Reputation Effects on the Floor of the New York Stock Exchange. Robert Battalio Mendoza College of Business University of Notre Dame

Transcription

1 This Draft: May 11, 2004 Preliminary & Incomplete Reputation Effects on the Floor of the New York Stock Exchange Robert Battalio Mendoza College of Business University of Notre Dame Andrew Ellul Kelley School of Business Indiana University Robert Jennings Kelley School of Business Indiana University Abstract: A growing empirical literature suggests non-anonymous, floor-based trading venues provide lower trading costs for large, difficult orders than anonymous, electronic trading platforms. Theory suggests reputations, developed through repeated face-to-face interactions, allow floor-based trading venues to attenuate the adverse selection problem associated with large orders. We identify instances when stocks listed on the New York Stock Exchange (NYSE) relocate. Although the specialist follows the stock to its new location, many floor brokers do not. We use this natural experiment to determine whether reputation affects trading costs. We find a discernable increase in the cost of liquidity in the days leading up to and immediately after a stock s relocation. The increase is more pronounced for stocks with higher adverse selection. Using NYSE audit-trail data, we find that the floor brokers that relocate with the stock obtain lower trading costs than those who do not move. Together, these results suggest the floor of the NYSE plays an important role in the liquidity provision process. Acknowledgments: We thank the NYSE, especially Katherine Ross, for providing data. The opinions expressed in this paper do not necessarily reflect those of the employees, members, or officers of the NYSE.

2 The rise to prominence of electronic trading venues in equities (Island, Instinet, and Archiepelago), equity options (International Securities Exchange), and futures (Globex) suggests that physical trading floors may soon be obsolete. Pirrong (1996) and Domowitz (2001) argue automated markets are attractive to investors because they are fast and because they are cheap to develop, operate, and monitor. American Century, a big mutual fund company, estimates the amount it spends on commissions would fall from $150 million to $25 million per year if all of its equity trading were done electronically (see Tully 2003). Perhaps more important, proponents argue electronic trading venues allow investors to reduce implicit trading costs associated with information leakage by giving traders direct control over their orders. 1 Harris (2003), however, suggests that floor-based exchanges will not disappear as long as floors provide valuable services to traders. A large empirical literature suggests floor-based exchanges excel at handling large, difficult orders. 2 Theoretically, Benveniste, Marcus, and Wilhelm (1992) and Weinstein (1993) argue the repeated dealings between floor brokers and specialists on floor-based exchanges allow specialists to penalize brokers who misrepresent their trading intentions (by providing less favorable prices in the future). Pagano and Roell (1992) argue that floor-based trading systems give participants the opportunity to observe who trades what with whom, how urgently they seem to want to trade, etc. These arguments suggest repeated, face-to-face interactions are one source of the floor s comparative advantage in handling institutional order flow. We use 1 For example, Tully (2003) notes in a November 10, 2003 Fortune article that, Buyers and sellers sacrifice their anonymity by divulging their orders to the Wall Street middlemen. Most securities firms zealously try to protect that information. But Wall Street is famously a sieve-- information leaks out from everywhere. And those whispers can dramatically move the price of the shares that the fund is in the process of accumulating. "When we're bulking up on a new stock," says American Century's Wheeler, "and we give the buy ticket to a bulge-bracket firm, they can trade ahead of us on a proprietary basis after they execute the first order. We might as well write them a check." 2 There is a growing empirical literature that suggests non-anonymous, floor-based trading venues provide lower trading costs for large, difficult orders than anonymous, electronic trading venues. For example, see Chakravarty (2001), Venkataraman (2001), Garfinkel and Nimalendran (2003), Barclay, Hendershott and McCormick (2003), Waisburd (2003), Bessembinder and Venkataraman (2003), Theissen (1999), and Handa, Scwhartz and Tiwari (2003). Boehmer, Saar and Yu (2004), Madhavan and Papaganeshan (199X), Corwin and Lipson (200x), Coval and Shumway (2001) each present evidence that suggests the floor is informationally rich. Pirrong (1996) notes that floor traders also argue that their ability to look other traders in the eye also provides valuable information about the motives of their competitors, and that this (to some degree) limits their vulnerability to being picked off by a more informed individual. Baker and Iyer (1992) examine trading crowds in a national option exchange and find that network structure influences price volatility and expected trading volume. They interpret their results as suggesting that the structure of the real communication network among investors may influence market behavior. 1

