information flow in a fragmented dealer market- three essays on price discovery

Using the Hasbrouck information share and Gonzalo-Granger common factor methodologies, I show that ECNs provide more than half of the price discovery for approximately one out of three N

INFORMATION FLOW IN A FRAGMENTED DEALER MARKET:

THREE ESSAYS ON PRICE DISCOVERY

DISSERTATION Presented in Partial Fulfillment of the Requirements for the Degree Doctor of Philosophy in the Graduate School of The Ohio State University

By Laura A Tuttle, M.B.A

*****

The Ohio State University

2004

Dissertation Committee: Approved by

Professor Ingrid Werner, Adviser _

Professor Andrew Karolyi Adviser

Professor René Stulz Graduate Program in Business Administration

Abstract

The 1990's were a period of rapid change in the trading of Nasdaq stocks Advances in network technology improved the market's ability to trade efficiently and disseminate real-time information Concurrently, regulatory changes mandated inclusion of alternate trading venues in the quote montage, and restricted the manner in which customer limit orders are handled by market makers This dissertation explores the price formation process in the Nasdaq market, examining how fragmentation and imperfect transparency affects price formation

The first essay, “Price Discovery in Nasdaq Issues”, investigates price leadership relationships between Nasdaq market makers and Electronic Communications Networks (ECNs) Using the Hasbrouck

information share and Gonzalo-Granger common factor methodologies, I show that ECNs provide more than half of the price discovery for approximately one out of three Nasdaq 100 stocks, although ECNs trades account for less than 33% of any Nasdaq 100 issue's trading activity

The second essay, “Hidden Orders, Trading Costs and Information”, explores non-displayed (reserve) depth in Nasdaq market maker quotes in SuperSOES While the presence of hidden depth at the inside has

no effect on effective half-spreads, the information content of a trade (as measured by the midquote adjustment in the 30 minutes post-trade) is lower when reserve size is quoted, suggesting reserve size signals short-term price movements Displayed depth does not predict daily returns, but the non-displayed orders of investment banks and wirehouses are indicative of daily price changes

In the final essay, “News or Noise: Is the Price Impact of Island Trades Persistent?”, I examine the trades

on the Island ECN to discover whether their information impact is transient or permanent I measure price

impact at a number of horizons, allowing for the possibility of price reversals from liquidity motivated trades Using simple regressions, I show Island trades are more informative than other trades only at short time horizons post-trade; at longer horizons, the price impact of an Island trade is not significantly different from trades in other venues Island trades can be shown to be more informative at longer horizons only when the experimental design controls for the endogeneity of the trading venue decision

ACKNOWLEDGMENTS

I would like to express my deepest gratitude to my advisor, Ingrid Werner, for her encouragement and guidance throughout my dissertation research She has been not only an advisor, but a mentor who inspires

me both professionally and personally

I am also indebted to Andrew Karolyi and René Stulz for their innumerable ideas and suggestions to improve and extend this research I am grateful to Karen Wruck and Spencer Martin for helpful

suggestions, ideas for future research, timely advice and indispensable encouragement I thank my family for their endless encouragement and support I am deeply grateful to my colleagues for their assistance and friendship while completing my dissertation, particularly Rodolfo Martell, Boyce Watkins, Christo

Pirinsky, and Christof Stahel Words cannot express my appreciation for the friendship and encouragement

of my colleague Nicole Boyson I thank Ralph Walkling for his faith in me

I am thankful for the opportunity to visit Nasdaq Economic Research and thank Nasdaq for access to proprietary data I gratefully acknowledge the assistance of Tim McCormick, Frank Hatheway, Jeffrey Smith and the staff of Nasdaq Economic Research I also wish to thank seminar participants at Nasdaq Economic Research, The Ohio State University, the University of Dayton, Oregon State University, Florida State University, the University of Kansas, the University of Georgia and the University of Missouri (Kansas City)

v

VITA

December 15, 1969 ……… Born – Kansas City, Missouri

1990……… ………B.S., Agriculture, Kansas State University

1995 ……… M.B.A., Northern Illinois University

FIELDS OF STUDY

Major Field: Business Administration

Concentration: Finance

TABLE OF CONTENTS

Absrtact ii

Acknowledgments iv

Vita v

List of Tables viii

List of Figures x

Chapters: 1 Introduction 1

2 Price Discovery Using Quotation Data 6

2.1 Introduction 6

2.2 Cointegration and Price Discovery Measures 9

2.3 Data and Methodology 17

2.4 Results 21

2.5 Conclusion 23

3 Imperfect Transparency: the Role of Non-Displayed Depth 25

3.1 Introduction 25

3.2 Motivation and Data Description 28

3.2.1 Motivation 28

3.2.2 Description of Data 31

3.3 Methodology and Results 36

3.3.1Market depth 36

3.3.2 Cross-sectional Determinants of Hidden Depth 38

3.3.3 Trading Costs and Informational Efficiency 41

3.3.4 Hidden Size, Information and Events: Earnings Releases 44

3.4 Conclusion 48

vii

4 The Persistence of Island ECN Trade Price Impact 50

4.1 Introduction 50

4.2 Trading of Nasdaq Securities 53

4.3 Data and Methodology 56

4.3.1 Data 56

4.3.2 Methodology 57

4.4 Results 61

4.4.1Simple Regressions 64

4.4.2 Endogenous Switching Regressions 67

4.5 Conclusion 69

5 Conclusion 70

List of References 73

Appendix A 78

Appendix B 82

Appendix C 84

Appendix D 110

LIST OF TABLES

Table 1: Descriptive Statistics for Sample 85

Table 2: Descriptive Statistics for Market Maker versus ECN Quotations 86

Table 3: Trace Test for Cointegration Rank 87

Table 4: Information Share and Common Factors with Alternative Normalization 88

Table 5: Price Discovery by ECNs and Market Makers 89

Table 6: Price Discovery by ECNs and Market Makers 90

Table 7: Characteristics of Sample Stocks 92

Table 8: Near-Inside Depth for Nasdaq Stocks 2001 and 2002 93

Table 9: Inside Market Participant by Type 94

Table 10: NYSE/Nasdaq Depth Comparison Between June 2001 and April 2002 95

Table 11: Cross-Sectional Regressions of Proportion of Hidden Size 96

Table 12: Determinants of Trading Cost and Price Impact 97

Table 13: Determinants of Trading Cost and Price Impact 98

Table 14: Returns for Portfolios Classified by Displayed Bid-Offer Depth Imbalances 99

