Transcription

1 Dealer liquidity in an auction market: evidence from the London Stock Exchange Sylvain Friederich Department of Economics, University of Bristol and Université deparis-i Richard Payne London School of Economics November 24, 2001 Acknowledgements: The financial support of INQUIRE-Europe is gratefully acknowledged. Thanks to Ron Anderson, Jim Angel, Bruno Biais, Greg Connor, Vassilis Hajivassiliou, Roger Huang, Matthew Leighton, Jamie Lebetkin, and Ian Rowell, as well as participants at the EFMA and NASDAQ/Notre-Dame 2000 conferences, and seminar participants at the LSE, Said Business School at Oxford and Université de Toulouse-ESC. Remaining errors are our own. Department of Accounting and Finance, London School of Economics, Houghton St, London WC2A 2AE, U.K. Tel: (+44-20)

2 Abstract Using recent data from the London Stock Exchange, we examine order flow interactions between an electronic limit order book and a network of upstairs firms voluntarily supplying dealership services. We analyse the determinants of the market share of the limit order book in a panel framework. Our results suggest that competing, off-book liquidity suppliers voluntarily perform at least some of the stabilisation functions normally assigned to designated market-makers, without informational or other privileges: the share of order flow executed against their inventory is greater at times when different dimensions of liquidity are lacking from the order book, at times of high trading activity, large average trade size, and unbalanced order flow. These findings appear robust to possible endogeneity of the measures of order book liquidity. 2

3 A pervasive feature of equity market structures today is that there are multiple mechanisms via which a liquid stock can be traded. More specifically, market structures appear to have converged towards hybrid systems in which a central role is given to order-driven trading. The two main equity market reforms of the late 1990 s, of the Nasdaq market and the London Stock Exchange (LSE) respectively, were moves in this direction in that limit order trading was introduced to both markets. 1 Previously, they had been organised as pure dealership systems. However, in all mature equity markets, order books are supplemented by the parallel supply of non-anonymous dealership services. On the NYSE the public limit order book is supplemented by the activities of the Specialist, upstairs brokers and NASD broker-dealers ( third market ). The Paris Bourse is widely regarded as the market closest to a pure limit order structure. However, off-book trading is available on-exchange for transactions over a given size and, more importantly, London-based dealers provide off-exchange liquidity in actively traded French stocks. Other European exchanges also have separate (off-book) arrangements in place for the execution of large orders. 2 Given the above, it would appear that a crucial issue in current applied microstructure research is to understand the way in which such hybrid markets work. While past research has attempted to identify pure real-world auction or dealership structures in order to evaluate whether they conform to some theoretical blueprint, such pure examples are becoming increasingly rare. Hence the aim of this study is to empirically investigate the operation of a hybrid market. We seek to understand the nature of the order flow routed off the order book, and what trade characteristics or market conditions lead to the dominance of one mode of trading over the other. Order flow may, for example, be executed off the order book as the result of preferencing and internalisation arrangements, or through brokerdealers simply free-riding on book prices and cream-skimming uninformed order flow (Easley, Kiefer, and O Hara (1996)). Alternatively, it may be executed with dealers because of low liquidity or high volatility of quotes in the book, in which case dealers are providing a valuable service. Our agenda is different from that in the well-known study of Madhavan and Cheng (1997), where the authors use a switching regression framework to compare price impacts in the 1

4 NYSE upstairs and downstairs markets. Their focus is on the role of non-anonymity in trading venues, specifically the potential upstairs brokers have for screening informationbased trades. They find evidence consistent with the hypothesis that an unobservable reputational variable matters in the quality of order execution. We do not attempt to model the factors that influence the price impacts of individual trades executed in one venue rather than another as do Madhavan and Cheng (1997), but are interested in characterising the periods in which and market conditions under which the majority of order flow is sent to one trading venue versus another. Our focus is on the competition between trading locales for transaction activity, hence our lower frequency, order flow analysis. We address these issues using two months of LSE data recorded at the end of The current microstructure of the LSE comprises a central limit order book (called SETS) which the exchange describes as the main price formation mechanism and alongside the book there exists a network of broker-dealers who supply quotes on a bilateral basis - similar to an upstairs market. The London market constitutes a particularly interesting venue for a number of reasons. First, the LSE s market structure exemplifies current trends in trading systems, where more freedom of choice of trading venue is left to the buy side, and more competition is encouraged in those venues among liquidity suppliers. Specifically, there are no forced interactions between the two trading mechanisms either in price or quantity terms: no part of an order must be executed in the order book and the prices of off-book trades are not constrained by book prices. This is not usually the case in other large markets for order flow executed off the book but on-exchange. On the NYSE, for example, the Specialist is constrained in his supply of liquidity services in a number of ways. 3 Also, upstairs brokers are compelled to expose any block order on the downstairs market and fill any downstairs demand at the block price or better, thereby constraining the conditions at which they may supply liquidity. In Paris, only trades over a given size threshold can be executed off-book and only then at or within weighted average book prices. On the LSE, the final customer or his broker decides on the type of execution he desires and dealers have complete control over the terms of trade they are willing to offer. 2

