Market Hypothesis

The local market hypothesis tests whether observed price dispersion can be attributed

to the pooling of data across several local markets. To test this hypothesis, two assumptions

are made. First, it is assumed that the dominance of the NYSE is great enough so that NYSE

73

cost and demand conditions can serve as a proxy for the entire industry.77 Second, it is
assumed that dealers on individual exchanges face downward sloping demand curves and
will adjust prices to reﬂect them.

To test the local market hypothesis, a set of regional and NYSE explanatory variables
are included simultaneously in regressions using local average liquidity premiums as the
dependent variable. The local market hypothesis will not be rejected if fringe coefficients
are statistically superior compared to their NYSE counterparts when explaining local
liquidity premiums. For this analysis, the sign, signiﬁcance and size (in order of importance)
of the respective coefﬁcients are used as the primary criteria to determine the extent of the
market. Conversely, the local market hypothesis will be rejected if NYSE explanatory
variables are as good or better predictors of local liquidity premiums. Tables 7a-f present
results ﬂom local market regressions. The discussion that follows is limited to hourly time
ﬂame results as these are considered to be the most pertinent.

The ﬁrst two variables in Tables 7a-f are average regional price and competition.
Both coefficients had the expected sign and were statistically signiﬁcant for all exchanges
using hourly time ﬂames. The average trade size coefﬁcients were positive and statistically
signiﬁcant for the regional (except for the Paciﬁc exchange) variables; the NYSE average
trade size coefﬁcient was negative and signiﬁcant. The regional volume coefficients were
negative and highly signiﬁcant for all exchanges; the NYSE volume coefficient was negative

and signiﬁcant in only one of the six regressions (Midwest).

 

77Obviously, this assumption may not hold in many instances. However, though
conditions may differ across exchanges, all that is required is that the differences are not great
enough to force a particular regional exchange to adjust its price. Given the discrete nature of
prices in this industry, this is not unrealistic.

655.34

338 and $38 2:5 88.3 80.33 82.3 623 $3.; 933 2:5:

53 3% 385 88d 22:. 98.? $2.9 $8.9 $8.0 $85 $8.? .wE M“3
gm mg .333

 

3.8.8... 825 v.85

 

553:5 owauo><

 

 

 

 

0823/ we..— onmm owe... owﬂo>< wou— owﬁo>< mod “coczomon
059E 2:? misc:
.Eoi .5
9.8.8 sad $2.: 33:3 AS _.8 :33 :35 :2 3 Sum: $8.3 3&3:
_:.o mo-mm wooed 886 m Sod wooed awed- owned. mnod 336 £86. .w>< we:
vogmsmu< move; x005 bags—.5 owEo>< oE=_o> mod onmm «we... owﬂo>< mo.— owﬂo>< mo; «covcaon
4
7
2:2,.— oEﬁ. mama—a:
.805 .5
9.8.3 c». a 38.: 3.33 82.8 $33 $3.5 $2.3 :3: 88.3 $8
won _ .o 8-1"; 886 good 386 vwood Nmbod- mf me. good _:.o m ﬂ 8.? .w>< mo.—
ﬁuﬁawum mg 353 ma E a g a 3.33 3543 a mag
woumamuxx mount. x85 bzuﬂo> owﬂo>< oE:_o> mod 85 092,—. owﬁo>< wou— owﬁo>< mod acovconon—

08.2..— o—EP Eun—

Ezmem each. «$.32 .aaemuom

ab «Bah

5
7

wowed

gamma
38.9...

good

3
8.8?

mowed

3
38.3.

 

 

 

 

 

 

52v and 838 83.8 9.93 823 $8.3 $3.3

mo-mw wooed wooed no-5. coed- mowed. mowed. mwwed

an 3 ma 333m ma Had—3 mg mg
895. x85 b=u£o> owﬁo>< 082o> we; oﬁm 03:. owﬂo>< wine‘—

oEauh 25h b.53—

§ we came dated $5.8 $8.3 A2 “.3 2.8.2 88$

3-me wooed Seed- Seed owwod owed- wewﬁo- efod.

mg 1833 a ﬂag a ﬂag a magnum
mount. :85 358:5 owﬂo>< oE:_o> mod BE 38,—. owﬁo>< mo..—

oEauh 25h. mane—a:

€33 32.5 25.8 33.9 :39 62.3 88.3 :33

mo-m_ wooed wooed wooed wwoed Seed- 5 _ ewd- ehmod.

Mada Ham a 333 Ma Hagan a ﬁg
mound. Joe—m base—3’ owﬁo>< oEBO> mod 05m 38,—. owﬁo>< wou—

oEanh 25H. Duan—

Eaaoso was .81“: .235:

as 2:»:—

ammo
£85

353%

owaeo>< mod

semi
285.

Egg

own—03.. meg

38$
£86-

gag

owﬂo>< mo.—

size
23....

$33
$8.?

823
$8.?

.805 .EA
.250:
.w>< we.—

wanna.»
Beaumon—

éem a:
banana
.w>< mo..—

aqua
«coocoeoQ

.805 .Eq
5.5
.w>< we;

ﬂag
“coda—omen

$3: 88.8 E _.8

 

 

 

 

.52.. 8:

$o.o8 $3.3 5.8-8 $33 $38 8§8 $33 88:

SS .e meme wooed meme wnwed _wS.e- mused- 33d- mowed vmtd owed- .w>< woq

335 $34 3:93 ma 3:33 maﬁa 333% a 3:35 i Na «Bug
3355. 83:. x85 b5§o> owﬁo>< 25:5 we.— 3% come... owmeo>< moi.— owEo>< woq use—Eamon

2:9...— 085. base:
.82.: .34
$25 23.8 $3.8 2:: $3.8 2a.. 3 $8.3 $3.8
mm» _ .e me-mw Seed wooed .

$38 $33 :88“:

 

 

 

 

 

 

$8 o 38.? 38.? £36- 3mm ”23 $8.? .w>< won
magi a ﬂag mg g ma mag mg 3:33 3543 a mg
6 332:? meow—H moo—m 353:5 owﬁo>< 25:5 wloq 05m 38,—. owﬁo>< we... owEo>< we..— Houseman
7
2:9:— oEﬁ. ham—Ea:
.85 a:
$3.8 $8.8 $8.3 $3.5 $2.3 $83 $3.3 9.2.8 :8. _ 8 $8.3 £8
Swwd 3A.: 3.9. come- Sved envod- ”Sod- m _ nod- wood Sbwd mhmed- .w>< woq
manna a ﬂag a tang ma magnum mg t E g Na «Eng
@8355. 8%... x85 553:5 owﬁo>< 082:5 we;

cam 03:. owﬁo>< mod owﬂo>< woq 303.030

2.5.:— 052. 5a:

.8382 .25 3:2 .22?—

ob 035.

7
7

336

3
83.32

boo—d

335
322

vow—d

gum
38?.

