Why do markets fragment? A panel-data analysis of off-exchange trading

Why do Markets Fragment? A Panel-Data Analysis of Off-Exchange Trading Kingsley Fong* Ananth Madhavan** and † Peter L. Swan School of Business, University of Sydney, Working Paper Current Version: March 2, 2001 Abstract We use unique data from Australia to analyze the nature and determinants of order flow fragmentation across all trades and every security traded. Our panel regression estimates shows that cross-sectional difference in off-market trading (ECNs, after-hours and upstairs trading alike) is driven by institutional trading interest (trading volume, indexation) and liquidity (bid-ask spread and market depth). At the transaction level, we study upstairs and primary downstairs block trades and find strong evidence that trade size, downstairs liquidity and a trader’s reputation affect his market selection decision. We conclude that there is significant competition between markets in highly liquid securities and their coexistence benefits those in a position to switch.

JEL classification: G14 Keywords: Fragmentation, Off-market trading, Transparency, Anonymity, Electronic Communication Networks (ECNs). * School of Banking and Finance, University of New South Wales, Sydney NSW 2052 Australia ** ITG Inc. and Department of Finance and Business Economics, Marshall School of Business, University of Southern California, Los Angeles, CA 90089-1427 † Finance Discipline, School of Business, Faculty of Economics and Business, University of Sydney (HO3), Sydney NSW 2006 Australia. Corresponding author, tel: 61 (0)2 9351 6466; Fax: 61 2 9351 6461; email: [email protected]. The comments and opinions contained in this paper are those of the authors alone. We thank the ARC Grant No. A79530183 and the Australian Stock Exchange (ASX) for financial support and the ASX and Securities Industry Research Centre of Asia-Pacific (SIRCA) for providing the data used in this study. We are also grateful to participants at the EFA Conference at LBS in 2000 for comments along with Alex Frino and Peter Skalkos. Of course, any errors are entirely our own. © K. Fong, Ananth Madhavan and Peter L. Swan, 2001.

1.

Introduction Financial economists have long recognized the importance of network externalities suc-

cinctly summarized by the Wall Street adage “liquidity begets liquidity.” In particular, the gravitation of order flow to venues with lower costs simply accentuates inter-market disparities in efficiency and liquidity, creating powerful incentives for securities markets to consolidate.1 Most established securities markets, however, experience some fragmentation2 through offmarket trading. Off-market trading, i.e., the diversion of order flow from the primary market, arises from a variety of sources. These include Electronic Communication Networks (ECNs) that operate proprietary electronic limit order books, crossing systems (e.g., POSIT) that match buy and sell orders using prices which may be determined in the primary market, the “upstairs” markets where block trades are negotiated and after-hours trading, where brokers or proprietary systems operate to match trades after the primary market has closed. In the U.S., recent growth of off-market trading, particularly that of ECNs, has attract much attention of the primary market; the New York Stock Exchange (NYSE) (2000, pp. 2-3) has drawn attention to the capturing of up to one-third of the Nasdaq Stock Exchange trade volume by ECNs and the fear that such

1

Pagano (1989a, 1989b) examines consolidation of multiple markets at a point in time, while Admati and Pfleiderer

(1988) examine intertemporal consolidation in a single market. See also Garbade and Silber (1979) and Cohen, Maier, Schwartz and Whitcomb (1982) for related arguments. 2

The NYSE (2000, p.3) defines fragmentation as “the trading of orders in several locations without interaction

among the orders”. While it is possible to place a pejorative interpretation on the term “fragmentation” such a connotation is not intended here. For example, “upstairs” and off-market trades may cease to exist if such markets were banned. Hence fragmentation can occur even in situations where the primary market gains from the existence of alternative markets.

fragmentation is not just healthy competition for the NYSE3. It could reduce liquidity on the NYSE and “thereby undermine the foundation of the equity markets”. Off-market trading is also of considerable importance to regulators and policy makers because of concerns regarding the lack of transparency away from primary markets, “creamskimming” of order flow, front-running, and the inability to expose public limit orders after the close of trading. Exchanges view off-market traders as “free-riding” off their price discovery, and as a threat to their existence because they could divert uninformed orders and increase the adverse selection problem in the primary market. From an academic viewpoint, understanding the nature of competition between established exchanges and off-market traders for the same securities is of considerable interest. As O’Hara (1995) notes, “multimarket linkages introduce complex, and often conflicting, effects on market liquidity.” In most markets, and including the U.S., empirical investigation of off-market trading is severely limited by the lack of data. Detailed information on trades away from primary markets is difficult if not impossible to obtain. Consequently, even the most basic empirical questions concerning the nature and extent of off-market penetration of an established, primary market’s order flow remains unanswered. Specifically, is off-market trading relatively more important in the more active securities or in less active securities? Why does off-market trading occur, can upstairs brokers effectively screen out informed customers? Does the co-existence of multiple markets result in more competition that compels markets to enhance its efficiency in order to survive?4

3

See also SEC (2000) and SEC (2001) for the regulator’s concerns regarding ECNs and fragmentation.

4

For theoretical issues see also Mendelson (1987), Chowdhry and Nanda (1991), Madhavan (1995), and Hender-

shott and Mendelson (2000).

2

This study uses detailed data from the Australian Stock Exchange (ASX) to examine these issues. Australia is in several respects ideal for such a study. First, like many markets, including the U.S., the ASX faces competition from a variety of off-exchange traders (crossing systems, after-hours and upstairs trading). Second, the data contains detailed information on the nature and extent of off-market5 trading across the entire range of traded securities. Third, the ASX operates an electronic limit order book as the primary market that is typical to many used around the world, and, there has been no study of off-market trading in a market with such a trading system. We find that off-market trading is concentrated in the most liquid stocks, suggesting that the gains from network externalities are capped above a certain level of trading activity. We show that cross-sectional difference in off-market trading is driven by institutional trading interest (trading volume, index inclusion), primary market liquidity (bid-ask spread, market depth, introduction of closing auction market) and the existence of derivative markets (options). At a transaction level we focus on the upstairs market, which runs in parallel with the primary downstairs market during normal trading hours, and study the market selection decision of traders and the differential price impact cost between the upstairs and downstairs markets. Our switching regression estimates show that a large trade in a stock is more likely to be executed upstairs, i.e., off-market, when the trade is large and when the primary downstairs market at the time of trade is relatively illiquid. Specifically, the probability of off-market execution is positively related to trade size relative to recent trading volume and the percentage bid-ask spread, and is negatively related to market depth and the mid-point of the bid-ask spread. Consistent with Madhavan and

5

We take the term “off-market” to include all trades not conducted in the primary market whereas “upstairs” trades

refers to a subset of off-market trades made up of large block trades. Thus smaller trades conducted after-hours are “off-market” but are not classified as “upstairs”.

3

Cheng (1997), we also find upstairs trading results in lower trade size induced price impact. Overall, we conclude that multiple markets can coexist with a primary market while offering close competition in highly liquid securities. Our results are closely related to three other concurrent pieces of empirical work in offmarket trading. Conrad, Johnson and Wahal (2001) study institutional order and execution data provided by the Plexus Group and conclude that ECNs and external crossing systems deliver lower execution costs than either traditional brokers or internal crossing systems can deliver. Barclay and Hendershott (2000) and Huang (2000) examine study after-hours trades and ECNs, respectively. They find that these markets contribute to price discovery. We proceed as follows: Section 2 reviews the theoretical explanations for off-market trading and some related research. Section 3 describes the relevant institutional features of the Australian market, our data and sample stocks. Section 4 presents an overview of the extent and pattern of off-market trading. Section 5 analyses the determinants of the cross-sectional difference in the level of off-market trading. Section 6 models upstairs versus downstairs trades using an approach that integrates market selection and price impact. Section 7 concludes.

2.

Screening versus Searching for Unexpressed Order Flow Models of trading activity (e.g., Pagano, 1989a, 1989b) suggest that markets enjoy net-

work externalities arising from the consolidation of order flow. As a result, new entrants find it difficult to compete with an established market. While many primary markets manage to retain a dominant position despite competition6, there are several cases where a primary market has lost

6

Davis and Lightfoot (1998) present evidence on the effect of competition from off-exchange traders on NYSE

stocks first listed after April 26, 1979. No reduction in bid-ask spreads or in the volatility of returns is detected. Thus

4

ground to new entrants offering better products, lower costs, and superior technology7. One advantage of off-market trading is trading hour flexibility since it is not confined to the limited opening hours of the primary market. However, trading hour flexibility, per se, does not automatically confer an entry opportunity for a new player. To the extent that traders have discretion over the timing of their trades, they may concentrate their trading (Admati and Pfleiderer, 1988) during times when liquidity is greatest, which is typically when the primary market is open. For those traders that need to trade urgently after-hours due to institutional reasons (e.g., try to trade at market closing price or are located in another time zone, etc.) or informational reasons (e.g., due to short-lived private information or overseas orders, etc.), they may have little choice but to trade off-market despite suffering potentially lower liquidity. Most primary markets, e.g. the NYSE, NASDAQ, Paris Bourse, etc., are continuous markets. An unavoidable characteristic of a continuous market, particularly that of a dealership market, is the existence of a positive bid-ask spread. Traders with a low demand for immediacy may choose to face the uncertain and even delayed execution offered via crossing systems in return for a lower bid-ask spread or none at all. Hendershott and Mendelson (2000) analyze such a

the authors do not find evidence that allowing additional competition from NASDAQ dealers or from other OTC markets made any real difference. However, as they point out, the internal NYSE exchange rules which prohibit members of the NYSE competing with specialists who make a market for a stock on the floor of the exchange for stocks listed prior to that date does not apply to NASDAQ dealers in NYSE stocks. Nor does it apply to NYSE members operating on regional exchanges who compete directly with the specialist on the NYSE. Hence all trading in stocks listed on the NYSE is “fragmented” and subject to external competition whether or not the SEC rule for stocks issued after April 1976 applies. 7

The London International Financial Futures Exchange (LIFFE), for example, eventually lost its dominant market

share in trading the German Bund futures contracts to the then newly established, automated DTB (now Eurex).

