The Impact of Architectural Features on Global Equity Market Performance: How Harmful is Opacity for Trading Success?

Transcription

1 The Impact of Architectural Features on Global Equity Market Performance: How Harmful is Opacity for Trading Success? Peter L. Swan Joakim Westerholm December 21, 2004 Abstract Utilizing a system of structural equations and a unique intra-day dataset for 33 major exchanges, we evaluate several types of market architecture and numerous architectural design features to explain relative trading costs, volatility, trade size and parcel numbers. We find that trading is sensitive to transaction costs and minimum tick size, the limit order book design is ideal for all but small stocks; transparency is generally preferable to opacity, and evidence of economies of scale and scope. We rank the performance of every exchange relative to predicted best practice to conclude that all exchanges have scope to improve performance drastically. Keywords: Exchange trading systems, Architecture, Performance, Transparency, Decimalization JEL Classification:G10, G15, G2 Peter Swan gratefully acknowledges financial support from the Australian Research Council (ARC) DP , and Australian Professorial Fellowship and Joakim Westerholm from OKO BANK Group Research Foundation and the Australian Capital Markets CRC. We gratefully acknowledge data provision from Reuters and SIRCA. We thank Jim Berry, Ekkehart Boehmer, Zhian Chen, Gerald Garvey, David Feldman, Doug Foster, Peter Ho, Ron Masulis, George Sofianos, Terry Walter and participants at the 16 th Australasian Finance and Banking Conference 2003, ASX seminar, Western Finance Association Conference, Vancouver, 2004, European Finance Association Conference, Maastricht, 2004, Economics School Seminar, 2004, the executive committee of the Securities and Derivatives Industry Association (SDIA), and 17 th Australasian Finance and Banking Conference, keynote session, 2004, for comments and Mats Grankvist for valuable assistance. The original title was The Impact of Market Architectural Features on World Equity Market Performance Copyright Peter L. Swan and Joakim Westerholm. All rights reserved. School of Banking and Finance, Faculty of Commerce and Economics, University of New South Wales, Sydney NSW 2052 Australia. peter.swan@unsw.edu.au. P. Joakim Westerholm, School of Business H69, University of Sydney, NSW 2006, Australia. j.westerholm@econ.usyd.edu.au.

2 In today s increasingly competitive global environment for stock exchanges the survivors are likely to be those exchanges that manage to improve their performance; creating markets with low transaction costs, a higher dollar value of investor trades for the typical stock, with more stock parcels trading. In short, the performance focus is on traded value, which is the product of the average dollar value of parcels of a given stock and the number of parcels traded. Fundamental to this approach is the notion that the demand for trading is downward sloping. Like any other product or service, if we offer investors better terms on which to trade, not just transaction costs but every conceivable dimension of service quality such as the appropriate degree of anonymity or transparency, then more trading services will be demanded with more stock parcels trading for a given parcel size. What then is the best strategy for the exchange or regulator to set to meet the investor welfare goal of maximizing traded value, reflecting the requirements of diverse investors in a global market? We attempt to address this question by establishing criteria for ranking the performance of all stock exchanges and benchmarking world best practice in terms of marketmicrostructure stock exchange architectural designs. One of the aims of our ranking is to encourage exchanges to adopt the most efficient architectural designs, under the control of the exchange or regulator, which we identify. Even the best currently performing exchanges should more than triple traded value, weaker ones a multiple of this and in some cases add trillions of dollars to trading activity currently discouraged by higher cost and less effective microstructure designs. Additionally, we establish a market microstructure modeling approach that distinguishes between the short- and long-run solutions to the system of equations describing the performance of all major exchanges. We utilize the Hausman statistical test to show that microstructure models should model relative transaction costs, volatility, the average value of trades and the number of trades as endogenous within a system of structural equations using a two-stage or three-stage estimating framework. In addition, we identify benefits from using daily panel data and a system of endogenous equations as a basis for examining both the cross-sectional and time-series (event study) implications of a whole host of market architectural design features. Transaction costs, volatility, trade size and the number of trades are normally endogenous in most microstructure models but existing econometric tests largely treat them as exogenous. Our simultaneous equation approach overcomes the usual endogeneity bias or errors in variables problem, and by utilizing a dynamic model with daily panel data overcome problems inherent in purely cross-sectional modeling. In addition to modeling the full gamut of stocks traded using this approach, we focus on the largest and smallest quintile to highlight the differences in required designs, and investigate controversial design features that may particularly advantage or disadvantage large institutional traders.