3 a unique natural experiment to empirically examine whether relationships between brokers and specialists on the floor of the NYSE are economically meaningful. NYSE trading is organized so that all orders in a particular security go to one physical location on the floor where exchange members wishing to trade that security gather. 3 One individual, the specialist, trades only the securities assigned to that location. Other individuals, floor brokers, might trade securities at several locations but usually handle all of their brokerage firm s orders in a particular security. The regular face-to-face interactions between a specialist and floor brokers allows relationships to form relationships that Benveniste, Marcus and Wilhelm (1992), Pagano and Roell (1992) and Weinstein (1993) argue play a roll in a security s trading process. If relationships are important in attenuating the adverse selection problem, then exogenous disruptions in these relationships provide an opportunity to examine their economic importance. In this paper, we investigate the joint hypothesis that trading relationships between specialists and floor brokers are important and that these relationships affect the trading process by examining the quoting and trading behavior of a sample of securities changing the location at which they trade on the NYSE floor. We begin by identifying reallocations of trading locations by individual specialist firms, reorganizations of the NYSE floor due to specialist firm mergers, and the opening of a new trading floor. None of these events are the result of an endogenous change in trading costs. If the imminent departure of the specialist from the current trading location suggests to floor brokers that the specialist will be unable to sanction undesirable and/or reward desirable behavior, then we might observe changes in trading behavior before the location change. Specifically, if the floor broker will no longer be subject to a particular specialist s power to sanction or reward (because the specialist has moved locations on the floor), then quoted and effective spreads may widen to compensate for the possibility of additional adverse selection risk. In theory, the old equilibrium unravels when the change in location is anticipated. If trading relationships 3 A schematic of the floor can be seen at 2

4 matter and these relationships are not established instantaneously at the new location, then the disruption in relationships caused by the change in location will manifest itself in the trading behavior of the security for some time after the change in location. In summary, if relationships are important, we expect to find a deterioration in trading costs leading up to and continuing for some time after the date securities change location on the NYSE floor. It is possible that the floor broker community is relatively unimportant to the specialist. For example, relationships with individual floor brokers might be irrelevant because much of the trading strategy is determined by traders at the brokerage firm s trading desk and not by the broker on the NYSE floor. That is, knowing that investors and the broker s traders are unaffected by NYSE space reallocation decisions, the specialist behaves no differently immediately before and after the change in location. In addition, because many floor brokers at the new location work for the same brokerage firms as those at the former location, the specialist can punish and/or reward the new brokers for the behavior of the other brokers from the same brokerage firm prior to a stock s relocation. These arguments suggest there should be no change in trading behavior when a security s trading location on the floor of the NYSE changes. We begin our analysis by using proprietary NYSE data to demonstrate that individual floor brokers typically do not follow a sample stock to its new trading location. This verifies our claim that relationships are disrupted by the move. Next, using the NYSE TAQ database, we examine relative effective spreads around the time securities change their trading location on the floor of the NYSE. Surprisingly, we find a discernible increase in effective spreads when a stock moves. Effective spreads begin to increase several days prior to the move and they remain high for some time after the move. Consistent with theory, the increase in trading costs is positively related to the measured adverse selection associated with the moving security and is negatively related to the number of floor brokers who follow the stock to its new location. Using more refined proprietary data, we examine the execution costs of the individual floor brokers who do and do not follow the stock to its new location. We find that post-move trading costs paid by brokers who move with 3

5 the specialist are significantly lower than those paid by the brokers who are new to the trading crowd. This result is especially strong in the first few days after the change when the specialist moves to a new floor location and is building reputation with a new trading crowd. We also find that the trading costs incurred for trades between two moving brokers (who have dealt with each other previously) is significantly lower than the costs incurred in trades between a moving and a non-moving (new) broker. Together, these results provide the first direct evidence that face-to-face relationships give floor-based exchanges an advantage over anonymous electronic trading systems in executing large, difficult orders. In the next section we develop our general hypotheses and introduce the data and variables we use to test our hypotheses. We also provide support for our premise that relationships are disrupted when the trading location of our sample stocks changes. Section II examines how NYSE trading costs are affected when stocks change location, Section III investigates the execution costs of individual floor brokers who do and do not follow a moving stock to its new location. Section IV concludes. I. Hypothesis Development, Data, and Microstructure Variables A. Hypotheses Development. The NYSE floor is organized to focus a security s trading at one physical location. Each location typically trades multiple securities, with the exact number being determined by the securities characteristics such as trading volume. One individual, the specialist, trades only at that location. The exchange charges the specialist with maintaining a fair and orderly market in the securities trading at that location. Other individuals, floor brokers, represent customers trading interests. 4 Because orders are time sensitive, the physical area covered by a single floor broker typically is limited to something less than the entire exchange floor. Depending on the size of the brokerage firm for which they work, a floor broker might trade securities at only a few locations on the exchange floor or might cover many locations. In most cases, a floor broker 4 Sofianos and Werner (2000) find NYSE floor brokers executed trades corresponding to 44% of the value of all buys and sells in January and February of