5

world and find that the introduction of such a crossing system to a dealership market has ambiguous effects on market liquidity8. Models emphasizing asymmetric information provide some rationale for the success of off-market competitors in attracting order flow from primary markets. Easley, Kiefer and O’Hara (1996) show that established markets could face new competition as a result of “creamskimming” orders that are most likely to originate from uninformed traders.9 Similarly, brokerdealers might internalize their order flow, passing on only the unmatched orders to the primary market, e.g., Roell (1990). To the extent that this is the case, cream-skimming and internalization directly compete with the primary market. Also “free-riding” off quotations from the primary market could thereby potentially degrade the price discovery process.10 “Cream skimming” removes the adverse selection risk inherent in the “pooling” of informed and uninformed orders. It enables uninformed traders to benefit from “purchased” order flow and better terms on execution. Hence uninformed traders gain better execution at lower cost. How can competition resulting in lower execution costs degrade the price discovery process? The spread in the primary market can widen as more “noise traders” and uninformed traders are diverted elsewhere. It is the uninformed traders who underwrite the process of price discovery because they are implicitly taxed by the informed. As the implicit subsidies are withdrawn by the

8

Specifically, orders of low immediacy, likely to be uninformed, are attracted to the crossing system. These cause

concentrated adverse selection and hence wider spread in the dealership market. However, the reduction in trading costs due to the introduction of the crossing system causes higher volume, hence it dilutes the adverse selection risk and reduces the bid-ask spread. 9

See Battalio (1997).

10

“Free riding” on the quotes can in principle be countered by placing a value on, and then charging for, the pricing

service while acknowledging possible practical difficulties. In fact, on the ASX upstairs brokers/traders pay the same fee per trade to the ASX, up to a maximum of $15 for very large trades, as do the brokers in the primary mar-

6

retreat of the uninformed, survival of the informed becomes more problematic, leading in the limit to the breakdown of primary markets. The reputational-based model of Seppi (1990) provides a similar justification for a major form of off-market trading, upstairs (block) trading. Exchange trading systems are generally anonymous, although floor traders and specialists on floor-based systems such as the NYSE may have some knowledge of their clients. By contrast, in off-market telephone brokerage systems, such as the upstairs markets, the broker knows the counter-party, even though the clients may remain unknown to each other. The reputation a trader gains for being uninformed may allow his large-block trades to be executed without much adverse price movement. The upstairs market is viewed as a screening device to eliminate informational-motivated trades. Madhavan and Cheng (1997) find support for this reputational model of upstairs trading using data for thirty Dow Jones stocks traded on the NYSE11. Aitken, Garvey and Swan (1995), however, find that upstairs broker-dealers are either unable or unwilling to fully detect informed traders and hence screen out informed trades. They argue that the role of the upstairs market is to make a market when the downstairs market fails to supply liquidity to the informed, instead of the uninformed, clients. They study the upstairs broker-dealers on the Sydney Stock Exchange12 and these broker-dealers consistently incur losses when facilitating trades, i.e., making a market, for long-term clients who make up the bulk of their brokerage businesses. These broker-dealers recover market-making losses on highly sig-

ket. Hence the scope for “free-riding” on the primary quotes is limited. The ASX charge for trading is exceedingly low relative to the other costs of trading such as the bid ask spread, stamp duty or brokerage costs. 11

In off-market proprietary trading systems such as Posit, the class of client is typically known (e.g., index funds)

even though the actual counter-party is anonymous. Again, the ability to screen out informed traders permit mutually advantageous off-market trades that might not otherwise occur. 12

Prior to its amalgamation with exchanges in other cities to form the ASX in 1987.

7

nificant volumes of agency trades from the same long-term clients transacted at higher than normal brokerage charges. In this way, markets function due to an implicit contract between the broker-dealers and the large clients: the clients guarantee to bring sufficient agency commissions to the broker-dealers to make up for the losses incurred in upstairs facilitating trades. Burdett and O’Hara (1987) and Grossman (1992) provide an alternative justification for upstairs markets. They argue that upstairs markets are more than screening devices to divert informed traders to the primary market. The main function of upstairs brokers is to search and match the unexpressed order flows of their clients. This unexpressed order flow can be informationally motivated as well as for portfolio rebalancing and other informationless reasons. Implicitly, it is assumed that customers refrain from the primary downstairs market for fear of exposing their trading intention to the public. This is a major concern for large traders as well as the informed. The function of upstairs brokers and dealers is to act as repositories of, and to infer and match, these unexpressed orders13. Other information models (e.g., Madhavan, 1995) suggest that in opaque markets, multiple centers of trading can coexist because large traders prefer to break up their orders across different dealers or markets. The trading hour flexibility, and the relative opacity that off-market trading confers, may also contribute to its continued survival. Thus, the literature suggests several potential reasons for the existence of off-market trading. To summarize, while the primary market may provide sufficient liquidity for a typical trader, “specialized” markets may serve some groups of traders better. These traders might de-

13

Discussions with market professionals indicate that one common technique brokers use to learn unexpressed order

is to declare that they have clients interested in trading in a particular line of stocks such that interested traders would call up the brokers to inquire. In doing so, brokers learn private information (not known by other brokers) about the source of unexpressed orders which can be useful for further searching and matching of orders.

8

mand extended trading hours, lower execution costs and the ability to trade large orders with minimal market impact. Off-market trading satisfies these desires by offering more flexible trading hours, relative opaqueness, screening for informed trades, the ability to avoid bid-ask spread, the ability to establish a trader’s reputation for lacking information and the facility to search for and match unexpressed orders. Some traders might even prefer to pay higher than normal brokerage to cross-subsidize market makers that provide liquidity for their large orders. The empirical significance of these explanations of off-market trading, however, has rarely been quantified. In the following, we use cross-sectional and transactional data from Australia to test some of these predictions. Now we turn to the institutional details.

3.

Institutional Details and Data

3.1

Trading on the ASX The Australian Stock Exchange (ASX) conducts all listed equities trading in Australia.

Since 1987 floor trading has given way to a fully computerized trading system, the Stock Exchange Automated Trading System (SEATS). SEATS opens trading in a stock with an auction algorithm between 10:00 AM and 10:15 AM14. Continuous trading commences after the initial auction and trading remains in this mode until 4:00 PM. Since May 1997 SEATS also perform a closing auction shortly after the closure of continuous trading, currently at 4:15 PM. SEATS allows both limit and market orders to be submitted. It is based on the Toronto CATS automated system and is typical of many markets around the world (see, e.g., Domowitz, 1990, 1993). Trading takes place through the intermediary services of brokers, who can trade as principal dealers to make a market or as agents in both SEATS (downstairs) and the upstairs market. Un-

9

like the NYSE where the specialist is granted both privileges and obligations to make a market downstairs but generally does not participate upstairs, ASX brokers are free to participate either upstairs or downstairs, and often in both, markets.15 As in the U.S., the primary market, SEATS, faces competition from a variety of offmarket traders including after-hours brokers, upstairs broker-dealers, and crossing system (e.g., POSIT Australia)16. SEATS does not execute trades outside normal trading hours which are, as noted, confined from 10:15 AM until 4:00PM. During the after-hours period brokers manually match after-hours orders by using SEATS as a bulletin board to post orders and negotiate with each other via telephone to confirm the terms of trade. Brokers also operate an upstairs market for block trades, known as “specials” in the ASX, both during and outside the normal trading hours of SEATS. There are two types of specials, “block specials” and “portfolio specials.” Currently, block specials are trades in one security for more than $2 million dollars in value. Portfolio specials are trades with value above $200,000 and as a component of a multiple securities order with aggregate value of more than $5 million dollars17. Specials may execute at prices outside the quotes prevailing in SEATS even during SEATS normal trading hours. However, under the rules of the ASX only specials may trade off-market during the SEATS operating hours. After the close of SEATS both specials and smaller trades may transact off-market.

14

The exact open time is stock specific. The system randomly picks a stock from an ascending alphabetic group of

stocks in different time intervals. 15

Other than maintaining an orderly market (which is not explicitly defined), and not charging brokerage to private

clients with whom they deal as principal, unofficial market makers, have no special rights or obligations. There are, for example, no affirmative requirements to make a market or provide price continuity (as are NYSE specialists) if they choose not to, and they will usually only make a market for known clients. 16

Note that the primary (main) market SEATS is already a centralized electronic limit order book therefore there is

no ECN, such as ISLAND, that operates in Australia. 17

This rule has been effective since October 14, 1996. Lower thresholds applied previously.

10

The ASX mandates that all specials executed during SEATS normal trading hours must be reported to the exchange via SEATS as soon as possible. Hence the reporting arrangements are in principle as timely and onerous in terms of transparency as in the downstairs market SEATS. There are no delays allowed as on the London Stock Exchange. However, the requirements after hours are less onerous, specials and other trades are required to be reported, but no later than the opening of business the next morning. One institutional feature that has already been alluded to requires special mention. The ASX introduced an anonymous closing auction in May 1997. There are no restrictions on who may trade in this market. Currently, the closing auction takes place at 4:15 PM and the auction algorithm is the same as the opening auction algorithm. Traders submit their limit orders and SEATS matches overlapping orders and execute them at volume weighted average prices. The introduction of the closing auction is of special interest to us because it is an exogenous change in the trading rule with the objective of improving downstairs market depth and it has the potential to divert off-market trades back onto SEATS. Significantly for our purposes, no attempt was made to match the non-anonymous nature of trading pertaining in the upstairs market. Similar changes were introduced to the Paris Bourse. Specifically, Hillion and Suominen (1998) study the closing prices of the CAC 40 stocks on the Paris Bourse and find evidence of reversals in the overnight period and also higher volatility and spreads at the close. These problems prompted the Paris Bourse to implement a closing auction.18 An auction that aggregates all limit orders and executes them at some market-clearing price at the close of trading could alleviate the liquidity pressure towards the end of day. Indeed, Thomas (1998) finds that orders are larger and

18

Meier (1998) surveys 49 leading stock markets including the NYSE, Nasdaq, London Stock Exchange, Paris

Bourse, and Frankfurt Stock Exchange. He finds that at year-end 1997, 35 (71%) exchanges used special procedures to open while 12 (25%) use special closing procedures.