Table 15: Returns for Portfolios Classified by Reserve Bid-Offer Depth Imbalances 100

Table 16: Returns for Portfolios Classified by Displayed Bid-Offer Depth Imbalances

on Earnings Release Days 101

Table 17: Returns for Portfolios Classified by Reserve Bid-Offer Depth Imbalances

on Earnings Release Days 102

Table18: Sample Stock Characteristics 103

Table 19: Trade Characteristics 104

Table 20: Mean Post-Trade Midquote Revision by Trade Size 105

Table 21: Simple Regressions of Post-Trade Midquote Revision 106

Table 22: Probit Model Estimates 107

Table 23: Second-Stage Price Impact Regressions 108

Table 24: Simulation Results 109

LIST OF FIGURES

Figure 1: QQQ Price: May-July 2001 111Figure 2: QQQ Price: Mar-May 2002 112

This dissertation adds to the finance literature concerning price formation in fragmented markets, as well as the literature documenting Nasdaq stock trading in the wake of market reform Barclay et al (1999) describe the dramatic drop in quoted and effective half-spreads in the Nasdaq market both in response to changes in SEC mandated Order Handling Rules and negative publicity associated with the odd-eighths study of Christie and Schultz (1994) A significant portion of this improvement can be attributed to the inclusion of Electronic Communication Networks (ECNs) in the quote montage The role of these alternate trading venues in price formation is described by Huang (2002), using a similar methodology to Hasbrouck

(1995) Huang finds that ECNs are substantial contributors to price discovery in Nasdaq issues, implying that Nasdaq’s alternate trading venues attract informed order flow This is in contrast to alternate trading venues for NYSE stocks which Hasbrouck describes as “cream skimming” uninformed order flow from the primary exchange, increasing the adverse selection costs borne by the primary market Barclay,

Hendershott and McCormick (2003) further document the price formation process by comparing the realized half-spreads incurred by trades with Nasdaq market makers and those occurring on ECNs They find that at short horizons post-trade, the trades of ECNs have a larger price impact, implying that ECN trades convey more information about future price movements

The first essay, “Price Discovery in Nasdaq Issues”, investigates price leadership relationships between Nasdaq market makers and Electronic Communications Networks (ECNs) Using both the Hasbrouck information share and Gonzalo-Granger common factor methodologies, I show that ECNs provide more than half of the price discovery for approximately one out of three Nasdaq 100 stocks, although ECNs trades account for less than 33% of any Nasdaq 100 issue's trading activity Furthermore, for the majority

of Nasdaq 100 stocks, ECNs contribute substantially more information to the market per unit of share trading volume than do market makers I also present simulation results that suggest both the Gonzalo-Granger common factor and Hasbrouck's information share measures understate the information content of ECN quotes due to noise levels in ECN quotation revisions

In the third essay, “News or Noise: Is the Price Impact of Island Trades Persistent?”, I examine trades on the Island ECN to discover whether their information impact is transient or permanent Unlike previous studies, I measure price impact at a number of horizons, allowing for the possibility of price reversals from liquidity motivated trades Using an experimental design akin to that in Barclay, Hendershott and

McCormick (2003), I show Island trades are more informative than other trades only at short time horizons post-trade; at longer horizons, the price impact of an Island trade is not significantly different from a trade

in another venues Furthermore, I demonstrate that the endogeneity of trading venue selection is an

important experimental design consideration Island trades can be shown to be more informative at longer horizons only when the experimental design controls for the endogeneity of the trading venue decision

This dissertation also ties into the literature on informational efficiency and liquidity in imperfectly

transparent markets Early theoretical work on market design often begins with the classic problem of an uninformed market maker facing a potentially informed trader The market maker’s quote is akin to an option: an informed trader exercises such an option by trading against a stale quote, creating an adverse selection problem for the market makers (Copeland and Galai, 1983) Under some circumstances the adverse selection costs market participants face may be mitigated by hiding the size of their trading interest Handa and Schwartz (1996) describe a security trading market in terms of a balance of the supply of liquidity (limit orders) and demand for liquidity (market orders) The authors conjecture in their discussion that increased transparency can reduce overall liquidity in a similar line of reasoning as that modeled by Foucault and Sandas (2002)

Empirical and experimental studies suggest there is a trade-off between liquidity and transparency Both Bloomfield and O’Hara (1999) and Flood, Huisman, Koedijk and Mahieu (1999) examine transparency in experimental market economies While Bloomfield and O’Hara show that both trade disclosure and pre-trade transparency of quotations increase informational efficiency while widening spreads, Flood et al report a contrasting result where opaque markets are less efficient and have lower spreads Other studies predict that market quality (as measured by spreads, liquidity and volatility) improves with transparency (Flood et al 1999, Harris 1996, Foucault, Moinas and Theissen (2002), among others); however this result

is not always supported empirically Madhavan, Porter and Weaver (1999) describe the natural experiment afforded by the Toronto Stock Exchange’s switch to an open limit order book structure; they report an increase in spreads and volatility, as well as a decrease in depth in the wake of the market structure change Boehmer, Saar, and Yu (2003) examine the NYSE’s introduction of OpenBook in 2002, which allowed market participants to observe limit orders (a move toward increased pre-trade transparency); they