 

 

 

 

 

 

 

 

:3: $2.8 34. e 82$ oi. E :23 $2.3 83:

me-m~ we-mm wooed Seed good- wooed- mowed- mowed
a 33.3 mg 333% a ﬁg ma Had—Mum
move; “:55 bios—n5 owﬂo>< 083:5 wow“ 05m ova; owﬂo>< woA

oEEh 25H. 3.5::

and: $8. 5 an”. : 38$ 63.3 $33 $3 3 God:

mo.mN meknﬁ wooed o _ ood m med embed- me _.e. embed
a ﬂag a 333m ma 3 ma Hg
ween—H zoo—m b530> owﬂo>< 089:5 mo..— 05m 38% owﬂo>< woq

2:9...— uEﬁ. man—ﬁn:

88.3 a _ he @2. o 68.3 a :3 £33 2 2 3 RS:

27m: mo.m: meoed oeeed honed mowed- amid- NNood
ma 3:33 a ﬂag a 353m a End—mud
noun—H goo—m b:s£o> owﬂo>< 085:5 we: 05m ova; owﬂo>< mo:—

oﬁauh 08E. man:

563 Z J88 gov—.32 .2833—

:o oEaH

sans
83o

0.0.5.43
owﬁo>< woq

56:
28.0

g
owﬂo>< mo..—

$5.:
983

3543
owﬂo>< mo»—

.EoE .31—

aﬁi 3:5:
2 Se. .w>< we;
Na «:33»
“codename—
..85 .5
$3.3 $9.2“:
38.? .w>< woo
a and»
33530
éem .5
$8.5 38
88d. .32 we.“

a 3
289595

78

5.3 Ens

 

62d 33%

 

$05.34
$88 :33 A863 28.: 393 $33 3.8:

 

 

 

 

good 21% 886 good mowed named 83.? ”god- $8.0 $26 «$9? .w>< we:
3333‘ nova; x85 3:329» owﬁo>< oEEo> mod 0N5 oust. owﬁo>< wax— owﬁo>< mo:— “cotangen—
oEub— 05:. 3.5::
.895 .55
$6.8 63% a2. 3 Emma 82.: $8.8 $2.3 $5.3 GEE anal $9.3:
3m _ .o 3.5 Sood 886 $86 good good 286- good- mind 336- .w>< woq
“6%:an 8cm; xooE baa—A; owﬂo>< oE=_o> wax— oumm 38h owﬁo>< woq owﬁo>< mod «newcoQoQ
25cm «SE. mania:
.EE .5
68$ €38 Gem. 3 :65 $2.3 $8.8 33.3 A363 5.2» $2.3 $3
326 cokmm Good 386 mNSd 356. owned omhod- hmoﬁo- vommd mmvod- .w>< wou—
vﬂwdqwum a magnum ma Hag ma gm mg g g NEH—dd 3
3533‘ mound. Moo—m Nam—«50> owﬂo>< 085:5 mod 35 on“; owﬂo>< woq owﬂo>< mod Houseman—

oEaah 05E. his

9525 Jack. 93—32 733%;—

uh charm

9
7

5.3.0

33:5
333

vowed

Hang
38.8....

«mm—d

33%

was?

33.3 83.8

:38 A83 $38
286 ~83. :85

$33
3.3-

 

3mm 8%
a 353%
83:. #85

23.3 3.2.:

bzuﬂo> amp—2:... oE=_o> mod

252m 085. base!

2%.: 9.8.: 38.8
385 885 $85

23.3
38.?

35 88... owﬂo>< woq owﬂo>< we:

33%

 

3”,: 3%
mg 353
88: .305

A35 9.8.8
23: 3mm

88¢ v.85

bzuﬂo> owmb>< oE:_o> me;

2:95 2:5. manna—am

88.: 3%.: $2.3
386 885 «.85-

3% 03¢. owm3>< we:

ASN.3
«Sod-

bm=§o> owEo>< o§_o> woq

252m 085. .52—

53235 #89 3:2 383%

:. 9...»:—

oﬁm ova; owﬁo><woq owﬂo>< woq

.895 .3;
$503
.w>< woq

ﬂung
602339

.55 .5

33:5
.w>< woq

3
2895qu

.895 .65
ran
.w>< we;

354% «mad 3%;
owﬁo>< woq

335““va

80

The regional average volatility coefﬁcients were positive and signiﬁcant in ﬁve out
of six regressions; the NYSE average volatility coefﬁcient was positive and signiﬁcant in
only two (Paciﬁc and Philadelphia) of the regressions. In these instances, the regional
average volatility coefﬁcient was slightly larger. The regional block trade coefﬁcients were
positive and signiﬁcant in four of the six regressions; the NYSE block trade coefﬁcient was
positive and signiﬁcant in all regressions. In the regressions where both coefﬁcients
obtained the expected sign and were signiﬁcant, the coefﬁcient values were roughly equal
in magnitude. However, because the average NYSE block trades value is much larger, its
elasticity will tend to be smaller given the semi-log speciﬁcation.

4. Summary of Local Market Hypothesis

These results do not provide strong support for the local market hypothesis: only the
behavior of the regional volume coefﬁcient suggests strongly that local conditions are
superior predictors of local prices.78 However, this result may be a simple manifestation of
cream skimming behavior (selective market making) and queuing on the NYSE. Non-NYSE
exchanges tend to become more active only when conditions are favorable: in active
markets, lower spreads and in smaller trades. Because they pick up relatively more volume
when inventory carrying and price risk costs are minimized, it is not surprising that regional
volume is strongly associated with lower liquidity premiums.

Still, it would seem that NYSE volume should have more of an impact on regional

liquidity premiums because it characterizes whether or not the industry is generally active.

 

78Regressions were also performed using regional and NYSE explanatory variables
separately to explain local average liquidity premiums. An F-test revealed that regressions using
both sets of variables had signiﬁcantly better explanatory power compared to regressions using
either set of variables alone. However, this result appears to be driven by the large number of
observations.

81

When the industry is more active, price information is disseminated, shrinking the NYSE
information advantage and reducing price risk and inventory carrying costs for those who
trade on the regionals. However, NYSE volume may not capture the type of activity that
reduces these costs because heavy volume may be comprised of a few large trades on the
NYSE. Heavy volume comprised of many smaller trades would help the competitive
position of the fringe more because they can beneﬁt from price information, queuing on the
NYSE and avoid information traders. Therefore, a more accurate measure of trade activity
and its corresponding effect on the regional exchanges may be the number of transactions
that take place within a time frame for a given size class. This is consistent with Hamilton

(1979).

CHAPTER 6
CREAM SKIMMING BEHAVIOR

This chapter explores the third explanation for statistically signiﬁcant price
differences in what appears to be a competitive industry environment. This explanation
holds that competition measures are inherently ﬂawed because they capture cream skimming
or selective market making activity. These terms describe the phenomenon where regional
trading activity and price performance relative to the NYSE are systematically dependent on
the activity of the trade environment. To examine the cream skimming hypothesis, this
analysis uses two variables (and a competition variable) to capture the conduciveness of the
trade environment to cream skimming behavior. These variables are then used to explain
hourly average price differences in order to quantify their impact given that one exchange
offers better or worse average execution (i.e., there exists a non-zero average price
difference).
A. Capturing Trade Activity: New Calculation of Liquidity Premium Diﬂ’erentials

In order to derive liquidity premium differentials, this paper used a scheme that
matched similar-sized regional trades to NYSE trades occurring one minute before or aﬁer
the regional trade. While this matching technique provides useful information, it has several
drawbacks, particularly when used in an analysis of cream skimming behavior. The two

main drawbacks are as follows.

82

83

First, in order to derive a single average liquidity premium difference for a given size
class, two averaging calculations were required. If more than one NYSE trade was matched
to a regional trade, then the NYSE liquidity premiums were averaged to derive a single
liquidity premium difference. These individual liquidity premium differences were then
averaged again to obtain yearly averages. The weakness of this approach is that it does not
capture the intensity of trading surrounding a regional trade. That is, the matching scheme
treats a regional trade matched to one NYSE trade the same as a regional trade matched to
ten NYSE trades. Clearly, the two scenarios may reﬂect different trading environments.

Second, the older matching scheme derived only one average liquidity premium
difference between two exchanges. Yet, one liquidity premium difference may not
accurately reﬂect execution quality on both exchanges simultaneously. For example, an
average liquidity premium difference of one cent indicates that on average, regional
execution is one cent higher than the NYSE. However, average regional execution that is
one cent higher than average NYSE execution need not imply that the average NYSE trade
also receives one cent better execution. This may occur if regional execution (relative to the
NYSE) is dependent upon whether the trade environment is active or slow, as measured by
the number of matched NYSE trades. The example below illustrates how this result may
occur.

To counter these drawbacks, a new matching scheme is employed in this chapter.
The most important feature of this new technique is that it does not average matched NYSE
trades. As a result, if regional execution (relative to the NYSE) is dependent upon trade
activity, then the price difference calculation will be weighted in the appropriate direction

to reﬂect this. In addition, the derivation of a price difference for trades occurring on each

84

exchange allows the analysis to identify the exchange that beneﬁts or suffers from this new
weighting of trades.