11

there are fewer cancellations after the closing call was implemented. She concludes that these findings provide evidence that the auction procedure leads to more efficient price discovery and less gaming at the close. Since downstairs liquidity is likely to be improved by such an auction procedure, we expect its introduction to the ASX to be associated with a reduction in off-market trading. 3.2

Data The data examined in this study is exceedingly comprehensive in that it consists of all

trades in all ASX stocks from January 1993 to December 1998. Transaction data is extracted from the SEATS database of the ASX, which contains a complete record of every trade entered into the SEATS system. The data are unusual in that they contain a field that identifies whether a transaction was executed off-market or not. The ability to directly identify off-market transactions across all stocks is the major distinguishing characteristic of this database. By contrast, previous studies (e.g., Keim and Madhavan, 1996, Madhavan and Cheng, 1997) focus on limited subsets of stocks or use indirect methods to identify trades matched away outside the primary exchange. Using the SEATS data, we screened every transaction record in every stock in the sample period. In order to compute trade size and trade frequency correctly, SEATS trades that were executed across multiple limit orders are aggregated using the corresponding order reference numbers. This identification of off-market trades and exact recreation of the underlying orders is not possible using public data for U.S. exchanges such as the TAQ data. The sample consists of a total of 25 million aggregated trades. Corrections were made for canceled or error records and checks made for data errors such as dropped or missing digits.

12

3.3

Sample Characteristics Between 1993 and 1998, approximately 1,771 stocks ranging in size from $5 million to

$37 billion are traded.19 There are 1,702 stocks for which we can obtain both transaction and market capitalization data. In any particular year, approximately 1,200 stocks are traded and there are 819 stocks that were continuously traded from 1994 to 1998. However, it is worth emphasizing the concentrated nature of trading volume in the market; the top 100 stocks ranked by dollar trading volume account for almost 87% of the market by trading volume. Table 1 presents descriptive statistics for the stocks in our sample period from January 1994 to December 1998. The stocks are grouped into deciles by dollar trading volume at the end of the previous year, where the decile cutoffs are calculated each year. Therefore statistics for 1993 are not reported in this table.20 Table 1 shows the average daily dollar trading volume, dollar trade size, stock price, percentage bid-ask spread, number of trades and number of days traded per year. These values are computed by averaging the stock specific value across all stocks within each decile. The sample encompasses a wide range of trading characteristics in terms of trading frequency, volume, average stock price and the percentage bid-ask spread. In particular, the average bid-ask spread is significantly negatively related to trading activity. The variation in spread is extraordinarily wide, ranging from an average of 0.75% for a top decile stock, to 3.05% for the third decile and 20.01% for the bottom decile. Hence the range of bid-ask spreads is extreme and reflective of unusual illiquidity for the stocks with the widest spreads. This extreme illiquidity for the smallest stocks may simply be a reflection of the large number of stocks listed in a relatively small market.

19

The total dollar volume across all stocks (market volume hereafter) has more than doubled over the six-year pe-

riod from less than 90 billion Australian dollars 1993 to more than 220 billion Australian dollars in 1998. 20

However, data for 1993 is used in all subsequent analysis. Hence we make full use of all six years of data.

13

4.

The Penetration of Off-Market Trading Table 2 reports statistics on the extent of off-market dollar trading volume over time

across stocks in different dollar volume deciles. Panel A is computed using both trading hours and after-hours off-market trades. Each cell represents the average (across stocks within each decile) percentage of dollar trading volume that took place off-market. The table reveals a strong cross-sectional pattern. The proportion of off-market trading is increasing in the total dollar volume. For instance, in 1996 the average proportion of off-market trading in the top volume decile group is 34.63% and the corresponding value of a stock in the tenth volume decile group is only 4.73%. This declining pattern of off-market involvement is observed every year. This represents cross-sectional evidence against the proposition that clients can credibly signal their lack of information via reputational means, especially in highly illiquid lower decile securities, which suffer from the endemic problem of informed trading and high bid-ask spreads. It would appear as if this problem is so severe for some stocks that off-market trading virtually ceases to exist for all practical purposes. If one can establish a reputation as an uninformed trader, the value of this reputation should rise progressively for lower decile stocks for which high bid-ask spreads and informational asymmetries become progressively higher. However, this cross-sectional pattern can also be due to the ASX trading rules on “specials” that was alluded to earlier. Specifically, only trades of large dollar value can be executed as specials, and few trades in the low decile stocks would meet this criterion21.

21

In order to explore the choice of trading venue across stocks further, we also divided the trading hour volume into

four categories each year: normal SEATS trades, SEATS crossings (the same broker act as both buyer and seller), the opening auction and the closing auction (table not shown). The most striking result was the very high ratio of dollar trading volume in the 10th decile that takes place at the opening auction. This ratio ranged from 5.97% in 1995 to 8.34% in 1997. By contrast, for the first decile the opening auction accounts for merely just over 1% in every year, except 1998 when it rose to 1.92%. These findings are highly supportive of Admati and Pfleiderer

14

The other pattern in the table is that there is a decline in the off-market trading ratio through time, particularly for stocks in the top few volume deciles that generate more than 95% of the dollar trading volume of the entire market. The proportions of off-market trading in 1997 and 1998 are lower than in previous years. This drop in the off-market ratio could be due to two factors. First, the ASX introduced a closing auction market in May 1997, which has a potential to divert orders from the upstairs market to the primary downstairs market. Second, prior to October 14, 1996, the threshold for upstairs trading was lower, $1 million for “block specials” and $100,000 for “portfolio specials”. The rule change directly reduces the number of trades eligible for upstairs trading during the normal SEATS opening hours22. This second factor also explains the relatively larger reduction in off-market trading in higher deciles stocks because off-market trading via the upstairs block market is more likely for this group. Panel B of Table 2 presents data on the cross-sectional and time-series pattern in offmarket trading conducted after-hours (both smaller trades and upstairs trades). The statistics show that after-hours off-market trading is a very significant part of off-market trading activity. For the top decile stocks, after-hours small trades and upstairs block trades account for 14.6% to 18.2% of annual total volume between 1994 and 1998, which represents 40% to 50% of all offmarket trades in each year (the ratio of the corresponding percentages in Panel B to those in

(1988) who predict that trading in stocks with considerable adverse selection problems will be concentrated at the one time when the market is most liquid. Batch trading in the form of auction seems to offer considerable advantages for such illiquid stocks. 22

Recall that only specials can be executed during SEATS trading hours. Smaller trades can be executed off-market

only after-hours. We also examine the change in off-market trading dollar volume ratio amongst the top decile stocks in more detail. The drop in the ratio is more pronounced at the tail end within the decile. This suggests that for larger stocks, the change in threshold upstairs trade size leads to a regrouping of large orders upstairs to meet the new rule. Whilst for the less actively traded stocks, where regrouping orders to create larger sized trades is more likely to be difficult, some previously upstairs orders are sent downstairs.

15

Panel A). It appears that trading hour flexibility appeals to both large and small traders in a significant way, which might be due to trade timing inflexibility, the ability to trade at a time with possibly lower adverse selection, and, the ability for the trader to condition on the pricing information for the whole trading day. In summary, our findings suggest that competition from off-market trading for the primary market is relatively uniformly concentrated in the most active stocks, but sharply decreases in less actively traded stocks. However, while tabulating off-market penetration across trading volume provides us with some interesting insights, a more detailed understanding of the determinants of off-market trading and assessment of theories calls for the examination of other factors simultaneously within the framework of a model.

5.

Determinants of the Level of Off-Market Trading

5.1

A Panel-Data Model of Off-Market Trading With a large cross-section of stocks with varying characteristics, it is tempting to exploit

this advantage in our study of off-market trading. Here we construct a panel regression model of the determinants of off-market trading. The literature posits off-market venues as competitors to the primary market with the ability to discipline poor performance. Based on our discussion in section 2, we expect the proportion of off-market trading within a stock in a year to be related to factors such as the demand for large trades, as measured by dollar trading volume (+), index inclusion (+) and market capitalization (+) (proxies for the trading demand by institutional investors). Primary market illiquidity as measured by a higher percentage bid-ask spread (+) and limit order book depth (–) should also be important. As should the negotiating and contracting diffi-

16

culty23, measured by intraday stock price volatility (–); competition from the derivative market, measured by the listing of options (–); and finally, the introduction of closing auction (–), and change in upstairs trading rule (–) at the ASX. Putting these factors together we have the following panel regression model: offi,t

=

ai + α1 lvoli,t + α2 lmcapi,t + α3 indexi,t + α4 pspdi,t + α5 ldepthi,t + α6 strpri,t + α7 optioni,t + α8 auctioni,t + α9 rulet + eit

(1)

where, offi,t

the ratio of off-market to total dollar trading volume of stock i in year t,

lvoli,t

the log of total dollar trading volume of stock i in year t,

lmcapi,t

the log of average daily market value of the issued shares of stock i in year t,

indexi,t

the proportion of trading days in year t that stock i is included in the index,

pspdi,t

the median daily average percentage quoted bid-ask spread of stock i in year t,

ldepthi,t

log of the median daily average dollar value of the best two limit bid orders and the best two limit ask orders of stock i in year t,

stdpri,t

the median daily ratios of the transaction price standard deviation to the average price of stock i in year t,

23

This is because intraday price volatility is expected to be positively related to the information revealed through

trading. It makes the negotiation of prices off-market, and the verification of informed and uninformed trades difficult. For instance, a trader or market maker can infer the information content of a trade by observing the subsequent price movements. A small post-trade price movement is prima facie an indication of an uninformed trade, while a large post-trade price movement would suggest the opposite. When prices are volatile, there is a higher chance of observing large price movement after an uninformed trade, making verification of actions in models like Seppi (1990) more difficult.