5 The fact that price and time priorities, although enforced on the order book itself, are not maintained between both trading locales doesn t make the London trading arrangements unusual: price and time priorities are routinely violated in other exchanges as well. 4 In the US, time priority is enforced within the NYSE but not between the NYSE and the regional exchanges or third market broker-dealers. Further, it is well known (see Amihud and Mendelson (1991), or Angel (1998)) that institutional trades in NYSE stocks are reported to other exchanges (such as London) specifically to avoid the priority rules of the upstairs market. Similarly in Europe, large volumes of trading in Continental blue-chips are executed in London to bypass the strict priority rules of the home Bourses. To an extent, it is simply because fragmentation is unimpeded in London that all of the trading activity remains on-exchange, and ends up in our data. Whether this is good or bad for market liquidity is a separate issue. Stoll (2000) argues against enforcing strict priorities across markets because it forces them to compete on price only, and prevents them from competing on other dimensions of an order such as size. The coverage and quality of the data set are another reason why examination of London data is both novel and interesting. The data provided by the exchange accounts for essentially all the trading activity in these liquid stocks. The only marketplace that competed with the LSE was a system called Tradepoint. 5 This competitor has handled a negligible fraction of order flow since the launch of SETS (Board and Wells, 2000). With regard to data quality, the information made available by the LSE covers all trades in the relevant stocks whether on or off-book. The data set also contains information on all orders placed on the SETS system. Again, this is unlike what is obtainable from most other exchanges. In certain cases, for example most NYSE data, the data does not normally allow one to discriminate between trades facilitated upstairs or downstairs, or trades which were executed against the Specialists inventory. For most European exchanges, the data available to researchers simply contains no information on trades executed off-exchange by the member firms. As the scale of this trading activity is large this is a serious deficiency. 6 Finally, the fact that the current structure of the LSE is broadly comparable to that observed in many other - non-u.s. - equity markets implies that it is possible to derive useful infor- 3

6 mation from analysis of the LSE data which is more generally applicable than that derived, for example, from the somewhat idiosyncratic Specialist-Order Book-Upstairs structure of the NYSE. Using the LSE data, we examine the determinants of the proportions of order flow sent to the book and the dealers respectively. This allows us to characterise times at which and stocks for which dealer services appear to be most valued. A secondary aim of the study is to provide, in terms of quotation and transaction activity, a clear picture of the manner in which the LSE currently operates. Since the October 1997 reform there has been little academic study supplying basic statistical evidence on the extent of order book usage, the types of trades typically routed to the book (and to the dealers) and information on the intra-day patterns in transaction and quote activity. Partial evidence has appeared in newspaper and other media reports but these are insufficient to gain an overall picture of current trading patterns. The main results from our analysis are as follows. First, we present basic trading and quoting patterns, showing that in the first couple of hours of the trading day, total traded volume and limit order submission to SETS are both at very low levels whilst spreads and return volatility are both high. In the final hours of trading all of the aforementioned activity and liquidity variables are above levels observed in the middle of the trading day. Hence, spreads and volatility on SETS follow the U-shaped intra-day pattern uncovered on most other exchanges. This is not true of trading and limit order activity. Our panel regression analysis of the shares of order flow sent to the two market segments uses a 30-minute sampling frequency. The key set of results from our study account for the potential endogeneity of explanatory variables in the specification by using a panel IV technique. Moreover, we demonstrate, using a number of diagnostic checks and alternative estimation techniques, that the main results are very robust. We show, amongst other things, that the market share of the order book tends to be low if traded volume is extremely high, if mean transaction size is high and if trade imbalance (i.e. the excess of buyer-initiated over seller-initiated or seller-initiated over buyer-initiated trades) is large. The liquidity service provided by dual-capacity firms is hence increasingly 4