document an increase in overall liquidity and decreased execution costs

The second essay, “Hidden Orders, Trading Costs and Information”, explores non-displayed (reserve) depth in Nasdaq market maker quotes in SuperSOES Non-displayed size represents 25 percent of the dollar-depth at the NBBO in the Nasdaq 100; this appears to be additional depth provided to the market, rather than a shift away from displayed depth to non-displayed depth Market participants tend to use reserve size more for firms with high idiosyncratic risk and high volatility While the presence of hidden depth at the inside has no effect on effective half-spreads, the information content of a trade (as measured

by the midquote adjustment in the 30 minutes post-trade) is lower when reserve size is quoted, suggesting reserve size signals short-term price movements Although this information impact is present at thirty second and five minute intervals post-trade for many classes of market participants, the presence of non-displayed depth by investment banks and wirehouses is predictive of price changes up to 30 minutes post-trade Displayed depth does not predict daily returns, but the non-displayed orders of investment banks and wirehouses are predictive of daily price changes This effect is strongest at earnings releases, where only investment bank and wirehouse non-displayed depth predicts returns of individual stocks in the wake of an earnings announcement

These essays offer several contributions to our understanding of how information is incorporated into prices when markets are fragmented or transparency is imperfect Like Huang (2002), I find that the quotations of ECNs suggest a substantial role for alternate trading venues in price discovery – their quotes contribute a greater proportion of the permanent innovation to stock prices than would be anticipated by their market share of trading However, I find that the longevity of this information may be short – it may

be information related to liquidity shocks and inventory effects which we would expect to precede price reversals Although I confirm that the trades of the Island ECN have a larger price impact at short time horizons, I find that the price impact at longer time horizons is not significantly different between markets using a simple regression experimental design akin to that in Barclay, Hendershott and McCormick (2003) Recognizing, however, that the choice of trading venue introduces a self-selection bias in the sample, I correct for this bias using a two-stage procedure akin to that in Madhavan and Cheng (1997); in this framework, I show that Island ECN trades have higher price impacts at both short and long horizons This

CHAPTER 2

PRICE DISCOVERY USING QUOTATION DATA

2.1 INTRODUCTION

The late 1990's were a period of rapid change in the trading of Nasdaq stocks A number of factors

contributed to the evolution of the over-the-counter market Advances in network technology improved the market's ability to trade efficiently and disseminate real-time information Concurrently, regulatory changes

to order handling rules mandated inclusion of electronic communication network (ECN) quotes in the quote montage, and further restricted the manner in which customer limit orders are handled by market makers The literature is only beginning to explore the impact of these changes on over-the-counter stock trading

With technological innovation occurring concurrently with regulatory and structural changes (such as decimalization) during a small-capitalization share bull market, the exact causality of microstructure changes in the Nasdaq market will likely remain unknown However, although causality is elusive, there is much we do know: since 1994, quoted and effective spreads have decreased dramatically (Barclay et al., 1999); trading volume has increased from approximately 7 billion shares per month in early 1997 to 20 billion shares per month in early 20011; and a growing proportion of Nasdaq trading is originating on proprietary trading systems without the participation of Nasdaq market makers2

1

NasdaqTrader web site, www.nasdaqtrader.com

Although it can be argued that competition between ECNs and market makers should improve aggregate market quality, it is also possible that this competition results in market fragmentation, as restrictions to information or trading access in one market erodes aggregate market quality One factor leading to the ECN Display Rule3 was concern that a two-tier market was developing: institutional traders and market

professionals with access to alternative trading systems could trade at better prices than were quoted to other investors (SEC, 2000) Furthermore, some ECN quotes were available only through Bloomberg terminals or required access to private computer networks, effectively limiting information contained in their quotes to market professionals Although incorporating ECN quotes into the montage alleviated this to some degree, many ECNs remain largely inaccessible to those outside the market professional community

The role of ECNs in Nasdaq market trading has generated a great deal of interest, yet relatively little academic work has found its way into the published literature Barclay et al (1999) documents changes in quoted and effective spreads for Nasdaq stocks from 1994 to 1997, when the Order Handling Rules

changes were implemented Although quoted and effective spreads declined by approximately 30%, almost two-thirds of this improvement occurred before the OHR revisions took effect, likely in response to negative publicity and government investigations in the wake of the Christie and Schultz (1994) odd-eighths study Barclay et al further document a sharp increase in quote revisions by market makers in the wake of these rule changes, though whether this is due to the Limit Order Display Rule4 or improved public price information and competition for order flow is unclear

The intraday pattern of spreads appears to have changed as well under the new OHR Before 1997, spreads tended to be high after the open, decreasing gradually during the day, and then narrowing dramatically at

the close (Chan et al., 1995) Barclay et al show that since OHR revision, spreads narrow soon after the open as they do for NYSE issues (McInish and Wood, 1992); however, spreads continue to decrease significantly during the final 30 minutes of trading It is possible that this results from increased trading between market makers on ECNs at the end of the trading day, as participants attempt to unwind inventory positions accumulated during the trading day (Simaan et al., 1998)

The decrease in spreads and changes in intraday spread patterns in the wake of the 1997 OHR revisions raises questions about the flow of information between markets One possible explanation for changes in intraday spread patterns is that information contained in ECN limit orders now enters the aggregate market more efficiently It is also possible, however, that the inter-market trading infrastructure links created during market reform increases liquidity to ECNs, as market makers may send order flow to ECNs for execution rather than trade at an ECN quotation The effect may be to make ECNs a more attractive trading venue to informed traders, who may have preferred to trade with market makers prior to the ECN Display Rule to realize immediate execution For some traders, the benefit of anonymity provided by ECN trading may outweigh the risk of delay in execution implicit in any limit order, in the wake of improved liquidity to ECNs following OHR revision