17

optioni,t

the proportion of trading days in year t that there are exchange traded options listed on stock i,

auctioni,t

the ratio of closing auction to total dollar trading volume for stock i in year t,

rulet

the proportion of year t that is after October 14, 1996,

ei,t

is the error term for stock i in period t.

We estimate this fixed effect panel data model with stock specific intercepts, ai, and common slope coefficients. The stock specific intercept captures the stock specific off-market trading activity due to stock heterogeneity and omitted variables that is not already captured by the explanatory variables. 5.2

Model Estimates for Off-Market Trading The sample used to estimate equation (1) consists of 1,573 stocks with 5,864 stock-year

observations where there is sufficient data to calculate all variables. Since some stocks are not frequently traded, the model is also estimated over a sub-sample of active issues, consisting of stocks traded for more than 240 days a year. There are 513 stocks that pass this criterion in various years with 1,334 stock-year observations. Table 3 contains summary statistics of the variables in equation (1). The dependent variable offi,t has a mean value of 0.149 across all observations and 0.263 in the sub-sample. This significant increase in mean value is expected since high dollar trading volume stocks have higher off-market dollar trading volume ratios (Table 2). The median and quartiles statistics further illustrates the difference across the two samples. The median value of offi,t for the active issues is more than five times that of the full sample, 27.1% and 5.1% respectively. Similarly, there

18

are also noticeable differences in trading volume, market capitalization, index inclusion, bid-ask spread, limit order book depth, and option listing. By contrast, the difference between the two samples in terms of intraday price volatility, stdpri,t, and the proportion of volume executed at the closing auction, auctioni,t, are similar. It should also be noted that the small mean value of auctioni,t (relative to the approximately 6.5% for after-hours trading with the 5th decile stocks) suggests that the closing auction does not appear to be particularly attractive. This is despite it potential for improving liquidity and market depth at the close of trading. Table 4 reports the estimation results for equation (1). With a diverse cross-sectional sample, heteroscedasticity of the errors, ei,t, is expected. Hence the t-statistics (in parentheses) are computed using White heteroskedastic-consistent standard errors. Panel A contains the results of the fixed effect model. The goodness of fit shown in Panel A is encouraging. The adjusted Rsquared indicate that 55% and 72% of the variations in offi,t in the full and sub-sample, respectively, are explained. The p-value reported is for the F statistic that tests the null hypothesis that the individual stock intercepts, ai, across all stocks are equal. This is rejected at the 1% level in both samples24. For the trading demand hypothesis, i.e., the demand for large trades, the coefficients on lvoli,t are consistently positive and significant at the 1% level. This suggests that the higher is the trading demand (which is known to be strongly positively correlated to trade size, see e.g., Harris (1994) and Brennan and Subrahmanyam (1998)), the higher the proportion of the dollar volume that is executed off-market. The coefficient estimates on indexi,t are positive, hence provide support for the trading demand hypothesis, but they are only statistically significant at the 5% level

24

The F-statistics for the two samples are F(1572,4282)=2.07 and F(512,812)=3.43. A model with plain OLS

pooling without the fixed effect adjustment would explains 40% and 44% of the variations in off-market dollar volume ratios for the full sample and active issues, respectively.

19

in the active issues sub-sample. The coefficient estimates on the market capitalization variable, lmcapi,t, however, are of the opposite sign to the prediction, but only statistically significant at 1% level in the active issues sub-sample. This perverse effect of lmcapi,t on offi,t and the weak statistical significance of indexi,t may be due to conflicting time-series and cross-sectional relationship between these variables and off-market trading. Firm size proxies and index inclusion are expected to be highly correlated with institutional and trading demand cross-sectionally (Brennan and Subrahmanyam (1998)). That is, cross-sectional institutional interest and trading demand are positively related to off-market trading. However, the change in these factors over time can have a negative effect on off-market trading since these changes can be positively related to liquidity in the primary market. For instance, a stock that has become more actively traded is also more likely to experience a lower bid-ask spread and greater depth. The net effect of these variables in a panel model could be ambiguous. There is strong evidence for competition in liquidity provision between the primary market and off-market venues. The results suggest that other factors being equal, the poorer the liquidity in a stock, the higher the proportion of dollar trading volume occurs off-market; the coefficient estimates on pspdi,t are consistently positive and those on ldepthi,t are consistently negative, and all are statistically significant at the 1% level. On the face of it, this suggests that the off-market trading should attract relatively illiquid stocks with high bid-ask spreads and low limit order book depth. However, as we have observed in the previous section, this tendency is swamped by the dollar volume effect, tilting the off-market trading towards highly liquid stocks. The empirical relationships between off-market trading, trading demand and primary market liquidity suggests the following: High trading demand is fundamental to a significant amount of off-market trading. Without the existence of a demand to trade large volumes, the

20

primary market supplies sufficient liquidity to satisfy traders’ need. When traders demand to trade large volumes, the primary market may not provide sufficient liquidity (i.e., high spread and shallow depth) to some traders. These traders seek alternatives and, consequently, trade offmarket. Poor liquidity in the primary market can be a symptom of an inefficient trading mechanism but also lack of trading interest25. Therefore poor liquidity per se is not sufficient for offmarket trading to take place. The coefficient estimates on stdpri,t are unanimously negative and statistically significant at the 1% level. This result supports the contracting difficulty hypothesis implied by reputation models like Seppi (1990). Specifically, large price volatility leads to difficulty in negotiating the price off-market, and verifying that a trader is uninformed. Hence high volatility reduces offmarket trading volume. There is weak evidence that derivative markets ease the liquidity pressure in the underlying stock market, hence reduces off-market dollar trading volume ratio. The coefficient estimates on optioni,t are negative and significant at the 5% level. The coefficients on the closing auction market variable, auctioni,t, are negative and significant at the 1% level. The magnitude of the coefficient for the full sample suggests that the substitution between trading at the closing auction and off-market is a ratio of 1:0.974, and statistically not different from 100% substitution at the 1% level. For the active issues sub-sample, the estimates suggests that a 1% increase in closing auction dollar volume ratio even leads to a reduction of 4.114% in off-market trading. The high substitutability between off-market trading

25

The lack of trading interest, i.e., low trading volume and infrequent trading, by definition means there are fewer

limit orders posted in an order driven system like the ASX SEATS. Even if there are market makers, lack of trading interest implies high inventory holding cost as well as high adverse selection costs, which also lead to high bid-ask spread and shallow depth.

21

and closing auction, however, should be interpreted with caution. It applies only to a small class of traders, since the mean percentage off-market dollar volume is 20 to >50 times the mean percentage of the dollar volume executed in the closing auction26. Finally, there is some evidence that the change in the upstairs trading rule in 1996 reduces the ratio of off-market trading to total dollar volume. The effect of this market wide event is more readily observed in the full sample than in the active issues sub-sample. The coefficient estimates on rulet are all negative but only statistically significant in the full sample. This is consistent with the intuition that the new rule of increased threshold upstairs trade size is more restrictive to the lower volume stocks than the active stocks. 5.3

Cross-Sectional Models of Off-Market Trading As noted there is a potential conflict between the cross-sectional and time series relation-

ship for offi,t and some variables such as lmcapi,t and indexi,t, therefore we also estimate panel models similar to equation (1) but using a purely cross-sectional specification. Panel B contains the estimates of such a model with only one observation per stock, where all regression variables are averaged across years. The estimates of cross-sectional models reinforce the strong results found in the fixed effect model. The adjusted R-squared of the model is very high, explaining 56% and 61% of the cross-sectional difference of offi,t in the full and sub-sample, respectively27. Panel B shows that the coefficient estimates on lvoli,t and pspdi,t are positive and those on ldepthi,t, stdpri,t, and auctioni,t are negative, and all of them are statistically significantly at the 1% level.

26

Adjusting for the shorter history of the closing auction still does not change the qualitative conclusion.

27

Note that these high R-squared are achieved without any stock specific intercept since this is a pure cross-sectional

model.

22

Confirming our suspicion of a conflicting cross-sectional and time series relationship between variables, Panel B shows that lmcapi,t and indexi,t are positive and statistically significant at the 1% level, as predicted by the trading demand hypothesis. Larger stocks that generally attract institutional interest, and stocks included in the index, have a higher share of their trading executed off-market. The effect of firm size on off-market trading among the already actively traded stocks is, however, not statistically significant. There are other differences between the fixed effect and cross-sectional model. The coefficient estimates on optioni,t have increased in statistical significance. This change suggests that the negative effect of option listing on offi,t is more related to stock specific factors, rather than a pure substitution (over time) between the trading volume in the options and underlying markets. Finally, the coefficient estimates on the rule change variable, rulet, are now positive, and statistically significant at 1% for the active issues. While the sign of the coefficient estimates on this variable has become perverse, this should not be a major concern since in a cross-sectional setting the difference in the average value of this variable also represents the difference in the surviving period. The significant positive effect of rulet in the active issues sub-sample could be interpreted as stocks that survived after the rule change have a higher off-market volume ratio. This still can be consistent with the rule change having a negative effect on off-market trading in a time series sense. As a further robustness check, we also estimate equation (1) using individual year data, as well as, a fixed effect model using year specific intercept. The qualitative results for the applicable28 variables are consistent with those presented, hence they are not reported here.