3 The most severe form of illiquidity occurs on non-trading days. Most analyses normally ignore non-trading days with nothing recorded due to illiquidity. Nonetheless, like the dog that did not bark in the Sherlock Holmes story, what does not happen can be more significant than what does. We address many of the features of stock exchange architecture that alter the probability of trade occurring to show that exchange features that drive successful trading in small stocks also promote trading when otherwise a no-trading day would have been recorded. As part of a series of robustness checks, we test the model s predictions for many design features that are almost purely cross-sectional and thus do not alter, against five major changes to architectural design that we examine as event studies. Our findings are mostly supportive of the largely cross-sectional predictions, meaning that there is a concordance between the cross-sectional and time-series impact of architectural features. We also replicate the cross-sectional methodology that has already received attention in the literature for examining traded value. Our findings utilizing 4,271 individual stocks summarized on a cross-sectional basis are supportive of our dynamic panel data analysis based on 1,268,188 daily observations and a larger number of stocks, totaling 4,631. Out-of-sample predictions of all variables of interest are also extremely accurate with explanatory power of up to 86% and relatively unbiased forecasts. Finally, the impact factors derived by solving for the reduced form of equations explicitly representing demand and supply for the components of traded value, incorporating all a priori information from microstructure and economic theory, are vastly superior to conventional leastsquares-no-restrictions (LSNR) reduced forms estimated via ordinary least squares (OLS) assuming no knowledge of the structural equations. Impact factors derived from the structural model are extremely accurate in that they explain over 93 percent of the variation in traded value expressed in levels across the 33 exchanges whereas the correlation between the predicted values using LSNR and actual values is negative. Choices that reduce transaction costs in the form of the effective spread are the basis of most market microstructure empirical work and market design recommendations. Goettler et al. (2004) point out, in the context of a simulated order book, that the unobservable true transactions cost correlates negatively with the observed effective spread. Welfare significance arises from efficient design that facilitates higher traded value. This is one of our motivations for going beyond the traditional exclusive reliance on transaction costs to include all the components of traded value in our model, with endogenous trading cost and volatility each feeding into the two components of traded value. The traditional focus on observable trading costs alone could be misleading. We find 3

4 that the observed relative effective spread is a good, but not perfect, negative predictor of traded value across 33 world exchanges. Our findings are also strikingly inconsistent with Kyle (1985) in which noise traders with an entirely inelastic demand for trading undertake the bulk of trading. Noise traders are more rational that they might first appear, because, unlike their more informed brethren, they are discouraged by higher trading costs. This explains why we would expect the proportion of relatively uninformed traders to be higher in large, low-transaction-cost stocks with a higher proportion of public to private information (Easley and O Hara, 2004). The present study analyzes the relationship between exchange performance and market architecture, including factors that exchanges can alter and institutional/environmental features that are outside their control. Utilizing the world s most comprehensive intraday database with coverage of about 240 exchanges, our analysis of 33 of the more major exchanges, capped at 200 companies per exchange, enables us to assess the impact of each design feature on exchange performance. Participants desire to approach long-run optimal values in a partial adjustment framework with geometric distributed lags. Hence, the long-run trading cost elasticity of demand for trading, after taking into account trade size, appears moderate at 0.47 while the short-run elasticity is relatively inelastic within our dynamic system. Our structural equation approach also elucidates the actual trading process: the way in which volatility detracts from trading in the long-run, with an elasticity of 0.54, which is similar in magnitude to the equivalent elasticity with respect to trading costs. A larger trade size in dollar terms is associated with both higher trading costs and lower volatility. These findings are in contrast to Brennan and Subrahanyam (1998), hereafter, B&Subra, although the results are closer when we replicate their cross-sectional methodology. Evidence is found of economies of scale and scope in the trading process with the relative trading costs incurred for each trade declining 22 percent for every doubling of the number of trades and largely unaffected by trade size since economies of scale are offset by more asymmetric information and also by volatility. These findings on economies of scale and scope are generally consistent with those of Hasan and Malkamaki (2001), and Hasan et al. (2003), even though they adopt a different approach based on exchange costs recognized in accounting statements. In conformity with theoretical predictions, if the number of shares traded in two stocks is the same but one trades far more frequently than the other does, that stock will be cheaper to trade (Madhavan, 1992). Smaller trades occur in lower-cost and more volatile markets, with both high 4

5 cost and high volatility inimical to the number of parcels traded as well as to traded value. There is other evidence of scope economies with a doubling in the number of stocks listed raising traded value by up to 92 percent. As we double the market capitalization of a company, a proxy for lower asymmetric information, trading costs fall by 11 percent and traded value climbs by 34 percent due to a larger endogenous trade size. Although it is not our primary focus here, the model suggests the possible eventual demise of individual domestic exchanges and the creation of a single integrated global market adopting world best practice to better reap these economies of scale and scope. Would domestic exchanges continue to have a role? The setting and enforcing of listing requirements could become their major focus. Conditional on a range of design features relating to transparency and other matters, the market design with the lowest realized volatility and highest traded value for the entire dataset is the electronic limit order book (LOB) market. This provides some justification for Glosten s (1994) prediction that LOB markets will dominate in that they appear impervious to competition from dealership markets. The electronic LOB improves on floor-traded markets such as the New York Stock Exchange (NYSE), stocks with affirmative dealers such as the NYSE and smaller Euronext stocks, and hybrid dealerships/electronic communication networks (ECNs) such as Nasdaq. Electronic LOBs supplemented by affirmative dealers for typically smaller stocks have the lowest overall trading costs. An exception to the out performance of LOBs is hybrid dealer markets for the largest stock quintile. Despite higher trading costs, dealership/ecns markets in this quintile reduce volatility sufficiently to raise traded value. This finding is supportive of the predictions of gametheoretic models such as Madhavan (1992) and Shin (1996) that emphasize the role of dealer competition in reducing margins and coping with asymmetric information by absorbing idiosyncratic risk. The possibility of coexistence between dealers and the LOB, with some degree of specialization according to size, and search-informational considerations, is emphasized in the theoretical models of Pagano (1989), Viswanathan and Wang (2002), Parlour and Seppi (2003) and Snell and Tonks (2003). The second exception to the out-performance of pure electronic LOBs is stocks trading in a LOB with affirmative dealers for the smallest quintile of stocks. In addition, markets with exchange floors, dealership markets and stocks with affirmative dealers are all better than electronic LOBs for discouraging extreme illiquidity in the form of non-trading days. However, we do find that affirmative dealers can absorb idiosyncratic risks associated with both large and small stocks and both hybrid dealers and affirmative dealers ameliorate the problem of non-trading days on which a 5