A number of theoretical studies examine the risks market makers bear in trading with individuals who may possess superior information In the absence of informed traders and assuming that stock prices follow a Martingale process, prices are as likely to rise as they are to fall following a small trade5 When informed traders enter the market, this no longer holds: prices are more likely to move against the market maker - increasing after she sells and decreasing after she buys Easley et al (1996) present a model of this type and show empirically that the probability of trading with an informed trader is a better predictor of spreads than absolute trading volume This suggests an important role for ECNs in determining market quality by attracting informed traders away from the market maker market and reducing their risks arising from informational asymmetry

2.2 COINTEGRATION AND PRICE DISCOVERY MEASURES

Price discovery is the process by which new information is revealed by market participants and

incorporated into observable asset prices In the case of an asset that trades in multiple markets, innovations

in price can be revealed in any trading venue Consider the case of a stock whose price can be represented

as a random walk and which trades without transaction costs in two venues:

−

− Ω

When a linear combination of I(1) variables is I(0), they are said to be cointegrated (denoted CI(d,b), where

d represents the integration order of each series and b represents the reduction in integration order achieved through cointegrating relationships) In the case of our simple two-market example, the difference between the two markets is governed by a long-run equilibrium relationship that prevents prices from diverging Because each price is I(1) and their difference is I(0), they are said to be CI(1,1)

If a process is CI(1,1) and we assume that the nonstationary vector of prices can be represented as a finite order autoregressive process, it can be represented through an error correction model (ECM) of the form:

2) ∆ = pt αβ pt− 1+ Γ ∆1 pt− 1+ Γ ∆2 pt− 2+ + Γ ∆ … k− 1 pt−k+ εt

where βpt-1 is a stationary combination of lagged price levels (E[βpt-1]=0 where E denotes the expectation operator) and the remaining Γi∆p,t-i terms represent a k th-order vector autoregression of first differences In

the case where the vector pt contains two elements, the ECM can be estimated using a two-step procedure

in which the cointegrating vector β is estimated in the first step through a cointegrating regression; the

remaining coefficients can be estimated with OLS (Engle and Granger, 1987) When the vector pt consists

of more than two elements, Johansen's reduced rank regression procedure can be used to identify the number of cointegrating relationships (Johansen, 1988); the system can then be estimated in one step using

maximum likelihood estimation6 Note that the system cannot be modeled as a VAR of differences without the βpt-1 error correction term; such a model is misspecified because it does not incorporate the long-run cointegration relationship, βpt, which prevents the elements of pt from diverging

6

In the simplest case, the n x 1 vector pt has two elements and only one cointegrating relationship When there are n markets, there may be up to n-1 cointegrating relationships In the case where all markets contribute some degree of innovation to the underlying true asset price, there are always n-1 cointegrating

relationships Furthermore, with minor adjustments to capture systematic differences in elements of pt

which arise from trading costs (bid-ask spreads in the case of price series based on quotations), the

participants cannot know with certainty whether market j's price change is due to information or noise Consequently, market i will react to market j over a number of lags (for example, as traders observe that the innovation persists), or will react to the disequilibrium βpt-1 itself (recall that in equilibrium βpt-1 = 0) to adjust market i's price for information originating in market j

Although a substantial literature exists examining long term relationships of cointegrated economic

variables, the case of one security trading in multiple markets is greatly simplified by prior knowledge of

the cointegrating vectors The Law of One Price demands that in the absence of trading costs, p1 = p2 = …

=pn Thus the equilibrium relationship βpt-1= 0 when there are three markets is

A number of different approaches to attributing price discovery using the ECM representation have been incorporated into the literature Harris, McInish, Shoesmith and Wood (1995) describe price discovery structure of β is known to be β = (ι | -I(n-1) )′, where ι is a (n-1)x1 vector of ones and –I(n-1) is the opposite of

occurring on the New York, Midwest and Pacific Stock Exchanges; they show that when a regional market's price differs from the NYSE's price (an out-of-equilibrium condition where βpt ≠ 0 in Equation 2), the regional exchange's adjustment is greater in magnitude than that of the NYSE: the regional exchange adjusts its price more than the NYSE does to bring prices back to equilibrium This approach's main advantage is its clear intuition: when market 1 and market 2's prices differ, the magnitude of α terms from the ECM suggest which market bears the price discovery burden However, this methodology ignores the adjustments captured in the VAR terms Γi∆pt-i, potentially discarding the information contained in

significant lagged reactions to innovations in alternate markets

An alternate approach involves identifying the common factors in pt Stock and Watson (1988) show that if

a series is cointegrated, there exists a common factor representation of the form

pt= Aft+ εt

where ft are common factors, A is a loading matrix which captures how each price series responds to an

innovation in the common factors, and εt represents transitory effects (Gonzalo, Granger, 1995) Harris, McInish and Wood (2000) apply the Gonzalo-Granger methodology to trade data for Dow Jones Industrial

a n x 1 identity matrix The equilibrium relationship is that all prices are equal

13

Average stocks on the NYSE and regional exchanges, documenting changes in information share and trading volume share over time This methodology has an attractive basis in permanent- versus transitory-effects decomposition, but requires a simplifying assumption7, leaving it with undesirable properties: the innovations in the common factor are generally highly autocorrelated and have a significantly larger variance than the innovations in the random walk described by the Stock and Watson common trend model (Hasbrouck, 2000)

A third approach to attribution of price discovery is presented by Hasbrouck (1995) Hasbrouck introduces the information share measure which captures the variation in the underlying random walk introduced by each market Assuming that each market's price is a random walk and that they share a single common trend8, prices at time t can be expressed as