An example illustrates the new matching technique and how it differs from the
technique used earlier. In this example, regional execution improves when many NYSE
trades surround a regional trade (constructed purposely to mimic the data). Figure 3 shows
a hypothetical situation where trading occurs at 9:30 and 10:30. For simplicity, regional (R)
trades are assumed to occur on the same regional exchange and all trades are assumed to be
in the same size class. Arrows indicate the trade that is used for matching purposes.

Assume a regional trade occurs at 9:30:10 (R1) at the midspread of 330.375. The
matching scheme records the price conditions (transaction price and midspread) on the
regional but does not consider this trade "matched" because the outstanding price conditions
on the NYSE are unknown.

The regional price conditions are then extended out one minute. If a NYSE trade
occurs within the minute, then the matching scheme records one matched trade that occurred
on the NYSE (assuming it is within the same size class). If a NYSE trade does not occur
within one minute, then the regional price conditions are discarded and the program again

looks to identify price conditions on both exchanges.

85

Dec.

5: one 3.2 $5 o3
_ _ _ _ _ 83$

 

 

madmmnEm .

03:0

/

mundane—mt.
32.0 >62

 

 

SJBIIOC]

 

 

l

i mhmdmw

 

 

 

 

 

 

mam—.35 wag—28 953 89:29::— mEEanm 5:55:
m «Suﬁ

86

Returning to Figure 3, assume four trades occur on the NYSE at 9:30:30 at $30.50.
The price conditions on both exchanges are now known and the program registers four
matched trades that occurred on the NYSE that paid an excess liquidity premium of 12.5
cents each compared to the most recent regional price conditions. A regional trade at 9:31
(R2) is also matched and the regional price conditions are updated. This trade received 12.5
cents superior execution compared to the most recent NYSE price conditions.

At 10:30, another regional trade occurs (R3). Again, the regional price conditions
are recorded, but the trade is not considered matched. The NYSE trade at 10:30:40 at the
new midspread is matched and receives 12.5 cents superior execution compared to the most
recent regional price conditions (which used the older quote). The regional trade at 10:31
(R4) is also matched (50 cents worse than the NYSE).

A calculation of the average liquidity premium difference using the two matching
schemes demonstrates the difference between them. Using the old matching scheme (where
NYSE trades before and after regional trades are matched and an average is taken if more
than one NYSE trade is matched), the average liquidity premium difference (regional minus
NYSE liquidity premium) would be calculated as:

Old Technique [(-12.5) + (-12.5) + (12.5) + (25)]/4 = 12.5/4 = 3.125 cents
This result does not recognize the fact that superior regional execution took place in a more
active trading environment because it averages the four matched NYSE trades. In addition,
the result implies that all NYSE trades received 3.125 cents superior execution, when only

a representative trade did.

87

Using the new matching scheme, an average difference can be calculated for matched
trades that actually occurred on both exchanges:

Matched Trades on Regional [(-12.5) + (25)]/2 = 12.5/2 = 6.25 cents

Matched Trades on NYSE [4(-12.5) + (12.5)]/S = -37.5/5 = -7.5 cents
This is a more accurate representation of the price differences for all matched trades that
actually took place on the respective exchanges. The new matching technique captures the
fact that superior regional execution takes place in a more active trade environment and
adjusts the NYSE average price difference to reﬂect this. That is why we derive the unusual
result that both exchanges seem to offer worse average execution.

To sum up, a matching scheme that does not average matched NYSE trades is
necessary to examine cream skimming behavior of fringe competitors because execution
quality on the regional exchanges may depend on NYSE trading intensities. Price difference
calculations must be allowed to reﬂect this. If this behavior is present, it will manifest itself
by making average NYSE execution look much worse. A matching scheme that derives a
single liquidity premium difference and does not identify where a trade occurred is not
helpful because worse average execution on one exchange automatically implies superior
average execution on the other exchange of an equal magnitude.

B. Comparison of Different Matching Techniques

Table 8 depicts average liquidity premium differences for 1992 using the new
matching scheme. The ﬁrst column of Table 8 lists the matched pair of exchanges, with the
ﬁrst exchange representing the location of the matched trade. For example, the ﬁrst pair of
ﬁgures in the upper left-hand comer of Table 8 are listed as 45,130 and -1.6819. This

indicates that 45,130 trades that occurred on the Boston exchange were matched to a NYSE

88
trade in the 100 - 400 size class.

On average, these trades received 1.6819 cents better execution compared to trades
on the NYSE using the most recent price conditions and a one minute time frame. Directly
below are trades that actually occurred on the NYSE and were matched to a similar Boston
trade. As shown, 67,281 trades occurred on the NYSE that had a match on the Boston
exchange. These trades paid 4.998 more on average. Table 9 displays earlier results using
the older matching technique for comparative purposes.

Comparing the matching techniques, two results become apparent. These results
provide strong evidence that regionals improve their relative execution when many similar
NYSE trades (in the same trade size category) surround regional trading:

1. Using the newer matching technique (Table 8), matched trades that actually occurred
on the regionals are fairly close to the values using the old matching scheme in Table

9. Allowing for trading intensity by counting all matched NYSE trades within the

time frame did not change relative execution for trades occurring on the

regional/Nasdaq exchanges.

2. Using the newer matching technique, matched trades that actually occurred on the
NYSE are much worse than the differences calculated using the old scheme. When
trading intensity was allowed and all matched NYSE trades were counted, average
NYSE execution suffered compared to average regional/Nasdaq execution. This
result mimics the above example. Only for the largest size class does this result not
hold; regional execution was still much worse.

89

 

 

.3385 005350 E3593: a Maw: 3505:»? Bermuda 8850 2.
49.85 853:8 E08365 a min: nonwoﬁcwﬁ 333me 88ch

 

 

 

 

.. 30m...” 9% Emma- new a 3.3.7 Stun . 83.7 :3. _. 33.... 09.2.
_. wand 5m ._ mmvo; Po .. 536 new; 336 _~w.m _. Gmmd 33v
. ”mam: an ._ Named- Shm .. SS... 34.: . ~82“- mamﬁ— .. wonN- Sign
.. anon; wmb _. 8:3 ~mw.~ _. 336. SN.» ._ mg... «mud— . vow—d- «$.3—
mwwhd mmm _. 336- Em; .. 33.7 03.5 _._ «QR- mung: ._ maxed- mom._~_
33.0- mam 2. $36 and; ”wood- .36 _. 386.. «m3. ._ med RN55
33”.? v3 . «NS... SEN . Nomad- a: .. wand- mg.» . can—d- mm 1am—
_. wave— mmw ._ «vane- 23.. _. 3:..- 09% .. .mwod- 3nd ._ 3126. 2Q;
_._ RS; owe a $36. $8.— .. ”$3.. v2.5 _. 9.3.? «3.5. ._ 3:”- Sink.
o .thd mwv $36- 3c.— . Sewn- veoﬁ _._ 33d- mmow 86.. 53.9.“
.. 3mm.~ o3 ._ 2.? 3m... . $5.? $3” . $3.? Kim a 38%? 33b
9 3.2.— ohm . «~36- $3 .. «922. main .. 33d- Sax” ._ 28.7 313.
owﬁo>< .«o “3:52 owﬂo>< no 52:52 oweo>< mo .6952 owﬂo>< mo Ban—=2 owﬂo>< .«o “3:52
83A 83.83 83.83 23.8.». 89-8—
880 35

8a .35 «am 3 82:9 85:: 25a

.8322 bun-.85 9:188 raunchy—«a 5330.5 ham—:95..— oqu><

w 03.;

eioﬁuﬁmgz
mm>zé=2§=§

82.2-3;
mm>z-§§z

ou_8m-mm>z
um>Zuoanm

~33v§¢m>z
mm>2¢m0322

mﬁgoﬁoﬁmg
mm>Zﬁu=£o£U

:3mom-mm>z
mm>2é3mom

add—Hun“

a

9O

 

 

:SCBE oocovuaoo 88.593: 5 wsmz oogcﬁwﬁ Guzman; museum 3..
:9:an 8:02:50 Scuba-one m wﬁmn 8:85am? Roxanna 8850 ..