6 stock is too illiquid to trade at all. Clearly, simple LOBs are deficient in terms of their ability to discourage extreme illiquidity/non-trading. Another striking finding from the study is that most transparent design features outperform opaque features by reducing information asymmetry. Green et al. (2003) find that intermediaries exercise more monopoly power in an opaque bond market, particularly with respect to smaller trades. An implication of the market power hypothesis is that in markets that are more transparent, smaller trades should receive many of the transaction cost benefits of large trades, since there is less opportunity to exploit smaller traders. When we include interactions between the (partial) marketdepth transparency variable and trade size, our findings are not supportive of the market power hypothesis. The execution costs of large trades relative to small trades improve in progressively more transparent markets. Reduced post-trade transparency, due to delayed reporting of block facilitating principle trades, is harmful overall, and for all quintiles investigated, because it directly increases asymmetric information, as well as promoting non-trading. A partial exception is an event study: the introduction of delayed reporting of very large facilitated principal trades by the Australian Stock Exchange (ASX) towards the end of our sample period. Its initial impact appears to have lowered transaction costs but with an offsetting rise in volatility. Pre-trade disclosure of broker IDs to other brokers and to investors is beneficial overall and for the largest stock quintile since it reduces volatility, but disclosure can harm smaller stocks and increase the probability of non-trading. An event study associated with the removal of pre-trade broker ID when Paris adopted Euronext rules supports these findings. Broker ID opacity raises transaction costs by between 30 and 42 percent, but leading to only a small decline in traded value. Our system of simultaneous equations provides an explanation as to why many reforms which reduce asymmetric information and lower trading costs do not impact significantly on traded value. An associated rise in volatility discourages trading. For the overall sample and the smallest quintile, revealing the depth of the LOB partially, leads to both lower transaction costs and volatility, boosting traded value by 27 percent and discouraging non-trading. For the largest stocks, partial revelation of depth appears to be harmful while there are benefits from moving to full disclosure. At least partial revelation of depth encourages trading, as opposed to days with no trades occurring. The only exception to the adverse impact of opacity is iceberg orders, a form of pre-trade opacity that does not reveal the full depth of the LOB. Surprisingly, these are beneficial overall and for encouraging trading when it might not otherwise 6

7 occur. An opening call auction confers benefits to trading in the largest stocks during the day by dramatically lowering volatility, despite the fact that it has an adverse impact on trading costs. This is because the opening attracts more uninformed trading away from the regular trading mechanism, raising asymmetric information during the day. The auction is also associated with a higher probability of non-trading. The auction leads to lower volatility for the largest stocks, presumably due to better price discovery. The strong overall adverse impact of the opening call simply reflects the transfer of trading from the main exchange to the opening. We control for the exchange environment in a number of ways. By classifying securities according to their effective spread plus exchange taxes and charges, volatility, average trade size, number of trades, the market to book ratio, and the market capitalization of each stock, we aim to determine what type of market architecture is the most suitable for each market segment. These characteristics also control for the nature of the stocks on each exchange. We introduce a range of proxies for trading demand: the market capitalization of equity for each country, the number of listed companies, aggregate income as measured by GDP, the population size each exchange draws on, and the ability to arbitrage cross-listed stocks, as indicated by opening hours shared in common with New York. We also control for shareholder rights conferred by the legal system and brokerage fees for institutional investors on a country and an exchange-by-exchange basis, respectively. The aim is to account for differences between exchanges in how well they have been able to establish their position in the competitive global market place, overall, in extreme size quintiles and their ability to discourage the severe problem of non-trading. Upstairs dealer markets grafted onto LOB markets may in general filter out less informed trades, giving rise to a higher spread in the LOB. For the largest stock quintile, which is most likely to be representative of upstairs trading, there is no reduction in traded value because of a reduction in volatility. The paper proceeds as follows: Section I provides a literature review while II introduces the performance measures and III the market design features. Section IV outlines the data used. Section V presents the model and empirical findings while Section VI concludes. I. Literature Review Perold and Sirri (1997) establish a variation in trading costs across international borders using information about institutional investors intents, as well as executed orders, and are thus able to measure implementation shortfalls and market impact costs. Domowitz et al. (2001) examine the 7