28

Not all variables are applicable in year specific model, e.g., rulet.

23

6

Transaction Analysis during Normal Trading Hours The panel model in the previous section provides an informative analysis of determinants

of the level of off-market trading across stocks and years. However, since tens of millions of market transactions are summarized into annual data, much valuable information about the timing of trades, market selection, price impact, etc., are lost. In order to exploit this detailed information so as to learn more about off-market trading, we estimate a transaction level econometric model based on Madhavan and Cheng (1997) to study the market selection and price impact cost of block trades. Since block trades are large trades and are potentially used by informed traders (see Easley and O’Hara (1987)), they can incur a large price impact cost. Hence focusing on them provides a strong test for reputation models of off-market trading. However, for a choice to exist between markets, traders must be able to transact in both the primary market and off-market venues. Therefore, in the following section we focus on a sub-set of off-market trades that could have been executed in the primary market, namely block specials that were executed during normal trading hours.

6.1

A Market Selection Model Based on Expected Price Impact Cost During SEATS trading hours, a trader with a large order can either submit the order to

either the primary downstairs (electronic) market or the upstairs (broker-dealer) market. As mentioned in section 3, block specials are single stock (as oppose to portfolio) trades in the upstairs market with a value of or over $2 million per trade. They can take place at any time during

24

SEATS trading hours (or even after-hours), and at prices outside the best quotes provided by the electronic limit order book29. Being able to trade a large volume at a lower execution cost is a major motivation to trade upstairs. In the following we assume that the difference in expected execution cost determine a trader’s market selection decision. Conceptually, execution cost consists of various types of trading cost, e.g., timing cost, opportunity cost and price impact cost. Empirically, only the price impact cost component, defined as the difference between the trade price and the mid-point of the pre-trade bid-ask spread, is directly observable. Easley and O’Hara (1987) suggest that price impact of a trade in the primary downstairs market is an increasing function of order size. Downstairs market realized price impact cost is modeled as follows: cdt = βd Xi + ε di, ,

(2)

where cdt is the price impact cost of the downstairs trade i, Xi is a k × 1 vector of explanatory variables consisting of a constant and trade size, and ε di is a disturbance term that captures idiosyncratic factors affecting the price impact of the trade. Studies of upstairs intermediation by Burdett and O’Hara (1987), Keim and Madhavan (1996) and Seppi (1990) suggest the price impact cost of a trade in the upstairs markets is also increasing in order size. Unlike the downstairs market, the identity of the initiator is revealed to the upstairs block broker through the negotiation process. Seppi (1990) shows that traders who can credibly signal to an upstairs broker that they are not informed will obtain a lower cost of trading in the upstairs market. Thus we model the realized market impact cost in the upstairs market as:

29

Except during a trading halt induced by takeover announcement. Since specials are large trades, if they were to be

executed in the limit order book, they would have caused price movements of at least a couple of ticks, therefore allowing them to be executed at outside quotes is not unreasonable.

25

cut = βu Xi –θ i+ ε ui ,

(3)

where cut is the price impact cost of the upstairs trade i, Xi is a k × 1 vector of explanatory variables, and ε

u

i

is a disturbance term that captures idiosyncratic factors affecting the price impact

cost of the trade. The signal θ i summarizes the broker’s information regarding the initiator’s reputation. Seppi (1990) suggests that, other things being equal, the price impact of a trade decreases with the probability that the trade is liquidity motivated, hence θ i is assumed to enter the equation negatively. However, θ i is not directly observable from the data, therefore econometrically equation (3) has a pooled error term, ξui , where ξui = ε di –θ i. Let Ωi denote the information set of the trader of trade i (for simplicity, we shall refer to this trader as trader i hereafter). The difference between the expected execution cost of trading downstairs relative to upstairs for trader i can be expressed as: u*i = α Wi + E [ cdt – cut | Ωi ],

(4)

where α is a vector of constants and Wi is a vector of state variables that capture other unobserved execution cost differentials. Substituting the price impact equations (2) and (3) into equation (4), and assuming for simplicity that E [ε di | Ωi ] = E [ε ui | Ωi ], we have u*i = α Wi + ( βd – βu ) Xi + θi.

(5)

Equation (5) is the criterion function for trader i, which can be written as u*i = γ Zi + θi,

(6)

where Zi = ( Wi , Xi ), and γ is a vector of coefficients. Note again that θi is a random variable from the econometrician’s viewpoint because the trader’s reputation is not observed. We assume that θi, ε

d

i

and ε

u

i

are jointly normally distributed with mean 0 and variance-covariance matrix

Σ. 26

Let ui represent the trader’s choice of venue, where ui takes the value of 1 if trader i selected upstairs and 0 otherwise. The trader’s decision rule is: ui = 1 if u*i > 0, ui = 0 if otherwise.

(7)

The econometrician observes ui = 1 if u*i > 0 and ui = 0 if otherwise. The market selection model consists of the price impact cost equations (2) and (3), the criterion function (5), and the decision rule (7).

6.2

From Market Selection to Switching Regression The estimation of the market selection model is complicated by the unobservable reputa-

tion signal, which affects both the choice of market as well as the price impact cost. Such missing variable and correlated error terms lead to inconsistent Ordinary Least Squares (OLS) estimates. Since we observe upstairs trades only when u*i >0, the error term ε ui does not have a zero mean conditional on there being an upstairs trade. Using the properties of the normal distribution and normalizing the variance of θi to be 1, the expected price impact cost conditional on observing an upstairs trade is E [ c u i | u i = 1] = β u X i + σ u [

φ (γ Z i ) ], Φ (γ Z i )

(8)

where φ (.) denotes the standard normal density function, Φ (.) denotes the cumulative standard normal distribution, and σu is the covariance between the composite disturbance term in the upstairs market price impact regression equation and the θi term in the criterion function. Similarly, E[ c d i | u i = 0] = β d X i + σ d [

− φ (γ Z i ) ], 1 − Φ (γ Z i )

27

(9)

where σd is the covariance between ε di and θi. The model places restrictions on the covariance in the above expressions, which can be used to test theories of upstairs trading. Using the definitions above, we have

σu = cov[ – θi + εui , θi ] = – var [θi ] + cov[εui , θi ].

(10)

The first term in equation (10) is unambiguously negative. Intuition suggests that the second term will also be negative. Specifically, a trader that can signal being uninformed, i.e., with θi > 0, is less likely to demand urgent execution. Keim and Madhavan (1997) show that traders requiring urgent executions (e.g., technical or momentum traders) incur a higher trading cost than traders that do not need to trade as urgently (e.g., index funds). Since the disturbance εui captures this unobserved urgency of the trade which is negatively correlated with the reputation signal, cov[εui , θi ] is expected to be negative. Since both the first and second terms are negative, it follows that σu < 0. As σu < 0, equation (8) shows that the expected upstairs price impact conditional on observing an upstairs trade is smaller than the unconditional expected upstairs price impact. Conversely, it is argued that traders with below average reputation signals are more likely to use more aggressive orders, choose downstairs trading and incur higher price impact costs, hence cov[εdi , θi ] = σd < 0. Therefore, the conditional expected price impact of a downstairs trade is higher than the unconditional expectation. Across the entire sample of upstairs and downstairs trades, observe that the expected price impact cost is: E[ci]

= E[ci | ui = 1] Pr[ui = 1] + E[ci | ui = 0] Pr[ui = 0] = βd Xi + (βu – βd) Xi Φi + φi (σu – σd),

(11)

where all terms are defined as above. Note that the second and third terms contain unobserved explanatory variables, the probabilities Φi and φi. Lee, Maddala, and Trost (1979) propose a sim-

28

ple two-stage procedure that produces consistent parameter estimates of this type of model with selectivity bias. It involves the use of the probit and maximum likelihood methods to estimate of the probabilities (probability density function and cumulative probability distribution) of upstairs trading. These probabilities are then used as explanatory variables to estimate the price impact cost equation (11), which is equivalent to simultaneously estimating equations (2) and (3) using all observations. 6.3

Estimation of the Probabilities of Upstairs Trading Estimation of the criterion function, equation (6), requires the specification of factors in-

fluencing the market selection decision. We write the variable Xi in Equations (2) and (3) as Xi = [1, sizei], where sizei is a measure of trade size. The expected price impacts in the upstairs and downstairs markets takes the form βu0 + βu1 sizei and βd0 + βd1 sizei, respectively. Other factors that can affect the market selection decision include market liquidity and the spreading of fixed negotiation cost. At times when the downstairs market is illiquid, which can be proxied by bidask spread and depth of the limit order book, it may not be possible to execute large trades downstairs. Madhavan and Cheng (1997) argue that the level of the stock price captures this effect, and the higher the dollar value of a trade, the smaller is the percentage cost of negotiation. The continuous criterion function now can be written in the form: u*i = γ Zi = γ0 + γ1 sizei + γ2 pspdi + γ3 ldepthi + γ4 mqi,

(12)

where sizei is the number of shares of trade i divided by the average number of shares traded in the immediate prior twenty trades (to control for heteroskedasticity across stocks and time), pspdi is the percentage spread immediately prior to trade i, defined as the ratio of the bid-ask spread to quotation midpoint times 100, ldepthi is the natural log of the number of shares available in the best two price steps of the limit order book (bid orders for sell trades and ask orders for buy trades), and mqi is the quotation midpoint immediately prior to trade i. 29