 

 

Svmd 3m .. EN: 5:.— .. aowmd 02...“. 39° 29> .. 3:6 mwvdw

2.2.N Sn; .. :86 9mm.» .1. de- cmndm .. 32.6- mvvdm .. $26- 81»:-

wabnd :k _. ~33 vamﬁ 3.. $2.? ”3.2 _.. nomad- SN;N .. 886 $33:

«Ev: m:.N 2. 3N6- coin .. «6.3.:- aeo.3 ._ maﬁa:- mow.2 .. Nemﬁo ”2.2.:

33.». Em; 2.. 3:3. ”aim .. 2.9..” NNQZ ._ Ekmd- 95.3 836- v5.2.

new: cg _. omovd- 39m .. emu-.7 2&6 .. mag- 93$ _. $3.7 magma
owﬂo>< anmwom owﬂo>< 3:033— owﬁo>< :aaomwom owﬂo>< 3.833— owﬂo>< ﬁcomwom
60:33: 3:83: 3:832 3:082 I 3:882

88A 8318:" 83.83 cam-8m 8n. - 8—
820 «Em

a 2:5.

2S asap :5 E 8.59 852.2 2::-
8o=ouoh=n 53895 38354 ou§o><

2203.2:
5.32
0&8:
“8322
325050
:Smom

gag

91

Although some of the liquidity premium differences in Tables 8 and 9 differ
signiﬁcantly, the only real difference between the two matching schemes is the weighting
of matched NYSE observations and the fact that the second technique uses a matching time
frame halfas long. However, price differences calculated using the older matching technique
were invariant between a one-minute, two-minute and ﬁve-minute time frame: different time
frames only changed the total number of matched trades. Therefore, the different results are
solely attributable to the equal weighting of all matched NYSE trades and using most recent
price conditions.

The results in Tables 8 and 9 demonstrate that regional exchanges tend to offer better
execution in an environment characterized by active NYSE trading activity. When all
matched NYSE trades were counted (Table 8), the average liquidity premium difference for
trades occurring on the NYSE grew noticeably worse, while the average liquidity premium
differences for trades occurring on the regional exchanges did not change appreciably. These
results strongly suggest some type of cream skimming behavior: the regional exchanges not
only choose to "compete" in active markets, but also tend to offer better execution at these
times. This behavior makes it possible for statistically signiﬁcant liquidity premium
differences to exist between exchanges because the fringe offers competition or better
execution only some of the time.

C. The Cream Skimming Hypothesis
To examine cream skimming activity, two variables are used to explain average

hourly liquidity premium differentials.79 These variables capture (1) market activity, (2) the

 

”Half-hour averages were also analyzed, but they did not change the general
conclusions.

92

ability of non-NYSE exchanges to free-ride from price information, (3) any beneﬁt arising
from the queuing of order ﬂow on the NYSE and (4) the lower spreads and risk associated
with active markets. They are as follows:

TRADES sum of number of matched NYSE trades during the hour (per stock)
TOTVOL sum of all trade volume on all exchanges, need not be matched (per stock)80

The TRADES variable captures speciﬁc activity on the NYSE exchange based on the
number of matched transactions that occurred there. It is anticipated that the TRADES
coefficient will be negative in most cases, reducing price differentials, no matter the
exchange offering better execution. As the number of TRADES grows large, the market
becomes very active, forcing spreads and liquidity premiums to their minimum values and
liquidity premium differentials to zero.

The TOTVOL variable captures all general trade activity over the time frame that can
be used for price information or inventory adjustment purposes. This measure includes
volume on all exchanges and in all size classes. It is hypothesized that the TOTVOL
measure will also have a negative impact on price differentials, particularly when regionals
can use this information to reduce information asymmetry. A competition measure
(ACTIVE2) is included to test the explanatory power of this measure relative to cream
skimming variables.81 The dependent variable is equal to the log of the average liquidity

premium difference (regional minus NYSE liquidity premium).

 

80This modeling assumes that these variables are exogenous. It is possible that smaller
liquidity premiums (and smaller liquidity premium differentials, potentially equalling zero if
spreads are at their minimum level) attract more trading. However, this analysis omits zero
liquidity premium differentials where this possibility may be most problematic.

81Regressions were also performed using the HHI competition variable. The results
were very similar to those using the ACTIVE2 competition measure.

93

l. Splitting of Data

Based on the results from Table 8, cream skimming by regional exchanges manifests
itself through poorer average NYSE execution: average liquidity premium differences for
trades occurring on the regionals were not affected greatly. Because regionals appear to offer
better execution in more active markets and the effects manifest themselves through worse
average execution for trades occurring on the NYSE, the complete price difference data set
for each exchange and size class was split into two smaller data sets: (1) matched trades
occurring on the NYSE that received better average execution and (2) matched trades
occurring on the NYSE that received worse average execution. This split allows the three
explanatory variables to have different impacts on the dependent variable depending on the
exchange offering better execution over the time frame.82

This split ignores observations where the average liquidity premium difference
equaled zero. While this will most likely result in omission of some useﬁrl information, it
is recognized that the data set is quite limited in its ability to characterize an environment as
conducive to cream skimming behavior. As a result, it is not rich enough to incorporate zero
liquidity premium differences in the analysis; there may be some relevant variables that were
omitted. In addition, the discrete nature of pricing in the industry causes a fairly wide range
of explanatory variables to be associated with zero liquidity premiums differences.
Therefore, the results that follow are conditional on non-zero liquidity premium differences.

When the data sets were split in this fashion, one of the data sets contained negative

liquidity premium differentials (when the NYSE gave worse average execution). To ease

 

82The data sets were also split in a similar manner for trades occurring on the regional
exchanges. Regressions using these data also tend to support the contention that regional trading
and execution is dependent on overall NYSE activity, though results are generally weaker.

94

the interpretation of the regression coefﬁcients, these differences were converted into
positive values. Using this conversion, all coefﬁcients may be interpreted in a similar
manner; in both regressions, a negative coefﬁcient reduces price differentials (given that one
exchange offered better average execution).

Tables 10a—f list average values for variables of interest. As shown, when the NYSE
offers worse average execution, the average liquidity premium difference tends to be much
larger, as are the TRADES and TOTVOL variables. (P-values indicate the probability that
these variables are not larger when the NYSE offer worse average execution.) The greater
number of matched NYSE TRADES that take place during time frames when the NYSE
offers worse average execution weighs down the average NYSE execution, as in the previous
example. This is especially true for the smallest trade size class, where the number of
TRADES is very different depending on the exchange offering better average execution.
However, even though the TRADES and TOTVOL variables tend to be larger on average
when the regionals offer better execution, a pertinent question is whether their effects are

statistically different depending on the exchange offering better execution.

95

dousooxo 3.83 auto mm; non? cowa— oa oouocogu 83an .0633— owﬁog use 4039—. .meE 05 35 837. £85093 :2. SE. .380? m_
£85092 :3 05 35 3:32.05 05 8865 mos—gi dag 358a 85 cotoéoo 8:308 Gm; - .maomwoc vantage 82an .323: owﬁo>< _.