8 cross sectional variation in total trading costs using information on institutional trading on 42 exchanges. Low and falling transaction costs promote higher and rising stock turnover. These studies do not directly associate market design and institutional features with exchange performance. Venkataraman (2001) compares the Paris and the New York exchanges to conclude that the New York floor trading system is superior to an electronic LOB. Jain (2002) examines the impact of various market designs on liquidity and finds, based on sampled observations of daily closing bid-ask spreads on 51 exchanges, that dealer-emphasis markets have higher transaction costs than LOB markets when emerging markets are included. Most previous literature analyzes one institutional characteristic at a time and compares two exchanges. La Plante and Muscarella (1997) examine market impact costs for block trades and find that liquidity provision for blocks is superior on the NYSE compared with Nasdaq. Chan and Lakonishok (1997) compare institutional trading on the NYSE and Nasdaq. They find that smaller stocks gain better execution on Nasdaq and larger ones on the NYSE. Bessembinder and Kaufman (1997) find that both transaction costs and volatility is higher on Nasdaq than the NYSE for comparable stocks. Two studies examine what at first blush appears to be a change in a single feature, pre-trade display of broker IDs to other brokers, on the Paris exchange (Foucault et al., 2003; Comerton-Forde and Frino, 2004). However, it is rare for an exchange just to change a single feature at one time. For example, while the introduction of the Euronext trading system occurred in Amsterdam, Brussels, and Paris on different dates, Paris moved to a more anonymous system with respect to broker IDs at the same time that it implemented the standardized Euronext trading system. Few studies take a wider cross sectional approach. Even if focusing exclusively on events gives more easily interpretable results closer to the ceteris paribus ideal, a cross sectional study across a wider array of exchanges should better address the problem that most exchanges differ by more than one architectural or institutional characteristic. Event study evidence alone, while often useful, is conditional on the complex array of policies in place on a given exchange and thus might be unreliable as a guide for designing an ideal exchange. Similarly, the determination of an optimal array of architectural features should account for complex interrelationships between architectural and institutional characteristics, together with macro-economic features that generate trading demand. The findings of traditional event studies may not always be robust with the altering of one or more architectural variables, or even a new trading platform introduced, due to endogenous changes in 8

9 asymmetric information, the size and number of trades, demand conditions and volatility. When embedded in a system of equations with endogenous variables and a complex array of exogenous controls correctly identified and modeled, event studies become more reliable. Trading on every exchange acts as a control for every other exchange with all the global relationships tied together by the underlying economic model and system of equations. Below, we illustrate with a series of five architectural design changes modeled with new event methodology, the strength of our approach. Occasionally the subject of event studies, and one of the most critical architectural decisions facing exchanges, is the question of transparency, both pre- and post-trade. Chowdhry and Nanda (1991), Forster and George (1992), Madhavan (1992, 1995) and Lyons (1996) address the transparency issue from a theoretical perspective. In a similar vein, Pagano and Röell (1996) establish that uninformed investors benefit from the greater transparency, which is inherent in auction markets such as pure public LOBs but not in dealer markets. Spreads should thus be lower in a transparent LOB with immediate reporting of all trades and revelation of broker IDs. Madhavan (1992) and Shin (1996) model differential information in a game-theoretic setting in which dealer markets are less prone to informational uncertainties than are decentralized order-driven markets. Bloomfield and O Hara (1999) find that spreads could be wider with greater transparency in an experimental approach. Likewise, Flood et al. (1999) adopt an experimental approach. Since both theoretical models and experimental markets are far from conclusive about the impact of transparency on financial markets, we now turn to empirical studies. Gemmill (1994), relying on several changes to post-trade transparency made by the London Stock Exchange (LSE), including a 90 minute delay and a 24 hour delay, found that delayed publication of block trades did not consistently reduce transaction costs and did nothing to improve liquidity. Grammig et al. (2001) find that uninformed traders prefer the non-anonymous traditional floor trading mechanism while informed traders prefer a relatively anonymous electronic LOB system. The adverse selection cost component of trading costs is higher in the system attracting more informed trading. If facilitating principle traders are given time due to delayed reporting of blocks to unwind a position, this introduces asymmetric information as a policy choice. Does such a policy give an unfair advantage to large brokers at the expense of other market participants and, consequently, have a detrimental effect on the liquidity of the entire market? Madhavan et al. (2004) find evidence of a decline in public liquidity of stocks on the Toronto Stock Exchange (TSE) following greater pre-trade transparency of orders in They attribute this to a greater propensity for picking off of orders viewed as free options. Boehmer et al. (2004) 9