Theoretical studies suggest that upstairs market may reduce the marginal price impact cost of order size relative to the downstairs market by mitigating adverse selection problems (Seppi (1990)), risk sharing (Keim and Madhavan (1996)), or locating potential trade counterparties (Burdett and O’Hara (1987) and Grossman (1992)). This implies that βd1 > βu1, and from equation (5), γ1 = βd1 – βu1 > 0. As the upstairs market may be preferred to downstairs trading when the downstairs market is illiquid or when dollar value of a trade is high, we expect γ2 > 0,

γ3 < 0, and γ4 > 0. The estimation is based on a sample of 62,177 upstairs and downstairs trades that are executed during the normal trading hour of SEATS. Mid-quote trades and trades effected by auction algorithms are excluded. We include upstairs and downstairs trades with a value of over $1 million as it is the minimum cut-off criterion for block specials during the sample period. There are 402 stocks during the period from 1 January 1993 to 31 December 1998 that have trades that meet these criteria. Table 5 contains the summary statistics of the distribution of the explanatory variables. The weighted trade size variable, size, has a mean of 3.51, i.e., an average trade has a size 3.51 times the total number of shares traded in its immediately prior 20 trades. The average bid-ask spread in the sample is 0.46 percent, with a median of 0.28 percent. The log of limit order book depth and the mid-quote variables have averages of 10.94 and 9.46, respectively. Table 6 presents the coefficient estimates (t-statistics in parentheses) of the model Pr [ui = 1| Zi] = Φ(γ Zi). The likelihood ratio test strongly rejects the null hypothesis that the explanatory variables have zero coefficients. All coefficient estimates are individually statistically significant at the 1% level. The coefficient estimate of sizei, γ1, is positive and supports our theoretical prediction that larger trades are more likely to be traded upstairs than smaller trades. This

30

positive effect of trade size on the market selection decision can be due to the relatively lower size related price impact cost, as we hypothesizes here, as well as any other size-related execution advantages of upstairs trading not explicitly included in our econometric model. The coefficients on liquidity variables γ2 and γ3 are also of the predicted sign. This provides strong evidence of the competition between the upstairs and downstairs markets. Specifically, when the downstairs market is illiquid, i.e., a high percentage bid-ask spread and low market depth, large trades migrate to the upstairs market. Previous research finds statistically insignificant effect of the bidask spread on the market selection decision, possibly due to the use of highly correlated variables and a smaller sample. The strongly negative effect of the quote midpoint level prior to the trade, mqi, is contrary to expectations based on Madhavan and Cheng (1997). The coefficient estimate, γ 4 , suggests that stocks or trades at higher prices are less likely to be executed off-market. Plausible explanations include price discreteness and the correlation between trading activity and price level. Specifically, higher price stock have lower percentage bid ask spreads relative to lower price stocks that are subject to the same minimum tick size, and a lower downstairs spread reduces the incentive to transact upstairs. In addition, there is well-documented positive price-volume relationship in the equity market (Karpoff (1987)). If higher mid-quote proxies more liquidity downstairs, higher mid-quote would be associated with lower probability of upstairs execution.

6.4

Estimation of Differential Price Impact Cost Based on equation (11) and substitute in our specification of Xi of the market impact cost

functions equation (2) and (3), we have, cut = βu0 + βu1 sizei –θ i+ ε ui , cdt = βd0 + βd1 sizei + ε di , and the reduced form regression equation:

31

ct = β0 + β1 sizei + β2 Φˆ i + β3 sizei Φˆ i + β4 φˆi + ε i,

(13)

where Φˆ i and φˆi are the maximum likelihood probability estimates from the structural probit model. These reduced form coefficients are related to the structural coefficients as follows: β1 =

βd1 > 0, β2 = ( βu0 – βd0 ), β3 = ( βu1 – βd1 ) < 0, and β4 = ( σu – σd ). Madhavan and Cheng (1997) argue that coefficient βu0, the size-unrelated price impact cost of an upstairs trade, contains compensation for the costly search and negotiation required by upstairs intermediation and find support for the proposition that βu0 > βd0. Therefore, we expect β2 < 0. It can be argued that the covariance of the trader’s unobserved signal with their price impact disturbances is the same in both the upstairs and downstairs market which implies β4 = – var[θi] < 0. Equation (13) can be estimated with OLS after adjusting for heteroscedasticity due to selectivity bias. Table 7 shows the coefficient estimates of equation (13) with heteroscedastic-consistent tstatistics in parentheses. The price impact cost variable, ct, is defined as absolute return from the mid-quote prior to a block to the trade price of the block. The regression explains an economically significant proportion of the price impact cost of upstairs and downstairs block trades, 14.83%, and the F-statistic of 2707.24 rejects the null hypothesis of the explanatory variables having zero slope at the 1% level. All coefficient estimates are also individually statistically significant at the 1% level. The coefficients on size related variables strongly support the theoretical predictions. Specifically, β1, which measures the expected size effect of downstairs price impact cost, is significantly positive as predicted, and, β3, which measures the expected differential price impact of a trade executed upstairs, is significantly negative as predicted. These coefficient estimates imply that price impact of a larger block is larger than the price impact of a smaller block, other things being equal, and this size effect is smaller in the upstairs market. This result is consistent with 32

models of Seppi (1990), Keim and Madhavan (1996), Burdett and O’Hara (1987), Grossman (1992), and U.S. finding in Madhavan and Cheng (1997). Specific support for the adverse selection and reputation model of Seppi (1990) can be found by observing the negative sign of β4 and its high t-statistic. This result provides strong evidence for the existence of trader’s reputation signal, θi, in the upstairs market. This finding is in stark contrast to the evidence of upstairs dealing found in Fong and Swan (2000) where they focus on the top twenty most liquid stocks traded on the ASX30. They find that upstairs brokerdealers are more likely to participate as a principal dealer when market impact costs are high, markets are thin, the block trade size in the upstairs market is relatively small, and when the trade is part of a larger portfolio trade. They also find that principal dealers are unable to selectively participate in more profitable uninformed trades as the reputation models of Seppi (1990) and Roell (1990) might suggest. Together, the evidence suggests that while there is significant reputation effect in reducing the price impact cost of a trade, broker-dealers in the upstairs market are not using this information for their brokering versus dealing decision. Such non-profit maximizing dealing behavior is supportive of the theory of Aitken, Garvey and Swan (1995) where upstairs dealing (market making), at least in the ASX, is used as a soft-dollar service forming part of a long-term implicit client-broker contract. Finally, the statistically significant negative sign of the coefficient of Φˆ i (estimated probability of upstairs execution), β2, is opposite to our prediction and findings in previous research. The negative sign of β2 suggests that the price impact cost of upstairs trades is not only less sensitive to trade size than downstairs trades, but it also has a smaller fixed component. This finding

30

These top twenty stocks were selected because they represent a very high proportion of the overall trading volume

and principal dealing on the ASX.

33

might be due to differences in institutional arrangement for upstairs trades between Australia and the U.S. Madhavan and Cheng (1997) note that on the NYSE, an upstairs trade still requires exposure to the public orders on the trading floor because the floor is where the upstairs trade is actually executed. In Australia, such a trade is actually executed upstairs and then the broker reports the trade to the market. The bypass of the exposure to public limit and market orders might allow upstairs trade in Australia to be more accurately conditioned on prior downstairs market prices than what is practical in the U.S. In addition, the relative size and proportion of upstairs trades to downstairs trades in Australia is larger than that in the U.S. The smaller fixed component of price impact cost of upstairs trades in Australia might also be due to this sample characteristic31. It should be pointed out again that price impact cost is but one of the reasons for choosing upstairs market. Even if both the fixed and size related components of price impact cost are lower upstairs, upstairs trading is not always preferred since traders still need to consider other execution cost factors such as opportunity costs, delay due to negotiation and the price movements in the market. We performed several robustness checks including transforming our trade size measure by taking the log and square root. The results are qualitatively equivalent. We have also defined the price impact cost as the absolute return from the mid-quote of the 20th trade prior to a block to the trade price of the block. The regression results are again qualitatively identical and therefore not reported here.

31

Most downstairs and upstairs blocks reported in Madhavan and Cheng (1997) are in the 10,000-20,000 shares

categories, with the downstairs trades out-numbering upstairs trades by a factor of approximately 10. The number of trades across the two markets in the 20,000-50,000 shares, and the over 50,000 shares categories are more similar, but there are still more trades in the downstairs market than downstairs market. In our sample the ratio of upstairs to downstairs trade is approximately 4:1, and the median weighted trade size of upstairs trades is twice that of downstairs trades.

34

7.

Conclusions Despite the presence of strong network economies, most established securities markets

experience some diversion of order flow through after-hours trading, electronic crossing networks, and upstairs trading. Such off-market trading represents competition for primary markets, but also gives rise to concerns regarding market fragmentation, the lack of transparency, the inability to expose limit orders after the close of trading, and front-running. Yet, in the absence of detailed information, even the most basic empirical questions concerning off-market trading remain unanswered. This paper uses unique and highly comprehensive data on off-market trading to analyze the nature and determinants of order flow fragmentation both over time and across all securities listed on the Australian Stock Exchange. Our panel regression estimates shows that crosssectional difference in off-market trading (crossing systems, after-hours and upstairs trading alike) is driven by institutional trading interest and liquidity. The level of off-market trading is particularly strongly related to the total dollar trading volume in a stock. The demand to trade a large volume and orders are the fundamental factors in determining the cross-sectional pattern in off-market trading. Primary market illiquidity is also a statistically significant factor in determining the level of off-market activity, but secondary to the trading demand effect. Specifically, while off-market trading in a stock is positively related to bid-ask spread and negatively related to limit order book depth, stocks with low trading demand and high bid-ask spread still do not have significant off-market trading activity. These results are consistent with the literature that views off-market trading as an alternative source of liquidity to supplement the primary market. This is particularly so when the primary market fails to provide a trading opportunity for some niche traders (including large, patient, or uninformed traders) at minimum cost, i.e., the theories

35

of Burdett and O’Hara (1987), Grossman (1992), and Seppi (1990). This competition between the primary market and off-market venues is also illustrated in the introduction of closing auction markets at the ASX, where it attracts trades at the expense of off-market venues. At transaction level, we study upstairs and downstairs block trades and find strong evidence that trade size and primary downstairs liquidity affect a trader’s market selection decision. The significantly positive effects of trade size and primary market bid-ask spread, and negative effects of electronic limit order book depth and price level, on the probability of a large trade to be executed upstairs further illustrate the competition between different market mechanisms. There is evidence that upstairs trading reduces both the fixed as well as size-related component of price impact costs relative to downstairs execution. We also find evidence of the existence of a reputational factor in the upstairs market that supports Seppi (1990). This is consistent with prior research by Madhavan and Cheng (1997). We conclude that there is significant competition between markets in highly liquid securities and off-market trading which benefits those in a position to switch by lowering their trading costs.