ε

=

Ψ ∑ captures the random walk common to all prices

in pt When there is a single underlying random walk, the rows of Ψ(1) must be identical The elements of each row quantify the impact of innovations in each market on the underlying shared random walk; if the

rows were different, the elements of pt (the prices in each market) would follow separate random walks After estimating the ECM, Ψ(1) (a sum of an infinite series of moving average coefficient) can be

7

The Stock and Watson common trend is a true random walk only when the common factor is a

combination of prices at all leads and lags In application, the Gonzalo-Granger common factor is a linear combination of contemporaneous prices (Stock and Watson, 1988 and Gonzalo and Granger, 1995)

8

The number of common trends is equal to the number of markets, n, less the number of cointegrating relationships When all markets contribute information, there are n-1 cointegrating vectors leaving a single common trend

approximated numerically from the error correction model's parameters Since each row of Ψ(1) is identical, Equation 3 can be expressed as

where ι is an n x 1 column of ones and ψj can be thought of as the proportion of market j's price

innovations impounded into the underlying random walk shared by all markets

Hasbrouck's information share measure is similar to a variance decomposition of the s step ahead forecast

of a stationary VAR process (Hamilton, 1994) Consider a zero mean covariance stationary vector

autoregressive process of order k with no unit roots

15

In the case of our ECM, the method of approximating the VMA representation is slightly different due to the error-correction terms included in the model The correspondence of the ECM to the level VAR is given by

Returning attention to our covariance stationary VAR process without unit roots (Equation 4), the

proportion of an s step ahead forecast of this process that is attributable to an innovation in the jth element

where p is the jth column of the Cholesky factorization of the covariance matrix of ε Hasbrouck's

methodology is closely related From Equation 3, the underlying random walk yt shared by all four prices

ε ψ ε ι

ψ ψ ψ ψ ψ ε

′

= Ω

∼

is simply the proportion of variance in the underlying, shared random walk attributable to market j's innovations Since Ω is generally not diagonal, however, we can only place upper and lower bounds on Sj This is accomplished by permuting ψ and Ω, placing the elements corresponding to each particular market

in the first and last position in turn, Cholesky factorizing each permutation The iterative Cholesky

factorizations ascribe the maximum and minimum fraction of total variance in pt to each market, allowing

us to bound the information share from above and below The range spanned by the maximum and

minimum of these factorizations is a function of what proportion of the variance of εt occurs in the diagonal elements of Ω In application, the range of the maximum and minimum information share is

off-smallest when pt is modeled with the finest feasible time resolution so the relationship between innovations

in different markets is identified in the greatest possible detail Time aggregation of data generally results in large off-diagonal elements of Ω, blurring the leadership relationships between innovations

While the information share methodology is somewhat more complex and computationally intensive than the common factor analysis suggested by Stock and Watson, it implies an underlying common trend with desirable properties The innovations in the implied efficient price incorporate the information at all lags

in the ECM and tend to have a lower variance than innovations to the Gonzalo-Granger common factor constructed as a linear combination of contemporaneous prices In this paper, I will report both measures; the conclusions regarding price discovery leadership are similar with either approach

17

2.3 DATA AND METHODOLOGY

My sample is the Nasdaq 100 during July, 1999; for brevity, I report detailed results and summary statistics for a sub-sample of twenty Nasdaq 100 stocks The sub-sample was selected by ranking the Nasdaq 100 by the fraction of trading volume occurring on ECNs, then taking the first and sixth stock from each decile Summary information is presented in Table 1 The stocks represent a wide-range of industries: technology, computer software, internet retailers, health care, and traditional retailers, among others Total market capitalization (based on closing price and total shares outstanding on July 1, 1999) ranges from $53.8 billion (Oracle Corporation) to $1.1 billion (First Health Group Corporation) The time-weighted average

of ECN participation in the inside market varies from a high of 51% (Amazon.com, Inc.) to a low of 4% (Oracle Corporation) All of the sample stocks average over twenty active market makers per day during the sample period10 Share price (which has been shown to be positively related to ECN trading volume (Smith, 1998) ranges from $12 per share (Food Lion, Inc.) to $151 per share (QUALCOMM,

Incorporated)

Quotation data is obtained from the NASD's NASTRAQ database of all quotations and trades For each stock, a continuous inside market is constructed using only ECN quotations; this provides a series of best ECN bid quotations and best ECN ask quotations A second inside market is created excluding ECNs; this provides the best market maker bid and ask quotations Quotations with a size element of zero are

excluded from consideration when identifying the best quoted price at a given time; market makers quotations with the "closed" flag are also excluded Regional exchanges and order-entry firms are not excluded from the market maker grouping, as separating quotations into ECN and non-ECN groups facilitates consideration of the impact of ECN trading on the market as a whole Furthermore, because

10

For the sample of all Nasdaq 100 stocks, all stocks average over twenty market makers quoting per day Share price ranges from $8.50 per share to $153.90 per share Market capitalization ranges from $885 million to $465 billion ECN participation in the inside market ranges from 4% to 53%

regional exchanges and order-entry firms rarely participate in the inside market, conclusions regarding price discovery are not sensitive to the inclusion or exclusion of order-entry firm and regional exchange quotations

In dividing the aggregate market into an ECN market and a market maker market, a number of question arise as to the comparability of these price series Table 2 summarizes descriptive statistics for the

quotation behavior of the market maker inside market versus the ECN inside market One readily

observable difference is the frequency of quotation updates; for all twenty stocks in the sub-sample, the best ECN quotation changes more frequently than the best market maker quotation11 This suggests ECN quotations may be noisier than market maker quotations

Because the methodological approach here is closely related to VAR variance decomposition, the unequal variances implied by the disparities in quote duration are of concern In Appendix A, I report results of a simulation in which two markets contribute known information, but have unequal noise levels The results suggest that, ceteris paribus, increased noise biases the price discovery results against the noisier market, probably the ECNs in this case