7 utilised in times of high execution risk (because of characteristics of the trade or market conditions) and, hence, these firms can be seen to smooth the trading process. We further show that the share of order flow sent to SETS is low if spreads in it are high or depth is low. Hence, when the terms of trade on SETS are poor, order flow migrates to the dealer segment of the market. The effect of increased return volatility on the market share of the limit order book is not consistently signed across our estimations. We suspect that this is due to the fact that the effects of return volatility on book trading activity will largely be felt via book liquidity (book liquidity suppliers withdrawing their orders in volatile times) and hence the direct effect from volatility to book share is inconsistent when one controls for liquidity. In sum, we feel that the evidence described above is suggestive of dealers performing a stabilisation role, in many respects similar to that assigned to dedicated market-makers. Dealers step in and intermediate an increased proportion of order flow when trading conditions are extreme or when the order book is sparse. Based on this assessment, there appears to be no clear need to re-introduce designated intermediaries in London. The rest of the paper is structured as follows. The next section provides some theoretical and empirical background to this study and sets out our research questions in more detail. Section 2 provides more information on the operation of the LSE, introduces and provides basic analysis of the LSE data and discusses how we process the data. Section 3 presents our empirical analysis of order flow and discusses the results derived. Finally, Section 4 summarises our findings and indicates directions for future research. 1 Related work 1.1 Theory Although differently structured from that at the NYSE, off-book trading in London can be thought of as constituting an upstairs market using the economic definition of Grossman 5

8 (1992): A downstairs market refers to an organized exchange where all members agree that trades take place publicly in a central place. An upstairs market refers to a market where trading takes place privately. Given the parallels between the dealer segment of the London market and the upstairs market at the NYSE, it is natural to examine this literature for potential theoretical predictions. Upstairs firms are typically not modelled in existing US literature as suppliers of immediacy (dealers), but as wholesale brokers or matchmakers. Early theoretical work such as Burdett and O Hara (1987) thus models the block trade syndication process, focussing on the trade-off between the benefits of searching for more counterparties and the potential price impact due to the leakage of information as this search goes on. Later studies focus on non-anonymity aspects, modelling the upstairs market as a screening device allowing for better price execution for uninformed block traders following Seppi (1990). The theoretical part of Madhavan and Cheng (1997) contains a similar intuition, leading to the testable implication that upstairs-executed trades should have a smaller permanent price impact. Implicitly, then, upstairs brokers supply a depth service, and, in the current context, we may interpret this as implying that the incentives to route an order off-book and take advantage of non-anonymous negotiation should be inversely related to book depth. (This is also consistent with empirical evidence on European stocks traded in London presented in de Jong, Nijman, and Röell (1995), which shows no clear relation between price execution and trade size for trades executed with London-based broker-dealers). It is important to note that even though the upstairs market at the NYSE is usually thought of as a brokerage or syndication process, Grossman and Miller (1988) argue that although search was the initial, and still remains the major function of the upstairs market, the amount of positioning and hence of market-making liquidity provided by upstairs firms has increased substantially in recent years. Stoll (1993) similarly evokes the dealerization of the NYSE, - although there is no direct evidence of this type of liquidity supply at the NYSE to our knowledge. We therefore also turn to the literature analysing the rationale for the existence of dealers (taken in this context to be synonymous with market-maker). An early task of microstruc- 6

9 ture theory was to explain the presence of market-making intermediaries within continuous financial markets, something not part of the standard Walrasian framework. The classic reference for this is Demsetz (1968) who views market-makers as suppliers of continuity or immediacy. The need for their services arises from the fact that buy and sell orders in continuous markets are not synchronised, such that order flow imbalances occur naturally. Some traders are then willing to pay a higher price in order to transact now rather than wait for matching orders to arrive on the other side of the market. An empirical implication of this analysis is that the demand for immediacy and hence for dealership services should be negatively related to a variable measuring trading activity in a stock. It also suggests that dealers should have greater value at times of order flow imbalance. Finally, trade size should also matter in determining the demand for dealer services since natural liquidity is less likely to be found on the other side of the market if the trade size is greater, for a given level of total trading activity. Building on this characterisation of market-makers as immediacy suppliers, Grossman and Miller (1988) model implications of the demand and supply of immediacy for market structure, in a framework of liquidity-motivated trading. At the heart of the model is execution risk, the notion that an adverse change in the equilibrium price may occur in the time it takes for an order to execute. Market-making intermediaries appear within continuous markets, and charge for the opportunity costs of maintaining a continuous presence in the market and for bearing execution risk. The model predicts that assets for which demand for immediate execution is high will tend be to traded on dealership structures. In turn, the demand for immediacy will depend to a large extent on the degree of price uncertainty (volatility). The authors illustrate this with futures, the leading examples of markets where the demand for immediacy is high, which are structured as competing market-maker systems. Another way to motivate the relationship between volatility and off-book trading is via the free option analogy of limit orders (Copeland and Galai, 1983): one would expect traders to switch to bilateral negotiation of trades rather than leave limit orders which can be hit by other participants at times of greater volatility (i.e. of greater informational asymmetries). 7