36

References Admati , Anat R., and Paul Pfleiderer, 1988, A Theory of Intraday Trading Patterns, Review of Financial Studies 1, 3-40. Admati, Anat R., and Paul Pfleiderer, 1991, Sunshine Trading and Financial Market Equilibrium, Review of Financial Studies 4, 443-81. Aitken, Michael J., Gerald T. Garvey, and Peter L. Swan, 1995, How Brokers Facilitate Trade for Long-term Clients in Competitive Securities Markets, Journal of Business 68, 1-33. Barclay, Michael, and Terrence Hendershott, 2000, Price Discovery and Trading Costs After Hours, working paper, University of Rochester. Battalio, Robert H., 1997, Third Market Broker-Dealers: Cost Competitors or Cream Skimmers? Journal of Finance 52, 241-52. Brennan, Michael J., and Avanidhar Subrahmanyam, 1998, The Determinants of Average Trade Size, Journal of Business 71, 1-25. Burdett, Kenneth, and Maureen O'Hara, 1987, Building Blocks: An Introduction to Block Trading, Journal of Banking and Finance 11, 193-212. Chowdhry, Bhagwan, and Vikram Nanda, 1991, Multi-Market Trading and Market Liquidity, Review of Financial Studies 4, 483-511. Cohen, Kalman J., Steven F. Maier, Robert A. Schwartz and David K. Whitcomb, 1982, An Analysis of the Economic Justification for Consolidation in a Secondary Security Market, Journal of Banking and Finance 6, 117-136. Conrad, Jennifer, Kevin M. Johnson, and Sunil Wahal, 2001, Alternative Trading Systems, working paper, University of North Carolina at Chapel Hill. Davis, Jeffry L. and Lois E. Lightfoot, 1998, Fragmentation Versus Consolidation of Securities Trading : Evidence From The Operation of Rule 19c-3, Journal of Law and Economics 41, 209-238. Domowitz, Ian, 1990, The Mechanics of Automated Trade Execution Systems, Journal of Financial Intermediation 1, 167-94. Domowitz, Ian, 1993, A Taxonomy of Automated Trade Execution Systems, Journal of International Money and Finance 12, 607-31. Easley, David, and Maureen O'Hara, 1987, Price, Trade Size, and Information in Securities Markets, Journal of Financial Economics 19, 69-90. Easley, David, and Maureen O'Hara, 1987, Adverse Selection and Large Trade Volume: The Implications for Market Efficiency, Journal of Financial and Quantitative Analysis 27, 185208. Easley, David, Nicholas M. Kiefer, and Maureen O'Hara, 1996, Cream-Skimming or Profit Sharing? The Curious Role of Purchased Order Flow, Journal of Finance 51, 811-833.

37

Fong, Kingsley and Peter L. Swan, 2000, “When is Making a Market for Your Client a Crime? The Determinants of Dealer Participation in the Upstairs Market” Working Paper, School of Business, University of Sydney. Garbade, Kenneth D. and William Silber, 1979, Dominant And Satellite Markets: A Study Of Dually-Traded Securities, Review of Economics and Statistics 61, 455-460. Grossman, Sanford, 1992, The information role of upstairs and downstairs markets. Journal of Business 65, 4, 509-529. Grossman, Sanford J. and Merton H. Miller, 1988, Liquidity and Market Structure, Journal of Finance 43 (3) (July), 617-637. Hendershott, Terrence, and Haim Mendelson, 2000, Crossing Networks and Dealer Markets: Competition and Performance, Journal of Finance 50, 1175-1199. Harris, Lawrence E., 1994, Minimum Price Variations, Discrete Bid-Ask Spreads, and Quotation Sizes, Review of Financial Studies 7, 149-78. Hillion, Pierre, and Matti Suominen, 1998, Deadline Effect of an Order Driven Market: An Analysis of the Last Trading Minute on the Paris Bourse, working paper, INSEAD. Huang, Roger, 2000, Price Discovery by ECNs and Nasdaq Market Makers, working paper, University of Notre Dame. Karpoff, Jonathan M., 1987, The Relation between Price Changes and Trading Volume: A Survey, Journal of Financial and Quantitative Analysis 22, 109-26. Keim, Donald B., and Ananth Madhavan, 1996, The Upstairs Market for Large-block Transactions: Analysis and Measurement of Price Effects, Review of Financial Studies 9, 1-36. Keim, Donald B and Ananth Madhavan, 1997, Transactions Costs and Investment Style: An Inter-exchange Analysis of Institutional Equity Trades, Journal of Financial Economics 46, 265-92. Madhavan, Ananth, 1995, Consolidation, Fragmentation, and the Disclosure of Trading Information, Review of Financial Studies 8, 579-603. Madhavan, Ananth, and Minder Cheng, 1997, In Search of Liquidity: An Analysis of Upstairs and Downstairs Trades, Review of Financial Studies 10, 175-204. Madhavan, Ananth and George Sofianos, 1998, “An Empirical Analysis of NYSE Specialist Trading,” Journal of Financial Economics 48, 189-210. Meier, R., 1998, Benchmarking Analysis of Stock Exchange Trading, International Federation of Stock Exchanges (FIBV). Monograph, SBF-Bourse de Paris. Mendelson, Haim, 1987, Consolidation, Fragmentation, And Market Performance, Journal of Financial and Quantitative Analysis 22, 189-208. NYSE, 2000, Market Structure Report of the New York Stock Exchange Special Committee on Market Structure, Governance and Ownership. O’Hara, Maureen, 1995, Market Microstructure Theory, Basil Blackwell, Cambridge, Mass.

38

Pagano, Marco, 1989a, Trading Volume and Asset Liquidity, Quarterly Journal of Economics 104, 255-274. Pagano, Marco, 1989b, Endogenous Market Thinness and Stock Price Volatility, Review of Economic Studies 56, 269-288. Roell, Ailsa, 1990, Dual capacity trading and the quality of the market, Journal of Financial Intermediation 1, 105-124. Security and Exchange Commission, 2000, Electronic Communication Networks and AfterHours Trading. Security and Exchange Commission, 2001, The National market System: A Vision that Endures, Speech by SEC Chairman Arthur Levitt at Standford University. Seppi, Duane, 1990, Equilibrium Block Trading and Asymmetric Information, Journal of Finance 45 (March), 73-94. Thomas, S., 1998, End of Day Patterns After Implementation of A Call Auction on the Paris Bourse, Unpublished Manuscript, SBF-Bourse de Paris.

39

Table 1 Descriptive Statistics This table presents descriptive statistics for the stocks in our sample period from January 1994−December 1998. The stocks are grouped into deciles by dollar trading volume at the end of the previous year, where the decile cutoffs are calculated each year. Therefore statistics for 1993 are not reported. The table shows the mean daily dollar trading volume, mean dollar trade size, mean price, mean percentage bid-ask spread, mean number of trades and mean number of days traded, per year. Decile Daily dollar Dollar Trade Price Spread (%) Trades Per Days Traded volume Size Year Per Year 1 (top) 2 3 4 5 6 7 8 9 10 (bottom)

5,133,576

37,523

8.71

0.75

18,942.67

243.05

352,627

12,714

2.86

1.88

2,856.67

229.98

151,947

8,461

1.89

3.05

1,534.45

211.88

75,888

6,031

1.60

4.32

1,087.61

193.67

53,282

4,708

1.19

5.22

897.21

179.76

43,381

4,359

1.09

5.51

738.56

163.05

31,782

3,802

0.92

6.82

546.32

140.96

25,880

3,305

0.75

8.35

412.45

116.95

21,996

2,985

1.04

10.38

229.40

79.75

14,604

2,834

1.21

20.01

107.73

41.60

40

Table 2 Off-Market Dollar Volume Ratio This table reports statistics on the ratio of dollar trading volume executed off-market for each year from 1994-1998, by year and dollar volume deciles. Each cell in Panel A represents the percentage of annual dollar trading volume that occurs off-market. Panel B reports the corresponding figures for the after-hours period only. For each stock, we first compute the total dollar trading volume in a given year. Using identifiers that flag offmarket trades, we repeat this exercise and compute the dollar trading volume off-market. We then compute the ratio of the off-market totals to the total dollar volume. These ratios are then averaged across all stocks within an annual dollar trading volume (previous year) decile. Panel A: Dollar Volume Ratio for All Off-Market Trades (%) Volume Decile 1994 1995 1996 1997 1998 1 (top)

36.97

33.33

34.63

34.30

29.80

2

30.06

29.64

32.42

30.85

25.27

3

21.13

17.60

24.02

19.62

19.73

4

15.23

14.78

16.83

13.08

14.36

5

9.53

11.82

15.13

8.18

8.86

6

10.17

9.23

8.19

11.28

8.11

7

8.80

8.10

8.90

6.92

9.75

8

11.83

9.02

7.59

8.13

6.35

9

8.52

6.13

8.64

8.39

4.70

10 (bottom)