6 executes all of their firm s orders in a given stock. The organization of trading on the NYSE suggests that specialists and floor brokers trade with each other repeatedly. Furthermore, because trading on an exchange floor is face-to-face, the trading relationship between a specialist and a floor broker is more likely to play a role in the trading process than in an anonymous electronic setting. 5 Theory suggests relationships between floor brokers and specialists have important implications for the trading process observed on an exchange. Benveniste, Marcus and Wilhelm (1992) argue that repeated face-to-face interactions on the floor of the NYSE allow specialists (and other floor brokers) to sanction floor brokers who exploit private information. 6 For example, a floor broker obtaining a favorable transaction price by misrepresenting a time-sensitive order as a liquidity trade to a specialist or fellow floor broker might find those parties less willing to trade in the future. Benveniste et. al. argue that the ability to sanction is important in constraining individual floor brokers from imposing adverse selection costs on the specialist and/or other brokers in the crowd. In the days leading up to a specialist s move, this constraint becomes relaxed for floor brokers who do not follow the specialist to the new location. Specifically, floor brokers who do not expect to follow the specialist will become increasingly willing to trade on private information. Anticipating this possibility, specialists and moving floor brokers will become more cautious when interacting with the floor brokers they expect to remain behind. After the move is completed, we expect it will take time and repeated face-to-face interactions for new relationships to develop. Although he focuses on borrower-lender relationships, Diamond (1989) presents a model in which reputation building affects how agents behave and, in turn, equilibrium prices. Diamond shows that borrowers alter their behavior to influence what lenders learn about them. Specifically, borrowers change their behavior to protect good reputations. If there is sufficient adverse selection, Diamond 5 Cao, Choe and Hatheway (1997), Corwin (1999), and Coughenour and Saad (2004) present evidence suggesting that a stock s trading characteristics are related to the specialist firm trading the stock. 6 Weinstein (1993) reaches similar conclusions, arguing specialists reward floor brokers that reveal private information regarding their orders by offering them tighter bid-ask spreads. 5

7 demonstrates that time is required before market participants build their reputations. In our setting, following the relocation of a specialist there typically are many new floor brokers in the crowd. Until new relationships are formed between the old and new members of the crowd, providing liquidity to new floor brokers is more risky than providing liquidity to those with established reputations. This suggests liquidity providers will offer less favorable prices for large orders handled by new floor brokers to compensate them for bearing the increased adverse selection risk. 7 Together, Benveniste et. al. (1992), Weinstein (1993) and Diamond (1989) suggest that we should see changes in the trading process in the days leading up to and immediately following the relocation of a specialist if crowd relationships are important in attenuating adverse selection. These theories suggest the effective spreads paid by large orders handled by floor brokers that leave (join) the crowd following a specialist s relocation will rise (fall) in the days leading up to (following) the move. While these changes should occur in all crowds that change sufficiently when a specialist relocates, they should be most drastic in stocks with the highest levels of asymmetric information. Following Huang and Stoll (1997), we use the difference between the effective and realized spread as a proxy for the level of asymmetric information in a stock to investigate this conjecture. 8 Finally, following Bacidore and Sofianos (2002) and Huang and Stoll (1996), we construct a matched sample of stocks that do not move to control for changes in effective spreads that are unrelated to the specialist s relocation. B. Data. Our analysis uses three primary datasets. We begin by identifying changes in location on the floor of the NYSE using post and panel data obtained from the Exchange. We next use the NYSE s Trade and Quote (TAQ) database to analyze NYSE trading costs around the specialist relocations. Although these tests 7 Bernhardt, et. al. (2003) note this prediction is robust across many theoretical settings. 8 Our conclusions do not change when we use the proxies for adverse selection used by Bessembinder and Kaufman (1997), Stoll (2000), Sarin, Shastri, and Shastri, (2000), and Brennan and Subrahmanyam (1996). 6

8 will not provide the sharpest tests of our hypotheses, they provide insight into their economic importance. Finally, we use proprietary consolidated equity audit-trail data (CAUD) obtained from the NYSE to examine whether effective spreads for large orders handled by floor brokers that leave (join) the crowd following a specialist s relocation rise (remain the same) in the days leading up to (following) the move. The post and panel data, the CAUD, and the TAQ data are described more fully below. The NYSE s floor is divided into 20 (17 active) trading posts. Posts are subdivided into as many as 30 panels, each with a specialist. Panels are so-named due to the flat-screen panel above the specialist that lists the stocks trading at that location and other pertinent data. Thus, the combination of post (numbered one through twenty) and panel (lettered beginning at A in each post) provides a unique location on the floor. For example, as of January 2004 General Electric stock trades at Post 13 Panel M. It is the only stock traded by that specialist. AOL and five other stocks are traded by a specialist at Post 3 Panel O. The NYSE s post and panel data provide daily information about the location at which stocks trade beginning in June We obtain these data from June 1999 through April To determine changes in location, we take the first difference of these data. We are not interested in all changes in post and panel. We need the distance a stock moves to be large enough to suggest some turnover in the crowd trading the stock. The floor of the Exchange is divided into five rooms. The Garage contains posts one through four. The Main Room consists of posts five through eleven. Posts 12 through 14 are in the Blue Room and posts 15 through 17 (currently inactive) are in the Extended Blue Room. Finally, Thirty Broad contains posts 18 through 20. To be included in our sample, a stock must move from one room to another. It is common for floor brokers to be assigned to a specific room if the brokerage firm is large enough to have multiple brokers. This prevents a broker from having to cover too much territory, which slows order placement. The broker typically works from a single booth (e.g., the firm s booth in the Main Room) regardless of the mix of stocks trading in that room. Thus, if a stock changes rooms, it is likely to be traded in a different floor broker crowd than it was previously as 7