 

 

 

odwﬁm mg ed .2: cogooxm Stem mm>z
<2 ﬁgﬁv 2N6 mm oovd o.o_ mm ooomA *cousooxm 38>» mm;
mémﬁm Nu— n.» of cousooxm “atom mm>z
366 56.0.3“ omod Nd Sod QQ mg ooov - ocom *cousooxm 883 mm>z
:bmﬁ a; Wm mum cousooxm Baum mm>Z
Sod "amt? Sod m.m Sod v.2 mom 8% - o2: *aousooxm 383 mm>Z
Wotan a; ﬁx 5N aousooxm Stem mm>z
Sod {omega Sod m.m Sod 5.2 o? ooa - com *cousooxm 883 mm>Z
v.8»; m.m m6 $3” noun—03m Stem mm;
Sod m.ocw.m Sod 3: Sod as 2 ﬁm oov - o3 *cousooxm 8.83 mm>Z
we pea—E: Z 35
A258 AO>HOH man—d9: 3:93.59 5380.5
3.5:: 3.5:: >32:qu 2.5::
owauo><

:82;— E. 8:332 .53 mm; :c wan—.30 863,—.
mnemmmouwom 3:20.59 Esmﬁoum bags—.5 .Su 835» «3.3.5.

«A: 93:8

96

52388 8.83 ﬂab: Mm; 855 892 Pa 388:6 8388: $25: 8:88 :5 4030:. .meg :5 85 88% $853.3 :3: 2E. :88? ﬂ
£85093 :3: :5 85 5:338: :5 8865 8287: .03? 038:: SE 8828: 3:82: Amm>z - 8:28.: ou88b€ 5382: 363:: ow~8>< _.

Sod

Sod

Sod

onto

Sod

«Esau—«88.2 «Suﬁ—88%

v.81:
oﬁoﬁw
Nvmﬁm
>636
N.moo.m
0.81:
5.5:..N
ﬁmmﬁw
w.cwm.m
NNOVH

Sod

Sod

Sod

Sod

Sod

0.:
ﬁm
m;
QN
a;
QM
a.—
5N
NS
ﬁm

Na
:93 §
3

so: a:
2

so: :2
2
so: 3
3
ﬂ so: 3

.33.: 88...»:

 

 

age 53.3
.#5::

mmgh 388.5: 5558...—
b5:: 3:55: b5::
8885‘

we:
hm

:8
$—

no:
08.:

Sum
www.—

maﬁm
:3:

gang

a: 8:55 Z

ooomA

ooav - ooom

oom— - coo—

oom - com

oov - 2:

mum—u

35

3.25850 a: 8:832 .53 mm; a: ”5.586 858,—.
2888.83— 88.8.5: 5558.5 3:55: .8: 825» owa8><

:3 «Ban.

8:83: 88: mm>z
*aowsooxm 88.3 mm>z

8:88AM: 88: mm>z
205.8%: 883 mm>z

noun—cox: 88m mmwz
Eousooxm 88:3 mm>Z

:Omuﬂouxm Hoﬁom WWVZ
*:Omwﬂooxm umuog mm>z

:Omwﬂooxm “atom mm;
.LuOﬂn—ODXW Dmuoa mm>z

Hang

97

.8888 0.983 0.0%.: mm; 8:3 .030. 0.... 00:08.5: 8808 2.2.8: 0888 E... AOEOH .meg 05 .05 8.80. 08.8.5. :8 05. 80.008. 0..
0.85093 :5. 05 .05 2:30.88 08 0.0065 820?: .038 028:: 0.5 8.8280 0.5805 Em; - 3830.. 00:08am... 8.508 2.28: 0w08>< ..

Sod

de

Sod

Sod

Sod

0808554 “8585.84

058.—
mswomw
wdmﬁm
adoﬂv
m.w$.m
mémmﬁ
oﬁmﬁm
w.wmm.m
Wag;
w.o$.m

 

mood

Sod

Sod

Sod

Sod

v;
a:
5.:
WM
:2
Mam
a;
5N
v...
m.m_

... :
8... m. ..
8
mm... m...
3
.8... .....
0.8
.8... 8.
E
2...... 2.
8 008

 

.8... 8:3
25::

magma.
25::

 

OOEQHOQA— 352:9:—

b....=..... 2.5....
09.83.

com
Nu

Sm
omm

gm
5%

gm.
of;

Kim
mmﬁm

888d

.8 8:552

ooomA

ocav - ooom

cog - coo—

oom - com

8: - oo—

83

05m

«8322 8 8:082 .53 mm; 8 mat—800 8.8.5.
088880: 0088.5: 5350...— 255—0: .8: 830.» 09.83..

03 03:9

:OmHSODunm Hoﬁom mm;
*ﬂOU—aooxm OmHOB mmtfz

:Omwﬂooxm uoﬁom mm>z
*Gomuﬂooxm Omanvg mm>z

8800.5 88m mm>z
$8808.: 0083 mm>z

8800.: 88m Mm;
*88000.m 00.83 mm>z

8800...: 88: mm > Z
*8u80xm 00.83 mm > Z

g

98

5.5.80.8 0083 80...... mm; 0055 .050. 0.0 00:08.05 53508 5.558: 0w0..0>0 .50 40:0... .meg 05 .05 00.0.0. 0.005005.— =5. 0:... 50.008. 0.
0.005093 :5. 05 .05 8:50.88 05 0.005... 00.53-.. .0...0> 03.80.. 0.8 00.00280 080005 Amm>z - .0880... 00008.05 52508 5.558: 000.03.. _.

wwmd

Sod

Sod

Sod

Sod

«505.5353 «302000.83

2.00.0
0.6010
5.mm5.m
0.wN5.v
TEQN
0.0555
..5mm.~
.0. 30;”
0.50..
5.05v.~

 

amod

Sod

Sod

Sod

Sod

0..
Q.
m.
wN
0N
0.m
m.N
QM
wd
Q0.

 

.88. .550...
5.0.8:

mag...
5.58:

 

5..
08... N...
0.0
.8... N...
a...
.8... 0....
.8
.8... 0...
N0
.8... .8
00.0.3. 000.004.
OUGOLOEQ aim—ﬂonm
8.0.0.... 2.5....
000.05.

05.
5m

50m
03

mm...
.3

w0m..
50m;

050.5
mm0.m

gang
.8 805.. Z

ooomA

oomv - ooom

ooa. - coo.

ooa .. com

000 - co.

0%

05m

0000.. Z .8 0.0.8.02 5.3 mm: .8 05.5000 0000:.

50.8053. 00:08.59 5550.5 5.5531— 8. 0030.» 000.03..

E: 0308

Gomwﬂooxm Hoﬁom mm>z
*Gomuﬂooxm OmuOB mm>z

.8580...”— ..0..0m mm>z
*505000xm 0083 mm>z

55:005. .0..0m mmwz
0.8.8005. 00.83 Mm>z

GOmwﬁoDXmm Huﬁom mm>z
*Qomwﬂooxm 09.03 WWVZ

505:00xm .050m umVZ
085.00%. 0083 mm > Z

g

99

2385 quB auto mm; 5:3 .8qu 2m uonouobmu Eaﬁoa £23: owﬂozw v5 ADP—bk .mmﬂé o5 $5 83% £353»: =2: 2? 6800?. a
€058»: =3: 05 $5 35339:. 05 8865 mo=_a>-m .02? 95:2— 85 39.8280 8:938 Amm>z - 3:295 ooceobmu 83an $23: owﬁo>< ..

 

 

 

~6me m; Q: mm cousouxm Stem mm>z
<2 N636 am ﬁo ON <2 2: mm ooomA *cowaooxm 3.83 mm>Z
33mm M: Q: ova nousooxm “atom mm>z
Sod ﬁwcmam m 56 ON Gﬁo wNﬁ H3 oomv - coon *aouzooxm 883 mm>Z
wémmnm Wm 02 En nousooxm 550m mm>z
Sod edema Sod ﬁn Sued v.2 o2. ooS - o2: *aounooxm 8.83 mm>z
mdvmam NM 2: com :ousooxm Stem mm;
Vmod 062$. Sod 5m Sod ~12 2m; com - com *uounooxm 88>? mm>Z
@6an cs m6 336 nouaooxm 85m mmVZ
Sod m.w$.m Sod 0.2 Sod 9w who.m cow - o3 *aounooxm 88>? mmwz
.«o hon—E: Z «Em
$2.8 AOEOH magma. :ouoba 83.5.5
2.5:! busoﬂ 3:553:— b.53—
owﬁo><

9525 .5 8:852 .53 mm; .8 ”5.530 mot—2H.
3329—qu 35.8.5: 8350.5 mum—.35: .Su mos—5» nausea

2: «Bab

100

5:583 883 ﬂute mm>2 5:3 532 2a 855:6 83.53 32:3: owm5>a :5 :OEOH £ng 05 3:: 5:5 50:89»: :3 SF .550? m_
£553»: :9: 05 $5 b59305 06 85:2: 529.: .02? 02:8: 85 55280 8558 am; - 3:25: 355%: 83:55 56:6: owﬁo>< _.