A second possible explanation for differences in quote revision frequency relates to actual depth (the number of shares that can be traded at the quoted price); the frequent price changes for ECNs reflect the limit-order-book nature of this market The third and fourth columns of Table 2 report the time-weighted quoted depth at the inside for ECNs versus market makers Market maker quotations are different in nature than ECN quotations, which are limit orders12: market maker quotations represent a commitment to trade at least as many shares as indicated in the quoted depth Market makers can and do trade in excess of this size

- thus the quoted depth in the market maker market is something of a lower bound With this caveat in

11

Quotation changes are defined as changes in the best quoted price on one side of the inside market, for the purposes of the duration statistic Automatic decrementing of quoted size does not impact this statistic

12

Some ECNs also allow participants to hide their order size, to avoid informing the market that the order

is in fact a large one Consequently, the ECN quoted depth may also underestimate the true market depth behind the order

19

mind, although there are differences across stocks, there is no clear pattern in Table 2 of one market having

a greater quoted depth than the other; for many stocks, the ECN quoted depth is as large or larger than that

in the primary market maker market

Finally, Table 2 reports the aggregated time for the month of July, 1999 during which ECNs did not post a quote for each stock13 Because the error-correction model requires both14 markets to be active, times when

no ECN was quoting the stock are excluded from the model15 Obviously, whatever price discovery

occurred during these times originated entirely in the market maker market For 67 of the Nasdaq 100 stocks, ECNs posted quotations continuously during regular trading hours for the month 11 stocks had less than 10 minutes of ECN inactivity for the month; 12 stocks had inactivity between 10 and 60 minutes 10 stocks had ECN inactivity for more than 60 minutes16 Any period of inactivity will bias the results against market makers in terms of their contribution to price discovery

I follow Hasbrouck (1995) in modeling the error-correction model with a one-second time resolution and

60 lagged VAR terms Although observing prices less frequently would greatly reduce the computational requirements and the number of ECM parameters that must be estimated, using a finer time resolution allows better identification of lead-lag relationships between quote revisions In the case of ECN versus market maker comparisons, using the finest time resolution possible is particularly important due to differences in the frequency of quote revision There are no structural restrictions (such as polynomial distributed lags) imposed on the VAR coefficients

13

Unlike registered market makers, ECNs are not required to post quotations continuously during market hours A period of inactivity indicates that the ECN did not have both a buy and sell order for the stock in question

14

More correctly, to model the ECM requires all four markets to be active: market maker bid, market maker ask, ECN bid and ECN ask If an ECN quotes only a buy order, it is for my purposes inactive as the ECN ask quotation is missing Furthermore, the ECN is inactive for sixty seconds after it resumes quoting

on both sides of the market, as sixty one-second lags are required in the error correction model

Specification of the error-correction model requires some assumption about deterministic trends in the level series and the error-correction process Osterwalk and Lenum (1992) describe an iterative fitting procedure that jointly tests cointegration rank with progressively less restrictive models, moving from a model with

no deterministic trends in the data or cointegrating relationship to models that allow for linear trends in the cointegrating vectors themselves Table 3 reports results of fit tests of a model that allows a linear trend in the data but no deterministic trend in cointegrating relationships Although tests of fit for a model with no deterministic trend in the level data are not rejected for some stocks, the less-restrictive model including a trend is not rejected at the 1% level for any stock in the sample For consistency, I include an intercept term in the ECM for all stocks

I employ two additional expedients in modeling the ECM First, as in Hasbrouck 1995, the structure of β (the matrix of cointegrating vectors) is taken as known, implying the equilibrium relationship

p ECNbid = p ECNask = p MMbid = p MMask

In reality, this relationship is violated to the extent that bid and ask quotations differ systematically due to trading costs (the spread) I assume that spreads are stationary over the sample period, and approximate the

out-of-equilibrium terms in the ECM by subtracting their sample mean Thus, I calculate the mean pMMbid

enters the ECM; doing so is akin to demeaning the data The structure of β also implies normalizing by one

of the four prices in the vector pt; I normalize by the market maker's bid quotation Normalizing by the ECN bid quotation produces almost identical results (Table 4)

In order to calculate Hasbrouck's information share measure, Ψ (1) (the matrix representing each market's contribution to the shared random walk) must be approximated This is accomplished by calculating many lags of an infinite-order vector moving average (calculated by inverting the VECM and integrating back to

GG results, the ratios of market maker and ECN common factors sum to unity by construction; the HIS measures do not18 To aid in interpretation, I report the information density as well, defined as

, 0.70

2.13 0.33

For all Nasdaq 100 stocks in July, 1999, the market maker market is the primary trading venue (the

majority of trades originate in the market maker market) Focusing for the time being on the GG

methodology, for nine of the Nasdaq 100 stocks, the ECN market is informationally dominant (the GGECN

measure is greater than 0.50) However, for stocks with GGECN measures less than 0.50, the information density measure characterizes the relationship between trading volume share and information contribution

If informative quote revision related to trading19 occurs with equal likelihood in all markets, the

information density measures should be close to one For 14 of 20 subsample stocks, the ID measure exceeds unity, suggesting that more information enters the aggregate market through ECN quotes than would be expected if informed trading occurred without regard to market For the entire 100 stock sample,

79 stocks have a GGECN information density measure greater than one This suggests that, in general, ECN trading tends to be more informed than primary market trading

The HIS measures support the informational dominance of ECNs as well The HIS lower bound measure exceeds 0.50 for 6 of 20 subsample stocks, and 34 of the 100 full sample stocks The information density