9 the brokers in the new room begin coverage. Our sample consists of all location changes from June, 1999 through April, 2003 involving an entire panel of stocks switching rooms. We require that the panel of stocks remain constant (i.e., the specialist trades the same stocks before and after the switch in location). The fact that the entire panel of stocks changes locations suggests that the specialist and stocks did not change, but that there is a potential change in the floor broker community. We find six occasions during our sample period in which one or more panels of stocks remains together after a room change. Table I summarizes our sample. [Insert Table I.] The July 1999, the June 2000, and the March 2002 events appear to be internal reorganizations by one or more specialist firms. Although there are many of these reorganizations, these appear to be the only ones resulting in a panel of stocks changing rooms. The majority of our sample location changes occur on three dates in 2000: November 20, December 11, and December 20. The November date is the opening of the 30 Broad trading floor, and the later two dates are apparently the result of specialist mergers. Specialist firms prefer to trade stocks at contiguous posts/panels due to efficiencies that can be gained in support staff. Thus, following specialist mergers, the Exchange typically reallocates space to accommodate the acquiring firm. To determine the broker turnover associated with these location changes, we acquire NYSE consolidated equity audit-trail data (CAUD). CAUD provides, among other information, the counter-parties to each trade. For electronically submitted orders (SuperDOT), only the member firm s name is provided. For trades involving a floor broker, however, the broker s badge number (both buyer and seller badge numbers if both are floor brokers) is part of the record. The badge number uniquely identifies an individual member. Another important feature of CAUD is that it classifies the counterparties to a trade as (a) a member of the crowd, (b) an electronically submitted (SuperDOT) order, (c) an order in the limit order book, (d) an order arriving from another trading venue via the Intermarket Trading System (ITS), or (e) an order to execute at the open (an OARS order). 8

10 By examining CAUD for a period of time before and after the switch in location, we can estimate how many of the floor brokers follow the specialist to the new location on the floor. We obtain audit-trail data for four weeks before and after the switch. We assume that a broker trades at least once in each of those periods. If that is not true, then we mis-classify that broker as one that did not move to the new location (trades before, but not after) or one that began trading after the switch (trades after, but not before). Table II provides some descriptive statistics regarding floor broker turnover around our sample relocations. [Insert Table II.] In the weeks prior to the switch date, the average sample stock has about 35 different floor brokers executing at least one trade (40 conditional on a stock having at least one broker making a trade before the move). On average, fewer than two brokers (4.7% of the 35.66) appear in both the pre- and the post-switch trade data. These brokers participate in about 2.7% of the trades, but trade 6.3% of the shares prior to the specialist s move. Thus, floor brokers following the specialist to the new location make fewer, but larger trades than the average floor broker. We compute effective spreads, realized spreads, and the corresponding adverse selection measure as in Huang and Stoll (1996) around specialist relocations using the NYSE s TAQ database, which contains intraday trades and quotes for all securities listed on the NYSE, the American Stock Exchange (AMEX), and the Nasdaq Stock Market. Each quote record indicates the underlying stock, the trading venue posting the quote, the date and time of the quote, the bid and ask prices and quantities, and a condition code indicating whether the quote is an opening or closing quote. Each trade record indicates the underlying stock, the date and time the trade was reported, the venue reporting the trade, the transaction size and price, and codes indicating whether the trade is subsequently cancelled or is made with other special conditions. 9 Following Bessembinder and Kaufman (1997) and others we eliminate trades with a Condition Code of Z or G and trades that have a Correction Code that is not equal to zero or one. The computation of effective (realized) 9 See the NYSE s TAQ2 User s Guide for an in-depth description of the TAQ database. 9