 

 

063% m: ad: mm:
ovvd :Nmﬁ <2 :2 M2: w:

@586 m: ed: mm:
Nwmd wébﬁm Sod Wm 92 mm

QmNQm w: 0.: Gm
Sod ﬂommﬁm Sod o.m v.3 mvm

@326 ON md: amv
Sod m.:mo.v Sod wN 5m: mam

maﬁa; w.m v.2. mmvav
Sod mdvdm Sod ad: 9w nmdm
358 326: mans: 53:5 east:

:58: 3.53.— bmcmaa: 3.5::
ou§o><

«En—03:5: .3 5:332 :33 mm; .8 warn—30 5.3:.
muommmoawom ounouoﬁn 83895 33:55 .5: 53.5 ow§o><

we 011$.

.3 5:8: 2

com:V - ooom

com: - 80:

5:53“: 55m mWwZ
Hog—cox”: 3.83 mm>2

cousooxm 55m mw>2
*cousooxm 883 mm > 2

ﬂomwﬂooxm notom mm;
*GOUDooxm 0303 WWVZ

cousooxm 55m mm>2
20:58:: 88>? mm;

GOSH—002m Hoﬁom mm;
shown—002m 09:03 mm>z

gag

101

2. Cream Skimming Regressions

The term cream skimming, as it is used in this analysis, describes the phenomenon
where regional execution relative to the NYSE improves in an environment characterized by
active trading. In other words, the quality of regional execution is directly dependent on
NYSE activity. Given this deﬁnition, the following hypotheses are tested:

Hypothesis 1 The number of TRADES reduces price differentials to a greater extent when
the NYSE oﬂers better average execution.

This result reﬂects the different impact that increased NYSE activity has on regional
exchanges. If the NYSE is able to offer superior execution (due to information
advantages or hidden liquidity), increasing NYSE activity signiﬁcantly reduces this
advantage. However, if for some reason the regional is already competitive with the
NYSE, increased NYSE trades disseminate little new information that is helpful. In
either case, the TRADES coefﬁcient will be negative simply because many TRADES
force liquidity premiums to their minimum level and price differentials down to zero.

Hypothesis 2 TOTVOL increases price diﬂerences in favor of regional exchanges, no
matter the exchange oﬂering superior average execution.

This variable may provide less information, but the information can be used
nonetheless. TOTVOL includes all volume in the stock during the hour, no matter
the size of the trade or the exchange. Given that TRADES is held constant, the fact
that TOTVOL actually improves regional execution relative to the NYSE is not that
surprising, especially for the smallest size class.

Hypothesis 3 The number of matched NYSE TRADES has the greatest impact on reducing
price differentials for the smallest sized trades.

Regional exchanges prefer to deal in smaller trades because these trades are less
likely to have originated from information traders, thereby lowering the traders risk
and costs to trade. In addition, smaller trades have less potential impact on
inventories.

Tables Ila-f list the results for the cream skimming regressions. As an example of

how to interpret the results, consider Table 11a. The ﬁrst entry is for the data set that

contains observations for trades matched between the Boston and NYSE exchanges where

102

the Boston exchange gave better execution (column 1). This data set includes observations
that actually occurred on the NYSE (column 2) for trades in the 100 - 400 size class (column
3). Columns that follow list coefﬁcient and t-statistic values for the TRADES, TOTVOL
and ACTIVE2 variables (the TRADES, TOTVOL and dependent variables are log
transformations). Adjusted-R2 statistics are listed as well. Due to space constraints, Durbin-
Watson statistics were omitted; values typically ranged between 1.6 and 1.9.

To test the three hypotheses, simple one-sided t-tests were used. These tests are
meant only to demonstrate that the variables of interest have a signiﬁcantly different effect
depending on the exchange offering better average execution. The hypotheses simply
compare regressions either within a particular trade size class (Hypotheses 1 and 2) or across
trade size classes (Hypothesis 3). Table 12 lists the results of the tests. A value of "1"
indicates that the ﬁrst hypothesis was conﬁrmed for that size class while a value of "2" or
"3" denotes that the second or third hypothesis held. (The third hypothesis only compares

regressions in which the NYSE offered better average execution across size classes.)

1 03
Table 1 la

Liquidity Premium Difference Regressions, Boston
Matches Between Boston and NYSE Exchanges
Log of Average Hourly Liquidity Premium Difference as Dependent Variable

Exchange Location
Giving Better of Matched Size Adjusted
Execution Trades Class TRADES IQIMQL ACIIYEZ W

 

 

 

 

Boston NYSE 100 0.3581 0.2842 0.2499 0.161
to (27.13) (14.82) (13.93)
400
NYSE NYSE -O.738 0.0527 0.032 0.592
(50.08) (3.281) (2.746)
Boston NYSE 500 0.032 0.0708 0.1381 0.021
to (-2.388) (1.181) (2.264)
900
NYSE NYSE 0.685 0.0410 0.1 1 0.402
(12.44) (0.985) (2.69)
Boston NYSE 1000 -0.085 0.0637 -0.038 0.002
to (1.417) (1.063) (-0.583)
1900
NYSE NYSE -0.5615 0.0410 0.0686 0.267
(9.249) (0.903) (1.417)
Boston NYSE 2000 0.3028 0.1963 0.1242 0.0574
to (3.019) (2.088) (1.133)
4900
NYSE NYSE -O.5088 0.0758 0.0029 0.1563
(-5.667) (1.316) (0.044)
Boston NYSE 5000 0.4072 0.1203 -0.058 0.002
and (1.545) (0.559) (0.305)
greater
NYSE NYSE -O.4065 0.1523 0.1318 0.0797
(3.374) (1.757) (1.516)

104
Table 1 lb

Liquidity Premium Difference Regressions, Cincinnati
Matches Between Cincinatti and NYSE Exchanges
Log of Average Hourly Liquidity Premium Difference as Dependent Variable

Exchange Location
Giving Better of Matched Size Adjusted
Exmticn Insists Class IRADES IQDLQL AQIIMEZ 1325511133751

 

 

 

 

Cincinnati NYSE 100 0.4168 0.307 0.1225 0.144
to (17.69) (8.82) (3.197)
400
NYSE NYSE 0.7165 0.0069 0.0057 0.546
(49.68) (0.346) (0.305)
Cincinnati NYSE 500 0.1438 0.0787 0.055 0.017
to (4.506) (2.810) (1.98)
900
NYSE NYSE 0.5861 0.0739 0.0219 0.293
(14.947) (2.562) (0.777)
Cincinnati NYSE 1000 0.1402 0.0093 0.01 15 0.016
to (4.035) (0.264) (0.276)
1900
NYSE NYSE 0.6977 0.0827 0.0359 0.365
(16.21) (2.434) (0.90)
Cincinnati NYSE 2000 0.1008 0.0489 0.1235 0.006
to (1.19) (0.566) (1.269)
4900
NYSE NYSE 0.5121 0.0558 0.0764 0.189
(6.99) (1.252) (1.591)
Cincinnati NYSE 5000 0.1 171 0.0381 0.3037 0.003
and (0.839) (0.268) (1.425)
greater
NYSE NYSE 0.5199 0.0145 0.01 1 0.186
(5.809) (0.229) (0.165)

105
Table llc

Liquidity Premium Difference Regressions, Midwest

Matches Between Midwest and NYSE Exchanges
Log of Average Hourly Liquidity Premium Diﬂ'erence as Dependent Variable