18

The HIS lower bounds sum to less than unity, the upper bounds sum to more than unity The midpoints

do sum to unity, but I report lower bounds to provide a conservative estimate

19

While there is clearly a link between trading and quote revision, the difference in nature between ECN quotes and market maker quotes comes into play here An ECN quote is a limit order, implying that the individual who placed the quote wants to trade A market maker quote reflects a willingness to trade, presumably to generate revenue as compensation for providing liquidity However, the Limit Order Display Rule (part of the Order Handling Rules revision of 1997) requires market makers to display any limit order that would improve their quotation Thus, the apples-and-oranges problem of ECN limit orders versus market maker quotations is mitigated to some degree

Referring to Table 2, recall that ECN market quotes tend to change more frequently than market maker market quotes A quotation update can be either news, noise, or a combination of both It seems safe to conclude that the ECN market is the noisy market in this application Since the GG results support the conclusion that ECNs are contributing relatively more information, it seems likely that the application is akin to the high-noise, high-news case in the simulation and thus the Hasbrouck information share measure

is a better measure of information attribution than the GG common factor measure In all likelihood, both measures underestimate the information originating in the ECN market, as we would expect as the ECN market appears noisier than the market maker market

This article has presented an analysis of informational innovation in an aggregate market when the

underlying asset trades in multiple markets Results are presented for two methodologies The Granger methodology constructs an implied efficient price that is a linear combination of contemporaneous prices in multiple markets; the innovations to the efficient price result from imposing orthogonality conditions on the error-correction terms in a Granger error-correction model In contrast, the Hasbrouck information share implies an efficient asset price constructed as linear combinations of price innovations The proportion of variance in the efficient price attributable to each market can be bounded above and below by iterative permutation and Cholesky factorization of the residual covariance matrix

Gonzalo-My results show that far more information enters the aggregate market through ECN quotations than would

be expected if information share was perfectly correlated with trading share For 30 of 96 Nasdaq 100 stocks, ECNs quotations contribute more information (more than 50% of all information) than all other Nasdaq quotations combined; none of these stocks has an ECN trading share exceeding 33% Using the information density measure (information share scaled by market share), I show that in the majority (at least 70 of 100) of Nasdaq stocks, ECN quotations contribute a fraction of information that exceeds their market share

25

CHAPTER 3 IMPERFECT TRANSPARENCY: THE ROLE OF NONDISPLAYED DEPTH

Although the Nasdaq market moved toward greater transparency with the revision of Order Handling Rules

in 1997, the introduction of SuperSOES in 2000 gave market makers the ability to post additional executable depth in their quotes that is not visible to the market as a whole This paper is the first to describe how Nasdaq market participants use this feature and measure its impact on market quality both at

auto-short horizons (up to 30 minutes post-trade), and at the daily level I also test the use of reserve size around earnings announcements to shed light upon how private information (be it fundamental to asset value or an artifact of order imbalance) is first captured in the quotes made to the market

Why would a market participant wish to hide depth? One reason might be to mitigate the adverse selection costs of the option that a market participant writes when he posts a quote Consider the classic problem of

an uninformed market maker facing a potentially informed trader The market maker’s quote is akin to an option: an informed trader exercises such an option by trading against a stale quote, creating an adverse selection problem for the market makers (Copeland and Galai, 1983) Nasdaq market makers are required

to maintain two-sided quotes during market hours, and to trade up to their quoted size when presented with

a willing counterparty Under some circumstances the adverse selection costs market participants face may

be mitigated by hiding the size of their trading interest

A second reason for market participants to hide depth may be to conceal information Although many early models began with the assumption that liquidity providers were uninformed and traded with liquidity demanders any of whom might be informed, this is surely an oversimplification A trader may acquire or unwind a position (for informational reasons or otherwise) via different order strategies and venues, depending on his desire for immediacy versus price certainty (Cohen et al 1981, Handa and Schwartz, 1996) He may submit a market order to demand liquidity, or submit a limit order and attempt to trade as a liquidity provider who earns rather than pays the spread He may also do so at different levels of

anonymity – using his quote in the Nasdaq montage or an anonymous ECN order The ability to hide size within SuperSOES is a vehicle to trade as a liquidity provider with some anonymity, albeit less than provided by an ECN The anonymity in ECNs does have a cost, however – the existence of substantial size quoted in an ECN is usually visible to the market as a whole, and is known to be informative of short-term market movements (Huang, 2002) Furthermore, ECN quotes are not auto-executable within SuperSOES:

to trade with the liquidity in ECNs requires special routing of the order Thus to some extent, the liquidity

in ECNs is less accessible to the market as a whole

27

Although using the hidden size feature of SuperSOES avoids the fragmented markets problem that may affect ECN orders, the market participant does sacrifice some execution priority in doing so In the case where multiple market participants share the inside, a market order that executes against their quotes will first exhaust all displayed depth at the best quote according to time priority Any remaining portion of the market order (that would now “walk the book” in a market with no hidden size feature) will execute against non-displayed depth in the market makers’ quotes Once the market order execution is complete, if there is additional hidden depth in a quote, it will replenish the displayed size and the market participant will have time priority (for the displayed size) for execution in preference to any market participant who posts a new quote at the inside

In this paper, I describe the use of hidden depth in the Nasdaq market and measure how it impacts the informational efficiency, overall liquidity and trading costs in the market I show that hidden liquidity accounts for 25 percent of the inside depth in Nasdaq 100 stocks Overall dollar depth in the Nasdaq market has increased 57 percent with the SuperSOES introduction (during a period when matched NYSE firms showed a decrease in displayed liquidity) The hidden depth feature is more likely to be used in stocks with a high probability of informational events and high volatility, supporting the idea of hidden orders as a vehicle for the mitigation of adverse selection costs to liquidity providers The use of hidden size has no significant effect on effective half-spreads incurred by trades; however, while displayed size conveys little information about future price movements, hidden size is predictive of future market price changes This effect is limited for most classes of market participants, but the reserve size use of

investment banks and wirehouses is indicative of price changes beyond 30 minutes post-trade