11 spreads requires trades matched with the National Best Bid and Offer at (five minutes after) trades are reported. At each moment in the trading day, a stock s National Best Bid and Offer (NBBO) is computed by taking the highest bid and the lowest offer (i.e., the best prices) quoted by venues on which the stock is traded. C. Microstructure Variables. Effective spreads are used to examine the impact of specialist relocation on trading costs. The effective spread, which is twice the absolute difference between the trade price and the midpoint of the NBBO when the trade is executed, provide estimates of the round trip cost of immediacy paid by liquidity demanders. Realized spreads, computed as twice the absolute difference between the trade price and the midpoint of the NBBO prevailing five minutes after a trade is executed, provide estimates of the gross trading revenue earned by liquidity providers. Finally, the adverse selection component of the spread as defined by Huang and Stoll (1996) is the difference between the effective spread (i.e., what investors pay) and the realized spread (i.e., what liquidity providers earn). Following Bacidore and Sofianos (2002) and others, we standardize each of our microstructure variables by the execution-time NBBO midpoint to control for variation in spreads due to residual differences in price levels. To ensure that any documented changes in trading costs are due to changes in the crowd (and not random market events), we compare spreads for stocks that change locations with spreads for a matched sample of stocks. For each stock that changes location on a given day, we find a similar, non-moving NYSElisted security. We match on the basis of price, trading volume, and number of trades occurring during the period beginning 120 days before and ending 80 days before each sample stock relocates. Following Bacidore and Sofianos (2002) and Huang and Stoll (1996), we find the matching, non-moving stock that minimizes the following: 10

12 where c i Moving denotes the value of the ith matching variable for the relocating stock, and where c i Non-Moving denotes the value of the ith matching variable for the stock that does not change locations. This minimization is done subject to the constraint that: for all i (i.e., for each of our three matching characteristics). Table III provides descriptive statistics for our sample and control stocks. [Insert Table III.] II. NYSE Trading Costs around Stock Relocations. Theory suggests relationships between crowd members who plan to follow a relocating stock and those who do not will begin to break down as the relocation date nears. Theory also suggests it takes time and multiple interactions for relationships between members of the new crowd to form. In each of these situations, the break-down of relationships creates greater adverse selection problems for liquidity providers in the crowd. Thus, if relationships on the floor of the NYSE help to attenuate adverse selection problems, we hypothesize that trading costs (effective spreads) will increase in the days leading up to a stock s relocation, will remain high on the day the stock relocates, and then will gradually decline in the days following the stock s relocation. Since theory does not provide guidance as to when the relationships will break down or evolve, we begin computing spreads for each relocating stock and its match 120 trading days before and end 100 trading days after the sample stock moves. We then test our hypothesis by examining the difference between the effective spreads of the moving and non-moving stocks through time using standard event study methodology. We include a stock pair (i.e., a relocating stock and its match) in our tests on an event day 11

13 when the sample and control stock each have five or more trades on the day and each have a trade-weighted average price that equals or exceeds $5.00. The price screen decreases the chance that our results are driven by large percentage spreads on low-priced stocks. The trading activity screen increases the reliability of our trading cost estimates. The imposition of these screens on a daily basis, which maximizes the sample size on a given day, implies that the sample size can change from day to day in our estimation period. The remainder of this section is as follows. We first investigate ENTIRE SAMPLE, then HI A.S. We examine realized spreads to see if liquidity providers are earning rents, we then do multivariate for blah blah blah. A. Relocation and Effective Spreads We begin by plotting the difference in relative effective spreads for sample and control stocks around the sample stocks relocation on the NYSE floor. Given the considerable volatility in relative spreads, for each stock-pair we compute the weekly average difference in relative effective spreads. In other words, for each stock-pair in the sample on a given event day, we subtract the matched stocks share-weighted relative effective spread for the day from the sample stock s share-weighted relative effective spread on the same day. 10 We next find the median difference across all stock-pairs in the sample on that day. 11 We then take a simple average of the daily median differences in a calendar week to produce the Figure 1. [Insert Figure 1.] Figure 1 presents weak evidence in support of our hypothesis. We find that the median difference in relative effective spreads between the relocating and control stocks is six to ten basis points in the period of time well before and well after the location change. Beginning about nine weeks before the switch, the spreads on the moving stocks increase relative to the non-movers. From event week -7 through event week +5, the median spread difference is 10 to 15 basis points. The median difference returns to its pre-relocation level by event week +5. With an average sample stock price of $26.58, a five basis point increase in relative effective 10 Since theory suggests larger trades are most likely to be affected by relocations, we aggregate effective spreads on a stockday by share-weighting effective spreads on individual trades. Results are qualitatively identical when we trade-weight. 11 We use median differences so the average is not overly influenced by outliers. 12