10 However, from an empirical perspective, the relationship between dealer/off-book activity and volatility may be somewhat clouded. In the prior theoretical literature volatility would lead to increased off-book activity as those supplying liquidity to the order book either refuse to do so or do so on poorer terms in volatile times. Hence the effects of volatility are felt through order book liquidity. Hence, if (as we do) one simultaneously relates order book activity to both book liquidity and volatility, the effects of the latter may well be small and inconsistent. One area of literature which does not provide much guidance for our analysis is the theoretical work examining the price-setting behaviour of market-makers, or contrasting dealership and auction architectures. This is because it focuses on the characteristics of price processes when order flow having certain characteristics is sent to one system or to the other, keeping other things equal. Therefore, this literature doesn t model a situation where both trading venues are available simultaneously and switching of order flow between them may occur. Besides, off-book trading in the London market does not represent a dealership market as such because quotes are not publically displayed and cannot be hit by any market participant. 1.2 Alternative hypotheses Our main aim is to determine whether the proportion of order flow routed to the network of broker-dealers is sensitive to the state of the market. There are, however, a number of reasons why off-book order flow could be of a fundamentally different nature to that executed on-book (and hence market-insensitive.) The widespread practices of order flow preferencing and internalisation documented in earlier research on the previous, pure dealership system SEAQ (Hansch, Naik, and Viswanathan, 1999) come to mind first, since they imply that part of the total order flow could be constantly routed off the order-driven system. Order flow that is preferenced is routinely sent by a large investor or her broker to the same dealer, often in exchange for other services (investment advice, stock research) which are not directly priced. Internalised order flow, also a feature of mature equity markets, is self- 8

11 preferenced in other words referred by a broker to the dealing desk of his own securities firm. Such order flow can be matched in-house by large firms drawing from their own pool of customers or executed against the firm s inventory, in both cases bypassing the order book entirely. This is more akin to the search/brokerage mechanism traditionally seen as the main function of the upstairs market and does not specifically involve the supply of dealership services in response to insufficient liquidity in the order-driven market. Third, it is also conceivable that certain types of traders having an investment strategy based on trading signals that require almost immediate execution (such as an active or momentum strategy) would never find the required degree of immediacy in the book and hence almost systematically bypass it. This implies a more mechanical complementarity between the two segments, related to the investment style of the final customer. Finally, another, very different reason why the order flow routed to dealers may be largely independent of the state of the order book is that member firms of the London exchange have to report any amount of shares temporarily taken in inventory. This will therefore include turnover related to the investment banking activities of member firms such as secondary equity issues or controlling blocks changing hands, rather than to their trading activities as such. This reported turnover will be of an economically different nature and essentially unrelated to instantaneous measures of book liquidity. We partly correct for this below in our filtering of the data. 1.3 Related empirical work There are few studies examining the functioning of the UK market since the introduction of order-driven trading. One of the main stated aims of the 1997 reform in London was to lower trading costs for public investors. Naik and Yadav (1999) examine changes in the cost of trading before and after the introduction of SETS, in the spirit of the study of the Nasdaq market by Barclay, Christie, Harris, Kandel, and Schultz (1999). They arrive, by and large, at comparable conclusions: public order exposure has produced a more competitive market, with less cross-subsidisation across different categories of traders, and 9

12 various measures of spreads indicate a decline in the cost of trading for liquidity suppliers. Another study using recent SETS data is Board and Wells (2000), who examine the amount and characteristics of order flow executed on Tradepoint, a proprietary trading system, since the launch of SETS. They report that it had a market share greater than 1% for only four stocks over 1998, and that this figure has declined to more or less zero since then. Turning to empirical studies of upstairs trading, the most well-known is the aforementioned paper by Madhavan and Cheng (1997). Based on the model discussed previously, the paper contains an important empirical contribution. The authors were to this day the only researchers to have been able to identify trades facilitated upstairs at the NYSE. They examine whether these trades are routed upstairs by block traders who want to make use of a lack of anonymity to signal that they are uninformed, and therefore obtain better price execution. Madhavan and Cheng (1997) report evidence consistent with this hypothesis, after correcting for the selectivity bias induced by the fact that traders who are able to signal they are uninformed will tend to select the upstairs venue. With UK market data comparable to ours, Ellul (2001) investigates patterns in volatility and examines trader choice using the selection model introduced in Madhavan and Cheng (1997). He reports much the same findings. In two recent studies, Booth, Lin, Martikainen, and Tse (2001) and Smith, Turnbull, and White (2001) confirm, for the Helsinki and the Toronto stock exchanges respectively, that upstairs trades have a lower permanent price impact than those executed downstairs, and also find that upstairs brokers are likely to execute order flow they identify as uninformed against their own inventory, consistent with the screening hypothesis. The analysis perhaps most closely related to ours is contained in two papers. Madhavan and Sofianos (1998) study the determinants of Specialist participation in overall NYSE trading activity, both across stocks and over time. An important contribution of this paper for interpreting the nature of the trading system in operation at the NYSE is to document large cross-sectional variation in Specialist market shares. This implies that very active stocks trade essentially in an auction market, with Specialist participation normally kept at a minimum, while illiquid stocks trade in a quasi-dealership market. They also report that the Specialist tends to step in and supply liquidity at times of high volatility, wide spreads, 10