 

 

 

 

Exchange Location
Giving Better of Matched Size Adjusted
Emmi] Trades Class IEADES m ACIIMEZ W
Midwest NYSE 100 -0.286 0.2122 -0.2418 0.128
to (-16.96) (9.001) (-1 1.82)
400
NYSE NYSE -0.6032 -0.0687 -0.0464 0.531
(-61.46) (-5.65 8) (-4.56)
Midwest NYSE 500 -0. 1246 0.0689 -0.0799 0.017
to (3.731) (2.252) (-2.913)
900
NYSE NYSE -0.5967 0.0249 -0.0617 0.278
(-15.829) (0.924) (-2.464)
Midwest NYSE 1000 -0.1834 0.0846 -0.0697 0.031
to (-4.966) (2.326) (-2.174)
1900
NYSE NYSE -0.541 1 0.0324 -0.0576 0.213
(-1 1.969) (0.917) (-1.795)
Midwest NYSE 2000 -0.0395 0.01 l 1 -0.0038 -0.008
to (-0.622) (0. 194) (-0.067)
4900
NYSE NYSE -0.5738 0.0895 -0.0496 0.177
(-8.483) (2.143) (-1.207)
Midwest NYSE 5000 -0.1517 0.0287 -0.0127 -0.029
and (-0.977) (0.229) (-0.088)
greater
NYSE NYSE ~0.3231 -0.0191 -0.0233 0.038
(—2.922) (-0.284) (-0.41 5)

1 06
Table 11d

Liquidity Premium Difference Regressions, N asdaq

Matches Between Nasdaq and NYSE Exchanges
Log of Average Hourly Liquidity Premium Difference as Dependent Variable

Exchange Location
Giving Better of Matched Size Adjusted

Exmimn Trades EMIRADESIQIXQLAEIIXEZRW

 

 

 

 

Nasdaq NYSE 100 0.3275 0.2945 0.2402 0.131
to (21.26) (13.42) (13.46)
400
NYSE NYSE 0.666 0.041 0.0323 0.553
(73.97) (3.426) (3.486)
Nasdaq NYSE 500 0.2479 0.158 0.1 128 0.049
to (7.87) (4.683) (3.61 1)
900
NYSE NYSE 0.6243 0.0041 0.0078 0.343
(24.28) (0.186) (0.423)
Nasdaq NYSE 1000 0.2091 0.0427 0.0324 0.042
to (5.945) (1.093) (0.958)
1900
NYSE NYSE 0.5319 0.0159 0.0466 0.275
(17.93) (0.634) (2.246)
Nasdaq NYSE 2000 0.2419 0.1 1 19 0.0488 0.035
to (3.648) (1.742) (0.739)
4900
NYSE NYSE 0.465 0.0666 0.0273 0.148
(9.647) (1.864) (0.852)
Nasdaq NYSE 5000 0.2644 0.0987 0.195 0.012
and (1.267) (0.720) (1.28)
greater
NYSE NYSE 0.3666 0.0089 0.0639 0.081
(3.189) (0.135) (1.019)

107
Table lle

Liquidity Premium Difference Regressions, Pacific
Matches Between Paciﬁc and NYSE Exchanges
Log of Average Hourly Liquidity Premium Difference as Dependent Variable

Exchange Location
Giving Better of Matched Size

Emhcnlmdcs

Adjusted
Class TRADES TCTXCL ACIIMEZ mm

 

 

 

 

Paciﬁc NYSE 100 0.3216 0.2004 0.1824 0.121
to (18.51) (7.909) (8.532)
400
NYSE NYSE 0.6591 0.0392 0.0669 0.489
(61.82) (2.808) (5.912)
Paciﬁc NYSE 500 0.175 0.0291 0.0097 0.025
to (5.786) (1.081) (0.372)
900
NYSE NYSE 0.562 0.073 0.039 0.219
(15.87) (2.547) (1.48)
Paciﬁc NYSE 1000 0.1967 0.0737 0.0023 0.028
to (4.911) (1.801) (0.60)
1900
NYSE NYSE 0.431 0.1278 0.1026 0.150
(1 1.22) (3.65) (3.21 1)
Paciﬁc NYSE 2000 0.0553 0.0454 0.0635 0.009
to (0.616) (0.571) (0.728)
4900
NYSE NYSE 0.5128 0.1609 0.0914 0.170
(-6.795) (2.965) (1.466)
Paciﬁc NYSE 5000 0.1361 0.3131 0.0492 0.024
and (0.593) (1.40) (0.304)
greater
NYSE NYSE 0.6502 0.1436 0.0252 0.1 1 1
(3.075) (1.265) (0.238)

108
Table 111'

Liquidity Premium Difference Regressions, Philadelphia
Matches Between Philadelphia and NYSE Exchanges

Log of Average Hourly Liquidity Premium Difference as Dependent Variable

 

 

 

 

Exchange Location
Giving Better of Matched Size Adjusted
Execution Tradss Class TRADES TQ'DLQL ACTDEZ W
Philadelphia NYSE 100 -0.2428 0.159] 01733 0.073
to (-10.92) (5.081) (-5.516)
400
NYSE NYSE -0.5769 -0.0333 -0.0402 0.456
(-50.54) (-2.412) (-3.344)
Philadelphia NYSE 500 -0.1 101 -0.0206 0.0222 0.013
to (-2. 163) (-1.425) (0.424)
900
NYSE NYSE -0.4263 0.0062 -0.004 0.155
(-9.528) (0.434) (-0.132)
Philadelphia NYSE 1000 -0.1667 0.1 1 13 -0.0452 0.010
to (-2.305) (1.572) (-0.612)
1900
NYSE NYSE -0.5874 0.1282 -0.0437 0.202
(-9.669) (2.929) (0868)
Philadelphia NYSE 2000 —0.0153 0.2551 -0. 1378 0.040
to (-0.139) (2.366) {-1.152)
4900
NYSE NYSE -0.4006 0.0894 -0.1 192 0.069
(-3.49) (1.40) (-1 .742)
Philadelphia NYSE 5000 0.3952 -0.1098 0.2299 -0.046
and (1.034) (-0.526) (0.933)
greater
NYSE NYSE -0.3904 -0.0176 0.061 1 0.076
(-3.293) (-0.221) (0.683)

109

Table 12

Summary of Cream Skimming Regressions

 

 

 

 

 

 

 

Endings 1M9 M W W200 25.000
Boston 1 2 na 1 2 3 1 2 3 1 2 3

Cincinnati 1 2 na 1 2 3 1 3 1 3 1

Midwest 1 2 na 1 2 1 2 3 1 3 1 2
Nasdaq 1 2 na 1 2 3 1 2 3 1 2 3 1 2
Paciﬁc 1 2 na 1 3 1 3 1 3 1
Philadelphi 1 2 na 1 3 1 1 2 3 1

 

 

 

 

 

 

 

 

Table 12 shows that 29 out of 30 pairs of regressions supported the ﬁrst hypothesis while 18
out of 30 supported the second hypothesis and 21 out of 24 supported the third hypothesis.
Based on these results, it appears that increased NYSE activity has different effects
depending on the competitive position of the ﬁinge and the size of the trade. Greater NYSE
matched trades quickly erode price differences in favor of the NYSE while increases in
TOTVOL actually increase price differences in favor of the regionals, given that the number
of TRADES is held constant. These results tend to be strongest for the smallest trades.
These results seem to conﬁrm the general hypothesis that relative regional execution is
dependent on the activity of the trading environment, as measured by the number of matched
trades and total volume.

No formal hypothesis regarding the behavior of the competition measure was
included in this analysis. However, based on the results in Tables Ila-f, its impact is quite
small (particularly when compared to the TRADES coefﬁcient) and trails off very quickly

after regressions for the smallest size class.