14 spreads represents over a penny per share. Given that the average dollar effective spread for the stocks we analyze is $0.0523, this represents an economically meaningful increase in the spread. To address the statistical significance of the spread change, we conduct standard statistical tests. We anticipate that the event (change in location) causes an increase in spread variance, which produces a test statistic rejecting the null hypothesis of zero spread differences more frequently than it should (see Brown and Warner, 1980, 1985). To solve this problem, we use the standardized cross-sectional test developed by Boehmer, Musumeci, and Poulsen (1991) and we use the nonparametric Wilcoxon test to detect whether there are significant differences in the daily mean and median relative effective spreads in sample and control stocks on each event date. 12 Table IV reports the differences in relative effective spreads and the results of the parametric and nonparametric tests. [Insert Table IV.] We find evidence of statistically increased effective spreads (relative to the control group) starting about four weeks before the stock relocates. Although the statistically significant increases in relative effective spreads are not consistent on a day-to-day basis in the pre-switch period, the number of large t-statistics exceeds what we expect by chance (13 of 20 using the t-test and 12 of 20 using the Wilcoxon). In the post-switch period, we find similar frequencies of significantly positive differences in relative effective spreads through day +45 (+20) with the t-test (Wilcoxon). This suggests that it takes a few trading months for effective spreads to return to normal after a sample stock changes location. Thus, we find weak evidence consistent with an upward dislocation in effective spreads in the days before the move and elevated spreads afterwards. These results empirical results suggest reputations and relationships within the crowd contribute to the NYSE s dominance in executing larger, more difficult orders. 12 The standardized cross-sectional test is the result of combining the standardized-residual technology developed by Patell (1976) and the ordinary cross-sectional methodology proposed by Charest (1978) and Penman (1982). 13

15 B. Relocating Stocks with High Adverse Selection. The prior analysis examines the entire sample. Theory, however, suggests that the stocks most affected by disruptions to their trading crowds are stocks with high adverse selection. In stocks with little private information, ending an old relationship or starting a new one should not matter much. We explore this hypothesis by dividing our sample into stocks with high adverse selection and stocks with low adverse selection. As noted earlier, we measure adverse selection as the difference between the sample stocks effective and realized spreads (normalized by price) on event days -120 through If the relative effective spread is much larger than the relative realized spread (i.e., the stock s price tends to move after a trade), then we conclude there is substantial adverse selection. We create our high and low adverse selection sub-samples by comparing each sample stock s adverse selection measure to the median measure. In Table V, we provide descriptive statistics for characteristics of high and low adverse selection sub-samples relative to their matched stocks. [Insert Table V.] Since adverse selection was not a matching criteria, it should not be surprising that the high adverse selection sub-sample of stocks have lower prices and fewer trades than their matches. We repeat the event study for the high adverse selection sub-sample. The median differences in the stock-pair s relative effective spreads are graphed in Figure 2. [Insert Figure 2.] The difference between the relative effective spread for sample and control stocks is between 50 to 70 basis points well before and considerably after the sample stocks change locations on the NYSE floor. On the days leading up to and immediately following the relocation date for our sample stocks, the difference in relative effective spreads increases to over 100 basis points. This change of more than 30 basis points, which is considerably greater than the five basis point increase for the overall sample, supports the hypothesis that 14

16 reputations within trading crowds on the floor of the NYSE are more important for stocks with greater adverse selection. As with the overall sample, we conduct statistical tests of differences in daily relative effective spreads for sample and control stocks before and after the sample stocks relocate. Table VI confirms the conjectures that have been drawn from Figure 2. [Insert Table VI.] The evidence of widening relative effective spreads for the high adverse selection sub-sample is more convincing than for the sample as a whole. Starting 45 days before the switch and continuing 45 days after the switch, the relocating stocks spreads exceed the control stocks spreads each day. C. Relocation and Realized Spreads. In the prior two sections we demonstrate that the trading costs paid by liquidity demanders (i.e., the effective spread) increases in the days prior to the relocation of a sample stock and remain high for several days after the move. We now investigate whether this increase in trading costs translates into additional trading revenue for liquidity providers by using the realized spread as a proxy for trading revenue. This analysis presumes that, on average, liquidity providers unwind their positions at the midpoint of the NBBO prevailing five minutes after they trade. Assuming that the effective spread measures the amount intermediaries charge for providing liquidity, the realized spread measures the amount that intermediaries actually earn from providing liquidity. If the stock price falls (rises) in the five minutes after an intermediary buys shares, the amount they earn from providing liquidity will be less (more) than the effective spread. Should the realized spread increase in the days surrounding a stock s relocation, we would conclude that liquidity providers (including the specialist) are able to maintain a portion of the increase in effective spreads documented in the prior two subsections. Should the realized spread remain constant across the change, we can conclude that the increase in effective spreads is just sufficient to offset additional adverse selection 15