13 or in large trade sizes, and in a way related to the Specialists inventories. Although related to ours, their aims and results of that paper are still significantly different since, as mentioned above, the Specialist has regulatory obligations to stabilise markets, implying that he must step in at times of greater volatility or illiquidity. We are examining whether freely competing liquidity suppliers tend to behave similarly, but without obligations or privileges. The second related paper (emerging around the same time as our work) is a study of on and off-book activity on the Australian Stock Exchange by Fong, Madhavan, and Swan (2001). In addition to a selection model, as in Madhavan and Cheng (1997), these authors estimate a panel regression model for the ratio of off to on-book trading activity, similar to that which we propose in Section 3. However, an issue with the Australian data is that trading frequency is so low relative to that on more established exchanges that the authors are forced to use an annual sampling frequency in the construction of their panel data set. This renders many interesting microstructure issues untestable. However, certain of their results from these annual data, notably those on the effects of book liquidity on off-book activity, corroborate those that we find using our intra-day data set. 2 The market and the data 2.1 The trading venues Prior to the 1997 reform, the London equity market was a pure dealership system. This structure was set up in the mid-1980s and explicitly modelled on the Nasdaq market. Key features of the trading system were its pre- and post-trade opacity, and concentration of order flow on the sell side. With regard to the latter, roughly five market-making firms handled between two-thirds and three-quarters of order flow in actively traded stocks. Although no claims of collusion were made as in Nasdaq, various constituencies became dissatisfied with these trading arrangements. The concentration of order flow left smaller market-makers or broker-dealers sidelined, while the institutional buy side complained 11

14 about the opacity of the market and retail investors felt they were cross-subsidising large traders. This ultimately led to the introduction of limit order trading in October Since then, an individual or institution wishing to trade equity on the London market has a choice of two fundamental trading methods. First, the investor could contact his dealer and instruct her to submit a limit or market order to the SETS order book. 7 Then, in the case of market order submission, he trades at the best prevailing book price. In the case of limit order submission, execution is not guaranteed but may occur with some price improvement. The second trading method is to call one s dealer and request an immediate trade with the dealer as counterparty. Here, the order book is not involved at all and any purchase or sale on the part of the investor has an opposite effect on the dealer s inventory. 8 Of course, an investor is not forced to choose solely between these two extremes and, typically, investors use both methods of trade. For example, an investor might call his broker with a request to buy a block of stock, specifying that 20% of the order is immediately executed against dealer inventory and that the rest of the order is worked through the SETS book over the remainder of the trading day. Finally, a dealer is not forced to take any part of any trade that he does not wish to. The two trading methods differ along a number of dimensions. The SETS system is described by the exchange as the main price formation mechanism. It is among the most transparent of all limit order books available in major equity markets as every outstanding limit order is displayed to member firms. There is also immediate publication of details of all book trades to participants. Finally, settlement is integrated into the SETS system. Shares traded on the book must be held in dematerialised form and have a t 5 day settlement cycle. All settlement details are handled by an outside firm called CRESTco. Off-book transaction activity (i.e. direct trade between dealers, direct matches of client orders by dealers or clients trading against dealer inventory) occurs in very different fashion. There is no pre-trade transparency as quotes are requested/provided on a purely bilateral basis. Post-trade transparency is high in that dealers are forced to report any trade consummated within 3 minutes of its occurrence. 9 Finally, there are no settlement restrictions in that paper shares may be traded and non-standard settlement cycles (i.e. not t 5) may be 12