2.62

5.57

4.73

3.85

5.26

Panel B: Dollar Volume Ratio for Off-Market Trades After hours (%) Volume Decile 1994 1995 1996 1997 1998 1 (top)

18.22

14.59

14.72

15.52

15.22

2

15.55

13.94

16.04

15.03

12.56

3

11.29

9.18

11.62

10.58

9.85

4

9.07

7.42

8.29

7.85

10.11

5

6.06

6.47

7.25

6.34

6.28

6

6.67

4.64

5.82

6.49

4.20

7

6.05

6.20

5.10

4.83

7.23

8

8.80

4.28

5.61

3.53

4.63

9

4.83

3.61

6.11

5.43

2.93

10 (bottom)

2.55

4.06

2.99

3.01

4.91

41

Table 3 Summary Statistics of Panel Regression Variables Definitions: offi,t is the ratio of off-market to total dollar trading volume of stock i in year t; lvoli,t is the log of total dollar trading volume of stock i in year t; lmcapi,t, is the log of average daily market value of the issued shares of stock i in year t; indexi,t is the proportion of trading days in year t that stock i is included in the index; pspdi,t is the median daily average percentage quoted bid-ask spread of stock i in year t; ldepthi,t is log of the median daily average dollar value of the best two limit bid orders and the best two limit ask orders of stock i in year t; stdpri,t is the median daily ratios of the transaction price standard deviation to the average price of stock i in year t; optioni,t is the proportion of trading days in year t that there are exchange traded options listed on stock i; auctioni,t, is the ratio of closing auction dollar volume to total dollar trading volume for stock i in year t; rulet is the proportion of year t that is after October 14, 1996; n is the number of stock-year observations. “All Observations” refers to the sample including all stock-year observations between 1993 and 1998, inclusive, with sufficient data to compute all variables. “Active Issues” is a subset that consists of observations where the stocks are traded for more than 240 days a year.

n off

lvol lmcap index pspd ldepth stdpr option auction rule

All Observations

Active Issues

5864 Mean 0.149 15.687 17.327 0.242 0.058 10.304 0.007 0.042 0.005 0.455

1334 Mean 0.263 18.595 19.462 0.755 0.015 11.350 0.007 0.160 0.005 0.495

Std Dev Median 0.186 0.051 2.416 15.571 2.005 16.970 0.400 0.000 0.067 0.039 1.048 10.157 0.005 0.006 0.201 0.000 0.010 0.000 0.463 0.208

1st Q 0.013 14.126 15.827 0.000 0.019 9.602 0.003 0.000 0.000 0.000

3rd Q 0.254 17.138 18.608 0.516 0.073 10.884 0.010 0.000 0.007 1.000

42

Std Dev Median 0.168 0.270 1.671 18.367 1.724 19.475 0.403 0.980 0.012 0.012 0.877 11.237 0.004 0.006 0.366 0.000 0.007 0.000 0.459 0.208

1st Q 0.119 17.324 18.218 0.756 0.007 10.706 0.004 0.000 0.000 0.000

3rd Q 0.368 19.700 20.556 0.988 0.019 11.929 0.009 0.000 0.011 1.000

Table 4 Cross-sectional and Panel Models for Annual Off-Market Dollar Volume Ratio This table contains the regression estimates of a fixed effect panel data model with stock specific intercepts given by offi,t = ai + α1 lvoli,t + α2 lmcapi,t + α3 indexi,t + α4 pspdi,t + α5 ldepthi,t + α6 strpri,t + α7 optioni,t + α8 auctioni,t + α9 rulet + eit, where offi,t is the ratio of off-market to total dollar trading volume of stock i in year t, lvoli,t is the log of total dollar trading volume of stock i in year t, lmcapi,t is the log of average daily market value of the issued shares of stock i in year t, indexi,t is the percentage of trading days in year t that stock i is included in the index, pspdi,t is the median daily percentage quoted bid-ask spread of stock i in year t, ldepthi,t is the log of the median daily average dollar value of the best two limit bid orders and the best two limit ask orders of stock i in year t, stdpri,t is the median daily ratios of the transaction price standard deviation to the average price of stock i in year t, optioni,t is the percentage of trading days in year t that there are exchange traded options listed on stock i, auctioni,t is the percentage dollar trading volume traded in the closing auction for stock i in year t, rulet is the proportion of year t that is after October 14, 1996, and, ei,t is the error term for stock i in period t. Figures in parentheses are t-statistics computed using White heteroskedastic-consistent standard errors. Panel A contains the results of the fixed effect panel-data model while Panel B contains the cross-sectional results (i.e., a regression of average variables across years) for: (a) All stock-year observations, (b) Active Issues, i.e., all stock-year observations where the stocks are traded for more than 240 days a year. Panel A Fixed Effect Model Panel B Cross-Sectional Model All Observations Active Issues All Observations Active Issues Coefficient t-statistic Coefficient t-statistic Coefficient t-statistic Coefficient t-statistic lvol 0.068 (16.78) 0.156 (9.76) 0.037 (10.86) 0.090 (9.68) lmcap -0.006 (-1.32) -0.065 (-3.66) 0.028 (9.96) 0.007 (1.09) index 0.019 (1.37) 0.063 (2.27) 0.045 (3.78) 0.104 (6.38) pspd 0.666 (8.75) 6.764 (6.54) 0.522 (8.44) 7.712 (8.32) ldepth -0.046 (-8.51) -0.101 (-6.54) -0.014 (-2.37) -0.066 (-5.53) stdpr -6.480 (-8.83) -23.764 (-6.04) -6.213 (-10.11) -25.479 (-10.03) option -0.044 (-2.17) -0.041 (-1.92) -0.141 (-8.68) -0.116 (-5.51) auction -0.974 (-3.99) -4.114 (-5.12) -1.289 (-3.30) -5.081 (-3.63) rule -0.021 (-3.55) -0.014 (-0.93) 0.004 (0.34) 0.068 (3.02) constant -0.751 (-14.93) -0.801 (-6.79) Adj. R-squared 0.56 0.73 0.56 0.61 P-value of F test 0.00 0.00 n 5,864 1,334 1,573 513

43

Table 5 Summary Statistics of Probit Model Explanatory Variables Definition: sizei is the number of shares of trade i divided by the average number of shares traded in the immediate prior twenty trades; pspdi is the ratio of the bid-ask spread to quotation midpoint times 100; ldepthi is log of the number of shares available in the best two price steps of the limit order book (bid orders for sell trades and ask orders for buy trades); mqi is the quotation midpoint immediately prior to trade i. n

62177

sizei pspdi ldepthi mqi

Mean 3.513 0.461 10.942 9.456

Std Dev 11.283 0.605 1.769 44.042

Median 1.604 0.281 10.846 5.330

1st Q 0.840 0.148 9.798 2.905

44

3rd Q 3.390 0.531 11.990 10.645

Table 6 Structural Probit Model Estimates This table contains the coefficient estimates of a structural probit model of the endogenous switching regression model: Pr[ui=1|Zi] = Φ(γ Zi), where Φ (.) is the cumulative standard normal distribution, ui is an indicator variable taking the value 1 if the ith trade was upstairs executed and 0 otherwise, Zi is a vector of independent variables, and γ is a vector of unknown coefficients. The linear combination γ Zi is given by

γ Zi = γ0 + γ1 sizei + γ2 pspdi + γ3 ldepthi + γ4 mqi, where sizei is the number of shares of trade i divided by the average number of shares traded in the immediate prior twenty trades, pspdi is the ratio of the bidask spread to quotation midpoint times 100, ldepthi is log of the number of shares available in the best two price steps of the limit order book (bid orders for sell trades and ask orders for buy trades), mqi is the quotation midpoint immediately prior to trade i. The likelihood ratio test reports the difference between the log likelihood of the unrestricted model and the restricted model with γ1 = γ2 = γ3 = γ4 = 0. The estimation is based on a sample of 62,177 upstairs and downstairs trades that are executed during the normal trading hour of SEATS. Mid-quote trades and trades effected by auction algorithms are excluded. We include upstairs and downstairs trades with a value of over $1 million. There are 402 stocks during the period from 1 January 1993 to 31 December 1998 that have trades that meet these criteria.

constant sizei pspdi ldepthi mqi

Coefficient 4.037 0.024 0.310 -0.303 -0.003

Likelihood Ratio Test (χ2)

Std Err 0.043 0.001 0.013 0.004 0.000

t-statistic 94.17 19.84 23.60 -81.52 -22.67

P-value [.000] [.000] [.000] [.000] [.000]

4423.03

[.000]

45

Table 7 Switching Regression Estimates This table contains the coefficient estimates of the second stage endogenous switching regression model of the price impact cost of block trades. The model estimated is ct = β0 + β1 sizei + β2 Φˆ i + β3 sizei Φˆ i + β4 φˆi + ε i, where sizei is the number of shares of trade i divided by the average number of shares traded in the immediate prior twenty trades; Φˆ i and φˆi are the maximum likelihood probability estimates from the structural probit model. The estimation is based on a sample of 62,177 upstairs and downstairs trades that are executed during the normal trading hour of SEATS. Mid-quote trades and trades effected by auction algorithms are excluded. We include upstairs and downstairs trades with a value of over $1 million. There are 402 stocks during the period from 1 January 1993 to 31 December 1998 that have trades that meet these criteria.

constant sizei Φi Φ i sizei φi

Coefficient 1.268 0.052 -0.413 -0.048 -2.400

F Adjusted R-squared

Std Err 0.034 0.006 0.026 0.006 0.061

t-stat 36.95 9.49 -15.92 -7.45 -39.43 2707.24 0.15

46

p value [.000] [.000] [.000] [.000] [.000]