17 costs. Finally, should realized spreads fall, we would conclude that liquidity providers are harmed by the change in trading location. Figure 3 contains a plot of the weekly average relative realized spread differences between movers and non-movers for the high-adverse selection subsample. [Insert Figure 3.] We see little in the way of a change in relative realized spreads around the relocation date; the realized spread differential fluctuates around zero. This suggests the increase in effective spreads is approximately sufficient to compensate liquidity providers for the higher levels of adverse selection in the market at these times. This conclusion is reinforced by statistical tests reported in Table VII. [Insert Table VII.] D. Multivariate Analysis of Effective Spreads of Relocating Stocks. Thus far, we have demonstrated that relative effective spreads for relocating stocks rise relative to their control in the days prior to their move, stay elevated for a period of time, and then fall back to normal levels. Realized spreads exhibit no such trend. To determine whether the effective spread result is robust in a multivariate setting, we estimate the following regression equation: Relative Effective Spread j, t = + 1 ( Trade Size j, t ) + 2 ( Post Relocation Dummy t )+ (1) 3 ( Pre-Event Adverse Selection j ) + 4 (High Adverse Selection Dummy j x Event Time Squared ) + 5 ( Low Moving Brokers j x Event Time Squared ) + j, t In equation (1), Relative Effective Spread j, t is the difference in the share-weighted relative effective spreads of the relocating stock j and its control on event day t, Trade Size j, t is the difference in the average trade size for relocating stock j and its control on event day t, Post Relocation Dummy t is equal to zero if the stock day is before the day on which the stock relocates and is equal to one otherwise, Pre-Event Adverse Selection is the difference in the adverse selection of the relocating stock and its control in the pre-event window, High Adverse Selection Dummy j is equal to one if relocating stock j has a greater-than-the-median 16

18 level of adverse selection in the pre-event window and zero otherwise, Time Squared is the event date squared, and Low Moving Brokers j is equal to one if relocating stock j has less than the 75 th percentile of brokers relocating (less than XX% of the floor brokers moving) and zero otherwise. We include fixed effects for each event day in the regression, but do not report those results. Based on theory, we expect 4 and 5 to be negative; consistent with a larger effect on Relative Effective Spread (more curvature in the time series of differences in effective spreads) for sample stocks with high adverse selection and few brokers moving. Trade Size is a control variable (we matched control stocks to relocating stocks on price, dollar volume, and trade volume). Since theory does not provide guidance as to whether effective spreads should be higher right before the relocation (due to the ending of relationships) or immediately after the relocation (when new relationships must be formed), we include the Post Relocation Dummy. We estimate a fixed effects (event time) regression model and report the results in Table VIII. [Insert Table VIII.] The difference in relative effective spreads between the relocating stock and its control stock is directly related to the moving stock s adverse selection during the pre-event period and the difference in the average trade size on the event day. The significantly positive coefficient on the Post Relocation Dummy suggests trading costs are higher when relationships must be initiated than they are when they are being terminated. Of particular interest, we see that the time series of differences in relative effective spreads is more extremely humped ( 4 and 5 are negative) for stocks with high adverse selection and/or for stocks that have few relocating brokers. These results demonstrate the robustness of the earlier univariate analysis. More generally, they provide compelling evidence that suggests reputations play a role in attenuating adverse selection costs on the floor of the NYSE. III. Floor Broker Trading Costs around Stock Relocations. While the evidence presented thus far suggests relationships on the floor of the NYSE are important, more direct tests of this hypothesis can be conducted by analyzing the trades of NYSE floor brokers around 17

19 stock relocations. In this section we use NYSE CAUD data to examine the effective spreads paid by liquidity demanders using NYSE floor brokers around the stock relocations in our sample. We begin with a simple univariate analysis of effective spreads for moving and non-moving floor brokers and find evidence consistent with our earlier results. We conclude our analysis with a multivariate analysis which confirms these results in a more controlled setting. A. Univariate Analysis [Insert Figure 4.] [Insert Table IX.] B. Multivariate Analysis [Insert Table X.] IV. Conclusions 18

20 References (Incomplete) 19

21 Table I Changes of Location on the NYSE s Floor On the dates listed below, the indicated number of stocks changed the location at which they trade on the floor of the NYSE. The move involves the entire panel of stocks moving to a new room on the floor, but continuing to trade as one panel. The apparent reason for the change in location is indicated. Our data begin in July of 1999 and end in April of Date of Location Change Number of Stocks Involved Apparent Reason for Location Change 07/28/99 11 Internal reallocation by Fleet 06/01/ /20/ /11/ Internal reallocation by LaBranche, Fleet, and Susquehanna Opening of new trading floor at 30 Broad Street Spear Leeds acquires Benjamin Jacobson 12/20/ Fleet acquires Meehan 03/25/02 28 Internal reallocations by Performance, Susquehanna, and Van Der Moolen 20

22 Table II Changes in the Crowd around Stock Relocations We capture the number of unique badge numbers the NYSE audit file indicates trade in the sample stocks in the weeks prior to the change in location. We examine the fraction of those brokers that also trade in the weeks after the relocation and compute the fraction of trades and shares the relocating brokers do prior to the move. Statistic of Interest Mean number of floor brokers making a trade prior to the specialist s relocation: Percentage of relocated stocks with one or more floor broker trades prior to the move: % Mean number of floor brokers making a trade prior to the specialist s relocation conditional on their being at least one floor broker trade prior to the specialist s move: Percentage of floor brokers that relocate with the specialist: 4.70 % Percentage of trades done prior to the specialist s move by brokers that follow the specialist: Percentage of shares done prior to the specialist s move by brokers that follow the specialist: 2.69 % 6.32 % 21