15 requested. Perhaps the most important feature of London s current market structure, though, is that there are no forced interactions between either prices or quantities of book and off-book trades. No minimum portion of any order must be executed on the order book and, similarly, off-book prices are not constrained in any way by prevailing book quotes. As discussed earlier, this situation is unlike that in many other major equity markets, although in practice, several ways exist for institutional traders to bypass these constraints on other markets. Some other features of the London market are relevant to our study and should also be explained at this point. First, there are non-standard reporting rules for a class of block trades known as Worked Principal Agreements (WPAs). 10 Such agreements involve a dealer guaranteeing a price for the execution of a block but then trying to better that price for the customer over the trading day. If any portion of the WPA is executed in the book then it is automatically reported as with any other book trade. However, any portion of a WPA executed off-book need not be reported by the dealer until the entire WPA has been filled. Then, the WPA is reported along with an average price. These arrangements are a legacy of the market-making system in place pre-october 1997 and have consistently declined in importance since the introduction of SETS. 11 A second feature of the market which impacts upon this study is the existence of three broker-dealer firms who supply retail order execution systems and are known as Retail Service Providers (RSPs). These firms provide dedicated terminals allowing the execution of retail orders without price negotiation. The prices at which such orders execute is guaranteed to be at or within the book spread. To an extent, these arrangements are reminiscent of the Nasdaq market s Small Order Execution System but with the important difference that RSPs voluntarily provide this service. Note that, as these systems provide book spreads at worst, retail investors have little or no incentive to use the SETS order book. Hence, in summary, the London Stock Exchange currently provides two distinct trading locales under one roof. Of course, the same agents operate in both of these segments but 13

16 there is no imposed relationship between the prices and quantities in each. The focus of the analysis contained later in this paper is the determination of the proportion of order flow which is routed to each segment of the market. However, prior to looking at this issue, we provide some basic statistical information on the workings of the two segments. 2.2 The data set and basic analysis The data we employ in our study cover the months of November and December The full data set comprises all trades for this period, whether executed on or off the order book, along with full information on orders submitted to the book. To keep our analysis manageable we focus only on data for constituent stocks of the FT-30 index. The FT-30 stocks are a convenient subsample of the entire universe of equities as they are drawn from various size/liquidity deciles. This allows us to examine how the use of the two trading systems might differ for stocks of different market cap and turnover. All stocks in the FT-30 index are actively traded and, in aggregate, the raw dataset comprises 1,060,000 trades. Basic information on our sample stocks is given in Table 1. The table presents firm names and identifiers, market values as of 1/12/1999 (the middle of our sample) and the final column indicates each security s normal market size (NMS) at the same date. The NMS is a measure of the average institutional trade size in a stock as computed (and regularly reviewed) by the Exchange. 12 We will use this measure repeatedly in our descriptive statistics as well as in the computation of our regressors. Table 2 gives basic information on the trading activity (in NMS terms) over the two months in question for each of our sample stocks. On a cross-sectional level, the total quantity traded ranges from less than 3,500 NMS (for Tate and Lyle) to over 22,000 NMS (for Vodafone Airtouch). Note also that, for the majority of stocks, the median quantity traded in a single deal is between 1 and 5% of an NMS. The mean quantities tend to be much larger than this indicating that the right tail of the quantity distribution is long. Finally for this section, we present some basic information on how trading activity and 14

17 liquidity evolve on an intra-day level for our sample of stocks. Figures 4 and 4 present the intra-day patterns in SETS spreads, SETS midquote return volatility, SETS quotation activity and total transaction activity computed for 30 minute bins of SETS trading hours (8am-4.30pm). 13 In the construction of these plots we have aggregated activity across all 30 sample stocks. Looking first at the intra-day patterns in spreads and volatility we see that both tend to be high at the beginning and end of the trading day. This is the intraday pattern observed in data from many other markets (see, for examples, the evidence on spreads from Paris in Biais, Hillion, and Spatt (1995) and that on volume and spreads on the NYSE contained in Foster and Viswanathan (1993).) The pattern in SETS spreads also conforms with recent evidence contained in Naik and Yadav (1999). Examination of the intra-day patterns in quote and transaction activity tells a different story, however. For these variables, activity tends to be at its lowest around the open of the market, is pretty flat for the majority of the trading day and then rises fairly dramatically towards the market close. The fact that volume is at its lowest level at the open is at odds with the evidence from a number of other markets (for example the NYSE and Paris Bourse.) It is possible that the design of the algorithm which is used to open SETS (a very short preopening period) discourages traders from supplying liquidity or transacting around the open. 2.3 Filtering the Data The goal of this analysis is characterisation of the determinants of the proportions of order flow executed on and off the order book. Quite clearly, then, we need to ensure that the transactions included in our analysis are those for which the two trading mechanisms might actually compete. This necessitates filtering a fairly large number of individual trades from the data set. The first category of trades we exclude from the data are those which occur outside the hours of order book operation. These trades represent about 1.3% of the entire sample by frequency and are generally very large. It is interesting to note that a large proportion of these trades occur right after the book closes. Around 25% are reported in the minute after 15