10 analyze the introduction of pre-trade transparency to LOB on the NYSE in January This was a response by the NYSE to the earlier introduction of decimalization of the quote size. Prior to this time, only the best bid and ask was visible. Contrary to the findings of Madhavan et al. (2004) for the TSE, they find an increase in liquidity with additional orders attracted to the LOB. Lee (1998) provides a discussion of transparency issues. Another controversial issue is the minimum tick size, or the minimum dollar difference in the price of a trade. A lower minimum tick may reduce market depth by as much as it lowers trading cost, either leading to no change, or an adverse impact on the value of trading. Nasdaq and NYSE have moved recently from one-eighth to one-sixteenth of one dollar and then, finally, to decimalization of the minimum tick size in response to regulatory demands. An early contribution was Harris (1994) who used simultaneous equation modeling to predict the effect of smaller tick sizes. The recent empirical literature includes Goldstein and Kavajecz (2000), Graham et al. (2003), Chakravarty et al. (2004a), and Bessembinder (2004). The consensus appears to be that quotes have fallen because of decimalization, facilitating a larger number of smaller trades, but Chakravarty et al. (2004a) find evidence that overall liquidity has fallen with reduced depth for larger trades and lower overall liquidity. A small minimum tick size reduces the importance of price and time priority and makes it possible for traders to front-run posted limit orders that may potentially have information content. By contrast, Bessembinder s (2004) findings support the earlier predictions made by Harris (1994) with no evidence of a liquidity decline. II. Trading Mechanisms and Performance Measures A. Trading mechanisms We consider the following types of trading mechanisms: (i) (ii) Dealer_Hybrid_Dummy i, a hybrid market with continuous dealer presence and the option of an order book (e.g., Nasdaq and associated ECNs) to which a value 1 is assigned to the four exchanges meeting this requirement and 0 otherwise; a pure public order driven electronic LOB (e.g., the Australian Stock Exchange, ASX) which has, in addition to the LOB, voluntary market-makers and possibly an upstairs dealer-market for exceptionally large institutional trades to which a value 0 is assigned to the 16 qualifying markets; 10

11 (iii) (iv) Stock_Affirmative_Dealer_Dum i is a variant on (ii), with designated dealers with an affirmative obligation to provide firm and continuous quotes and to trade against the wind to limit volatility (e.g., some Euronext stocks in several European countries and all NYSE stocks). A value of 1 is assigned to qualifying stocks in the 13 exchanges falling into this category and 0 otherwise. Market_with_Exchange_Floor_Dum i, a traditional floor trading system (exclusive to NYSE and Frankfurt) to which a value of 1 is assigned and 0 otherwise. The pure electronic LOB assigned a value of 0 thus becomes the standard of comparison for each alternative market type. In a LOB market with designated market makers in some or all stocks, entry of dealers is controlled but incumbents obtain privileged status such as the absence of trading fees in exchange for obligations. Of course, the NYSE is unique in a number of ways. Every stock has a designated dealer and that dealer is unique to that stock as the specialist. The specialist also operates visibly on the trading floor. Under a pure electronic-order-book trading mechanism, entry of non-designated market makers is free, but there are no concessions granted to or obligations imposed on broker-dealers acting in this role. The NYSE is the only exchange to qualify in two categories, as both an affirmative dealer and as a floor-based system. B. Performance Measures We measure exchange performance from four main aspects: transaction costs, volatility, average dollar value of trade size and the number of trades, the counterpart of trade duration or gap between trades. As mentioned, the product of the last two variables generates the dollar value of trades, which is of welfare significance to traders. Deflating traded value by market capitalization yields the stock turnover rate, a standard measure of liquidity. The more a trading system facilitates the trading desires of participants due to superior design, the greater will be traded value expressed in dollar terms. Apart from being of critical value to investors whose objective is to trade, the dollar value of trades is of particular concern to most exchanges since, apart from listing fees, levies on traded value are typically the primary source of exchange income. Exchanges that are cooperatives of broker/dealers will not necessarily wish to maximize traded value since the exchange owners may be able to extract rents from traders via designs that discourage trading relative to the optimal trading design. Differing pressure from brokers may help to explain the diversity we find in market architectural solutions, together with departures from best practice. When an exchange employs an opening or closing auction mechanism, a single large trade is 11

12 recorded which distorts measures of average trade size and trading costs in the continuous market. We thus exclude such trades from our trade dataset, which we compute from trade-by-trade data over the period of normal trading, excluding both opening and closing calls. The measures of transaction costs are calculated using intra-day, trade-by-trade data. Every time a trade occurs, a bid-ask spread is observed, either as the difference between best bid and ask in a LOB environment or as the difference between the quoted buy and sell price in a dealership environment. 1 As the primary transaction cost measure, we use the effective spread, which measures how far from the mid-point of the spread trade execution occurs. It also includes the benefits of any price improvement over the quoted spread. We add information to the effective spread measure by weighting it by the number of shares in the parcel when we calculate the daily average. See Lee (1993) and Chalmers and Kadlec (1998) for earlier applications of the effective spread. We calculate the trade volume weighted relative effective spread as: Askt + Bidt t ABS t c Trade Price 2 trade volumet 2 t= t Ask 0 t + Bid t total trade volume, (1) 2 where t o is the time when regular trading commences during a trading day following an opening algorithm, t is time when a trade is executed, t c is the time when trading ceases for the day, and trade volume refers to the number of shares traded rather than the traded value. We follow convention by doubling the effective spread on a single trade to compute the round-trip cost. A smaller spread indicates lower transaction costs. There are, however, five major components of transaction costs: brokerage, bid-ask spread, market impact, exchange fees and taxes (stamp duty), as well as the short-fall cost of an inability to make a desired trade. If we knew the entire size of each share parcel, together with time-stamps for the first and last shares in the parcel, we would be in a better position to measure accurately market impact costs. The effective bid-ask spread is one way to take into account the market impact effect. Since we calculate the trade weighted effective spread, the size of the executed trade has an impact on the spread. Because government imposed exchange taxes and transaction fees paid to the exchange all add to trading costs, we obtain these for exchanges where they are significant (e.g., the LSE and China) and add the round-trip cost to the 1 In dealer markets, quotes are often only indicative to provide a degree of protection to the dealers themselves. 12