Furthermore, while displayed size has no predictive value for the current day’s returns, the aggregate displayed depth imbalance does This effect is largely attributable to the reserve size of investment banks and wirehouses, and is strongest in the wake of an earnings announcement, suggesting some degree of superior information on asset value

3.2 MOTIVATION AND DATA DESCRIPTION

3.2.1 Motivation

There are several models suggesting testable hypotheses regarding the usage and effect of hidden depth Foucault and Sandas (2002) present a model in which a risky security is traded in a market with discrete prices and a time-priority rule for execution, similar to the models of Glosten (1994), Sandas (2001) and Seppi (1997) Risk neutral traders arrive sequentially and can place a single limit order with both visible and hidden depth Noise traders place market orders which execute against the aggregate book using both time priority and displayed-depth priority (all displayed depth at a given tick is exhausted before hidden depth is filled) Later, an information event may happen, in which case informed traders arrive

instantaneously and place a market order which executes against the liquidity providers’ aggregate book

In the market described in the model, if a news event occurs, an informed trader arrives at the market He would like to buy (sell) an infinite number of shares at the best offer (bid), but can only trade up to the depth in the book The informed trader must decide how many shares he wishes to buy with a market order (the model disallows use of marketable limit orders21); if the quantity of his market order exceeds the depth

at the best price, he must purchase or sell shares at an inferior price His criterion function in deciding how many shares to purchase considers the equally-weighted average price per share The trader does not know how many total shares are available at ticks with stale prices (where he can profitably trade)

Consequently, he submits an order for fewer shares than he would if he could see the hidden depth in the book, considering the possibility that he may be purchasing (selling) some of those shares for a price that exceeds (is less than) the current security value In this manner, the liquidity providers reduce their adverse selection costs by hiding the size of their orders, since they are less likely to trade with an informed

counterparty in the wake of an informational event

21

See Foucault and Sandas (2002) for a discussion of the restrictiveness of this assumption In a market in which there is no cost to placing a marketable limit order, traders would not use hidden size in equilibrium However, execution priority rules may impose an opportunity cost on these orders, still allowing for the use

of hidden size in equilibrium

29

Foucault and Sandas describe an equilibrium in which the displayed depth in the book is the same whether hidden orders are allowed or disallowed; any hidden depth in the book is additional liquidity provided to the market because of the allowance of hidden orders There is always a strictly positive probability of hidden depth in the book, and the proportion of hidden orders (relative to displayed orders) increases with the probability of a news event Handa and Schwartz (1996) describe a security trading market in terms of

a balance of the supply of liquidity (limit orders) and demand for liquidity (market orders) The authors conjecture in their discussion that increased transparency can reduce overall liquidity in a similar line of reasoning to that of Foucault and Sandas

Rindi (2002) presents a second model of pre-trade transparency based on the models of Grossman and Stiglitz (1980) and Kyle (1989) The model features two groups of risk-averse agents, some of whom may

be informed insiders These agents submit limit orders to hedge their endowment of risky assets and possibly to speculate on information Uninformed traders observe the book and try to infer the information contained in informed traders’ orders Noise traders submit a randomly determined market order against the aggregate limit order book

Rindi characterizes the equilibrium in this model under three regimes of transparency In the

low-transparency setting, only market clearing prices are observed In the medium low-transparency setting, limit and market orders are observable, but the identity of traders is not In the full transparency setting, both orders and trader identities are observed

In characterizing the equilibria in these three transparency regimes, Rindi shows that when information acquisition is endogenous, enhanced transparency can actually reduce market liquidity, unlike in previous models in which the uninformed increase the liquidity they provide when transparency is high Because the uninformed traders can infer the informed traders’ information by observing the book, they will trade as if they were informed when transparency is high Anticipating this, traders invest less in information

acquisition activities; fewer informed traders enter the market and engage in costly information acquisition, since their information will be revealed in the transparent book, diminishing their trading profits Without

their information acquisition activities, the equilibrium has fewer informed orders, less information can be inferred from the limit order book and liquidity providers – both informed and uninformed – offer less liquidity overall Thus in equilibrium, liquidity in the fully transparent market is lower than that in less transparent setting

Both Bloomfield and O’Hara (1999) and Flood, Huisman, Koedijk and Mahieu (1999) examine

transparency in experimental market economies Glosten (1999) discusses the two papers and differences in experimental design in introductory comments While Bloomfield and O’Hara show that both trade disclosure and pre-trade transparency of quotations increase informational efficiency while widening spreads, Flood et al report a contrasting result where opaque markets are less efficient and have lower spreads Differences in experimental design, including the type of test subjects, provisions for intradealer trading, and the use of continuous trading versus discrete trading rounds may explain the differences in results and implications for what may be expected in empirically

Other studies predict that market quality (as measured by spreads, liquidity and volatility) improves with transparency (Flood et al 1999, Harris 1996, Foucault, Moinas and Theissen (2002), among others); however this result is not always supported empirically Madhavan, Porter and Weaver (1999) describe the natural experiment afforded by the Toronto Stock Exchange’s switch to an open limit order book structure; they report an increase in spreads and volatility, as well as a decrease in depth in the wake of the market structure change Boehmer, Saar, and Yu (2003) examine the NYSE’s introduction of OpenBook in 2002, which allowed market participants to observe limit orders (a move toward increased pre-trade

transparency); they document an increase in overall liquidity and decreased execution costs

These studies suggest testable hypotheses regarding the effect of the hidden depth provision of SuperSOES upon the Nasdaq market:

Hypothesis 1: Liquidity providers will commit to trade more shares if they are not obligated to reveal the

complete size of their order Foucault and Sandas’s model suggests that uncertainty about depth at the