18 the book closes and 75% in the 30 minutes following book closure. This is consistent with these trades being prearranged with London dealers to take place at the closing price, as found in US data by Barclay and Hendershott (2000). Second, for reasons of sheer size, we remove all individual trades larger than 8 NMS in quantity (0.06% of trades by frequency). 14 Finally, we remove all trades with non-standard settlement cycles as they would not have been eligible for book execution. This represents 26.5% of all trades, most very small and, obviously, all executed off the order book. In aggregate, this filtration of the data removes 27.6% of the trades by frequency and 21.8% by value, leaving about 770,000 trades in the data. Descriptive statistics on trading activity by market segment are given in Table 3. A first point to note from Table 3 is that almost exactly the same number of (filtered) trades occur in the two market segments. In traded quantity terms the share of the book in activity is lower at around 45%. We also see that book trades tend to be smaller than non-book trades (on average) and that the dispersion in book trade size is only around one third that of non-book trades. The empirical quantiles of the book and off-book trade size distributions are given in the final rows of Table 3. As in the stock-by-stock analysis from Table 2, the average trade size is much larger than median trade size for both book and off-book trades. The key thing to note from these quantiles, however, is that for all up to the 95th, the quantile of the book distribution is an order of magnitude larger than its non-book counterpart. On the other hand, the 95th percentile of off-book trade size is over 50% larger than that of book trades. The implication of these results is that off-book trades are generally very small but also contain a very large share of very big transactions. We would expect many of the very small trades to be executed off-book due to the RSP arrangements discussed in Section 2.1. Also, we would expect a high proportion of very large transactions to be handled off the book due to the likelihood that the terms on which they are executed would need to be pre-negotiated. Finally for this section, in Table 4 we supply stock-by-stock evidence on the extent of order 16

19 book usage. For the vast majority of the sample stocks, the share of the book in quantity traded (NMS) is between 40 and 55%. When looking at the number of trades routed to the book the comparable range is between 50 and 70%. This clearly indicates that the book is an important source of liquidity for all of the equities in question. From visual inspection of Table 4, the share of the order book in turnover does not appear to be clearly related to market cap. However, a simple cross-sectional regression of book share (in quantity terms) against market value demonstrates that a significant positive relationship exists between the two. 15 Hence, as economic intuition would suggest (and to the extent that value proxies for trading activity) there appears to be a greater need for dealership services (to ensure market continuity, etc.) in less actively traded stocks. The final columns of Table 4 give empirical quantiles for book and non-book trade sizes on a stock-by-stock basis. This evidence simply confirms that drawn from Table 3 on the type of trades routed to the dealers. Dealers tend to execute a very large share of both very small and very large trades. As such, if we compare quantiles for book and non-book trades we can see that for only 1 stock does the median non-book trade size exceed the median book trade size and for only 3 stocks is the 75th percentile of the book trade size distribution smaller than it s non-book counterpart. 3 Order flow interactions: panel analysis 3.1 Specification We now turn to examination of the determinants of the market share of the limit order system. The dependent variable in our analysis is the market share of the limit order book by number of shares traded, which lies in the 0 1 interval. We compute this market share over half-hour intervals of all trading days. This gives 17 intervals for each trading day (as SETS is available from 8:00 to 16:30 in our sample period). In cross-sectional terms, the final sample is made up of 28 stocks after removing two of the original sample. 16 The sample period covers 40 trading days, after removing the 24 th and 30 th of December during 17

20 which the market was formally open but when trading activity was exceptionally low. This leaves 19,040 observations in the panel data set used for estimation. To detect whether order flow migrates from the book to the dealers in response to changing market conditions, we use a panel regression framework. As indicated in the Introduction, our focus is on the competition between on and off-book liquidity suppliers for order flow, hence the specification of the dependent variable. We employ a 30 minute sampling frequency, rather than anything higher, for a couple of reasons. First of all, we are interested in when trading activity, in general, tends to be routed to one venue or another (rather than when a single trade might be routed to a given venue). Second, since off-book trades are potentially reported with a delay of up to three minutes, transaction level analysis of the hypotheses in which we are interested is not possible as the relevant variables cannot be accurately measured. For example, due to the slight reporting delay of off-book trades, at a given point in time the total trading activity over the last k minutes is not available to market participants as some of the trades occurring in this interval may not have been reported. For these two reasons, we conduct our analysis at a higher level of aggregation than is done in the switching regression studies of Madhavan and Cheng (1997) or Ellul (2001), which make use of transactions-level data Choice of explanatory variables A number of these regressors are expressed in Normal Market Sizes, which, as mentioned above, is a convenient way of making them comparable across stocks. Based on theory, evidence from descriptive statistics, and our own discussions with broker-dealers, we consider the following variables as potential determinants of switching of order flow between the two market segments: 1. A measure of order book depth, as less order flow is likely to be sent through the book when it is less deep, other things equal. This depth proxy is constructed as the time-weighted average of the total quantity available on the bid and ask sides of the book, for each time bin, expressed in NMS terms. 18