13 effective spread to obtain our estimate of trading cost. Brokerage charges will clearly impinge adversely on traded value and are thus included in the equations explaining the components of this value. We exclude brokerage directly from our transaction cost measure, partly because we lack information on a trade-by-trade or stock-by-stock basis but, more importantly, because we are interested in how architectural design features determine the effective spread. Transaction costs are important for the performance of an exchange since lower transactions cost induce a higher level of trading activity. The responsiveness of trading to trading costs and the impact of taxes such as stamp duty on trading activity is an important and controversial issue. Increasingly, global fund managers have discretion about where trade execution occurs. Pulatkonak and Sofianos (1999) show that the allocation of trades in US cross-listed stocks responds to the relative transaction costs in the different global exchanges markets. As global competition intensifies, exchanges are motivated to lower execution costs. As our volatility measure, we use daily volatility, calculated as the squared daily continuously compounded close-to-close return: Volatility i,t = pt c 2 [(ln( )], (2) p ( t 1) c where t c is the time when trading ceases at the end of the trading day and ( t 1) c is the time trading ceased the previous day. This measure is a proxy for realized volatility computed from intraday data described by Andersen et al. (2001). We also experimented with the five-minute and 15- minute standard deviation of returns computed from the intraday trades and quotes. These measures proved unsatisfactory relative to our volatility measure, especially for relatively illiquid stocks subject to thin trading. We could attempt to allow for some double counting of trades in dealer markets such as Nasdaq, but do not do so because of the difficulty of making reasonably precise estimates across a number of markets where the degree of double counting is declining over our sample period. However, even with some limited double counting, markets such as Nasdaq are not rated highly. Hence, we do not believe that any double counting unfairly biases our results in favor of dealer markets. By breaking up traded value into its two natural and distinct components, we are able to analyze the impact of asymmetric information on trading costs since information is more likely to be contained in larger trades. Moreover, the impact of market architectural features is likely to be quite different on these two components of traded value. For example, lower trading costs will result in more 13

14 trades but the impact on trade size could go either way. While we do not specifically incorporate factors such as execution speed and rapidity of price discovery, for example, floor markets are typically slower than dealer markets; we believe these are implicit in our traded value method. Ceteris paribus, faster execution systems are likely to encourage a greater value of trading. III. Architectural, Governance and Institutional Features Market architectural features included in this study focus on the type of trading mechanism used and on the features provided to the market participants using the mechanism. Apart from categorizing exchanges according to the four basic types described in Section II above, there are a large number of other architectural features related to trading systems and rules. We include estimation with a selection of such variables to assist stock exchanges in improving their market architectural design and to test a variety of theoretical models. In this study, our focus is the efficiency of trading systems with different designs. Many microstructure studies actually or potentially encounter the problem of simultaneity bias as most of the explanatory variables such as transaction costs, volatility, size, and numbers of trades are endogenous and thus cannot, or should not, be used as explanatory variables in standard regression models. This may make it virtually impossible for empirical studies that encounter these problems to appropriately test theoretical models and provide policy guidance. However, several studies address issues arising from endogenous variables using two stage least squares (2SLS) structural equation methods; see for example, Harris (1994) and B&Subra. These methods, like ours, are similar to an instrumental variable model in which the instruments are exogenous variables. It therefore helps to mitigate the errors-in-variables problem. Some special variables that relate to issues of governance seem to belong naturally in all four structural equations as they affect both supply and demand. The most notable of these is the La Porta et al. (1998a) and (1998b) definition of countries that support shareholder rights. Shareholder rights, including supportive legal systems, should, by improving the climate in which trade takes place, lower transactions costs, lower volatility, raise the dollar value of trade size and possibly increase the number of parcels traded over and above the increase due to lowered trading costs. A country with shareholder rights is assigned a dummy variable 1, and one that does not is assigned a value of 0. We incorporate a second governance variable, the market to book ratio defined at the individual stock level rather than the exchange or country level. The higher is this ratio the better managerial ability for the firm concerned and the more it qualifies as a growth stock. It is also the 14