110

D. Conclusions

This paper examined why statistically signiﬁcant price differences exist in an
industry where competition, as measured by an HHI variable or the number of exchanges that
actively trade a given stock, appears competitive. Along the way, some interesting results
were established. First, data from 1992 conﬁrmed the general results of earlier studies: both
competition measures still appear to effectively reduce NYSE spreads and liquidity
premiums. Second, when a liquidity premium measure was used in place of the spread as
a dependent variable, it did not change the general impact of the explanatory variables.
Third, explanatory variables appear to have similar impacts on prices across all exchanges,
except for the average trade Size variable. Fourth, when time frames were shortened, the
explanatory power of the regional volume and average trade size variables grew markedly
while the NYSE volume and average trade size coefficients were not affected. In addition,
the activity competition measure became superior to the alternative HHI measure. This
ﬁnding was consistent with the view of non-NYSE exchanges as a single fringe competitor.

Statistically Signiﬁcant price differentials were also found to exist. This analysis
rejected a local market explanation for this phenomenon: except for the regional volume
variable, regional explanatory variables were not systematically superior to NYSE
explanatory variables when explaining local liquidity premiums. To explain the price
differentials, a new matching technique was used. This technique captured NYSE trade
activity surrounding regional trades and established that superior regional execution occurs
in a more active trade environment. This result suggests that regionals free-ride from NYSE
activity and beneﬁt from the queuing of order ﬂow when the market is active.

This result led to an examination of cream skimming in these markets. Regressions

111

demonstrated that regional exchanges use speciﬁc NYSE trade activity (TRADES) and
general trade activity (TOTVOL) to improve their execution relative to the NYSE: their
relative execution systematically depends on market activity. This cream skimming or
selective market making behavior allows price differentials to exist while Simultaneously
permitting competition variables to appear effective in reducing prices.

Finally, this analysis also found that competition measures have relatively weak
explanatory power compared to an activity measure, such as the number of trades that are
occurring on the NYSE, when explaining price differences. This result calls into question
the traditional method of measuring competition within the securities industry. Regional
exchanges tend to trade or “compete” only in smaller Hades or when the market for the stock
is active.

Given this behavior, it is questionable whether or not the positive effects attributed
to competition between exchanges is accurate, at least in terms of price measures. It is
conceivable that order ﬂow need not be fragmented to obtain the positive impacts attributed
to competition between exchanges; volume effects and competition on the NYSE itself (from
ﬂoor traders) may be enough to reduce the price of liquidity services to a magnitude similar
to when order ﬂow is fragmented. Evidence suggests that competition does not cause prices
to fall, but rather regionals tend to trade when the market is active and spreads are tight. By
trading on the regional exchanges in this environment, traders eliminate the possibility that
they may beneﬁt from hidden liquidity on the NYSE, which appears to be the one of the
driving forces behind superior NYSE execution. If this observation is correct, then the costs
and beneﬁts associated with the fragmentation of order ﬂow Should be reexamined or

reconsidered.

BIBLIOGRAPHY

BIBLIOGRAPHY

Branch, Ben and Walter Freed, "Bid-Asked Spreads on the AMEX and the Big
Board," W95, Volume 32, No. 1, March 1977, pp. 159 - 163.

Coffee, John C., Jr., Comment on "Competition, Fragmentation and Market Quality,"
I1 1--11=..0°-.__1_ notii 111:”-.. 110 _1‘-‘ __1 ' 11.1; AndrewLo, editor,
The University of Chicago Press, 1996, pp. 78- 83.

Cohen, Kalman and Conroy, Robert M., "An Empirical Study of the Effect of Rule
19c-3." WWW; Volume 33, April 1990. PP- 277 - 305-

Cohen, Kalman J., Maier, Steven F. ,Schwartz, Robert A., and Whitcomb, David K.,
"An Analysis of the Economic Justiﬁcation for Consolidation in a Secondary Security

Market "WWVOIMC6 1982 PP 117- 136

Demsetz, Harold. "The Cost of Transacting." Wes,
Volume 82, 1968, pp. 33 - 53.

Engel, Louis and Hecht, Henry R., W, Eighth Edition, Little, Brown
and Company, Toronto, 1994.

Hamilton, James L., "Marketplace Fragmentation, Competition, and the Efﬁciency
of the Stock Exchange," W Volume 34, No. 1, March 1979, pp. 171 -
187.

"Off-Board Trading of NYSE-Listed Stocks: The Effects of Deregulation
and the National Market System," Ihglgmmalﬂiname, Volume 42, No. 5, December
1987, pp. 1,331 - 1,344.

 

Harris, Lawerence, "Minimum Price Variations, Discrete Bid/Ask Spreads and
Quotation Transparency," Working Paper 91 - 42, University of Southern California, June
1992.

Jang, Hasung and Venkatesh, P.C., "Consistnecy between Predicted and Actual Bid-

Ask Quote-Revisions," 1111mm, Volmne 46, No. 1, March 1991, pp. 433 -
446.

113

114

Lee, Charles M., "Market Integration and Price Execution for NYSE-Listed
Securities," WW, Volume 47, No. 3, July 1993, pp. 1,009 - 1,038.

Lin, Ji-Chai, Sanger, Gary C. and Booth, Geoffery G., "Trade Size and Components

of the Bid-Ask Spread," Ihe Review ef Fingeial Studies, Volume 8, No. 4, pp. 1,153 -
1,183.

Mann, Steven and Robert W. Seijas, "Bid- Ask Spreads, NYSE Specialists, and
NASD Dealers "IhelilumainfﬁcrtfdhdManagementFall 1991 PP 54- 58

McInish, Thomas H. and Wood, Robert A., "An Analysis of Intraday Patterns in
Bid/Ask Spreads for NYSE Stocks," WW, Volume 47, No. 2, June 1992,
pp. 753 - 779.

. "Competition, F ragrnentation and Market Quality,"
working paper, University of Memphis.

. "Hidden Limit Orders on the NYSE," W

 

Wm. Spring 1995, pp. 19 - 26.

. and 0rd, Keith J ., "An Investigation of Transactions
Data for NYSE Stocks, " WW Volume 40, No. 3, July 1985, pp. 724 -
739.

Mulherin, Harold, Comment on "Competition, Fragmentation and Market Quality,"
Th .10. .111, 0-31-1411, 1.1! i' 9 .01 f __1‘ ' ...I e U - in , Andrew LO, editor,
The University of Chicago Press, 1996, pp. 74 - 78.

Neal, Robert and Reiffen, David, "The Effect of Integration Between Broker/Dealers
and Specialists," The Bureau of Economics, Federal Trade Commission, March 1994.

Parrillo, Douglas E, editor, MW Probus Publishing Company,
Chicago Illinois, 1987.

Petersen, Mitchell A. and Fialkowski, David, "Posted Versus Effective Spread, " Elie
mummleume 35 1994 PP 269- 292

Reilly, Frank and Slaughter, William C., "The Effect of Dual Markets on Common
Stock Market Making," 11 - . . . -. . . ° ,March 1973.

 

Reiss, Peter C. and Werner, Ingrid M., "Transacting Costs in Multiple Dealer
Markets: Evidence from the London Stock Exchange," NBER working paper number 4,727,
May 1994.

115

Scherer. F. M. and Ross, David, W
Eezfqmenee, Houghton Mifflin Company, Boston, 1990.

Stigler, George J. and Sherwin, Robert A., "The Extent of the Market," Ihelgumel
We; Volume 28, October 1985, pp. 555 - 585.

Stoll, Hans, from Comments on "An Analysis of the Economic Justiﬁcation for

Consolidation in a Secondary Security Marke " by Cohen et al., WW
Engage, 1982, pp. 137 -140.

The New York Institute of Finance, Wing, Second Edition, Toronto,
1992.

Tinic, Seha M., "The Economics of Liquidity Services," WM
Eemmiee, Volume 86, 1972, pp. 79 - 93.

__and West, Richard F, "Competition and the Pricing of Dealer Service 1n the
Over-the Counter Stock Market," I1 - . .
1972, pp. 1,707- 1,723.

 

 

"llllllllllllllllllllllls