21 2. A measure of the cost of trading in the order book. We compute this regressor as the (time-weighted) average percentage inside spread from book quotes, and expect the estimated coefficient on this regressor to be negative if order flow is switched away from the order book when the cost of trading in it is relatively high. The theory in Demsetz (1968) suggests that dealership services are valued when liquidity is not present on the other side of the market, which could occur either in times of high order flow imbalance, large average trade size, or low activity. Also with respect to size, traders wanting immediate execution of large blocks are likely to find it cheaper to trade through the dealers if they can credibly signal that they are uninformed, which similarly suggests that off-book activity and trade size will be positively related (once the fact that much retail order flow is sent to the RSPs has been taken into account). These arguments suggest the inclusion of the following regressors: 3. A proxy for order flow imbalance. We compute the net quantity bought (the quantity of buyer-initiated trades less the quantity of seller initiated, again in NMS terms) and take the absolute value. This simple measure is similar to that used in Blume, MacKinlay, and Terker (1989) The average trade size for each stock-interval, computed in NMS terms. 5. A measure of total value traded: following the reasoning in Demsetz (1968) we include a regressor for total value traded in NMS terms. However, we expect that there will be non-linearities in the relationship between traded value and order book activity. Specifically, due to the reasons outlined above, we expect the largest trades to systematically go to the dealer segment of the market. If this is the case then, at high traded value levels, the relationship between book share and value might be negative. To investigate possible non-linearities we include two extra regressors. Dummies are created that pick out the intervals that lie in the lowest and highest quartiles of the traded value distribution taken across all stocks, respectively, and these dummies are then interacted with value itself. 19

22 The predictions in Grossman and Miller (1988) and the free option analogy of limit orders (Copeland and Galai, 1983) suggest inclusion of the following: 6. A measure of volatility, computed for a given half-hour segment as the square root of the sum of the squared 5-minute transaction price returns within that 30-minute interval. This measure is motivated by the research on realized volatility initiated in Andersen, Bollerslev, Diebold, and Labys (1999). When including this variable in our panel regression specification we lag it to avoid picking up a mechanical correlation between book activity and book midquote return volatility. We also add an interaction term to account for the overnight breaks in the sample. 18 Theory suggests that return volatility and the book s share in order flow should be negatively related. However, as noted previously, the fact that we also include book liquidity measures in our empirical specification means that the relationship between volatility and book share may be hard to pick up. Finally, we include several dummies to take into account some features of the market and the data documented above: 7. Two small trade dummies to account for the large numbers of very small trades mechanically routed to the Retail Service Providers. Hence, if a particularly high proportion of the volume traded in a given half-hour interval is executed in trades of very small size, we would expect the market share of the book to be relatively low. We use two dummies because the fraction of trades sent off the book decreases with trade size (among retail-sized trades). For all stocks and time bins together, we examine the distribution of trade sizes for trades smaller than or equal to one-tenth of an NMS (the first size bin in figure 1). We use the median value (0.005 NMS) and the 95th percentile (0.05 NMS) as cutoff points to define tiny and very small trades, respectively. If, within a half-hourly bin, the proportion of the value traded accounted for by each of these two categories of trade exceeds a usual level (the mean, computed across days and time bins, within stocks) by one standard deviation, the dummy equals one for that bin. We expect the estimated coefficients to appear 20

23 with a negative sign, and the coefficient for the very small dummy to be greater than the one for the tiny dummy. 8. We add an opening dummy variable dummy, to account for the relatively low proportion of trading done in the order book in the first hour of trading (see Figure 4) Choice of estimator The selection of a panel estimator is done in very standard fashion. The first step is to determine whether persistent individual effects are present in the data or if the time series can simply be pooled. The null hypothesis that a single constant term is adequate is rejected by an LM test. Having detected individual effects, we test whether they can be modelled as random or whether they are correlated with the regressors. A Hausman test rejects the hypothesis that the difference in estimated coefficients between random and fixed effects is not systematic, and we therefore fit a fixed-effects model as our baseline specification. We extend the baseline model in a number of directions subsequently. 3.2 Results Results are presented in table 5 (where the reported intercept is an average of the stockspecific estimated intercepts.) A first general comment is that all explanatory variables appear to be significant, most of them strongly so, since the significance levels given in the table are for two-tailed tests, whereas we have clear expectations regarding the signs of most of these coefficients. The only coefficient that is not significantly different from zero is that on the volatility interaction term (and hence in all further estimations we drop this variable). A second point to note is that all coefficients for which we had clear priors from theory and intuition have the expected sign. It appears from our results that the share of the order book in transaction activity drops 21