15 inverse of the Fama-French risk factor and bankruptcy risk is low. We predict that transaction costs should be lower for these reasons, and volatility should be higher as the valuation of growth stocks is more difficult. The direct impact on trade size and trade numbers is harder to predict a priori. We experimented with country variables representing the quality of regulatory regimes, anticorruption measures and the enforcement of the rule of law from World Bank sources (Kaufmann, et al., 2003) but these were not included because better-managed economies according to these criteria appear to have more poorly performing stock markets with higher transactions costs. They also do not relate directly to the functioning of stock markets, unlike, for example, shareholder rights. We define a structural model with a supply of transactional services, which we invert to express trading costs as a function of control and market architectural variables, together with a downward sloping demand for transactional services that reflects demand factors impinging on an exchange. We describe institutional features such as: (i) the impact of the aggregate equity market capitalization of all exchanges affiliated with the World Federation of Stock Exchanges in each country on the nature and magnitude of trading; (ii) similarly, the impact of income, as measured by GDP; and (iii), the impact of the overall market size, as measured by population. We also consider (iv), the number of trading hours overlap with the New York Stock Exchange (NYSE) to capture the ability to arbitrage American Depository Receipts (ADRs) and provide additional liquidity for European and North American exchanges. None of the Asian or Oceania exchanges has concurrent trading times and hence we assign a value of zero. Other institutional factors include, (v), if it is an electronic LOB market, does it have an upstairs dealer market facility for block trades? Another included size control variable is the total number of stocks listed on the exchange, although the number of listed stocks will reflect more than simply size. IV. Data The original data provided by Reuters to SIRCA 2 contains intra-day trade, quote, and volume information for all securities listed on all world stock exchanges. We choose the 33 exchanges used in the study to provide a generalized cross sectional picture of world stock exchanges. Three are selected from North America, 15 from Europe, 10 from Asia, two from Oceania, two from South 2 This is an exclusive arrangement with SIRCA, Securities Industry Research Centre of Asia-Pacific, which represents a consortium of 25 universities, to receive and store all Reuters trading information. 15

16 America and one from Africa. We used NYSE Trades and Quotes (TAQ) data for NYSE and Nasdaq stocks. We collected consistent exchange specific information from the International Federation of Stock Exchanges, Annual Year Books, the official Internet home pages of the exchanges, and exchange rulebooks published by the stock exchanges. Demarchi and Foucault (2000) is the source of the European market designs while Naik and Yadav (2003) is the source of the latest changes in market design for the London Stock Exchange. The Australian Stock Exchange (2003) provides a microstructure survey of major exchanges for the period adjacent to our data sample. We confirm as much exchange information documented in the publicly available information sources through direct correspondence with the exchanges as we can. The most difficult part of the exercise was not only identifying architectural features that changed over our sample period, but also the actual date of the change and the stocks to which the change is applicable. We obtain brokerage fees for institutional investors on an exchange-by-exchange basis from Chakravarty et al. (2004b) and add-in comparable estimates for North American exchanges based on a Plexus AIMR report. We also collect institutional and country-based information on population and GDP from World Bank sources. Table I lists the investigated exchanges, the country, the full name of the exchange, the number of stocks listed on each exchange, the market capitalization of each exchange as of the start of 2000, and the classification of the exchange according to our schema. The included exchanges represent 96 percent of the capitalization of stock exchanges that are members of the World Federation of Exchanges. We collected a consistent set of exchange information regarding market architecture and institutional feature variables for all 33 exchanges. We also collected data on a number of smaller exchanges but in the end, we excluded these because we could not be certain of our coverage and data quality. The largest included exchange is the NYSE, a floor-trading system with a LOB and affirmative dealers ( specialists ), followed by Nasdaq, Tokyo, which is a pure LOB system, and LSE. INSERT TABLE I ABOUT HERE Reuters intra-day trading and bid-ask spread data is extracted for the period between start of March 2000 to end of October We selected up to 200 common stocks with the highest value of securities traded during the period selected from 31 exchanges, excluding the two US exchanges. We obtained Thomson Financial Datastream data on share numbers, stock splits and close-to-close returns for as many stocks for which the data is available, within the limitation of the top 200 stocks by traded value. To prevent over-representation of two of the world s largest exchanges, NYSE and 16