23 Table III Descriptive Statistics for the Relocating Stocks and their Controls Trade Price is the trade-weighted average price for a stock on a day (i.e., on a stock-day). # of Trades is the number of trades in a stock on a day. # of shares is the number of shares traded in a stock on a day. Each variable is computed on a daily basis from event day -120 through event day -80. An observation for a stock-day is included in this analysis if the Trade Price exceeds $5.00 and there are at least 5 trades. The mean Trade Price for the Sample is computed by taking the average Trade Price of all stock days surviving the screens for all sample stocks from event day -120 through event day -80. Variable Mean Median Std. Dev. Minimum Maximum Trade Price Sample $14.08 $12.86 $6.30 $5.00 $94.10 Control $15.72 $13.91 $8.67 $5.00 $99.88 # of Trades Sample ,459 Control ,723 # of Shares Sample 149,743 33, , ,079,700 Control 141,091 37, , ,844,300 Trade-Weighted Effective Spread Sample $ $ $ $ $ Control $ $ $ $ $ Share-Weighted Effective Spread Sample $ $ $ $ $ Control $ $ $ $ $

24 Table IV Differences in Relative Effective Spreads Between Relocating and Matched Stocks Control stocks are selected based on market capitalization, share price, trading volume, and stock price volatility. For each sample stock on each event day, the mean difference in relative effective spreads is computed by taking the difference between the shareweighted relative effective spread of a relocating stock and the share-weighted relative effective spread of its matched, non-moving stock. The parametric difference in means test is conducted using the Boehmer, Musumeci, and Poulsen (1991) t-statistic, using event days -120 through -80 as the base period. The nonparametric Wilcoxon test is used to detect differences in the medians. Days Relative to Relocation Difference in Effective Spreads Days Relative Difference in Effective Spreads Mean (bps.) Median (bps.) to Relocation Mean (bps.) Median (bps.) Event Day ** ** * ** ** ** ** ** ** * ** * * * ** ** * * ** * * * ** ** ** * * * ** * * * ** ** ** * * ** ** * ** * * * * ** ** * * ** ** * * ** ** * Indicates the difference is statistically significant at the 0.95 confidence level. ** Indicates the difference is statistically significant at the 0.99 confidence level. 23

25 Table VI Differences in Relative Effective Spreads Between Relocating Stocks with High Adverse Selection and Their Matching Stocks Control stocks are selected based on market capitalization, share price, trading volume, and stock price volatility. We only consider sample stocks with above-median levels of adverse selection. For each of these sample stocks on each event day, the mean difference in relative effective spreads is computed by taking the difference between the share-weighted relative effective spread of a relocating stock and the share-weighted relative effective spread of its matched, non-moving stock. The parametric difference in means test is conducted using the Boehmer, Musumeci, and Poulsen (1991) t-statistic, using event days -120 through -80 as the base period. The nonparametric Wilcoxon test is used to detect differences in the medians. Days Relative to Relocation Difference in Effective Spreads Days Relative Difference in Effective Spreads Mean (bps.) Median (bps.) to Relocation Mean (bps.) Median (bps.) Event Day ** ** * ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** * ** * ** ** ** ** ** ** ** * ** * ** ** ** * * ** ** ** ** ** ** ** ** ** ** ** ** ** * ** ** ** ** * ** * * * ** ** * Indicates the difference is statistically significant at the 0.95 confidence level. ** Indicates the difference is statistically significant at the 0.99 confidence level. 24

26 Table VII Differences in Relative Realized Spreads Between Relocating Stocks with High Adverse Selection and Their Matching Stocks Control stocks are selected based on market capitalization, share price, trading volume, and stock price volatility. We only consider sample stocks with above-median levels of adverse selection. For each of these sample stocks on each event day, the mean difference in relative realized spreads is computed by taking the difference between the share-weighted relative realized spread of a relocating stock and the share-weighted relative realized spread of its matched, non-moving stock. The parametric difference in means test is conducted using the Boehmer, Musumeci, and Poulsen (1991) t-statistic, using event days -120 through -80 as the base period. The nonparametric Wilcoxon test is used to detect differences in the medians. Days Relative to Relocation Difference in Effective Spreads Days Relative Difference in Effective Spreads Mean (bps.) Median (bps.) to Relocation Mean (bps.) Median (bps.) Event Day ** ** * ** * ** ** ** ** ** ** * Indicates the difference is statistically significant at the 0.95 confidence level. ** Indicates the difference is statistically significant at the 0.99 confidence level. 25