17 Nasdaq, we capped our inclusion at 200 of the largest stocks by traded value from each exchange, with NYSE Trades and Quotes (TAQ), University of Chicago Center for Research into Security Prices (CRSP) and Standard & Poors North American data Compustat as our sources. For smaller exchanges, we included the total number of available listed companies. The average number of included companies per exchange is 140. This selection process resulted in a relatively balanced portfolio representing world common stock markets while still giving representation to smaller exchanges. We obtained intra-day trade-by-trade prices, numbers of trades and average trade size expressed in dollars, and best bid-ask quotes or orders, whichever is applicable, for all included stocks, calculated comparable exchange rate adjusted measures using intra-day data, and present them as a daily time-series for each company. We expressed the series for the average value of each daily trade in each stock in all 33 exchanges in USD of the day. We added transaction taxes and exchange fees, expressed as relative measures on a round-trip basis, to the effective spread calculations based on equation (1) above. 3 This final sample represents approximately 74 percent of the market capitalization of the included exchanges at the start of the investigated period or 71.8 percent of world market capitalization. For further details, see Table AI in the Appendix. Since our market capitalization control might not fully capture the effect of size differences between stocks on large and small exchanges, we carry out a range of robustness checks on sample sets with more uniform stock sizes such as global quintile size rankings of stocks. The quintile of the largest stocks consists of 1,046 stocks and the smallest quintile, 1,075 stocks. See Table AI for further details. V. Model and Empirical Findings A. Descriptive Statistics Table II reports means and ranks for each of the 33 exchanges for six daily averages per stock: the trade weighted relative effective spread with the addition of exchange fees and taxes, the realized volatility, the trade size per stock expressed in USD, the number of trades, the traded value per exchange and the average relative minimum tick. The latter is defined as the minimum tick size for 3 Jim Thames of Arrowstreet Capital provided information on exchange fees and taxes for international exchanges, as well as the overlap in trading hours with the NYSE. 17

18 each stock within a given price range deflated by its closing price. The reported measures are average stock level measures of our included stocks for each exchange. While some of the large exchanges such as NYSE, Nasdaq and London figure prominently in the rankings, so do LOB hybrids with affirmative dealers such as Amsterdam and even electronic LOB markets such as the ASX. The large trade size value for exchanges such as New York, with 200 of the largest stocks, reflects the absence of smaller NYSE stocks in our capped sample. Interestingly, the highest volatility exchanges include the largest and apparently most successful dealership exchange, Nasdaq, and also a smaller LOB exchange, Korea while some of the lowest include countries such as New Zealand and Singapore. The ranking by relative minimum tick size puts volatile countries such as Korea with the lowest relative tick size. It is no coincidence to find that exchanges that perform relatively poorly based on purely architectural features, such as New Zealand and Singapore, have the highest relative tick size. While these countries with high relative tick sizes have decimal pricing, an unwillingness to reduce tick size in the face of low stock prices per share hampers exchange performance. These rankings incorporate a whole host of demand and environmental variables as well as market design. Hence, further analysis is required to identify and explain good architectural design. INSERT TABLE II ABOUT HERE Table III provides descriptive statistics for the 13 non-dummy variables and the correlation matrix for the same continuous variables, but replacing the market to book ratio that has extremely low correlations with all the other variables with the shareholder rights dummy. The high negative correlations between transaction costs and trade size, trade numbers, market capitalization for stocks and for the country as a whole not only indicates a strongly declining demand curve for traded value, but also why transactions costs and the traded value variable need to be treated as endogenous. The correlation matrix for the architectural dummies (not-reported) shows far lower levels of correlation for most of the design items. The high significance levels for most of the estimated coefficients in subsequent tables would suggest that multicollinearity is not a significant problem. INSERT TABLE III ABOUT HERE 18

19 B. Cross-Sectional and Time-Series Analysis We start with a pooled cross-sectional and time-series analysis on daily data, which we have aggregated from intra-day data. We thus include daily observations for the period March 2000 to end of October 2001 for each company for all 33 exchanges. The size of the dataset depends on the number of analyzed companies available and the number of included trading days, or 1,268,188 observations. Since some smaller stocks on smaller exchanges do not trade every day, the number of observations is smaller than the theoretical maximum. Thus, there are 365,139 trading days for the top quintile but only 208,680 for the bottom, indicating that illiquidity due to non-trading is a serious problem confronting exchange design. We compute the relative minimum tick size based on the stock s minimum tick deflated by the stock s closing price from Datastream for that stock/day, capturing both the days and stocks in which both the New York Stock Exchange (NYSE) and Nasdaq adopted decimalization of tick sizes during our sample period. We also compute the skewness and kurtosis of the four endogenous variables, relative trading cost, daily volatility, daily average trade size, and daily number of trades. We do the same for six exogenous continuous variables, the average stock market capitalization of each stock over the sample period, number of listed companies, total equity market capitalization for the country in which each exchange is located, income (GDP) for the country, together with its population and the minimum relative tick size. Subsequently, we carry out statistical specification tests, which confirm that, indeed, our four endogenous variables are truly endogenous. Taking logs of all 12 continuous variables reduces both skewness and kurtosis and enables us to create a simple linear in logarithms structure for our system of equations that is easily solved for the set of reduced-form impact factors eliminating all endogenous variables. A Box-Cox test confirms that the log specification for the endogenous variables describes them better. For each of the four endogenous variables, y, j ( 1,..,4), described below, we begin with a partialadjustment geometric distributed lag model (e.g., Greene, 2003): j t *, ( 1 λ )( ) j j j j j j t t 1 t t 1 t y y = y y + ε, (3) in which the adjustment of the actual level is a proportion of the difference between the desired level, *, j y t, and the actual level in the previous day. The equations to estimate become: ( 1 ) 1 ( 1 ) y = a λ + λ y + β λ x + ε, (4) j j j j j j j j j t t t t 19